Patent Grant
12082140
The present invention relates generally to sensor networks, and particularly to methods and systems for synchronization, localization and navigation using sensor networks.
Source localization systems based on Time of Arrival (TOA) are used to locate emitters or reflectors using sensors distributed in an area of interest. The area can be as large as the earth (as in Global Satellite Navigation Systems, GNSS, including the well-known GPS) to very small, such as a room in an office building. Each sensor typically receives signals from emitters in its range, records the times of arrival of the signals, and then reports the results to a computer (at a single location or distributed), which constructs an estimation of the emitter locations. In many systems, the signals are processed to produce Differential Time of Arrival (DTOA) measurements, based on the differences between times of arrival of a signal at different locations. For convenience in the description that follows and in the claims, the term “TOA” is used generically to refer to all forms of measurement of signal times of arrival, including DTOA, unless specifically indicated otherwise. The term “emitter” is used generically to refer to all forms of point energy sources, including bodies that reflect energy.
Sensor synchronization is a key factor in the ability of the computer to find the emitter locations. Synchronization may be maintained by equipping each sensor with a very accurate clock (such as an atomic clock) or a GPS receiver, but these solutions are costly, pose hard constraints on system design, and require extra electrical power. In some systems, synchronization is maintained by two-way messaging or two-way TOA ranging between the sensors, but these approaches also increase the cost, complexity and power consumption of the sensors.
Critical infrastructures such as wireless communication networks, stock trading systems, smart power grids, airport landing guiding systems, and digital broadcast networks, are heavily dependent on GNSS signals for time and positioning. Since the GNSS signal is weak, however, even the smallest jammer can do significant damage to infrastructure, and many jamming events are detected daily around the world. Furthermore, if the operator of the GNSS constellation chooses to shut down its services (or perhaps limit them to certain authorized military users, for example), GNSS-dependent infrastructures will fail.
Embodiments of the present invention that are described hereinbelow provide methods and systems that use measurements made by a network of unsynchronized sensors to perform accurate synchronization of the sensor clocks, as well as localization and navigation using this sort of synchronization.
There is therefore provided, in accordance with an embodiment of the present invention, a method for sensor operation, which includes deploying a network of sensors, the sensors having respective clocks that are not mutually synchronized. At least a group of the sensors receive respective signals emitted from each of a plurality of sources, and record respective times of arrival of the signals at the sensors according to the respective clocks. Location information is provided, including respective sensor locations of the sensors. The respective clocks are synchronized based on the recorded times of arrival and on the location information.
In a disclosed embodiment, synchronizing the respective clocks includes estimating offsets and skews between the respective clocks.
The method may also include computing the source locations based on the sensor locations and the recorded times of arrival. In a disclosed embodiment, computing the source locations includes applying an estimator to a set of equations relating the recorded times of arrival and the source and sensor locations. Applying the maximum likelihood estimator may include applying an iterative optimization process to the set of the equations, wherein the optimization process derives a set of linear constraints on offsets and skews of the respective clocks based on the received signals. Additionally or alternatively, applying the iterative optimization process includes performing a convex optimization using maximum volume inscribed ellipsoid centering.
In one embodiment, the method includes detecting a fault in the network of the sensors based on the location information and the recorded times of arrival.
Typically, receiving the respective signals includes receiving radio signals, wherein the sources may include at least one satellite source, or possibly three or more satellite sources. Alternatively, the plurality of the sources may include the at least one satellite source and at least one terrestrial source. Synchronizing the respective clocks may then include using the times of arrival of the respective signals emitted from only a single satellite source and a single terrestrial source in order to synchronize the respective clocks of the sensors.
In some embodiments, the method includes finding a location of the at least one terrestrial source based on the recorded times of arrival and/or finding a direction of the at least one satellite source based on the recorded times of arrival.
Typically, the at least one satellite source is not a Global Satellite Navigation Systems (GNSS) satellite. Synchronizing the respective clocks may include detecting an attempt to spoof a satellite source, and discarding the signals received from the spoofed satellite source.
In a disclosed embodiment, synchronizing the respective clocks includes applying an orthogonal decomposition to a measurement space of the satellites.
Additionally or alternatively, receiving the respective signals includes, upon detecting a loss of signal from one of the sources, selecting a new source, and recording the respective times of arrival of the signals from the new source.
In one embodiment, the method includes navigating based on the recorded times of arrival and the location information.
In another embodiment, the sensors are associated with respective computers, and the method includes synchronizing operation of the computers based on synchronization of the respective clocks.
In an alternative embodiment, receiving the respective signals includes receiving acoustic signals.
There is also provided, in accordance with an embodiment of the present invention, a method for transmitter operation, which includes deploying a network of sources, having respective clocks that are not mutually synchronized. At least a group of the sources transmit, at times determined according to the respective clocks, respective signals, which are received by a plurality of sensors, and record respective times of arrival of the signals at the sensors. Location information is provided, including respective source locations of the sources. The respective clocks are synchronized based on the recorded times of arrival and on the location information. The sensors may be located accordingly.
In one embodiment, the network of sources includes base stations in a cellular communications network, and the sensors include user equipment in the cellular communications network, and synchronizing the respective clocks includes synchronizing operation of the base stations based on the signals received by the user equipment.
There is additionally provided, in accordance with an embodiment of the present invention, a method for sensor operation, which includes deploying a network of sensors, the sensors having respective clocks that are not mutually synchronized. At least a group of the sensors receive respective signals emitted from each of a plurality of sources, and record respective times of arrival of the signals at the sensors according to the respective clocks. Location information is provided, including respective sensor locations of the sensors. The source locations are computed based on the recorded times of arrival and on the location information.
There is further provided, a network system, including a network of sensors, which include respective clocks that are not mutually synchronized, and which are configured to receive, in at least a group of the sensors, respective signals emitted from each of a plurality of sources, and to record respective times of arrival of the signals at the sensors according to the respective clocks. A processor is configured to process the recorded times of arrival, using location information including respective sensor locations of the sensors, so as to synchronize the respective clocks.
There is moreover provided, in accordance with an embodiment of the present invention, processing apparatus, including a communications interface, which is configured to receive, from at least a group of sensors in a network of the sensors, which have respective clocks that are not mutually synchronized, times of arrival recorded by the sensors according to the respective clocks of signals emitted respectively from each of a plurality of sources. A processor is configured to process the recorded times of arrival, using location information including respective sensor locations of the sensors, so as to synchronize the respective clocks.
There is furthermore provided, in accordance with an embodiment of the present invention, a computer software product, including a non-transient computer-readable medium in which program instructions are stored, which instructions, when read by a computer, cause the computer to receive, from at least a group of sensors in a network of the sensors, which have respective clocks that are not mutually synchronized, times of arrival recorded by the sensors according to the respective clocks of signals emitted respectively from each of a plurality of sources, and to process the recorded times of arrival, using location information including respective sensor locations of the sensors, so as to synchronize the respective clocks.
Other embodiments provide systems, apparatus and software products that operate in accordance with the other methods laid out above.
The present invention will be more fully understood from the following detailed description of the embodiments thereof, taken together with the drawings in which:
Embodiments of the present invention that are described herein derive accurate synchronization from networks of sensors, without requiring that the respective clocks of the sensors themselves be mutually synchronized. In the disclosed embodiments, a group of sensors receives respective signals emitted from multiple sources, which may be terrestrial, space-based (i.e., satellites), or a combination of the two. The sensors record the times of arrival of the signals at the sensor locations according to their respective, unsynchronized clocks.
A processor (which may be centralized or associated with one or more of the sensors) receives and uses the recorded times of arrival, together with known location information, to synchronize the sensor clocks. Typically, this known location information comprises the sensor locations, while the source locations are not known a priori, although some of the source locations may be known. “Synchronizing the clocks,” in the context of the present description and in the claims, typically means calculating the offsets and skews between the various unsychronized sensor clocks, whether or not this synchronization is actually applied by the sensors themselves or only as a part of the computations performed by the processor. In some cases, some of the sensor clocks may be synchronized in advance, while others are not; and the term “synchronizing the clocks,” as used in the present description and in the claims includes cases in which only some of the clocks need be synchronized. As a by-product of the synchronization computation, the processor may also compute the hitherto-unknown source or sensor locations.
In alternative embodiments, the same methods of synchronization and localization may be applied, mutatis mutandis, to a network of unsynchronized emitters. Such techniques may be applied, for example, in synchronizing a cellular communication network, as explained in detail hereinbelow. For simplicity and clarity, the methods described below will refer in detail to the case of sensor synchronization, but the modification of these methods for emitter synchronization will be straightforward for those skilled in the art and is considered to be within the scope of the present invention.
The processor thus synchronizes and calibrates the time offsets and skews of the sensors (or, in alternative embodiments, the emitters) in the system passively, based only on the information that the sensors receive by wireless reception from source transmissions. As a result, the sensors can use inaccurate internal clocks and need only to be capable of measuring the time of arrival of the emitter signals relative to their own clocks and to relay this information to the processor. The sensors can thus be made small and inexpensive, while enjoying very low power consumption. The sources are not generally required to cooperate with the system in any way, thus allowing the use of sources of convenience—such as existing satellite or terrestrial sources, or very simple dedicated signal emitters—which operate in parallel without any interconnection.
Only limited location information need be known in advance for operation of the synchronization methods that are described herein. For example, the known locations (of the sensors or the sources) may be given in absolute terms or only relative to one another. Furthermore, if the locations of a certain number of the sensors (typically at least three) are already known, the remaining sensors can be located if they are allowed to transmit until their locations are estimated to the required accuracy. Alternatively, the sensors of unknown position can be considered navigators, and their positions and clock offsets can be resolved without necessarily ranging transmission on their part.
In some of the embodiments that are described hereinbelow, the processor carries out a process of estimation and optimization, such as maximum likelihood estimation, in order to compute both the offsets (and the skews if any) between the respective clocks and the unknown sensor locations. The possible values of the clock offsets and skews can be linearly constrained by the time of arrival measurements of the source signals. The clock offsets and skews and the unknown locations are estimated by applying an iterative optimization process to a set of equations and constraints relating the recorded times of arrival and the source and sensor locations. For terrestrial sources in particular, convex optimization using maximum volume inscribed ellipsoid centering has been found to give a good starting point for this estimation process. When satellite sources are present, the convex optimization can be replaced by other means, as described in detail hereinbelow.
Embodiments of the present invention are capable of operating with different numbers and configurations of terrestrial and/or satellite signal sources. For example, synchronization is possible in all of the following combinations:
Precision of synchronization in sensor networks according to embodiments of the present invention is limited only by the available processing capability and the sensor position accuracy. The precision can be enhanced by increasing the bandwidth, signal/noise ratio, observation time and number of sources processed. The sensor positions may be determined by any suitable means known in the art, such as surveying, localization (if the sensors also transmit signals), or navigation by the methods described herein, assuming that the positions of certain sensors are known and can be used as a reference.
Embodiments of the present invention that use satellite sources of signals provide notable benefits over existing GNSS systems, which are prone to jamming, spoofing and limitations on availability, particularly in crowded urban areas. (Spoofing is the act of a malicious transmitter that transmits a signal mimicking the GNSS signal, causing GNSS receivers to output a different location and/or time from the actual location and/or time.) By contrast with GNSS systems, embodiments of the present invention are capable of exploiting signals from substantially any of the thousands of satellites that roam the heavens, and permit satellite signals to be chosen for this purpose arbitrarily, on the basis of convenience at the time of use. The set of satellite signals that is used may also be changed at any time if and when required, for example if one or more of the signals are lost. The disclosed embodiments provide techniques for satellite-based timing synchronization, localization and navigation that are more robust and less prone to jamming and spoofing than GNSS signals.
Unlike GNSS satellites, whose locations are precisely known and whose signals are tightly controlled, the exact orbital parameters of most satellites in space are not known to great accuracy. The signals they emit have different spectral characteristics and are typically unsynchronized with one another, and their transmission times are not known in advance to the receivers. Embodiments of the present invention overcome these limitations, while taking advantage of the superior geometrical diversity and availability of arbitrary satellite constellations relative to the GNSS constellations. These embodiments provide methods that use the radio-frequency (RF) signals received from the satellites at a network of terrestrial sensors to provide timing synchronization and navigation for a sensor in the area of such a network and provide a superior alternative to GNSS for critical operations. Because the sensors can choose any desired group of satellites, the network is largely immune to jamming, and spoofing can be detected easily due to the deviation of the spoofed signal from the normal signal model. For sensor networks that also process terrestrial signal sources (of known or unknown positions), the synchronization accuracy can be further improved by incorporating measurements of signals from terrestrial sources, such as broadcast signals or wireless user equipment or emitters localized by the system, in the synchronization process.
In some embodiments, localization accuracy finer than the RF carrier wavelength can be achieved, and carrier phase measurements may then be used to refine the location and synchronization estimates still further. Furthermore, the large number of different satellite signals (as well as terrestrial sources) received by the network can be used to derive a stable clock frequency in all of the sensors, thus enabling the sensors to use longer coherent integration times. One of the benefits of extending the coherent integration time is the ability to detect weak signals in high-attenuation environments, such as indoors.
In another embodiment, the methods described in the present patent application are used to estimate the direction (azimuth and elevation) of received satellite signals or other distant sources. These estimates can be applied in satellite orbital tracking, as well as ranging to the satellite when Doppler measurements are also available.
Sensors 22 are connected by a communication network 30 to a synchronization controller 28, which processes TOA data that is measured and conveyed by the sensors over network 30 in order to perform synchronization, localization, and other related functions that are described herein. Network 30 may comprise any suitable sort of communication network that is known in the art. For example, network 30 may comprise a dedicated, wireline or wireless network. Alternatively or additionally, network 30 may comprise a public network, such as the Internet or a cellular communication network. Although controller 28 is shown and described here as a centralized unit, the functions of the controller may, additionally or alternatively, be performed in a centralized or distributed fashion by processors at different locations on network. 30, including processors that are coupled to and associated with one or more of sensors 22. For purposes of the basic methods described below, it is sufficient that sensors 22 transmit their respective TOA data over network 30 to controller 28, without return transmissions of any sort. Return transmission may be useful in some applications, however, for conveying control, synchronization and/or navigation information, for example, from controller 28 to sensors 22.
Although sensor network 20 and the embodiments described hereinbelow operate on the basis of RF signals transmitted by sources 24 and 26, in alternative embodiments (not shown in the figures), sensors 22 may receive and measure times of arrival of signals of other types. For example, sources 24 may comprise acoustic emitters, and sensors 22 may comprise acoustic receivers, which are used in localizing the emitters using the methods described below.
As noted earlier, local clock 36 need not be particularly accurate and is generally not synchronized with the local clocks of other sensors. Any suitable sort of oscillator, as is known in the art, may thus be used in generating the clock signal according to which processor 34 makes its TOA measurements. If one or more of sensors 22 has access to absolute-time clock, then this information can be used to propagate the absolute time among the sensors in sensor network 20 after synchronization is established.
Processor 42 and the other components of controller 28 may be parts of a general-purpose computer, which is programmed in software to carry out the functions that are described herein. This software may be downloaded to controller 28 in electronic form, over a network, for example. Alternatively or additionally, the software may be stored in a tangible, non-transitory computer-readable medium, such as optical, magnetic, or electronic memory (which may be incorporated in memory 44). Further alternatively or additionally, at least some of the functions of processor 42 may be carried out by dedicated or programmable hardware logic.
To begin the procedure, the clock offset vector and cost function values are initialized, at an initialization step 50. In each iteration n, controller 28 estimates the location of each source t=1,2, . . . ,N based on the estimated clock offsets and the current value of the cost function, at a position estimation step 52. Then, given the estimated source locations, controller 28 updates the clock offset estimates, at a clock update step 54.
To determine whether the estimates have converged sufficiently to terminate the computation, controller 28 compares the new cost value to the cost value from the previous iteration, at a cost evaluation step 56. Specifically, the controller may evaluate whether the change in cost since the previous iteration is less than a certain threshold. If not, n is incremented, and control returns to step 52 for a further iteration.
If the computation is found to have converged at step 56, controller 28 synchronizes the clocks of sensors 22 based on the current value of the estimated clock offsets, at a synchronization step 58. Depending on application requirements, the controller may either send actual synchronization values to the sensors, or it may simply record and apply appropriate corrections to the clock readings received from the sensors. The controller may also output or otherwise apply the current estimate of the source location values, at a location output step 60.
In dealing with satellite sources, it is assumed that sensors 22 are sufficiently close together (typically not more than a few tens of kilometers apart) so that ionospheric and tropospheric effects on the satellite signals can be considered to be uniform. Otherwise, measurements should be corrected by tropospheric and ionospheric correction models or measurements. If the satellite signals occupy a large bandwidth or are scattered over a wide frequency band, the ionospheric correction can be estimated accurately, as is known in the art. Alternatively or additionally, comparison of signals may be limited to measurements from sets of sensors that are close to one another, thus allowing the entire sensor set to be deployed on a wider area.
The subsections that follow will demonstrate a number of different implementations of the method of
Consider M sensors and N sources. The s-th sensor has an internal clock offset denoted by δs. We use sensor number 1 as a reference sensor and therefore we set δ1=0. Let mt(s) be the TOA measurement performed by the s-th sensor for the t-th source. Denote by qs,pt the known coordinates vector of the sensor and the unknown coordinates vector of the source, respectively. Let τt be the unknown transmit time. Then
where etsis a random measurement error with zero mean and a standard deviation of σt(s). Here, for reasons of clarity, we assume σt(s)=σt, or more precisely, the noise is assumed to be independent and identically distributed (i.i.d). Under this model the problem of self-synchronized localization can be defined as follows: Using all measurements mt(s) made by the sensor network estimate δs,pt∀s,t where s=1,2, . . . M; t=1,2, . . . N.
The self-synchronized localization problem, in a planar geometry, has 3N+M−1 unknowns: 2N unknown source coordinates, N unknown transmit times and M−1 sensor time offsets w.r.t. the reference sensor. If all sensors receive the signals of all the sources, we have MN TOA measurements. Thus, the problem is well posed if MN≥3N+M−1. Define the vectors associated with source t
Using the above definitions the measurements defined in (1.1) become,
mt=rt+1Mτt+δ+et. (1.2)
When the error vector is Gaussian i.i.d the Maximum Likelihood cost function is
The cost function is minimized w.r.t. the time offsets by
Substituting this result back in (1.3) yields
where
Collecting all vectors for t=1,2, . . . ,N we get
m[{tilde over (m)}1T,{tilde over (m)}2T . . . {tilde over (m)}NT]T
r[{tilde over (r)}1T,{tilde over (r)}2T . . . {tilde over (r)}NT]T
e[{tilde over (e)}1T,{tilde over (e)}2T . . . {tilde over (e)}NT]T
C[IM,IM . . . IM]T. (1.7)
Thus, the cost function for minimization is now
Q=∥m−r−C{tilde over (δ)}∥2 (1.8)
The minimum w.r.t.
Substituting back in (1.8) we get
Thus, the search is “only” over {pt}t=1N (2N unknowns). To ensure that the number of equations is not smaller than the number of unknowns, at least two sources and five sensors are required. Thus, solving the above equation for two sources requires a search within a four-dimensional space.
Reconsider equation (1.5) and note that if δ is given the cost function is separable into independent components, one for each source. This enables us to locate each source separately, given the assumed δ vector. Therefore, the problem can be solved by successive optimization of two-dimensional problems. Using this fact, we can devise the following iterative algorithm for the estimation of the clock offsets and the sources position:
Note that if the position of one source is known, then the clock offsets of the sensors that sense it can be estimated directly at the accuracy of the raw measurement. Additional refinement can then be achieved by using that estimation as the initial guess for the δ vector. A possible source with known location can be a sensor that is also allowed to transmit.
The above algorithm can be used for network-wide synchronization and localization as long as the number of measurements is not smaller than the number of unknowns.
Define the M(M−1)/2×M difference matrix, D, where each row consists of a single 1, a single −1 and zeros, such that Dmt is the vector whose entries are all the possible differences of the entries of mt. Define δ′ such that, δ=[0 δ′T]T and {tilde over (D)} as the difference matrix, D, with the first column omitted. Applying D to both side of equation (1.2) we get
since D1M=0, and δ1=0 by definition. Note that Drt is the vector holding all possible Time Difference of Arrival (TDOA) measurements. Using the triangle inequality we have |rt(i)−rt(j)|<d(i,j), where d(i,j) is the distance between sensor i and sensor j. Collecting all possible pairs, (i,j), the triangle inequality becomes the componentwise inequality −dDrt
d, where d is a vector whose entries are d(i,j). Applying this result to (1.11) we get,
{tilde over (D)}δ′ Dmt−Det+d
−{tilde over (D)}δ′ −Dmt+Det+d. (1.12)
The set of δ′ vectors satisfying the above constraints form a polyhedron in M−1 space.
Assuming that Det is negligible, one can find a δ vector that satisfies the constraints by solving the Linear Programming (LP) feasibility problem:
Program 1.1 Clock Offset Feasibility Test
Find: δ
Subject to: ±{tilde over (D)}δ′±Dmt+d, t−1, . . . ,N
This problem will return an arbitrary point in the feasible set defined by the constraints. A center point in the set would be favorable since its distance to the real vector will be on average less than any point on the boundary. One way of getting a center point of the polyhedron is finding the Maximum Volume Inscribed Ellipsoid (MVIE), which finds the center coordinates δ′c of the biggest ellipsoid ε defined as:
that can fit inside of the polyhedron defined by the constraints. Here SM−1++ stands for positive definite M−1×M−1 symmetric matrix. Then the centering program is defined by:
Program 1.2 Maximum Volume Inscribed Ellipsoid Centering on Clock Offsets
Minimize: logdetG−1
Subject to: ∥GaiT∥+aiδ′c≤bt
ai is the i-th row of A, and bt
The measure of uncertainty of the estimation of the δ vector is useful in a number of ways. As noted earlier, given a source of known location, the clock offsets are readily found at raw measurement accuracy. Therefore, if we have a subset of more than three clock offsets estimates with high certainty, we can get a good source position estimation and thus a high-certainty estimation of the clock offsets for all the sensors that received a transmission from this source.
When the noise is significant we cannot discard the term Det from the constraints in (1.12). If we collect all the constraints related to a given sensor pair we get:
δi−δj≤min({D(i,j)mt−D(i,j)et+d(i,j)}t=1N)
δi−δj≥max({D(i,j)mt−D(i,j)et−d(i,j)}t=1N)
where D(i,j) is the row in D that relates to the i,j delta pair. The measurement error vector, et, is unknown. We can ignore the measurement error and still use the min and max operators, but this approach will result in large errors.
Instead, we introduce a function that encourages small errors, the Pseudo-Likelihood function, which provides the likelihood of Δijδi−δj given the constraints imposed on d(i,j) for each sensor pair using all measurements made by the pair. The Pseudo-Likelihood function fΔ
where Tij is the set of sources that were measured by both sensor i and sensor j, which here include all sources.
Using the pseudo-likelihood function we can obtain the pseudo-probability that Δij∈(lij,hij). We select Δ∈(lij,hij) so that Δij is within the interval with pseudo-probability 1−α. Since the pseudo-likelihood is normalized, it has the properties of a probability density function, therefore:
As there is a degree of freedom in the choice of lij and hij, these elements will be chosen such that ∥hij−lij∥ is minimized.
When noise is introduced, the ML estimator is such that it is not necessary to estimate the full δ vector, but only its projection Bδ. Estimating Bδ cannot be achieved directly by using MVIE since the domain of Bδ has M−1 non-singular dimensions, making its volume in M always zero for any set of constraints. Therefore any inscribed ellipsoid volume will always be zero, making any volume maximization impossible. To overcome this limitation, we apply the MVIE only to the projection subspace defined by B. This subspace is the span of the eigenvectors of B corresponding to the unity eigenvalues.
We define the matrix V, whose columns are the eigenvectors of B whose corresponding eigenvalues are different from zero. Collecting the likelihood functions fΔ
Program 1.3 Maximum Volume Inscribed Ellipsoid Centering on Clock Offsets Under Pseudo-Likelihood Imposed Bounds
Minimize: logdetG−1
Subject to: ∥GaiT∥+ai{tilde over (δ)}′c≤bi
Return: {tilde over (δ)}c←V{tilde over (δ)}′c
Where:
bi is the i-th element of b; hΔT(α) and IΔT(α) are the upper and lower bound vectors imposed by α, via the pseudo-likelihood function on all delta pairs; {tilde over (δ)}c is the MVIE estimate of Bδ; and {tilde over (G)} describes the uncertainty ellipsoid of the estimation.
Note that A in Program 1.3 is not sparse. If δ′c were to be estimated using the MVIE instead of {tilde over (δ)}c, then A would be sparse, and would therefore enable lower complexity computation without loss of accuracy.
If Program 1.3 returns the non-feasible token, then the value of α is decreased, and the solution is repeated.
MVIE gives an initial estimation of the target position by embedding the estimated vector in the source measurement model, which gives:
where
is the MVIE estimation error of {tilde over (δ)}. If we model the MVIE estimation error as an independent noise with the covariance matrix GTG, then the initial source position estimate can be found by solving the following weighted least squares (WLS) optimization problem:
where Σ=GTG+Bσt2, and σt2 is the measurement noise variance of the source. After locating all of the sources using the above estimator, the computation can jump to step 4 of the iterative ML algorithm presented above and continue iterating until convergence is achieved. Two other possible ways to proceed are
Using GTG as a covariance matrix is conservative, since it describes the ellipsoid within which the clock offsets vector almost always lie. Therefore adding a constant factor to the matrix to convert it from a certainty ellipsoid to a covariance matrix is appropriate. Defining the constant factor as κ>1, the covariance matrix for the MVIE clock offset estimation is
Methods are known in the art for computing κ to convert certainty ellipsoid to a covariance matrix of a multi-variate normal distribution, but κ can also be determined empirically.
The MVIE/ML estimator performs well under the imposed constraints. The MVIE/ML estimator even after only one ML iteration comes very close to the converged ML performance. This feature makes the combined MVIE/ML estimator very attractive in terms of processing time. Comparing the results of the combined MVIE/ML estimator to the known clock offset case, we observe that even in the first iteration, the penalty of not knowing the clock offsets is not very large. Moreover, even when σn is on the order of the distance between the sensors (which is a highly nonlinear configuration), the MVIE/ML estimator converges successfully. Moreover the MVIE can also be used as a priori information in a MAP estimator.
To simplify the description above, it was assumed that all sensors intercept all the sources. It is straightforward to extend the above results to the more general case.
Note also that the ML estimator cannot Use measurements of sources that are received by three sensors or less, but MVIE can.
One further point to note is that clock bias is actually the measurement bias, meaning that it contains any propagation delays inside of the sensor. For example, a sensor that has an antenna connected with a long cable will include the cable propagation time as part of the bias. In sensing systems, this effect should generally be calibrated. The method described above provides sensor self-calibration of such delays.
In the previous section, the MVIE used constraints based on measurements of a source by a pair of sensors. This section will present constraints based on measurements of a source by a sensor triplet. Methods to incorporate these constraints into the MVIE will also be introduced.
Let there be three sensors with position coordinates qi. Set direction vectors vij as:
vijqj−qi (2.1)
Define two angles:
α∠(v01,v02)
β∠(v12,−v01) (2.2)
where the convention is that for vectors μ,ν the angle ∠(μ,ν) is positive if the ν direction is restored by counterclockwise rotation of μ by less than a π rad turn. The noiseless DTOA measurement made by a pair sensors on a target positioned at p is defined, in metric units, as:
Ψij(p)=∥p−qj∥−∥p−qi∥ (2.3)
Measurement space is defined as the R2 space (Ψ01,Ψ02). A nine-segment {ui}i=08 path in space is established to map to the bound in measurement space. The path will traverse the plane spanned by vectors v01,v02, as follows:
u0(t0)=q0−t0v01
u1(t1)=q0−t1v02
u2(t2)=q1+t2v01
u3(t3)=q1−t3v12
u4(t4)=q2+t4v02
u5(t5)=q2+t5v12
u6(φ0)=q0+rT(φ0)v01,φ0∈[0,α]
u7(φ1)=q0+rT(φ1)v01,φ1∈[π−β,π]
u8(φ2)=q0+rT(φ2)v01,φ2∈[π+α,2π−β]
ti>0,∀i (2.4)
where r goes to infinity, and T(φ) is the rotation matrix for a counter clock wise rotation of φ degrees.
Traveling the path segments in space translates to the following path in measurement space:
h0(l0)=l0[|v01|,|v02|]T+(1−l0)[|v01|,|v02|cos(α−π)]T
h1(l1)=l1[|v01|,|v02|]T+(1−l1)[|v01|cos(π+α),|v02|]T
h2(l2)=l2[−|v01|,|v12|−|v01|]T+(1−l2)[−|v01|,−|v02|cos α]
h3(l3)=l3[−|v01|,|v12|−|v01|]T+(1−l3)[−|v01|cos β,−|v02|cos(α+β)]T
h4(l4)=l4[|v12|−|v02|,−|v02|]T+(1−l4)[−|v01|cos α,−|v02|]T
h5(l5)=l5[|v12|−|v02|,−|v02|]T+(1−l5)[|v01|cos β,|v02|cos(α+β)]T
li∈[0,1]∀i
h6(φ0)=−[|v01|cos φ0,|v02|cos(α−φ0)]T
φ0∈[0,α]
h7(φ1)=−[|v01|cos φ1,|v02|cos(α−φ1)]T
φ1∈[π−β,π]
h8(φ2)=−[|v01|cos φ2,|v02|cos(α−φ2)]T
φ2∈[π+α,2π−β] (2.5)
The elementary bound, derived from the triangle inequality, −|νij|≤Ψij≤|νij|, is directly responsible for the linear part of the bounding path. Therefore the additional information gained by considering two-dimensional bounds is found in the nonlinear part of the bounding path (namely h6,7,8).
To avoid the complexity of incorporating the nonlinear bound in the convex optimization described above, it is possible to find the tangents to the nonlinear part of the bound. Each tangent will act as a half-space bound on the offset values. The more tangents we use, the tighter the bound will be, but at the cost of greater complexity.
The nonlinear part of the bound is composed of three sections of the ellipse:
Ψ02=−|ν01| cos(α−φ)
Ψ01=−|ν01| cos(φ)
φ∈[0,α]∪[π−β,π]∪[π+α,2π−β] (2.6)
Taking the partial derivative of Ψ01 and Ψ02 by φ gives:
Dividing the two expressions results in:
The tangent equation is then:
As noted above, there are three segments of the ellipse that are part of the bound. We take K tangents for each segment, evenly spaced in terms of φ so that:
The segments h6,7 are always in the lower part of the bound, so that their tangents pose a lower bound, while h8 is always on top, so that its tangents pose an upper bound.
The bounds on a pair of DTOA measurements made on a single source are related to the set of possible clock offsets. Let the measurement made on a source t by sensor s be:
mt(s)=rt(s)+τt+δs+et(s) (2.11)
Define {tilde over (m)}(k,s,p)mt(p)−mt(k)−q1(φ)(mt(s)−mt(k)) and Δij=δj−δi. Expanding this expression gives:
where {tilde over (e)}tet(p)+(q1(φ)−1)et(k)−q1(φ)et(s) and Ψt(ks)=rt(s)−rt(k). The last equation defines the relation between the bound on the DTOA measurements and the clock offsets. The direction of the inequality is determined by the value of φ: For φ∈[0,α]∪[π−β,π] it will be ≥ and for φ∈[π+α,2π−β] it will be ≤.
To fuse all of the measurements made by all the sensors we again turn to the pseudo likelihood function. The pseudo likelihood function
of {tilde over (Δ)}ksp(φ) is defined by:
where {tilde over (σ)}t2=2σt2(1+q1(φ)), and Tksp is the set of all sensor triplets. The direction of the inequality is determined, by the value of φ as before. The pseudo likelihood function yields the constraint
w.p. 1−ε. This is a linear constraint on the clock offsets that can be incorporated into the MVIE optimization problem.
Since this is a one-sided constraint, some a priori knowledge is needed on the bound of {tilde over (Δ)}ksp(φ) on its other side. For this purpose, the bounds can be derived from the elementary bounds, or the lowest feasible value can be used.
In this section, the above methods are expanded to include clock skew estimation by convex methods. With slight approximations, the skew estimation is independent of the offset estimation and can be achieved by similar methods.
Consider M sensors and N sources. The s-th sensor has an internal clock with offset denoted by δs and skew denoted by θs. We use sensor number 0 as a reference sensor, and, therefore we set δ0=0 and θ0=1. Let mt(s) be the TOA measurement performed by the s-th sensor for the t-th source. Denote by qs, pt the known coordinate vector of the sensor and the unknown coordinate vector of the source, respectively. Then
is the propagation time between source t and sensor s. Let τt be the unknown transmit time. Then
mt(s)=θs(rt(s)+τt)+δs+et(s) (3.1)
where ets is a random measurement error with zero mean and a standard deviation of σt(s).
It can be reasonably assumed that all clock skews can be described by a slight deviation from unity that is denoted by η:
θs=1+ηs (3.2)
With reasonable values of η, it can be shown that
θsrt(s)≈rt(s) (3.3)
is a good approximation since ηsrt(s)«σt(s). For the same reason we will assume that multiplying the noise by any clock skew does not change the noise statistics.
Given this model, the problem of self-synchronized localization can be defined as follows: Using all measurements mt(s) made by the sensor network, estimate δs,θs,pt∀s,t where s=1,2, . . . M−1; t=1,2, . . . N.
The self-synchronized localization problem, in a planar geometry, has 3N+2(M−1) unknowns: 2N unknown source coordinates, N unknown transmit times, M−1 sensor time offsets, and M−1 clock skews w.r.t. the reference sensor. If all sensors receive the signals of all the sources, there are MN TOA measurements. Thus, the problem is well posed if MN≥3N+2(M−1). Define the vectors associated with source t,
where ∅ is defined as the element-wise division operator. Using the above definitions, the measurements defined in (3.1) become,
When the error vector is Gaussian, the Maximum Likelihood cost function is
The cost function is minimized w.r.t. the time offsets by:
Substituting this result back in (3.5) yields:
where
Collecting all vectors for t=1,2, . . . ,N gives:
m[{tilde over (m)}1T,{tilde over (m)}2T . . . {tilde over (m)}NT]T; r
[{tilde over (r)}1T,{tilde over (r)}2T . . . {tilde over (r)}NT]T;
e[{tilde over (e)}1T,{tilde over (e)}2T . . . {tilde over (e)}NT]T; C
[A1,A2 . . . AN]T. (3.9)
Thus, the cost function for minimization is now:
Q=∥m−r−C{tilde over (γ)}∥2 (3.10)
The minimum w.r.t. {tilde over (γ)} is obtained for
{tilde over ({circumflex over (γ)})}=(CTC)−1CT(m−r) (3.11)
Substituting back in (3.10) gives:
As a result, the search is over only {pt}t=1N (2N unknowns). To ensure that the number of equations is not smaller than the number of unknowns, at least three sources and seven sensors are required (for the minimal number of sources). Thus, solving the above equation for three sources requires a search within a 6 dimensional space.
Reconsider equation (3.7) and note that if {tilde over (γ)} is given, the cost function is separable into independent components, one for each source. This separation enables controller 28 to locate each source separately, given the assumed {tilde over (γ)} vector. Therefore, the problem can be solved by successive optimization of two-dimensional problems. Using this fact, the following iterative algorithm can be applied for the estimation of the clock offsets and the source positions:
Controller 28 starts by collecting all measurements made by a pair of sensors k,s. For a source t the measurement of sensor k is given by Eq. 3.1. Taking the difference between the measurements made by sensors k and s on source t gives:
mt(k)−mt(s)=
θk(rt(k)+τt)+δk+et(k)−θs(rt(s)+τt)−δs−et(s)≈
rt(k)−rt(s)+τt(θk−θs)+δk+et(k)−δs−et(s)=
ϕt
where ϕtrt(k)−rt(s) is the DTOA measurement on source t by sensors k,s.
We define:
Δksδk−δs
dmtmt(k)−mt(s) (3.26)
The unsynchronized DTOA measurement for source t by sensors k and s is then:
Taking the difference between two unsynchronized DTOA measurements of different sources gives:
where {tilde over (e)}t,p(k,s)et(k)−et(s)+ep(k)−ep(s). When considering a sensor pair, the relative clock skew may be defined as
Applying this concept gives:
Since in Eq. 3.28, τt−τp is multiplied by the clock skew deviation η, only a crude estimation of τt−τp is needed to get an accurate clock skew estimation. The following approximation can be used for the Tx time difference:
The summation is over the common sensing sensors of t and p.
Dividing by the estimated time difference gives:
Finally we get:
The triangle inequality gives bounds on ηks:
where dks∥qs−qk∥.
The pseudo likelihood function provides the likelihood of a value of ηks given the constraints imposed by Eq. 3.32 for each sensor pair using all measurements made by the pair. Assuming all measurement noise is i.i.d and has variance σt, the pseudo likelihood function fη
Here Tk,s is the set of source pairs that were measured by both sensor k and sensor s, which is assumed here to include all sources. It is assumed that ddmt,p(k,s) are uncorrelated with one another, but this approximation can easily be eliminated. From the pseudo likelihood function, we obtain E[ηks], var[ηks], and ηks∈(lks,hks) w.p. 1−α.
The estimation for the relative clock skew is given by {circumflex over (Θ)}ks=1+{circumflex over (η)}ks. Therefore, bounds that were found on ηks can be translated into bounds on Θks. If ηks∈(lks,hks) w.p. 1−α, then Θks∈(1+lks,1+hks) w.p. 1−α. Taking the natural logarithm of the inequalities imposed on Θks gives the following set of linear constraints on lnθk:
lnΘks=lnθk−lnθs∈[ln(1+lks),ln(1+hks)]w.p. 1−α. (3.34)
Collecting the linear constraints for all k,s sensor pairs, Maximum Volume Inscribed Ellipsoid (MVIE) centering can be used to find the estimate , and the related error covariance matrix can be approximated by GlnθGlnθT.
Program 3.1 Maximum Volume Inscribed Ellipsoid Centering on Clock Skews Under Pseudo Likelihood Imposed Bounds
Minimize: logdetGlnθ−1
Subject to: ∥GlnθaiT∥+ailnθc≤bi
Where:
bi is the i-th element of b, and hlnΘT(α) and llnΘT(α) are the upper- and lower-bound vectors imposed by α, via the pseudo likelihood function, on all relative clock skews.
If the program returns the non-feasible token, be value of α can be decreased, and the solution repeated.
Note that Gθ=Glnθ, since the error is much smaller then unity, and Taylor approximation holds firmly. Therefore, the estimator of the θ vector is:
{circumflex over (θ)}=exp(); Gθ=Glnθ (3.36)
After establishing an estimation of the clock skew vector {circumflex over (θ)} and deriving an approximation for its error covariance matrix: GθGθT, the next step is to estimate the clock offset vector δ. For this purpose, the measurements made by the sensors are normalized by their respective clock skew estimates as follows:
Applying Taylor approximation to
gives
where eθ
The above equation states that the clock skew uncertainty is amplified by the source transmit (Tx) time. The Tx time can be referenced to a selected point in time. To minimize the noise amplification effect, the reference time can be set to the mean of the Tx times of all of the sources. Then τt may be estimated by:
Using the above first-order approximation of τt, the equivalent measurement difference noise for sensors k and s is:
when vksT is a difference vector holding 1 and −1 in the k−1, s−1 rows respectively, and containing zeros elsewhere. (If s=1, then the −1 is missing from the vector.) Given σks, the pseudo likelihood measure can be applied to give bounds on
for every s,k. The {tilde over (Δ)} pseudo likelihood measure is defined as follows:
where Tij is the set of sources drat were measured by both sensor i and sensor j (taken to include all sources, as noted earlier). The pseudo likelihood function also gives E[{tilde over (Δ)}ij], var[{tilde over (Δ)}ij], and {tilde over (Δ)}ij∈(lij,hij) w.p. 1−α.
Collecting the likelihood functions f{tilde over (Δ)}
Program 3.2 Maximum Volume Inscribed Ellipsoid Centering on Clock Offsets Under Pseudo Likelihood Imposed Bounds
Minimize: logdetG{tilde over (δ)}−1
Subject to: ∥G{tilde over (δ)}aiT∥+ai{tilde over (δ)}c≤bi
Where
h{tilde over (Δ)}T(α) and l{tilde over (Δ)}T(α) are the upper- and lower-bound vectors imposed by α, via the pseudo likelihood function, on all {tilde over (δ)} pairs.
From {tilde over (δ)}c, an estimate of δ can be computed by multiplying it element-wise by {circumflex over (θ)}:
{circumflex over (δ)}={circumflex over (δ)}c⊙{circumflex over (θ)} (3.42)
This multiplication cancels the prior normalization and therefore introduces no additional errors.
Having estimated {circumflex over (θ)} and {circumflex over (δ)} with their respective approximations of the error covariance matrices, GθGθT, and GδGδT, the skew and bias can be removed from the sensor measurements as follows:
Defining {tilde over ({tilde over (m)})}t=[{tilde over ({tilde over (m)})}t(0),{tilde over ({tilde over (m)})}t(1), . . . ,{tilde over ({tilde over (m)})}t(M−1)] and calculating the covariance,
the optimal source position in the WLS sense reduces to the following optimization problem:
Thus, controller 28 can synchronize a sensor network with skewed, asynchronous clocks by measuring the TOA of a set of source signals. Using MVIE to get the initial guess helps avoid local minima in the cost function.
If the position of one source is known, the clock offset can be measured at the time of transmission at raw measurement accuracy. If a source transmits twice without changing position, then the clock skew of the sensors that sense it can be estimated directly at the accuracy of the raw measurement. Incorporating such measurements into the system model can increase the system accuracy. A source with known location can also be a sensor that is allowed to transmit. Periodic sources (such as sonic transmitting satellites) usually have a transmission period that is very short, so that adjacent transmissions have essentially the same position,
As noted earlier, the MVIE estimate together with its accuracy estimation can also be used as a priori probability of the unknowns in a MAP estimator.
The MVIE of the normalized clock offset can be estimated at a few points in time to increase accuracy if needed.
Even without assuming that all sources are received by all sensors, the above algorithm will still perform well with only slight modifications.
In the description that follows, system 20 is assumed (with full generality) to comprise a set of M sensors 22 at given locations {qi}i=0M−1, a set of Ns satellites 26 in orbit of unknown position {pi(s)}i=0N−1, and a set of Nt terrestrial sources 24 of unknown position {pi(t)}i=0NNt+Ns, and set the origin of the coordinate system at the center of the sensor network deployment area. The orientation of the coordinate system is set to North West Up.
The time offset of sensor s is δ(s). Therefore, the TOA measurement of source t (satellite or terrestrial in origin) by sensor s is
mt(s)=rt(s)+τt+δ(s)+et(s), (4.1)
wherein τt is the unknown transmission time of the signal, rt(s) is the time of flight from source t to sensor s, and et(s) is zero mean Gaussian error whose standard deviation is σt(s).
By defining the vectors:
the measurements of source t can be written in vector form as:
mt=rt+1Mτt+δ+et. (4.2)
We use the same notation for both terrestrial and satellite sources, wherein the indexes t=0 . . . Nt−1 are used for terrestrial sources and the indexes t=Nt . . . Nt+Ns−1 are used for satellite sources.
For the sake of simplicity in the formulation we make the assumptions:
1. All of the terrestrial sources are intercepted by all of the sensors with the same error standard deviation, σT,
2. All of the satellites are intercepted by all of the sensors with the same error standard deviation σS.
The problem to be solved is twofold:
1. Estimate the sensors clock offsets.
2. Estimate the source location using the available measurements.
For satellite signals the range rt(s) can be written as:
Therefore the vector rt is approximated by
wherein
and
This means that:
G is referred to as the geometry matrix. This approximation of the projected range vector is the reason that the satellites location can be described by two variables. For example, longitude and latitude, instead of three that are needed for a general point in space.
The second-order term in the Taylor expansion determines the approximation error. As an example, for GPS satellites at an orbital height of 20180 km, and a sensor network of 1 km diameter, the error term maximum value is:
In terms of clock synchronization, this is equivalent to an error of 18 ps. Looking at the approximation from the opposite direction, we may ask what is the maximal radius of the sensor network such that the approximation error is less then 1 m, for the same satellite orbital radius. We calculate:
Neglecting the second order term is possible, as long as the possible deployment area of the network is limited. The results can be extended to wider-area networks, given crude information on the orbital radius of the satellites.
Returning to eq. (4.7), the domain of the vector {tilde over (r)}t is part of a subspace of M. Due to the structure of G, {tilde over (r)}t resides in a two- or three-dimensional subspace of
M. The subspace will be two-dimensional for planar sensor configuration, and three-dimensional in the general case.
For ground-based sensor configurations, the domain of
wherein
The vector {tilde over (r)}t is a linear transformation of 3 onto
M, which is a half ellipsoid in
M. Singular Value Decomposition (SVD) of −BG reveals the exact shape of that half ellipsoid. We define the SVD of −BG as:
wherein U and V are unitary square matrixes of size M×M, and 3×3 respectively; ΛR is the non negative singular value diagonal matrix of size 3×3 at most; R rows hold the left singular vectors related to the nonzero singular values; and the rest of the left singular vectors form the rows of H. The nonzero singular values define the lengths of the semi-axis of the ellipsoid, while the related left-singular vectors define the semi-axis of the ellipsoid.
When the sensor formation is planar, there are only two non-zero singular values in ΛR. In this case the domain of {tilde over (r)}t is a two-dimensional full ellipsoidal disk in M.
As stated earlier for three-dimensional ground sensor formations, the domain of
This plane is not guaranteed to be aligned with the semi-axis of the ellipsoid. When the height variation of the sensor configuration is much smaller than the planar position variations, however, the plane normal is approximately aligned with the semi-axis of the ellipsoid.
In the above description, there was an implicit condition on the number of sensors. If there are only two sensors, then rank (BG)=1, and the domain of BG
The ML estimator for terrestrial sources without clock skew was derived above for the localization of a general source using TOA measurements. The cost function to minimize in that case was:
where rt is the vector of distances of the source t to all sensors, and τt is the unknown source transmission time. The configuration containing satellites and terrestrial sources of unknown position and transmission time fits this cost function. Assuming the satellite and terrestrial sources are measured with the same error statistics yields:
where rt(xt,yt,zt) is the distance vector for a terrestrial source located at pt[xt,yt,zt]T, and rt(ϕt,θt) is the distance vector for a satellite source at azimuth θt and elevation ϕt. The transmission times τt that minimize the cost function are given by:
We now define:
Substituting the transmission time estimates into the cost function and using these definitions gives:
Collecting all vectors for t=0,1, . . . ,NS+NT−1 now gives:
Q′=∥m−r−Cδ∥2, (4.18)
where
Substituting the clock offsets into the cost function gives:
Calculating the elements of r is straightforward for terrestrial sources. For satellite sources, we can use the fact that they are far away to write:
where R is the orbital radius of the satellite sources. Thus, finding the ML estimate requires a search in 3NT+2NS dimensions: three position coordinates for each terrestrial source and two bearing angles for each celestial source. As noted above, for ground-based sensor formations, the satellite directions are above the horizon. Therefore:
This criterion limits the search interval.
To avoid a brute-force search over so many dimensions, an estimate of {tilde over (δ)} can be used to decouple the problems of estimating the terrestrial source positions and satellite directions. The following iterative procedure can then be used to solve the ML estimator:
If the computation is found to have converged (step 56 in
For every iteration of this algorithm, controller 28 performs NS+NT nonlinear least square optimizations. To avoid convergence of the cost function to a local minimum (rather than the true minimum corresponding to the actual clock offsets and source positions), a good initial estimate of the source locations and/or clock offsets is useful. Such an estimate can reduce the computational load considerably by reducing the number of iterations required for convergence. in tracking applications, a previous estimate may be used as a good starting point, but in the absence of a previous estimate, other information may be used.
The following sections describe methods that can be used to find an initial point so that no more than a few iterations are needed. Moreover, given a good initial point, controller 28 can solve the original cost function using gradient methods, which are very effective.
In the iterative method of
We define:
yBδ−Bmt
ABG. (4.24)
Then the Lagrangian is:
Equating
to zero gives:
(aTA−Iλ)
Substituting the solution for
and equating to zero gives:
yTA(ATA−Iλ)−1
ATA can be represented by SVD as:
ATA=VTΛV
yTAVT(Λ−Iλ)−1(Λ−Iλ)−1VATy=1. (4.28)
Noting that Λdiag (γ0,γ1,γ2), and defining VATy
α=[α0,α1,α2]T we get:
This equation is equivalent to finding the roots of a sixth degree polynomial. An efficient numerical solution for the roots can be found by seeking the eigenvalues of the companion matrix of the polynomial. One way to reduce the complexity in runtime is by realizing that the γ values are constant as long as the sensors are static, they are needed only once, at initialization.
After calculating the six possible values of λopt and discarding any complex solutions, we place them the equation for
A weight matrix W may be used in the original problem:
To solve the problem, y is replaced with Wy and A with WA. The use of the weight matrix extends the results to include correlated measurements and can be applied to reduce the dimension of the measurement vector, as explained below.
This section shows that the log likelihood cost function for signals from space can be divided into two separate cost functions. One is linear, while the other is nonlinear but is only three-dimensional. Controller 28 may decompose the cost function into such components in order to solve the problems of clock synchronization and satellite source direction more efficiently.
H and R are defined above as the null and range-space projectors of BG
T[HT,RT]T. (4.31)
The matrix T is a unitary matrix, since it is a square matrix with orthonormal rows. Multiplying the source measurements mt by T does not reduce the information in the measurements. Proceeding with the measurement of the satellite signals gives:
Tmt=Trt+T1Mτt+Tδ+Tet (4.32)
Solving the ML estimator then requires that controller 28 minimize the following cost function:
Transmission time is estimated as usual resulting in:
since HBrt=0 by definition of H.
We now define:
δ⊥HBδ
δRRBδ. (4.35)
Therefore:
The two parts of the cost Q can be solved independently, thereby making the ML estimator simpler to compute. Because the measurement vectors HBmt and RBmt are uncorrelated, the estimation error covariance matrix is block diagonal. The ML estimators related to the two cost functions can thus be computed separately, as described below.
The ML estimator of δ⊥ can be derived from the partial cost function:
For this purpose, we define:
wherein ⊗ is the Kronecker Product. Now we can write:
Q⊥=∥m⊥−C⊥δ⊥∥2. (4.39)
The estimator of δ⊥ that minimizes the cost function is:
It can be shown that HBHT is idempotent, and H can be chosen such that:
Furthermore, regardless of the source locations, the estimation accuracy depends only on the number of signals measured. The positions of sensors 22 also do not affect the estimation accuracy, as long as the sensors are confined to an area such that the linear approximation of the satellite ranges holds.
A single source is sufficient to establish an estimation of the null space projection of the clock offsets (although this is not the case for the range space). Estimating the clock offsets in null space is sufficient to localize terrestrial sources with reasonable accuracy. For a non-planar sensor configuration, sources confined to a plane can be localized, if at least six sensors observe the signals, and for sources that are not in a plane at least seven sensors are required. For a planar sensor configuration, sources confined to a plane can be localized if at least five sensors observe the signals, whereas for sources that are not confined to a plane, at least six sensors are required.
The fact that we can choose H such that HBHT has zero off-diagonal elements can speed up the estimation process.
For satellite signals, the nonlinear part of the cost function can be expressed as follows:
Controller 28 can estimate δR in the presence of RBrt as a nuisance parameter. In this calculation, since R=RB, the measurement vector RBmi elements are independent and identically distributed Gaussian random variables. The mode of calculation depends on the sensor formation: If the sensors are confined to a plane, then RBrt can be any value inside an ellipse, whereas if the sensor distribution is in 3D space, RBrt is distributed on a 3D ellipsoidal half shell, which enables better accuracy.
First we consider the case of a planar sensor configuration. In this case the problem of estimating δR is reduced to finding the possible centers of an ellipse of known dimensions given a set of points in the ellipse domain corrupted by noise. The set of possible centers can be found, for example, by calculating the mean of the points (which is guaranteed to be inside of the ellipse) and then expanding the set of possible points by propagating out from that point. Controller 28 can determine whether a point is in the domain of possible offsets by checking whether an elliptical disk with its center at that point include all of the measured points in its domain. Regardless of how the set of points is found, its mean can be written as {circumflex over (δ)}R, and its empirical covariance matrix as ΣR.
For the case of a non-planar sensor configuration, because RBrt resides on a 3D ellipsoidal half-shell, estimating {circumflex over (δ)}R is reduced to finding the ellipsoid center given a set of points that are located on the ellipsoid half shell in the noiseless case, or are near this shell in the noisy case. This problem can be solved in a number of ways, including high-dimensional ML estimation, low-dimensional iterative ML estimation, and ellipsoid surface fitting.
The high-dimensional ML estimator takes the range space cost function and minimizes it for all satellite directions
The low-dimensional iterative ML estimator is similar to the iterative ML algorithm described above, but is applied only to minimize the cost function with respect to satellite directions and a 3D δR vector. Each iteration solves two or three dimensions at a time and is inherently fast. This method uses a two-step iterative process: Until convergence do ML Estimation of δR given the current estimation of
The surface fitting method uses the fact that an ellipsoid center is to be fitted to noisy points, which is again a 3D problem with fast convergence properties. It is based generally on techniques described by Ahn et al., in “Orthogonal Distance Fitting of Implicit Curves and Surfaces,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):620-638 (2002), which is incorporated herein by reference. This fitting method is a good approximation of the ML estimator and is relatively simple to apply in the present case since the ellipsoid to be fitted is of known shape and orientation, so the only parameters left to fit are the ellipsoid center coordinates.
The problem of 3D ellipsoid fitting can be trained as follows: Given the set of noisy points {RBmt}t=0N
C(a)=dTPTPd
d=[d1,d2, . . . ,dN
di=∥xi−x′i(a)∥ (4.43)
Here a is the proposed ellipsoid center, x is the noisy point, x′ is the closest point to x that is on the ellipsoid centered at a, PTP is a weighting matrix, and di is the Euclidean distance between xi and x′i. It is easier to consider x, x′ in a coordinate system that is centered at a and aligned with the eigenvectors given by the rows of R. In that coordinate system the ellipsoid is given by the following implicit function:
wherein x′=[x′1,x′2,x′3]T, and λi i=1,2,3 are the non-zero eigenvalues of BGGTBT, related to the eigenvectors defined by the rows of R.
The ellipsoid center estimation may be carried out using the following orthogonal distance lining algorithm to find be ellipsoid center â:
End While
In step 2.2, x′ can be found using a generalized Newton-Raphson method, starting from the initial guess x′0=x:
wherein:
In step 2.2, Δa is found using:
JTPTPJΔa=−JTPTPd
Δa=−(JTPTPJ)−1JTPTPd|k
âk+1=âk+ηkΔa (4.47)
wherein
and η∈(0,1) is a constant that helps convergence in noisy situations. In simulations, the inventors found it useful to set η=0.9.
In step 3, the covariance is found using
{circumflex over (δ)}R=â
ΣR=cov({circumflex over (δ)}R)=σS2(JTJ)−1. (4.48)
The above algorithm for ellipsoid center estimation requires inverting matrices only of size 3×3. Because this algorithm uses the steepest descent at the ellipsoid center update, as defined by Δa, whereas the previous algorithm uses ML estimation of the center under the assumption that the satellites directions are known, the present algorithm may converge in fewer steps, although each step is a bit more complex.
Regardless of the method used to estimate the range space, it is important to start the iterative process with a good initialization of â (or {circumflex over (δ)}R) in order to avoid reaching a local minimum. The points to be fitted, assuming a small sensor height variation, reside on the upper or lower half of an ellipsoid, and the sign of −RBG·
By combining the above estimates for the null space and range space of the clock offset vector Bδ, controller 28 can now establish full synchronization of sensors 22 using satellite sources 26. For this purpose, the estimations of the range and null projections may be treated as the measurements:
The error elements nH,nR have the covariance matrices Σ{circumflex over (δ)}
wherein Σblkdiag(Σ{circumflex over (δ)}
To estimate the directions of the satellite sources, is inserted into the ML cost function:
Qt∥B(mt−rt)−
∥2, (4.52)
and controller 28 finds the solution that minimizes the cost for every satellite t. Alternatively, the Lagrange multipliers described above may be used for this purpose.
Thus, controller 28 is able to synchronize a set of sensors 22 deployed over a wide area using satellites having unknown orbital parameters and time references. The directions to the satellites are estimated as a by-product. The accuracy of the synchronization is only limited by the radio bandwidth, signal/noise ratio (SNR), observation time, and the number of sources processed.
Given only the null-space offset projection {circumflex over (δ)}⊥, controller 28 can establish synchronized time-based localization of terrestrial sources 24. This approach can be useful since estimating {circumflex over (δ)}R is more demanding, and cannot be accomplished by the above methods with fewer than three celestial signals. Moreover, this sort of localization enables estimation of the full clock offset vector Bδ, as explained below.
Although HBrt=0 for celestial sources, this is not the case for terrestrial sources 24. Therefore, if there is a sufficient number of sensors 22 receiving terrestrial signals, it is possible to get a synchronized localization of the terrestrial sources by solving:
HBmt=HBrt+HBδ+HBet
HBmt−{circumflex over (δ)}⊥=HBrt+e{circumflex over (δ)}
wherein mt is now the vector of TOA measurements of terrestrial source t, and e{circumflex over (δ)}
Since the error e{circumflex over (δ)}
Alternatively, controller 28 may locate terrestrial sources 24 using the fully-estimated offset vector and the terrestrial source sensor measurements:
Bmt=Brt+Bδ+Bet (4.55)
In this case, pt is found by minimizing:
argminp)T{tilde over (Σ)}t†(Bmt−Brt−
)
{tilde over (Σ)}t=TTΣT+BσT2 (4.56)
The expected estimation error under linear approximation is given by:
cov(pt)=(LtT{tilde over (Σ)}t†Lt)−1 (4.57)
This estimation will thus provide better localization performance than the solution based only on null-space offset projection.
The accuracy of the clock offset estimation for sensors 22 may be improved by considering the localization of terrestrial sources 24. After refining the clock offsets, controller 28 can improve the localizations of the terrestrial sources and continue iterating to convergence. Convergence is assured, as improving the clock offset estimation accuracy increases the accuracy of localization of the terrestrial sources and of the directions of the satellite sources, and vice versa.
The ML estimator for Bδ may be derived in this case as follows: The error of {circumflex over (p)}t is a vector in two or three dimensions, whereas {tilde over ({circumflex over (r)})} is a vector in M dimensions. We define the error vectors:
eppt−{circumflex over (p)}t
e{tilde over (r)}{tilde over (r)}t−{tilde over ({circumflex over (r)})}t (4.58)
There is a linear relationship between the error vectors, assuming the error is small:
e{tilde over (r)}
The error covariance matrix of {tilde over (r)}t is given by:
cov({tilde over ({circumflex over (r)})}t)=Ltcov(pt)LtT, (4.60)
which is clearly singular and of the same rank as cov(pt) (which is rank 2 or 3 depending on the sources distribution in a plane or in three dimensional space). Eigenvalue decomposition of cov({tilde over ({circumflex over (r)})}t) gives:
cov({tilde over ({circumflex over (r)})}t)=UtΛtUtT
Ut=[RtT,HtT]
Λt=diag(λt
RtRtT=I
HtHtT=I, (4.61)
where the rows of Rt are the eigenvectors of cov({tilde over ({circumflex over (r)})}t) corresponding to the non-zero eigenvalues, while the rows of Ht are the eigenvectors corresponding to the zero eigenvalue, and λt
Now we calculate the covariance of Hte{tilde over (r)}
This expression means that the first two moments of the projected error Hte{tilde over (r)}
Hte{tilde over (r)}
Since pt is unknown, the evaluation of Lt and cov(pt) is performed using the estimate of pt.
Coming back to the source measurement equation, multiplying from the left by Ht gives:
HtBmt=Ht{tilde over (r)}t+HtBδ+HtBet (4.64)
This equation is modified to compensate by {tilde over ({circumflex over (r)})}t:
The result is a linear measurement of a projection of the sensor clock offsets vector. Concatenating measurements from all of the terrestrial sources gives:
which is completely independent of the localization errors.
Writing all of the equations we have on Bδ at this point gives:
The ML estimator is established using WLS with the following definitions:
enables writing the estimator for Bδ as:
A similar formulation can be applied when the range space measurements are not used as part of the estimation process. In this case, the range measurements are simply omitted from the WLS formulation, and the position error covariance is modified accordingly. In the case of low SNR, performing a few iterations of this process provides a solid estimate even though the assumption in Eq. (4.63) does not hold perfectly at first.
At this point the only information that is left unused to improve the estimate of the clock offsets is the terrestrial measurements RtBmt. These measurements can be expressed as:
Rt(Bmt−{tilde over ({circumflex over (r)})}t)=RtBδ+RtBet+Rte{tilde over (r)}
This equation is solved iteratively, as the uncertainty in the values of the clock offsets is dependent on the uncertainty in the terrestrial source position, which is again dependent on the uncertainty in the clock offsets. Therefore, it is advantageous first to re-estimate the terrestrial source locations using the refined clock offsets. It is simpler to consider the original measurement for the second iteration:
Bmt−{tilde over ({circumflex over (r)})}t=Bδ+Bet+e{tilde over (r)}
It is a good approximation to say that these measurements are relatively uncorrelated to the clock offset estimations derived from the satellites signals, Since the little correlation that exists originates from the current position error, which is only partially correlated to the clock offset measurements derived from the satellites signals.
An expression may be developed as follows for the covariance matrix of:
mT[(Bm0−{tilde over ({circumflex over (r)})}0)T,(Bm1−{tilde over ({circumflex over (r)})}1)T, . . . ,(BmN
For this purpose, the error element may be expressed as follows:
Now we can write the covariance matrix of mT:
cov(mT)ΣT
(ΣT)i,i=σT2(I+{tilde over (L)}i)T+{tilde over (L)}iΣBδ{tilde over (L)}iT
(ΣT)i,j={tilde over (L)}iΣBδ{tilde over (L)}Tj. (4.71)
Combining all measurements of the sensor clock offsets finally gives:
The optimal estimator is established using WLS. Defining the following:
yf[mTT,{circumflex over (δ)}⊥
Af[(1N
Σfblkdiag(ΣT,Σ⊥,ΣR)
the estimator is given by:=(AfTΣf†Af)†AfTΣf†yf
cov()=(AfTΣf†Af)†.
After estimating the clock offsets, controller 28 may again re-estimate the locations of terrestrial sources 24, and repeat until convergence as described above.
in some embodiments it is possible to skip the steps that use RBmt and return, or alternately do the final refinement using gradient methods on the original cost function and measurement vector.
Based on the above estimators, it can be shown that even with only one satellite source 26, it is possible to achieve accurate synchronization and localization as long as there, is a terrestrial source 24 that is received by at least six or seven sensors (depending on whether they are distributed in a plane or in 3D space). In other words, for any combination of satellite sources 26 and terrestrial sources 24, as long as they are received by sufficient number of sensors 22, the estimators defined above are well posed and have performance directly related to the measurement error.
The problem of navigation of a moving object in system 20 is equivalent to assuming that a sensor of unknown position is added to the network. This addition does not change the form of the measurement model. Considering only celestial sources, for example, the only difference is that now one row of the G matrix is unknown, and another clock offset is added.
Defining sT as the unknown sensor position and δs as its clock offset, the estimator may be written as:
Assuming 1Tδ=0, as is expected from the estimation process of the clock offsets, leads to:
Rearranging the equation gives:
Collecting the equations for all t gives:
wherein:
Γ({[Γ0T,Γ1T, . . . ,ΓN
A({[A0T,A1T, . . . ,AN
Substituting for the satellite directions and the sensor clock offsets with their estimates (based on the methods described above without the use of the sensor of unknown position) gives:
The navigation problem is thus reduced to a weighed least squares estimation of the values:
This estimation can be used as an initial point for the minimization of the full ML cost presented in Eq. (4.13), with the only difference being that the position of one of the sensors is unknown. Starting from the initial estimate given by Eq. (4.75), controller 28 can use gradient methods to find the optimal point in terms of ML cost.
The problem framed above can be extended to the case of multiple sensors of unknown position by solving the LS estimation for each of these sensors. The ML estimator can use these initial estimates to find the optimum of the full ML cost, and in so doing also refine the estimates of the known sensors clock offsets and the satellite directions. Terrestrial sources can also be incorporated into the ML cost function.
There are two possible modes of navigation in system 70:
a. Standalone mode: The measurements of sensor 76 are used only to resolve its location and time offset. This mode is linear in the sensor position and clock offset, and therefore fast, and could be carried out by a processor (not shown) on board airplane 74.
b. Joint mode: The measurements of sensor 76 are also used to improve all other system estimates, including the positions and time estimates of other airplane sensor.
The arrangement shown in
Trading time is of great importance in such systems, and time-stamping is regulated by law. In systems that rely on GPS-based synchronization, if the GPS receiver is jammed and accurate time stamping is not available, trading must stop until it is recovered. Spoofing of the trading system time could cause trades to be recorded at a loss.
In contrast to existing solutions, servers 82 use sensors 22 to implement spoof-proof and jamming-proof time-stamping, based on the methods of clock synchronization that are described above. As time transfer is enabled without the use of GNSS, the absolute time can be maintained by all servers 82, which are unlikely to be GNSS-jammed at the same time, or the servers as a group could perform time holdover until the total jamming has passed. System 80 is thus able to maintain normal operation even under jamming attacks and to detect and reject attempts at source spoofing quickly and robustly.
The precise synchronization that is achieved in system 90 can be used in various ways to improve network performance, such as reducing interference, increasing network capacity or bits per Hz, improving resource allocation, enabling Cooperative Multipoint (CoMP) transmission in small and macrocell deployments, and supporting time-based localization techniques such as OTDOA (observed time difference of arrival), and UTDOA.
Assuming system 90 operates in accordance with the Long-Term Evolution (LTE) family of standards, clock synchronization in the system can be implemented as follows:
Although the above example, relates specifically to LTE, other mobile network technologies (such as UMTS) that provide OTDOA-like measurements can use the methods described herein.
One such measurement type is UTDOA (Uplink TDOA), which is performed by dedicated sensors called LMUs (Location Measurement Units). The LMU can be an integral part of the base station or a separate unit, installed on the base station or in another position. When using UTDOA measurements, the emitters are the user equipment (UE), and the LMUs are considered the sensors of known position. Using the methods described herein, the LMUs clocks are synchronized, while the UE locations are also resolved. Moreover, if the LMU clock is correlated to the base station clock, then the base stations and LMUs are also synchronized, enabling all of the benefits stated above and eliminating the need for a dedicated external synchronization unit in each LMU.
In another embodiment (not shown explicitly in the figures), sensors are used to gather information on the environment, and the locations of the sensors need to be determined (as in the case of Real Time Location Services, for example). These environmental sensors may be regarded as “sources” in the context of this embodiment, and localization of these sensors may be achieved based solely on the signals that they emit. The sensor measurements of these signals constitute the information gathered by the sensors that is relayed to a central node, so that no special transmissions are required for achieving sensor localization.
In such a network, in other words, there are two types of sensors:
1. Localization sensors 22 (which measure TOA), preferably of known position.
2. General sensing, sensors, which gather sensing information on their environment (such as temperature, vibration, humidity, etc.) and serve as sources 24.
The synchronization that is achieved among the type 1 sensors can be used to efficiently manage the network.
In some embodiments, synchronization may be the major interest and not localization. This situation can arise when, for example, a wireless network needs to manage wakeup times of its nodes. An embodiment of the present invention provides a good solution for such cases, since emitters are usually present in any environment. Thus, even if only pairwise distances of the sensors are known, synchronization can be achieved to some extent.
Additionally or alternatively, the techniques described herein can be used to check whether an existing system is synchronized and calibrated properly with respect to time offsets and skews, i.e., to detect the occurrence of synchronization faults. If the clock offsets and skews derived using the methods described above do not match those in the existing system, there may be a malfunction in the hardware or software of the existing system. In this case, the time offset and skew of the system can be recalibrated based on the synchronization described by the above methods, and this synchronization can also be used to provide a backup clock or even a set of such clocks for the system, thus affording maximal availability and reliability.
Although the embodiments disclosed above relate specifically to problems of sensor synchronization and emitter localization (along with navigation and spoofing detection), other problems that can be formulated in an equivalent manner can be also solved by the methods that are described herein. For example, the principles of the techniques described in the present patent application may be extended to equivalent embodiments that arise from the time/frequency duality between DTOA measurements and differential Doppler measurements. All such solutions are considered to be within the scope of the present invention.
It will thus be appreciated that the embodiments described above are cited by way of example, and that the present invention is not limited to what has been particularly shown and described hereinabove. Rather, the scope of the present invention includes both combinations and subcombinations of the various features described hereinabove, as well as variations and modifications thereof which would occur to persons skilled in the art upon reading the foregoing description and which are not disclosed in the prior art.
It is the intent of the Applicant(s) that all publications, patents and patent applications referred to in this specification are to be incorporated in their entirety by reference into the specification, as if each individual publication, patent or patent application was specifically and individually noted when referenced that it is to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present invention. To the extent that section headings are used, they should not be construed as necessarily limiting. In addition, any priority document(s) of this application is/are hereby incorporated herein by reference in its/their entirely.
This application is a continuation of U.S. patent application Ser. No. 14/388,808 filed on Sep. 28, 2014, which is a National Phase of PCT Patent Application No. PCT/IB2013/052020 having International Filing Date of Mar. 14, 2013, which claims the benefit of priority of U.S. Provisional Patent Application Nos. 61/617,186 filed on Mar. 29, 2012, and 61/755,485 filed on Jan. 23, 2013. The contents of the above applications are all incorporated by reference as if fully set forth herein in their entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 20060187866 | Werb | Aug 2006 | A1 |
| 20110185216 | Zhao | Jul 2011 | A1 |
| 20140016621 | Zhang | Jan 2014 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20220039036 A1 | Feb 2022 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61755485 | Jan 2013 | US | |
| 61617186 | Mar 2012 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 14388808 | US | |
| Child | 17492773 | US |
