1. Field of the Invention
The present invention relates to a method of locating a mobile telephone unit within a cellular service area, and more particularly to a method of predicting the location of a CDMA mobile unit based upon the probability of its being at a particular location of the service area using an algorithm providing a likelihood estimation of the mobile unit's location in response to a sequential set of attributes observed by the mobile unit and reported back to a base station.
2. Description of Related Art
A cellular telephone system must be able to locate a mobile unit within a cellular service area under various RF propagation conditions such, for example, when an E911 call is made from the mobile unit. Conventional methods for locating a mobile unit are typically based on either a triangulation technique which requires signals from three or more base stations within a designated service area, or an angle of arrival technique which requires at least two base stations. In many areas, the number of base stations the mobile unit can detect is less than two. Furthermore, both the triangulation and angle of arrival techniques inherently suffer from inaccuracies and signal fading which result from multi-path propagation.
In the above-noted related patent application U.S. Pat. No. 6,496,701 entitled “Pattern Recognition-Based Geolocation”, RF characteristics pertaining to one or more pilot signals radiated from a base station and specific to a particular location within the service area are detected by a mobile unit and transmitted back to a base station where they are matched to a known set of RF characteristics and other information obtained from making attribute information measurements at all the grid points (sub-cells) in a cellular service area and which are then stored in a database located, for example, in a base station server.
In the above-noted related patent application U.S. Ser. No. 09/294,997 entitled “A Bayesian-Update Based Location Prediction Method For CDMA systems”, the invention is directed to a method of estimating, by a Bayesian probability algorithm, the location of a mobile unit in the service area of a CDMA cellular telephone system using a model based approach which, among other things, simplifies the generation of a database containing a pilot signal visibility probabilities. This eliminates the need for repeated attribute measurements at all of the grid points in the service area.
In the above-noted related patent U.S. Pat. No. 6,263,608 entitled “Geolocation Estimation Method For CDMA Terminals Based On Pilot Strength Measurements”, the invention is directed to a method of estimating the location of a mobile unit in the service area of a CDMA cellular telephone system also using a model based approach, but which now eliminates the need for a stored database containing pilot signal visibility probabilities for all of the grid points or sub-cells in the cellular service area. The estimation procedure is based entirely on analytical results involving one or more key approximations derived, for example, from an integrated model of the wireless communications system, its RF environment, and attribute measurement.
The subject invention is directed to predicting the location of a mobile wireless communication unit in the service area of a CDMA communications system utilizing two likelihood functions that define maximum likelihood estimators of the mobile unit's location, based on attribute measurements, such as but not limited to pilot signal strength, being made at the location of the mobile unit and reported back to a base station. One of the likelihood functions comprises a frequentist likelihood function and the other comprises a Bayesian-modified likelihood function. The likelihood functions are based on the assumption that there is an RF model which provides the probability a mobile unit is able to detect one or more attributes associated with an arbitrary base station, given it is located at an arbitrary location within the service area. The frequentist likelihood assumes the RF model provides exact visibility probabilities. In contrast, the Bayesian-modified likelihood assumes the RF model only provides reasonable approximations to the true visibility probabilities, and uses the approximations to construct a Bayesian prior distribution for the true values. Each of the likelihoods can be used in an iterative fashion to produce a maximum likelihood estimator for the location of the mobile unit by determining the coordinates within the service area which maximize the respective likelihood function. Alternatively, or in addition to, each of the likelihoods can be incorporated into a sequential Bayesian procedure which outputs a posterior distribution for the location of the mobile unit.
Referring now to the drawings and more particularly to
Turning attention now to
In the invention described in the referenced related patent, U.S. Pat. No. 6,496,701 entitled “Pattern Recognition-Based Geolocation,” each sub-cell 181, . . . , 18n of the service area 10 is identified by a set of observable characteristics which are referred to as attributes. Examples of attributes are pilot signal strengths (Ec/Io), phase-offsets, angles of arrival, and pilot round trip delays. The invention of U.S. Pat. No. 6,496,701 includes a database which contains attribute information which differentiates one sub-cell 18 from another and is generated by making a repeated and exhaustive survey which involves taking repeated measurements at all the sub-cells 181, . . . , 18n (
During the operation phase, after the database has been set up and the location service has been deployed, the mobile unit 20 detects and measures attribute values from its actual location in sub-cell 18i and reports them via a message, e.g., a pilot signal strength measurement message (PSMM), to the base station(s) 141, . . . , 14n (FIG. 3), which can be one or more of the base stations with which it is in communication. The base station(s) forward their respective reported measurements to the geolocation server 22. The digital computer apparatus 23 associated with the server 22 statistically compares the measured values with the known attribute values stored in the database (memory) 24 of all the sub-cells 18 in the service area 10. The sub-cell 18i whose attribute values as stored in the database provide the best match for the measurements reported by the mobile unit 20 is considered to be the best estimate of the mobile unit's location.
In the above-referenced related application, Ser. No. 09/294,997, entitled “A Bayesian-Update Based Location Prediction Method For CDMA systems”, a database is also used to assist the process of location estimation. However, in contrast to the first referenced patent application, i.e. U.S. Pat. No. 6,496,701, it uses a model based approach to generate a database containing pilot visibility probabilities for different sub-cells 18 in the service area 10. The model-based approach requires that a limited number of pilot strength measurements be carried out along a few representative routes in the service area 10. These measurements are then used to identify the parameters of the model that characterizes the service area and its RF environment. Once these parameters are identified, simulations are then carried out to populate the database containing the pilot visibility probabilities, which are used in the computation of the location distribution of a mobile unit requesting location service. An iterative procedure based on a Bayesian probability computation is then used to obtain improved estimates of the mobile unit's location in response to multiple sets of attribute measurements being reported by the mobile unit 20. The model-based approach eliminates the need to carry out extensive measurements required by the first named invention, U.S. Pat. No. 6,496,701.
In the above-referenced related patent U.S. Pat. No. 6,263,208, entitled “Geolocation Estimation Method For CDMA Terminals Based On Pilot Strength Measurements”, the model-based approach embodied in Ser. No. 09/294,997, “A Bayesian-Update Based Location Prediction Method . . . ” to characterize the RF environment is used, as is the iterative procedure for computing the Bayesian posterior distribution for the location of the mobile. However, the database containing pilot visibility probabilities is replaced by analytical formulas that can be evaluated in real time. The evaluation procedures are compact and can typically be evaluated in the digital computer apparatus 23 shown in FIG. 3.
Considering the present invention, the analytic formulation for the pilot visibility probabilities taught in the above-referenced patent, U.S. Pat. No. 6,263,208, “Geolocation Estimation Method For CDMA terminals Based On Pilot Strength Measurements”, now serve as the starting point for the derivation of two likelihood functions, hereafter referred to as the frequentist and Bayes-modified likelihood functions, respectively. Each of the likelihood functions is derived based on the assumptions and mathematical formulations described in attached Appendix A. In as much as the likelihood functions depend on the analytic evaluation of the pilot visibility probabilities, attached Appendix B provides a self-contained development of the relevant details of these formulas. Each likelihood function is a function of (x,y), an arbitrary location of the mobile unit 20 in the x and y grid shown in FIG. 2. Accordingly, each likelihood function is used in a first method to obtain a maximum likelihood (ML) estimator of the location of the mobile unit 20 by finding the (x,y) coordinates which maximizes the value of the respective likelihood function. An iterative technique for sequentially updating each ML estimator with additional pilot signal strength measurements is utilized. In a second method, each of the two likelihood functions are also incorporated into a sequential Bayesian procedure, which outputs a posterior distribution for the location of the mobile unit.
The Bayes-modified likelihood function, whether it is used in the context of ML estimation or a sequential Bayesian procedure, is a substantial deviation from the inventions disclosed in the second and third above referenced related applications in the following way. Both of these previously disclosed inventions use an RF model to estimate pilot visibility probabilities, the former via simulation techniques, the latter via analytical formula evaluation, and implicitly assume that the model holds precisely. The present invention, however, uses the same RF model only to determine the means of beta distributions that are used as Bayesian priors for the true (unknown) pilot visibility probabilities. A beta distribution is completely determined once its mean and variance have been specified. Subject to the fixed mean values, each of the beta distributions is fully specified by maximizing their variances. Maximizing the variance of the beta priors, subject to the specified mean values, is consistent with a non-informative (vague) prior specification.
A summary of the derivations which appear in Appendix A will now be given as prefatory remarks to the description of the overall mobile unit 20 location prediction process depicted in FIG. 4. In the example shown in
For each (x,y)εA, let θij(x,y) denote the true probability that the mobile unit 20 is able to see the pilot in sector j of base station i when it is located at (x,y). From hereon, the notation ij will be used to exclusively reference pilots from the set K. Let {tilde over (θ)}ij(x,y) denote an approximation of θij(x,y) based on an RF model described in Appendex B. For each pilot ij, let μijs equal one or zero depending on whether the mobile unit 20 can see pilot ij at the s-th measurement epoch or not, respectively. The frequentist likelihood through the first s measurement epochs has the following recursive form, starting with the definition LML0(x,y)≡1:
For the Bayes-modified likelihood, the prior for θij(x,y) is a beta distribution with parameters:
The Bayes-modified likelihood function through the first s measurement epochs has the following recursive form starting with the definition LBML0(x,y)≡1 for all (x,y)εA:
where
is the number of times pilot ij was visible amongst the first s−1 measurement epochs.
Each of the likelihood functions (1) and (4) are functions of (x,y)εA, an arbitrary possible location for the mobile unit 20. The ML estimator for the location of the mobile unit 20 is obtained by evaluating (1) for all (x,y)εA and selecting the values, say (xMLs, yMLs), which gives the largest value of (1). An updated ML estimate is produced at each measurement epoch. In a similar way, using function (4) rather than function (1) generates a sequence of Bayes-modified ML estimates, say (xBMLs,yBMLs).
Utilizing functions (1) or (4) with a Bayesian sequential procedure can generate an alternative sequence of predictions for the location of the mobile unit 20. In each case, the initial prior distribution for the location of the mobile unit 20 is assumed to be a discrete uniform distribution of the form:
where ∥A∥ is the number of grid points 181, . . . , 18n contained within A. The posterior distribution for the location of the mobile unit 20, through s measurement epochs, based on the frequentist likelihood function (1) is, up to a constant of proportionality:
Alternatively, the posterior distribution of the location of the mobile unit 20, through s measurement epochs, based on the Bayes-modified likelihood function (4) is, up to a constant of proportionality:
A Bayesian sequence of predictions on where the mobile unit 20 is located follows from functions (6) or (7) by using the mean or mode of the posterior distribution obtained at each measurement epoch. When function (7) is used, the sequence of prediction involves two distinct prior distributions, beta and discrete uniform, and the methodology is referred to as doubly-Bayesian. This completes the summary of Appendix A.
A description of the mobile unit 20 location prediction process depicted in
The process, referred to hereinafter as the geolocation process, begins at step 30 (
In the example shown in
First, suppose path (2) of
Next, suppose path (3) of
Next, suppose path (4) of
Finally, suppose path (5) of FIG. 4A and further shown in
Each of the iterative procedures described above with respect to paths (2)-(5), operates to provide an improved estimate of the location of the mobile unit 20 as more and more sets of pilot measurements are reported. Each of the procedures can be readily extended to include other measured quantities such as phase offsets by including those data in the respective likelihood formulations. Also, when desirable, the method of the present invention can be modified so that the actual signal strength of the visible pilots can be used rather than a binary representation of the fact that the pilots are or are not visible.
The foregoing description of the preferred embodiment has been presented to illustrate the invention without intent to be exhaustive or to limit the invention to the form disclosed. In applying the invention, modification and variations can made by those skilled in the pertaining art without departing from the scope and spirit of the invention. It is intended that the scope of the invention be defined by the claims appended hereto, and their equivalents.
The notation used in this Appendix is the same as what has been introduced in the main portion of this application. In what follows, we make the following assumptions:
(i) The visibility of a particular pilot at (x,y)εA at a given time t is independent of whether the pilot was visible at any time s<t, and is independent of t as well.
(ii) The visibility of pilot i at time t is independent of the visibility of pilot j (j≠i) at time t.
(iii) Throughout the duration of the prediction process, the mobile is relatively stationary, i.e., the mobile unit remains in its initial sub-cell.
Defining L0(x,y)=1, ∀(x,y)εA, a recursive form for the exact likelihood function through the first s measurement epochs for the unknown location of the mobile based on (i)-(iii) is
The functions θij(x,y) in (A1) are unknown, so the exact likelihood function is not calculable. Replacing the θij(x,y) with their approximations {tilde over (θ)}ij(x,y), defined by equation (B9) in Appendix B, gives the definition of the frequentist likelihood function LMLs(x,y) shown in expression (1).
The frequentist likelihood function (1) was derived under the assumption that a family of functions {tilde over (θ)}ij(x,y) accurately approximate the unknown functions θij(x,y). The approximations {tilde over (θ)}ij(x,y) are based on the RF model described in Appendix B. The RF model is an approximation to a very complex stochastic process. Bayesian methods allow the uncertainty associated with {tilde over (θ)}ij(x,y) to be incorporated into the model framework. Rather than assuming {tilde over (θ)}ij(x,y)≡θij(x,y), we can model the parameters θij(x,y) with independent prior distributions that have mean values equal to {tilde over (θ)}ij(x,y). For each (x,y)εA, we use a prior distribution for θij(x,y) that is a beta distribution with probability density function given by
where B[·,·] is the complete beta function. Note that the two parameters, αij(x,y) and βij(x,y) to be determined in what follows, depend on both the possible location (x,y)εA and the pilot ijεK. Whereas the frequentist likelihood is defined by implicitly assuming Pr[θij(x,y)={tilde over (θ)}ij(x,y)]=1, the Bayes-modified likelihood is defined by taking the expected value of the exact likelihood (A1) with respect to the beta distributions defined by (A2). The recursive form of the exact likelihood (A1) does not amend itself to evaluating the expected value as easily as the non-recursive form of the exact likelihood which is
where nijs is the number of times pilot ij was visible through the first s measurement epochs. Evaluating the expected value of (A3), with respect to the beta distributions defined by (A2) gives the Bayes-modified likelihood function
It is easily verified that equation (A4) has the recursive form shown in the functional expression (4).
The definition of the Bayes-modified likelihood function will be complete once values for αij(x,y) and βij(x,y) are specified. Determining what values to use for αij(x,y) and βij(x,y) is an example of the classic Bayesian dilemma—how to specify a prior distribution? An underlying principle in specifying a prior distribution is that any available information about the parameter of interest, in this case θij(x,y), should be reflected in the prior distribution that is used. In this case, we have an approximation of θij(x,y) from the RF model, namely {tilde over (θ)}ij(x,y), which we will use as the mean of the prior distribution for θij(x,y). Accordingly, we have the following constraint:
Since we have two unknowns, αij(x,y) and βij(x,y), we need one additional constraint to uniquely determine both αij(x,y) and βij(x,y). To obtain the second constraint, we impose that the prior be as vague as possible, subject to the first constraint. To implement this second constraint, we will maximize the variance of the prior, subject to the fixed mean value. The second constraint is consistent with the quantity of the prior knowledge we have about θij(x,y) and except for anticipating a value near {tilde over (θ)}ij(x,y), we know nothing else about its potential value. We also restrict attention to beta distributions which have bounded density functions, implying αij(x,y)≧1 and βij(x,y)≧1. The restriction to bounded density functions only rules out “U-shaped and “Slide-shaped” beta distributions which would not typically be intuitive representations of the prior information on θij(x,y). The problem of selecting αij(x,y) and βij(x,y) has now been sufficiently constrained so that it can be solved by using the following Lemma.
Lemma
Consider the family of beta distribution with parameters a≧1 and b≧1. The values of a and b which maximize the variance, subject to the distribution having a given mean value μ, are: a=1 and b=(1−μ)/μ, if μ≦½; a=μ/(1−μ) and b=1, if μ>½.
Proof of Lemma
The variance of the beta distribution is given by σ2=ab/[(a+b+1)(a+b)2]. Since we have μ=a/(a+b), we can write b=a(1−μ)/μ, and hence σ2=μ2(1−μ)/(a+μ). To maximize the variance, we need to minimize a, subject to the constraint that both a≧1 and b≧1. Clearly, if μ≦½ then we can choose a=1 and b=(1−μ)/μ≧1 will be ensured. Likewise, if μ>½ then the smallest we can choose a and still have b≧1 is a=μ/(1−μ)≧1.
Applying the Lemma with a=αij(x,y), b=βij(x,y) and μ={tilde over (θ)}ij(x,y) gives the formulas in equations (2) and (3) for αij(x,y) and βij(x,y).
In this Appendix, we drive the approximations {tilde over (θ)}ij(x,y) of the pilot visibility probabilities. The notation used in this Appendix is the same as what has been introduced earlier in this specification. Consider a grid point (x,y)εA. As in the above-referenced related patent U.S. Pat. No. 6,263,208, entitled, “Geolocation Estimation Method For CDMA Terminals Based On Pilot Strength Measurements”, the RF power received by the mobile unit 20 from sector j of base station i is modeled by the expression
Rij(x,y)=TijGij(x,y)Lij(x,y)Fij(x,y)Mij(x,y) (B1)
where Tij is the transmit power associated with the sector, Gij(x,y) is the antenna gain for the sector along the direction pointing towards the location (x,y)εA, Lij(x,y) is the distance loss between the base station associated with the sector and the location (x,y)εA, Fij(x,y) is the shadow fading factor and Mij(x,y) is the measurement noise factor, all in absolute, not dB, units. The measurement noise factor is meant to include the effects of fast fading (e.g., Rayleigh/Rician) as well as inaccuracies in the measurement process. If γ denotes the fraction of Tij that is used for the pilot channel, then γRij(x,y) is the pilot channel power received by the mobile unit 20 when it is located at (x,y)εA.
The model assumes that the distance loss Lij(x,y) can be expressed as
Lij(x,y)=CP[dij(x,y)]−α (B2)
where dij(x,y) is the distance between the base station associated with the sector and the location (x,y)εA, and CP and α are constants. Typically, CP takes a value in the range 10−15 to 10−10 and α is between 3 and 5 when dij(x,y) is expressed in miles. The two parameters, CP and α, moreover are environment specific.
The shadow fading factor, Fij(x,y), models the impact of terrain and large structures (e.g., buildings) on signal propagation which create deviations around the signal attenuation predicted by the deterministic path loss factor. We assume that Fij(x,y) is lognormally distributed and thus can be written as
where φij(x,y) is a zero mean Gaussian random variable with standard deviation σφ.
The measurement noise factor, Mij(x,y) is also assumed to have a lognormal distribution so that we can write
where μij(x,y) is a zero mean Gaussian random variable with standard deviation σμ.
Assume now that the pilot strength measurement carried out and reported to the base station by the mobile unit 20 is the Ec/Io value of the corresponding pilot channel signal. This value is the ratio of the pilot channel power from the concerned sector received by the mobile unit 20 to the total power received by the mobile unit including thermal noise, and possibly external interference.
By letting Pij(x,y) denote the strength of the pilot channel associated with the sector as measured by the mobile unit 20 located at (x,y)εA, it follows that
where N0 denotes thermal noise and external interference and (as before) kl is shorthand notation for the pilot associated with sector l of base station k.
For convenience, we define
Cij(x,y)=TijGij(x,y)Lij(x,y). (B6)
Observe that the expected value of Rij(x,y) is equal to Cij(x,y). It is implicitly assumed that the shadow fading and measurement noise factors are uncorrelated. We approximate Pij(x,y) (B5) by the following expression
The difference between (B7) and (B5) is that for kl≠ij, the Rkl(x,y) quantities have been replaced by their mean values. Using (B3) and (B4), we can write
where
The probability that the mobile unit receives pilot ij with a signal strength in excess of T is
θij(x,y)=Pr[Pij(x,y)>T]
≅Pr[Zij(x,y)>T]
≡{tilde over (θ)}ij(x,y).
It can be easily shown that
Note that the expression for {tilde over (θ)}ij(x,y) depends on the following unknown parameters: σφ2, σμ2, N0, CP and α. As in the above noted related application entitled, “Geolocation Estimation . . . ”, in the present invention, pilot strength measurements are first carried out along a small number of representative routes in the service area. The measurements so obtained are then used to select values of σφ2, σμ2, N0, CP and α which provide the best fit between the predicted pilot visibility probabilities, using equation (B9), and the observed pilot visibility probabilities.
This application is related to U.S. Ser. No. 09/139,107, now U.S. Pat. No 6,496,701 entitled “Pattern Recognition-Based Geolocation”, filed in the names of T. C. Chiang et al on Aug. 26, 1998; U.S. Ser. No. 09/294,997 entitled “A Bayesian-Update Based Location Prediction Method for CDMA Systems”, filed in the names of K. K. Chang et al on Apr. 20, 1999; and U.S. Ser. No. 09/321,729, now U.S. Pat. No. 6,263,208, issued on Jul. 17, 2001, entitled “Geolocation Estimation Method For CDMA Terminals Based On Pilot Strength Measurements”, filed in the names of K. K. Chang et al on May 28, 1999. These related applications are assigned to the assignee of the present invention and are meant to be incorporated herein by reference.
Number | Name | Date | Kind |
---|---|---|---|
5933462 | Viterbi et al. | Aug 1999 | A |
6249252 | Dupray | Jun 2001 | B1 |
6263208 | Chang et al. | Jul 2001 | B1 |
6496701 | Chen et al. | Dec 2002 | B1 |
6564065 | Chang et al. | May 2003 | B1 |
20010022558 | Karr et al. | Sep 2001 | A1 |