The present invention relates to the monitoring of physical systems, processes or machines, and more particularly to systems and methods for discerning signal differentiation using an operator to detect deviations in one or more sets of signals from monitored sensors.
Generally, there is a need to detect when one or more of a set of signals from sensors monitoring a system, whether a machine, process or living system, deviates from “normal.” Normal can be an acceptable functioning state, or it can be the most preferred of a set of various acceptable states. The deviation can be due to a faulty sensor, or to a change in the underlying system parameter measured by the sensor, that is, a process upset.
While threshold-type sensor alarms have traditionally been used to detect when parameters indicate that a component has strayed away from normal, acceptable or safe operating conditions, many deviations in sensor or underlying parameter values go unnoticed because threshold detection can only detect gross changes. Often such detection may not occur early enough to avoid a catastrophic failure. In particular, there is a critical need to detect when a component, as indicated by a signal or underlying parameter value, is deviating from an expected value, given its relationship to other system components, i.e., in the monitored machine, process or living system. This detection should occur even though the corresponding signal in question is still well within its accepted gross threshold limits.
A number of methods exist that try to use the relationships between sensors, signals, data or the underlying parameters that correspond thereto, to detect notable component changes that otherwise would be missed by “thresholding.” Such methods are often data-intensive and computationally demanding. There is a need for accurate empirical modeling techniques that provide computationally efficient and accurate system state monitoring.
The present invention provides an improved monitoring system and method for ultrasensitive signal differentiation that achieves the detection of changes to and faults in one or more sensor signals in a set that characterizes an underlying “process or machine.” The invention achieves accurate results with improved computational efficiency without relying on thresholding. Therefore, less memory and CPU power are needed to perform the necessary calculations. Also, because of the improved computational efficiency, more data “time-slices” can be processed with a given CPU speed. This is useful particularly in systems where signals or data must be sampled at a high rate, or in an application where the CPU or micro-controller is limited.
An empirical model is developed of the process or machine to be monitored, and in real-time sensor data from the monitored process or machine is compared to estimates of same from the model. The results of the comparison are statistically tested with an ultrasensitive difference test that indicates alerts on a sensor-by-sensor basis, thereby providing early warning of process upsets, sensor failures, and drift from optimal operation, long before these would be noticed by conventional monitoring techniques. According to the invention, an improved similarity operator is used in generating the estimates.
The invention provides an improved operator that can be implemented in software on a variety of systems. A typical implementation would be in a programming language like C or C++ on a Unix or Windows workstation, either as a standalone program sampling data via a National Instruments-like pc-card, or as a part or module of a broader process control software system. The program can be a Windows-like DLL or callable routine, or can comprise an entire suite of presentation screens and settings modification screens. The software can be reduced to a PCMCIA card for addition to a computer with such a port. Then, the data being input to the computation could be fed through a dongle attached to the PCMCIA card. An alternative would be to put the program into microcode on a chip. Input and output of the requisite data to a broader system would depend on the design of the overall system, but it would be well known to circuit designers how to build in the microchip version of this invention. In yet another embodiment, the computation could take place on a server remote (WAN) from the sensors that feed to it. As contemplated in another embodiment, the internet could be used to deliver real-time (or semi-real-time) data to a server farm which would process the data and send back either alarm level data or higher-level messages. In that case, it would become necessary to ensure that the asynchronous messaging “delay” of the internet was of sufficiently unobtrusive to the semi-real-time monitoring taking place over the internet/WAN. For example, bandwidth could be guaranteed so that the asynchronicity of the messaging was not any kind of delay. Alternatively, the sampling rate of the system could be slow enough that the delivery time of Internet messages was negligible in comparison.
Briefly, the present invention relates to a computationally efficient operation for accurate signal differentiation determinations. The system employs an improved similarity operator for signal differentiation. Signals or data representative of several linearly and/or non-linearly related parameters that describe a machine, process or living system are input to the inventive system, which compares the input to acceptable empirically modeled states. If one or more of the input signals or data are different than expected, given the relationships between the parameters, the inventive system will indicate that difference. The system can provide expected parameter values, as well as the differences between expected and input signals; or the system can provide raw measures of similarity between the collection of input signals and the collection of acceptable empirically modeled states.
The novel features believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as the preferred mode of use, further objectives and advantages thereof, is best understood by reference to the following detailed description of the embodiments in conjunction with the accompanying drawings, wherein:
FIG. 7 and
The present invention is a system, method and program product for monitoring operation of a machine, process, living system or any other such system to accurately predict individual component failures, i.e. component end of life, such that failure or process upset can be anticipated to reduce planned downtime, reduce unplanned downtime and provide for process or product quality optimization.
It is directed to the employment of an improved similarity operator within an empirical modeling engine for use in generating estimates of sensor values for a monitored process or machine based on input of the actual sensor values. Advantageously, the similarity operation carried out in the modeling according to the invention has comparatively low computational needs as compared to other similarity operators that can be used in such an empirical model-based monitoring system, such as is described in U.S. Pat. No. 5,987,399 to Wegerich et al., wherein is disclosed a similarity operator employing computationally intensive trigonometric variables. In particular a representative “training” set of data is compared against monitored signal data in a similarity engine and a statistical significance engine. Techniques for achieving ultrasensitive signal differentiation are similar to what is generically described in U.S. Pat. No. 5,764,509, directed to the application of a sequential probability ratio test (“SPRT”). Parameter data are gathered from signal sensors monitoring a system such as a machine, process or living system. The number of sensors used is not a limiting factor, generally, other than respecting computational overhead. The present invention is highly scalable. Preferably, the sensors should capture component parameters of at least some of the primary “drivers” of the underlying system. Furthermore, all sensor inputs to the system are best interrelated in some fashion (non-linear or linear).
Turning to
The machine 102 may be connected to and controlled by a control system 104, generally comprising a microcontroller- or microprocessor-based digital system with appropriate analog digital and digital/analog inputs and outputs as are known to those skilled in the art. Machine 102 is instrumented with sensors monitoring machine components or reactions thereto (e.g., chamber temperature or pressure) and providing resultant sensor values along outputs 106. During training, the machine 102 is operated through an expected range of operations, and data acquisition system 108 records values of all sensors 106 with which machine 102 is instrumented. Additionally, control signals from control system 104 may also be recorded by data acquisition system 108, and may be used as “sensor signals” that correlate with the other sensor signals.
Data acquired by data acquisition system 108 can, accordingly, be processed using a computer module 110 for producing a distilled training data set representing the operational ranges of machine 102, using the preferred training method, or other appropriate such methods as may be known in the art.
The monitoring system described herein includes an empirical modeling engine and a statistical decision-making engine supported by a suite of software routines for data preconditioning, training, and post-decision reporting. This system is modular and can be applied separately depending on the requirements of the particular monitoring application. Typically, process monitoring equipment employs sensors having some common characteristics. A sensor data set is acquired as being representative of the normal or desired operation range of the system. The parameters monitored by sensors should be chosen for the model to be correlated, either linearly or nonlinearly. Generally, multiple sensor inputs may be necessary, however, the described algorithms may apply to single sensor applications by using signal decomposition of the single sensor signal into components that can be treated as multiple, correlated inputs for modeling and monitoring. The identification of small deviations in signals from normal operation is provided as indicative of the status of the sensor's associated physical parameter.
The inventive monitoring system can be adapted to provide not just monitoring of existing sensor data, but also generation of derived “virtual” sensor data based on other actually-monitored sensors. Thus, an example 120 of an embodiment of the present invention in an on-board processor is shown in
Processor 128 can also be a part of the control system 124, and in fact can be the processor on which the control system routines are executed, in the event the control system is a digital computed control system. Ideally, the processor 128 and memory 132 are powered by the same power source as the control system 124. However, under certain circumstances, it may also be preferable to provide for a processor 128 and memory 132 independent from another processor and or memory (not shown) of the control system 124, in order to provide for example virtual signals 130 in a timely fashion, as though they were truly instrumented parameters. For example, processor 128 may operate at a higher clock speed than the control system processor.
Furthermore, a display 156 may be provided at the remote location for displaying data descriptive of the process 142, i.e., sensor signals 144 and any virtual signals derived therefrom or both. The virtual signals generated by processor 152 can also be transmitted from link 150 back to link 148 and input over leads 158 to control system 146 for advantageous control of the process. Data from original sensor signals and or virtual sensor signals (if included) can also be transmitted to a third remote communications link 160, located at yet a third distant place, for display on display 162, thereby providing valuable information concerning the process to interested parties located at neither the physical process site nor at the monitoring system site.
The remote communications links can be selected from a variety of techniques known in the art, including internet protocol based packet communication over the public telecommunications infrastructure, direct point-to-point leased-line communications, wireless or satellite. More specifically, remote links 148,152 and 160 may be internet-enabled servers with application software for accumulating, queuing and transmitting data as messages, and queues for receiving and reconstituting data arriving as messages. Alternatively, communications can be synchronous (meaning in contrast to asynchronous, message-based communications) over a wireless link.
The embodiment of the invention shown in
A preferred method of employing the improved similarity operator of the present invention is shown in
Real time monitoring stage 190 is indicated at step 188, whereupon the model generated at 186 is employed in the steps 192 for estimation. Real-time data is acquired in 192 and an estimate of what the sensor values should be is generated based thereon in view of the model. For each snapshot of data acquired in 192, the similarity operator of the invention is used to compare the actual real-time sensor values to vectors of sensor values in the training set.
The differences between the expected values and real-time sensor signals, i.e., the residual signals formed in 192, are sent to a decision-making engine in the process deviation decision-making step 182. The decision-making engine continuously renders a reliable decision at 194 based on a SSCADI index calculated at 198 over a moving window of successive residual observations, determining whether or not the residual signals reveal a statistically relevant abnormality.
The described embodiment uses a platform with an operating system for real-time monitoring 190 and sampling of the real-time sensor data via a National Instruments feed, with a monitor including an estimation module 192 with a windowing environment to provide feedback in the form of alarms 194 or other high-level messages 196 to an operator of equipment. The program embodying the invention can be written in C or in LabView, a product available from National Instruments Company. The operator portion itself comprises a callable Windows DLL that the LabView software calls. The software may be written as components, including: (1) a training component 172 (for selecting the D matrix from a set previously obtained of “normal” input data), (2) an estimation module component 192 for modeling (to provide the similarity operations both on matrix D and on the real-time input vector, and also to provide an estimated output or an output of the degree of similarity of the vector to the states recognized in the D matrix), and (3) a statistical test component 198 (which tests data such as using SPRT on the similarity measurement or residual between the input and the estimate).
Monitoring begins by providing signal data as vectors, each with as many elements as there are sensors. A given vector represents a “snapshot” of the underlying system parameters at a moment in time or time correlated. Additional pre-processing can be done, if necessary for time correlation to insert a “delay” between cause and an identified effect between sensors. That is to say, for example, if sensor A detects a parameter change that will be reflected at sensor B three “snapshots” later, the vectors can be reorganized such that a given snapshot contains a reading from sensor A at a first moment, and a corresponding reading from sensor B three snapshots later. Each snapshot is representative of a “state” of the underlying system. Methods for time-correlating in this way are known to those skilled in the art.
Turning to
In this example, each snapshot represents a vector of five elements, one reading for each sensor in that snapshot. Of all the collected sensor data from all snapshots, according to this training method, only those five-element snapshots are included in the representative training set that contain either a global minimum or a global maximum value for any given sensor. Therefore, the global maximum 216 for sensor 202 justifies the inclusion of the five sensor values at the intersections of line 218 with each sensor signal 202, 204, 206, 208, 210, including global maximum 216, in the representative training set, as a vector of five elements. Similarly, the global minimum 220 for sensor 202 justifies the inclusion of the five sensor values at the intersections of line 222 with each sensor signal 202, 204, 206, 208, 210. Collections of such snapshots represent states the system has taken on. The pre-collected sensor data is filtered to produce a “training” subset that reflects all states that the system takes on while operating “normally” or “acceptably” or “preferably.” This training set forms a matrix, having as many rows as there are sensors of interest, and as many columns (snapshots) as necessary to capture all the acceptable states without redundancy.
Turning to
In step 234, if the sensor value of sensor i at snapshot t in X is greater than the maximum yet seen for that sensor in the collected data, max(i) is updated to equal the sensor value and Tmax(i) stores the number t of the observation in step 236. If not, a similar test is done for the minimum for that sensor in steps 238 and 240. The observation counter t is incremented in step 242. In step 244, if all the observations have been reviewed for a given sensor (t=L), then t is reset and i is incremented (to find the maximum and minimum for the next sensor) in step 246. If the last sensor has been finished (i=N), step 248, then redundancies are removed and an array D is created from a subset of vectors from X.
First, in step 250, counters i and j are initialized to one. In step 252, the arrays Tmax and Tmin are concatenated to form a single vector Ttmp having 2N elements. These elements are sorted into ascending (or descending) order in step 254 to form array T. In step 256, holder tmp is set to the first value in T (an observation number that contains a sensor minimum or maximum). The first column of D is set equal to the column of X corresponding to the observation number that is the first element of T. In the loop starting with decision step 258, the ith element of T is compared to the value of tmp that contains the previous element of T. If they are equal (the corresponding observation vector is a minimum or maximum for more than one sensor), it has already been included in D and need not be included again. Counter i is incremented in step 260. If they are not equal, D is updated to include the column from X that corresponds to the observation number of T(i) in step 262, and tmp is updated with the value at T(i). The counter j is then incremented in step 264. In step 266, if all the elements of T have been checked, then the distillation into training set D has finished, step 266.
Once the D matrix has been determined, in the training phase 172, the preferred similarity engine may begin monitoring the underlying system (122, 142 or 178) and through time, actual snapshots of real sensor values are collected and provided to the similarity engine. The output of the similarity engine can be similarity values, expected values, or the “residual” (being the difference between the actual and expected values).
One or all of these output types are passed to the statistical significance engine, which then determines over a series of one or more such snapshots whether a statistically significant change has occurred as set forth hereinbelow. In other words, it effectively determines if those real values represent a significant change from the “acceptable” states stored in the D matrix. Thus, a vector (Y) of expected values:
{right arrow over (y)}out={overscore (D)}·{right arrow over (W)}
(i.e., estimates) is determined for the sensors. The expected state vector (Yout) is equal to contributions from each of the snapshots in D, which contributions are determined by the weight vector W. The multiplication operation (·) is the standard matrix/vector multiplication operator. W has as many elements as there are snapshots in D. W is determined by:
D is again the training matrix, and D′ is the standard transpose of that matrix (i.e., rows become columns). Yinput includes the real-time or actual sensor values from the underlying system, and therefore it is a vector snapshot.
The similarity operator {circle around (x)} can be selected from a variety of mechanisms for providing a measure of the similarity of one snapshot as compared to another. One general class of similarity operators performs an element for element comparison, comparing a row of the first matrix with a column containing an equal number of elements from the second matrix. The comparison yields a “similarity” scalar value for each pair of corresponding elements in the row/column, and an overall similarity value for the comparison of the row to the column as a whole, e.g., for one snapshot to another.
The similarity operator has been improved in the present invention to define the similarity si between ith elements as:
where:
Accordingly, vector identity yields si=1.0 and complete dissimilarity yields si=0.0. Typically si falls somewhere between these two extremes.
Overall similarity of a row to a column is equal to the average (or some other suitable statistical treatment) of the number N (the element count) of si values:
or alternatively the similarity S of a row to a column can also be calculated:
that is, the average of the θi, divided by ρ, raised to the power of λ, and subtracted from one. Accordingly, the S values that are summed from θi or si are the S values depicted in the result matrix above. This operator, having been described generically above, works as follows in the determination of W. The first factor, D′{circle around (x)}D, is referred to herein as matrix G, i.e.,
{overscore (G)}={overscore (D)}′{circle around (x)}{overscore (D)}
So, for example, with D and its transpose for 4 sensors a, b, c, d and for n training set snapshots are:
then the matrix G is:
where, for example, element g12 in the matrix G is computed from row 1 of D-transpose (a1, b1, c1, d1) and column 2 of D (a2, b2, c2, d2) as either:
where the elements si and θi are calculated as:
The resulting matrix G is symmetric around the diagonal, with ones on the diagonal. To further calculate W, the computation of (D′{circle around (x)}Yinput) is performed in a similar fashion, except Yinput is a vector, not a matrix:
The G matrix is inverted using standard matrix inversion, and the result is multiplied by the result of D′{circle around (X)}Yinput. The result is W:
These weight elements can themselves be employed, or expected values for the sensors can be computed, applying the training matrix D to the weight vector W:
Turning to the analysis or tuning parameters λ and ρ, while either value can remain set to its default value of 1, analysis parameter ρ can be varied to all values greater than or equal to 1. The sensitivity analysis parameter λ can be varied to any value and, preferably, to any positive value and more particularly, can be selected in the range of 1-4.
The computation results of the expected sensor values is subtracted from the actual sensor values to provide a “residual”. The residual is also a vector having the same number of elements as the sensor-count, with each element indicating how different the actual value is from the estimated value for that sensor.
So, for example, set forth below is an example of C code for the overall functionality of the inventive similarity operator based upon a modeling technique for the calculation of the similarity between two vectors.
First user specified selectable parameters include:
Then, inputs include:
Outputs are defined as:
Before real-time monitoring can begin, the global range R for each of the i=1,2, . . . , M variables in the model must be calculated using a representative data set as described hereinabove with reference to
“Ri=max(X(n,i))−min(X(n,i)), over all n”
for (column=0; column<M; ++column)
Then after determining the ranges Rt for each variable, the OP operator model matrix C can be constructed from an appropriate training matrix D. The model matrix is an N by N square matrix calculated using the OP operator, as follows:
“G=t{circle around (x)}SSCOPD, (where,{circle around (x)}SSCOP is the SSCOP operator)”
The SSCOP operator measures the similarity between vectors. When calculating C, the similarities between all pairs of column vectors in D are calculated. The SSCOP operator uses the ranges Ri, non-linearity analysis parameter λ, and the scaling analysis parameter ρ to calculate similarity measurements. Performing the following procedure carries out the similarity between two vectors, d1 and d2 in D using SSCOP.
If D contains N column vectors, each including M elements,
D=[d1|d2| . . . |dN]
dk=[d1(t)d2(t) . . . dM(t)]
then vector similarity is measured as follows:
A statistical significance test is applied to the elements of the residual, or to the residual as a whole. More particularly, a test known as the sequential probability ratio test (“SPRT”) is applied to the residual for each sensor. This provides a means for producing an alarm or other indicator of statistically significant change for every sensor in real-time.
The SPRT type of test is based on the maximum likelihood ratio. The test sequentially samples a process at discrete times until it is capable of deciding between two alternatives: H0μ=0; and HA:μ=M. It has been demonstrated that the following approach provides an optimal decision method (the average sample size is less than a comparable fixed sample test). A test statistic, Ψt, is computed from the following formula:
where ln(·) is the natural logarithm, fHs( ) is the probability density function of the observed value of the random variable Yi under the hypothesis Hs and j is the time point of the last decision.
In deciding between two alternative hypotheses, without knowing the true state of the signal under surveillance, it is possible to make an error (incorrect hypothesis decision). Two types of errors are possible. Rejecting H0 when it is true (type I error) or accepting H0 when it is false (type II error). Preferably these errors are controlled at some arbitrary minimum value, if possible. So, the probability of a false alarm or making a type I error is termed α, and the probability of missing an alarm or making a type II error is termed β. The well-known Wald's Approximation defines a lower bound, L, below which one accepts H0 and an upper bound, U beyond which one rejects H0.
Decision Rule: if Ψt<L, then ACCEPT H0;
To implement this procedure, this distribution of the process must be known. This is not a problem in general, because some a priori information about the system exists. For this purpose, the multivariate Normal distribution is satisfactory.
For a typical sequential test
where M is the system disturbance magnitude and {overscore (Y)} is a system data vector.
Thus, the present invention provides a superior system modeling and monitoring tool for ultra sensitive signal differentiation accurately detecting system parametric deviations indicative of more subtle underlying system or system component changes, whether the system is a machine, a process being carried out in a closed system, a biological system or any other such system. The tool differentiates between anomalous results caused by defective sensors and component related failures, allowing adjustment therefor. Further, the signal differentiation sensitivity of the tool far exceeds prior art thresholding results using less memory and computer power to consider larger volumes of data, interactively if desired, irrespective of available computer power.
It will be appreciated by those skilled in the art that modifications to the foregoing preferred embodiments may be made in various aspects. The present invention is set forth with particularity in the appended claims. It is deemed that the spirit and scope of that invention encompasses such modifications and alterations to the preferred embodiment as would be apparent to one of ordinary skill in the art and familiar with the teachings of the present application.
This application claims the benefit of priority under 35 U.S.C. § 119(e) to U.S. Provisional application Ser. No. 60/193,228 filed Mar. 30, 2000.
Number | Name | Date | Kind |
---|---|---|---|
3045221 | Roop | Jul 1962 | A |
4336595 | Adams et al. | Jun 1982 | A |
RE31750 | Morrow | Nov 1984 | E |
4480480 | Scott et al. | Nov 1984 | A |
4639882 | Keats | Jan 1987 | A |
4707796 | Calabro et al. | Nov 1987 | A |
4761748 | Le Rat et al. | Aug 1988 | A |
4796205 | Ishii et al. | Jan 1989 | A |
4823290 | Fasack et al. | Apr 1989 | A |
4841456 | Hogan, Jr. et al. | Jun 1989 | A |
4931977 | Klemes | Jun 1990 | A |
4937763 | Mott | Jun 1990 | A |
5003950 | Kato et al. | Apr 1991 | A |
5005147 | Krishen et al. | Apr 1991 | A |
5025499 | Inoue et al. | Jun 1991 | A |
5052630 | Hinsey et al. | Oct 1991 | A |
5093792 | Taki et al. | Mar 1992 | A |
5113483 | Keeler et al. | May 1992 | A |
5119287 | Nakamura et al. | Jun 1992 | A |
5123017 | Simpkins et al. | Jun 1992 | A |
5195046 | Gerardi et al. | Mar 1993 | A |
5210704 | Husseiny | May 1993 | A |
5223207 | Gross et al. | Jun 1993 | A |
5251285 | Inoue et al. | Oct 1993 | A |
5285494 | Sprecher et al. | Feb 1994 | A |
5311562 | Palusamy et al. | May 1994 | A |
5327349 | Hoste | Jul 1994 | A |
5386373 | Keeler et al. | Jan 1995 | A |
5420571 | Coleman et al. | May 1995 | A |
5421204 | Svaty, Jr. | Jun 1995 | A |
5446672 | Boldys | Aug 1995 | A |
5455777 | Fujiyama et al. | Oct 1995 | A |
5459675 | Gross et al. | Oct 1995 | A |
5481647 | Brody et al. | Jan 1996 | A |
5500940 | Skeie | Mar 1996 | A |
5502543 | Aboujaoude | Mar 1996 | A |
5539638 | Keeler et al. | Jul 1996 | A |
5548528 | Keeler et al. | Aug 1996 | A |
5553239 | Heath et al. | Sep 1996 | A |
5586066 | White et al. | Dec 1996 | A |
5600726 | Morgan et al. | Feb 1997 | A |
5629878 | Kobrosly | May 1997 | A |
5638413 | Uematsu et al. | Jun 1997 | A |
5657245 | Hecht et al. | Aug 1997 | A |
5663894 | Seth et al. | Sep 1997 | A |
5668944 | Berry | Sep 1997 | A |
5671635 | Nadeau et al. | Sep 1997 | A |
5680409 | Qin et al. | Oct 1997 | A |
5682317 | Keeler et al. | Oct 1997 | A |
5699403 | Ronnen | Dec 1997 | A |
5708780 | Levergood et al. | Jan 1998 | A |
5710723 | Hoth et al. | Jan 1998 | A |
5727144 | Brady et al. | Mar 1998 | A |
5748469 | Pyötsiä | May 1998 | A |
5761090 | Gross et al. | Jun 1998 | A |
5764509 | Gross et al. | Jun 1998 | A |
5774379 | Gross et al. | Jun 1998 | A |
5787138 | Ocieczek et al. | Jul 1998 | A |
5790977 | Ezekiel | Aug 1998 | A |
5796633 | Burgess et al. | Aug 1998 | A |
5822212 | Tanaka et al. | Oct 1998 | A |
5845230 | Lamberson | Dec 1998 | A |
5864773 | Barna et al. | Jan 1999 | A |
5895177 | Iwai et al. | Apr 1999 | A |
5905989 | Biggs | May 1999 | A |
5909368 | Nixon et al. | Jun 1999 | A |
5913911 | Beck et al. | Jun 1999 | A |
5930779 | Knoblock et al. | Jul 1999 | A |
5933352 | Salut | Aug 1999 | A |
5933818 | Kasravi et al. | Aug 1999 | A |
5963884 | Billington et al. | Oct 1999 | A |
5970430 | Burns et al. | Oct 1999 | A |
5987399 | Wegerich et al. | Nov 1999 | A |
5995916 | Nixon et al. | Nov 1999 | A |
6006260 | Barrick, Jr. et al. | Dec 1999 | A |
6023507 | Wookey | Feb 2000 | A |
6026348 | Hala | Feb 2000 | A |
6049827 | Sugauchi et al. | Apr 2000 | A |
6104965 | Lim et al. | Aug 2000 | A |
6119111 | Gross et al. | Sep 2000 | A |
6128543 | Hitchner | Oct 2000 | A |
6181975 | Gross et al. | Jan 2001 | B1 |
20020087290 | Wegerich et al. | Jul 2002 | A1 |
20020091499 | Wegerich et al. | Jul 2002 | A1 |
20030109951 | Hsiung et al. | Jun 2003 | A1 |
Number | Date | Country |
---|---|---|
WO 0067412 | Nov 2000 | WO |
Number | Date | Country | |
---|---|---|---|
20020055826 A1 | May 2002 | US |
Number | Date | Country | |
---|---|---|---|
60193228 | Mar 2000 | US |