Industrial process surveillance system

Description of equivalent WO 9749011 A1

Industrial Process Surveillance System

This invention was made with U.S. Government support under Contract No. W-3 1-109-ENG-38 awarded by the Department of Energy. The U.S. Government has certain rights in this invention.

The present invention is related generally to a method and system for carrying out surveillance of industrial processes using sensor or data source outputs. More particularly, the invention is concerned with a method and system for processing sensor data and using virtual data as an improved methodology over basic statistical approaches to industrial process surveillance. Further, the invention involves use of a plurality of techniques coupled for enhanced analysis of industrial process data.

Conventional parameter-surveillance schemes are sensitive only to gross changes in the mean value of a process or to large steps or spikes that exceed some threshold limit check. These conventional methods suffer from either large numbers of false alarms (if thresholds are set too close to normal operating levels) or a large number of missed (or delayed) alarms (if the thresholds are set too expansively). Moreover, most conventional methods cannot perceive the onset of a process disturbance, sensor deviation or data anomaly which gives rise to a signal below the threshold level for an alarm condition. Most methods also do not account for the relationship between a measurement by one sensor relative to another sensor measurement.

In another monitoring method, a conventional sequential probability ratio test ("SPRT") technique has found wide application as a signal validation tool in the nuclear reactor industry. The SPRT method is a pattern recognition technique which processes the stochastic components associated with physical process variables and has high sensitivity for the onset of subtle disturbances in those variables. Two features of the conventional SPRT technique make it attractive for parameter surveillance and fault detection: (1) early annunciation of the onset of a disturbance in noisy process variables, and (2) the SPRT technique has user-specificable false alarm and missed-alarm probabilities. SPRT techniques are primarily directed to the analysis of data from paired or multiple pairs of sensors in contrast to a large number of different process sensor data points. SPRT is also typically dependent on assumptions of the data being independent of other data sources and being Gaussian distributed data. The SPRT technique used alone therefore has certain shortcomings in identifying anomalies in processes.

Other types of statistical techniques also have been developed for industrial process monitoring and analysis but have other insensitivities for certain classes of sensor data.

It is, therefore, an object of the invention to provide an improved method and system for surveillance of industrial processes and apparati.

It is another object of the invention to provide a novel method and system for on-line surveillance of industrial processes and apparati with multiple sensors.

It is also an object of the invention to provide an improved method and system for evaluation of process data, on-line or off-line, from sensors or data accumulation sources.

It is a further object of the invention to provide a novel method and system for performing preliminary analysis of data for alarm conditions prior to data input to a SPRT system.

It is an additional object of the invention to provide an improved method and system for masking selected sensor data and substituting virtual data to perform tests to determine whether abnormal process conditions or abnormal sensor conditions exist and whether or not to halt or modify the process under scrutiny.

It is still another object of the invention to provide a novel method and system using training data characteristic of normal system and/or sensor and/or data source operation to compare with ongoing industrial processes and/or data accumulation.

It is yet a further object of the invention to provide an improved method and system for processing data from a process to determine training data for normal operation, storing such training data on a computer storage media and analyzing real process data relative to the normal training data using a plurality of mathematical methodologies stored on a ROM or PROM storage medium.

It is also an additional object of the invention to provide a novel method and system utilizing a virtual signal characteristic of normal state operation derived on the basis of correlation with a plurality of other process data values to compare with a real process data signal set for deriving the likelihood of an abnormal process or operation of data sources.

It is yet another object of the invention to provide a novel method and apparatus to accumulate training data to recognize any one of a plurality of specific states of operation and thereby identify a particular type of fault or condition present in a process or other system.

It is also a further object of the invention to provide a novel method and apparatus for monitoring a process using training data to identify slowly changing operational sensor data characteristic of normal process changes.

It is still an object of the invention to provide an improved method and system for determining whether a system or data source abnormality can be ignored without undesirable effects.

Other advantages and features of the invention, together with the organization and manner of operation thereof, will become apparent from the following detailed description when taken in conjunction with the accompanying drawings described below.

Brief Description of the Drawings

FIGURE 1 illustrates a schematic functional flow diagram of a preferred embodiment of the invention;

FIGURE 2 illustrates a functional flow diagram of a time lead-lag correlation methodology;

FIGURE 3 illustrates a functional flow diagram of a method of determining a full range of data by searching normal state training data;

FIGURE 4 illustrates a functional flow diagram of a method for modeling behavior of commercial system operating states;

FIGURE 5 illustrates a functional flow diagram of a method for peifoπTiing pattern recognition;

FIGURE 6A illustrates sensor signals having a four second delay before applying a lead-lag method, and FIG. 6B illustrates the sensor signals after applying the lead-lag method;

FIGURE 7A illustrates sensor signal data from pump 1 power with an SMSET estimate superimposed thereon; FIG. 7B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 7C illustrates a histogram of the error;

FIGURE 8A illustrates sensor signal data from pump 2 power with an SMSET estimate superimposed thereon; FIG. 8B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 8C illustrates a histogram of the error;

FIGURE 9A illustrates sensor signal data from pump 1 speed with an SMSET estimate superimposed thereon; FIG. 9B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 9C illustrates a histogram of the error;

FIGURE 10A illustrates sensor signal data from pump 2 speed with an SMSET estimate superimposed thereon; FIG. 1OB illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. IOC illustrates a histogram of the error;

FIGURE 11A illustrates sensor signal data for reactor outlet flow rate; FIG. 11B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 11C illustrates a histogram of the error;

FIGURE 12A illustrates sensor signal data for primary pump 2 flow rate; FIG. 12B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 12C illustrates a histogram of the error;

FIGURE 13A illustrates sensor signal data for subassembly outlet temperature 1A1; FIG. 13B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 13C illustrates a histogram of the error;

FIGURE 14A illustrates sensor signal data for subassembly outlet temperature 2B1; FIG. 14B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 14C illustrates a histogram of the error;

FIGURE 15A illustrates sensor signal data for subassembly outlet temperature 4E1; FIG. 15B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 15C illustrates a histogram of the error;

FIGURE 16A illustrates sensor signal data for subassembly outlet temperature 4F1; FIG. 16B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 16C illustrates a histogram of the error; FIGURE 17A illustrates sensor signal data for reactor outlet temperature 1534CF; FIG. 17B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 17C illustrates a histogram of the error;

FIGURE 18A illustrates sensor signal data for primary tank sodium level 530 Float; FIG. 18B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 18C illustrates a histogram of the error;

FIGURE 19A illustrates sensor signal data for primary tank sodium level 531 induction; FIG. 19B illustrates the SMSET estimation error between the SMSET estimate and the sensor signal data; and FIG. 19C illustrates a histogram of the error;

FIGURE 20A illustrates standard deviation of SMSET errors for each of the data in FIG. 7-19; and FIG. 20B illustrates the mean value of SMSET errors for each of the data in FIG. 7-19;

FIGURE 21 shows subassembly outlet temperature ("SOT") and SMSET estimates and in particular FIG. 21 A illustrates time dependent normal SOT for 3F1 in the EBR-II nuclear reactor; FIG. 21B illustrates normal SOT for 3C1; HG. 21C illustrates normal SOT for 5C2 and FIG. 21D illustrates normal SOT for 7A3;

FIGURE 22A-D illustrates SMSET estimation error for each of the data of FIGS. 21A-D, respectively;

FIGURE 23A-D illustrates SPRT results for each of the data of FIGS. 21A-D, respectively;

FIGURE 24 A corresponds exactly to FIG. 21 A; FIG. 24B includes a linear drift component compared to FIG. 21B; and FIGS. 24C and 24D correspond exactly to FIG. 21C and 21D, respectively;

FIGURE 25A corresponds exactly to FIG. 22A; FIG. 25B includes the effect on SMET estimation error of the linear drift of FIG. 24B; and FIGS. 24C and 24D correspond exactly to FIGS. 22C and 22D, respectively;

FIGURE 26A corresponds exactly to FIG. 23A; FIG. 26B illustrates the SPRT results for the linear drift error of FIG. 24B; and FIGS. 26C and D corresponds exactly to FIG. 23C and D, respectively;

FIGURES 27A and 28B corresponds exactly to FIGS. 21A and 21B, respectively; FIG. 27C includes a temporary amplitude pulse of 0.25% of the signal magnitude; and FIG. 27D corresponds exactly to FIG. 21D; FIGURES 28 A and 28B corresponds exactly to FIGS. 22 A and 22B; FIG. 28C illustrates SMSET estimation error for the amplitude pulse effect of FIG. 27C and FIG. 27D corresponds exactly to FIG. 22D;

FIGURES 29A and 29B corresponds exactly to FIGS. 23A and 23B; FIG. 29C iUustrates SPRT results of the amplitude pulse in FIG. 27C; and FIG. 29D corresponds exactly to FIG. 23D;

FIGURE 30 A illustrates EBRII subassembly temperature data 3F1 but includes a uniform gain change compared to FIG. 21 A and FIGS. 30B-D correspond exactly to FIGS. 21B-D;

FIGURE 31 A illustrates the SMSET estimation error for the gain change of FIG. 30A; and FIGS. 31B-D correspond exactly to FIGS. 22B-D, respectively; and

FIGURE 32A illustrates the SPRT results for the gain change of FIG. 30A and SMSET analysis of FIG. 31 A; and FIGS. 32B-D correspond exactly to FIGS. 23B-D, respectively.

Detailed Description of Preferred Embodiments The system 10 herein includes a methodology (see FIG. 1) and apparatus for surveillance of sensor or data accumulation configurations in industrial, utility, business, medical, investment and transportation applications. The system 10 is useful for sensitive identification of the onset of sensor or data source degradation, process or system anomalies, or the onset of change to a different operational state. The most preferred form of the system 10 comprises a synergistic integration of four techniques to provide enhanced surveillance capabilities compared to conventional approaches (including neural networks), and also provide enhanced reliability and improved computational efficiency. The four elements that make up the most preferred surveillance form of the system 10 are embodied in four different methodologies generally characterized as a time correlation module 20, a training module 30, a system state estimation module 40 and a pattern recognition module 50.

Many attempts to apply conventional neural networks to nuclear, petrochemical or any industrial process surveillance applications have met with poor results in part because they fail to take into account lead-lag relationships (lack of proper time correlation of the data sets) between the various sensors or data sources. In one example, a pipe is instrumented with a sequence of N thermocouples ("TCs") which measure the instantaneous temperature of the fluid passing through the pipe; and the signals from these TCs are displaced in time along the fluid stream flow path. If the sampling interval for the sensors is smaller than the transit time for fluid passing through the pipe, any attempt to monitor these signals with a neural net will produce very high uncertainties in the estimated signals for the TCs. For example, if it takes ten seconds for the fluid to pass through the pipe, and the sampling interval is one second, then at any given instant in time, TC(N), at the outlet end of the pipe, is seeing fluctuations that passed TC(1) ten seconds ago. These TCs may still contain a small degree of correlation due to gross changes in fluid temperature from a heat source or sink that is upstream of the pipe; however, the more valuable intersensor correlation that arises from local temperature perturbations carried along the pipe will be lost. This same phenomenon degrades the performance of neural networks and other pattern-recognition paradigms applied to any processes wherein the physical sensors or data sources are displaced in time across the process they are monitoring. Other examples of time delays in correlated systems include: systems with slow data flow rates and/or large physical distances (oil refineries, power plants, HVAC systems, and financial systems), delays due to analog or digital electronics (noise filters and large capacitors) or transmission delays (satellite communications, or transmitting data over different BUS systems.

In a preferred form of the time correlation model 20, a Leadlag component of the invention (see FIG. 6) performs dynamic, real-time intersensor lead-lag adjustments. The Leadlag module 20 performs adjustments so that the output signals, which are then input subsequently into the SMSET routine (the system state estimation module 40), are optimally correlated and impart the maximum information content to the pattern recognition module 50. The Leadlag module 20 is attached hereto as a computer software Appendix A. The Leadlag module 20 accomplishes the adjustment function by performing, for each pair of signals, an iterative regression procedure that generates a vector of correlation coefficients with respect to lag time. This vector of correlation coefficients is a ummodal concave function of lag time. Thus, the optimal lag time between the pair of signals is identified simply by searching for the zero-crossing of the first derivative of the vector with respect to the lag time.

In other forms of the invention it may be unnecessary to utilize the Leadlag module 20, as noted in FIG. 1 wherein the option exists to skip use of the module 20. This could occur for systems in which there is adequate time correlation or if time shifting is not needed to achieve correlation. If the Leadlag module 20 is not utilized or the data has already been processed by the Leadlag module 20, the data is preferably input to a training module 30.

In a preferred embodiment this training module is a MiniMax module 30 which searches through all the observations for all signals or data during a training time period to construct training vectors that include the highest point and lowest point for each signal or data space under surveillance. A computer software Appendix B sets forth the MiniMax module 30. The MiniMax module 30 produces an "optimal" training set. It is optimal in the sense that it contains only, at most, 2N vectors, where N is the number of signals or data points in the system; and these vectors span the full range that all sensors or data sources have noted during the available training period. Wherever two or more sensors or data sources simultaneously attain maxima or minima, the resulting number of training vectors will be less than 2N.

In another form of the invention both the Leadlag module 20 and the MiniMax module 30 can be skipped, and the data can be input directly to the system state module 40.

Once the MiniMax module 30 has constructed a system model (or been skipped as noted above), the system state estimation module 40 (such as the preferred Subsystem Multivariate State Estimation Technique ("SMSET") module) models the behavior of a system through examples of the operating states of the commercial system being modeled. A computer sofflvare Appendix C sets forth the SMSET module 40. In general, the system state estimation module 40 can be any one of a variety of modeling methodologies, such as auto regressive moving average, a neural network, or a Kalman filtering technique or an empirical methodology.

The SMSET module 40 utilizes its memory of the learned states of the commercial system in conjunction with a single new observation to provide an estimate of the current "true" system state. States of the system are represented by vectors whose elements are comprised of direct values of system parameters (measured signals) as well as any transformation of these system parameters that produce scalar values, e.g. calculated parameters based upon measured data. The SMSET module 40 does not require the state vector elements to be linearly independent as do most other types of estimation techniques. The learning process, which results in a "learned-state" matrix, is performed according to the MiniMax module 30 and the Leadlag module 20 described hereinbefore.

The basic methodology of the SMSET module 40 involves the input of a new observation of the behavior of a system that is compared with the "memory" of previous system behavior embodied in the learned-state matrix. A series of mathematical operations are performed that generates an estimate of the states in the system's memory that is "closest" to the new observation. The definition of "closest" that is used by the SMSET module 40 is the state that is lying closest to the new observation from the point of view of a set of rules that determine the association of two vectors. From this closest state, an estimate of the "true" state of the system is performed for each and every element of the state vector. Thus, given a set of current observed parameters of a system, the SMSET module 40 provides an estimate of the current true state of the system. The value of this method is that an estimate of all of the values of the system parameters in the state vector can be provided even if the current observation vector is incomplete (e.g. some sensors or data sources may have failed or are no longer available), contains erroneous or faulty elements (some sensors may have drifted, become uncalibrated, become contaminated with high noise levels, etc.), or even if the new system state does not coincide with previous operating states. However, the new system state must, in a general sense, be bounded by the domain of the states used to develop the system memory (learned-state matrix).

This estimation of the true current state of the commercial system, including estimated values of all system parameters, is used in conjunction with the actual measured system parameters to ascertain the operability of sensors (or other data sources) and disturbances in the system state. This state estimation process can further be described as an inference engine that accepts as input a set of learned states and a new observation of the commercial system. After a series of operations are performed by the inference engine on this input, the result is an estimate of the learned state "closest" to the new observation. The definition of "closest" used here is the state lying closest to the new observation from the point of view of a set of rules that determine the association (overlap) of any two vectors. Another result is the estimation of the "true" value of each and every element in the new observation vector in the form of an estimated state vector. The series of operations performed in the inference engine consist of various matrix operations. First, all pairs of learned states are preferably associated two at a time using a rule set to create the elements of a recognition matrix. Next, the new observation is associated with each learned state using the rule set to produce a vector that has the same number of elements as the number of learned states. The largest element value in this vector identifies the "closest" learned state to the new observation. Finally, the normal matrix product of this vector with the recognition matrix produces a set of linear combination coefficients for combining the learned states into the estimated state vector. This methodology, when applied to any true state of a commercial system that is a combination of the learned states, yields a very close approximation to the true state. The actual closeness achieved depends most upon nonlinearities arising from the rule set and physical and/or random fluctuations in the variables and is demonstrated by direct testing. General experience with use of this method for real operating commercial systems has indicated predictive capabilities typically no worse than +/- 0.5% ad normally +/-0.1%.

Once the SMSET module 40 has modeled the data as described hereinbefore, the data is input to a pattern recognition module 50, such as the Sequential Probability Ratio Test ("SPRT") module. The computer software for the SPRT module 50 is in Appendix D. This SPRT module 50 is a sensitive pattern recognition method that can detect the onset of subtle degradation in noisy signals with high reliability, and with quantitative false-alarm and missed-alarm probabilities. Output from the SMSET module 40 is provided as a set of estimated signals (also called "virtual signals") for each sensor under surveillance. These virtual signals are fed into a network of interacting SPRT modules 50 together with the actual sensor readings. Each of the SPRT modules 50 receives one sensor-signal, virtual-signal pair. If any sensor degradation or process disturbance starts to affect the output of one or more signals under surveillance, the SPRT module(s) 50 provide an annunciation to the operator and an actuator signal to the control system, which can selecuvely as needed automatically swap in the virtual signal to replace the degrading sensor signal, or data source. Further details of the SPRT module 50 are descπbed in USPN 5,459,675, which is incorporated by reference herein.

The above-described combination of methodologies enables identification of a faulted process, a particular type of fault, a faulted sensor or data source or faulty data itself and enables actions to be taken to correct or modify the process being monitored.

In some cases when a failed sensor, or improper data stream source, is not important to the continued operauon of a commercial system, the user can continue operaung the commercial system or process if the sensor or data source were operaung normally. For example, the system 10 can operate to substitute in a modeled estimate into a actual commercial system or process as input to replace a failed sensor or failed data source. This allows the commercial system or process to keep operating.

Since the system 10 does not rely on analytical modeling by itself, it is applicable to a wide vaπety of processes and systems, such as petro-chemical, power generauon, automotive, manufactunng, medical, aeronauUcal, financial and any system in which signals are available for processing that are related to the commercial system process operation or performance. The only requirement of the system 10 is that there is some type of cross-correlation, be it linear or nonlinear, between the signals used as input to the system 10. The signals can be linear, nonlinear, stationary, nonstationary, clean or noisy (with an arbitrary distribution). The system 10 uses a database of histoπcal operation data to model the commercial system or process. The database is assumed to contain data from all relevant operaung modes of the system; however, if a new mode of operation is encountered and is determined not to be a result of commercial system or sensor failures, a new vector can be added to the existing training matnx to incoφorate the unanticipated operaung mode in the system model. The following nonlimiting examples illustrate various aspects of the invention described herein. The data used is all taken from the EBR-II reactor at Argonne National Laboratory (West).

Example I

The effect of time delay correlation is described in this example, and FIGS. 6A and 6B show two voltage signals with a four second delay between them. Before the signals are processed by the Leadlag module 20 (see FIG. 6A), the correlation coefficient is 0.0182 which implies no correlation versus processing through the Leadlag module 20 to obtain a correlation of 0.9209 (see FIG. 6B). When the set of signals, or data, being used is more than two, all the possible pair combinations are used to calculate maximum possible correlation coefficients so all signals can be properly correlated.

Example II

An experiment to determine the accuracy of the invention (the "SMSET" methodology generally) was carried out using sensor data from the Experimental Breeder Reactor II (EBR-II) at Argonne National Laboratory (U558). The sensor data set contained 13 signals from sensors monitoring EBR-II. Table I shows the SMSET Estimation accuracy for EBR-II Data. Table I includes the channel numbers and descriptions for each of the sensor signals used in the analysis. The experiment was conducted in three steps; first the SMSET module was trained using two days worth of EBR-II data, next the trained SMSET module was used to estimate the state of approximately 110 hours worth of EBR-II data, and then the accuracy of the estimates was analyzed. For each of the sensor signals listed in Table I, FIGS 7-19, respectively, show the sensor signal (top plot) and SMSET estimate superimposed, the middle plot shows error between the SMSET and the sensor signal (in percent of the signal magnitude), ad a histogram (bottom plot) of the error. The histogram plots are compared to a Gaussian distribution with the same mean ad variance to give an idea of how Gaussian the error signals. FIG. 20 provide a summary of the data of FIGS. 7-19.

A methodology entitled MiniMax (Appendix B) was used to train the system using the two days of training data cited above. After the MiniMax method was applied, a training matrix was constructed consisting of twenty-five unique vectors constituting an empirical model of the overall system. After creating the model, the methodology was then applied to the signals listed in the accuracy table. Each signal in the system has its own estimation error signal that is a measure of how close the pattern recognition model is representing the system relative to the sensor readings. The second column of Table I lists the standard deviation of the estimate error for all of the signals in the experiment in terms of each of the signals' magnitude. The magnitude of the signal is defined by its mean during normal operation. The third column in Table I lists the mean of the estimate error for all of the signals also in terms of the signal magnitude. In general the estimate error standard deviations are in the range of 0.01 % to 0.1 % and the estimate error means are centered around 0. Bar graphs of the tabular information are shown in FIGS. 20A and 20B as graphic representation of the accuracy information.

Table I

SMSET Estimation Accuracy Table for EBRII Data:

<img file="WO9749011A1_D0001.tif" />

<img file="WO9749011A1_D0002.tif" />

Example III

In FIGS. 21-32 examples of different sensor failure modes are shown along with how the system reacts to the failures. The preferred method of FIG. 1 is applied to the data. The sensor signals used in these examples are from a subset of 22 sensor signals used in the system. The 22 sensors monitored the EBR-II subassembly system at Argonne National Laboratory (West). Each of FIGS. 21-32 contains four subplots in which the upper most plot is related to Subassembly Outlet temperature ("SOT") 3F1, the upper middle plot is related to SOT 3C1, the lower middle plot is related to SOT 5C2, and the bottom plot is related to SOT 7A3. The system applied in each of the examples uses the same training matrix, which consists of 83 vectors selected from a training data base containing almost a weeks worth of data taken once every minute.

In FIGS. 21-23 are shown the results of using the system 10 during approximately 5.5 days of normal operation of EBR-II. FIG. 21 shows the SOT signals with their corresponding SMSET estimates (signal being the circles and the lines being the estimate). FIG. 22 shows the respective raw estimate errors (not in terms of the signal magnitude) derived by taking the difference between the SOR signals and corresponding SMSET estimates. Finally in FIG. 23 the results are shown from applying the decision making module of the system 10 (the SPRT module 50-- see Appendix D) to the SMSET estimation errors of FIG. 22. The SPRT plots show a total of only three false alarms which is a false alarm rate of 9.4 x 10^"s, and this is well within the specified false alarm rate of 1.0 x 10^'3.

One type of failure mode that is common among sensors is a slow drift. This type of failure can be difficult to detect early on especially when the sensor signals contain noise. FIGS. 24-26 illustrate a comparative example of processing data from this type of failure and failure identification. Signal #2 (FIG. 24B) has a 0.2% linear drift in it's mean over the 2.75 day period starting at 4000 minutes into the signal. The other sensors are operating normally. FIG. 25 shows the resulting SMSET estimation errors for each sensor signal. The error plot for signal #2 (FIG. 25B) shows evidence of drifting after the sensor signal has drifted approximately 0.05%. In FIG. 26 the SPRT method has determined that #2 (FIG. 26B) is drifting after approximately 0.05 % of drift and that all other sensors are operating normally.

Another type of failure that can occur is a step change in the sensor signal. This can be the result of a short in the sensor or DAS, a calibration error or for a variety of other reasons. FIGS. 27-29, show a example of this type of failure for the SOT measurements. In this example sensor signal #3 (FIG. 27C) contains a pulse with an amplitude of 0.25% of the signal magnitude. The pulse starts at 4000 minutes and lasts for 2000 minutes. FIG. 27 shows the sensor signals and the SMSET estimates for the four SOT signals. FIG. 28 shows the resulting SMSET estimation errors. The error signal for #3 (FIG. 28C) shows that there is a problem starting at 4000 minutes and ending at 6000 minutes. The error signals are fed through the SPRT module 50, and the results are plotted in FIG. 29. Clearly, there has been a disturbance in sensor #3 (FIG. 29C) beginning at time 4000 minutes and ending at 6000 minutes.

In FIGS. 30-32 an example of a failure mode related to the sensor gain is shown. In this example the gain of the sensor signal changes over tune, i.e., the amplitude is increasing over time. The gain begins changing linearly over time from a beginning value of 1 to a final value of i +0.075% of the sensor magnitude. The system 10 for the estimation error is applied to the signals, and the results are shown in FIG. 31. A human operator would most likely not be able to tell that there is a problem even after 8000 minutes by looking at the sensor signal. In FIG. 31 A, it is apparent that signal ϋ\ is operating abnormally. This is confirmed in FIG. 32A by the SPRT results, showing a steadily increasing number of SPRT alarms over the 8000 minute period.

While preferred embodiments of the invention have been shown and described, it will be apparent to those skilled in the art that various changes and modifications can be made without departing from the invention in its broader aspects as set forth in the claims provided hereinafter.

Appendix A Computer software for Leadlag module which performs dynamic, real-time intersensor lead-lag time correlation adjustments.

<img file="WO9749011A1_D0003.tif" /><img file="WO9749011A1_D0004.tif" /><img file="WO9749011A1_D0005.tif" /><img file="WO9749011A1_D0006.tif" /> Appendix B Computer software for producing an optimal training set derived by searching signal information during a training period to construct training vectors including highest and lowest points for signals under surveillance.

<img file="WO9749011A1_D0007.tif" /><img file="WO9749011A1_D0008.tif" /><img file="WO9749011A1_D0009.tif" /><img file="WO9749011A1_D0010.tif" /><img file="WO9749011A1_D0011.tif" /><img file="WO9749011A1_D0012.tif" /><img file="WO9749011A1_D0013.tif" /><img file="WO9749011A1_D0014.tif" /> Appendix C Computer software for modeling behavior for examples of the operating states ommercial system.

<img file="WO9749011A1_D0015.tif" /><img file="WO9749011A1_D0016.tif" /><img file="WO9749011A1_D0017.tif" /><img file="WO9749011A1_D0018.tif" />

<img file="WO9749011A1_D0019.tif" /> end

Appendix D

Computer software for performing pattern recognition by detecting onset of degradation in noisy signals.

<img file="WO9749011A1_D0020.tif" />

<img file="WO9749011A1_D0021.tif" /><img file="WO9749011A1_D0022.tif" />