A METHOD FOR MONITORING THE OPERATIONAL STATE OF A SYSTEM

Abstract

In the present invention signals are obtained from a plurality of sensors S.sub.1, S.sub.2, Sn and fed into an encoder (12). The encoder (12) is operable in use to receive input signals from each of the sensors S.sub.1, S.sub.2, S.sub.n and translate said signals into one or more vectors characterising the state of one or more of the operational parameters of the monitored system, hereinafter referred to as an encoded vector V.sub.E. The signals from the sensors S.sub.1, S.sub.2, S.sub.n may relate to one or more different operational parameters of a connected system. The encoded vector V.sub.E is fed into a translation engine 13 which translates the encoded vector V.sub.E into feature space to form a feature vector V.sub.F. The feature vector V.sub.F is subsequently fed into a residual vector generator (16) which compares the feature vector V.sub.F with a predicted vector V.sub.P generated by a prediction engine (14) and thereby output a residual vector V.sub.R which characterises any differences between the feature vector V.sub.F and the predicted vector V.sub.P, In addition, the feature vector V.sub.F is also fed directly into the prediction engine (14). In this way, the current operational state of each of the variable parameters of the monitored system can be input into the prediction engine (14) to update subsequent predictions made by the prediction engine (14). The formed residual vector V.sub.R is then input into a computation unit (18) for analysis, such as to determine whether the differences identified between the predicted and feature vectors V.sub.P, V.sub.F indicate that there is a fault in the monitored system.

Claims

1. A method for monitoring the operational state of a system having one or more variable parameters comprising the steps of: (a) forming an input vector containing one or more values for the or each variable parameter of the monitored system; (b) encoding the input vector to form an encoded vector; (c) translating the encoded vector into feature space to form a feature vector; (d) generating a predicted vector containing a corresponding number of predicted values for the values of the formed feature vector, wherein the predicted values are generated by using a Markov chain to statistically model the probability of various spatial or temporal transitions occurring on the basis of one or more previously generated feature vectors; (e) subtracting the feature vector from a predicted vector to form a residual vector; and (f) using the residual vector to identify one or more discrepancies in spatial and/or temporal changes in the or each variable parameter of the monitored system, so as to thereby identify a fault in the operation of the monitored system.

2. A method as claimed in claim 1 wherein subtracting the feature vector from the predicted vector comprises subtracting the or each real-number variable within the feature vector from a corresponding value in the predicted vector and wherein the corresponding value in the predicted vector is the equivalent vector component of the feature vector.

3. (canceled)

4. (canceled)

5. A method as claimed in claim 2 wherein each time a feature vector is generated, the feature vector is additionally incorporated into a statistical model for generating a further predicted vector for use in analysis of a subsequently formed feature vector.

6. (canceled)

7. (canceled)

8. (canceled)

9. A method as claimed in claim 1 comprising assigning each feature vector to an individual system state.

10. A method as claimed in claim 9 wherein feature vectors deemed to correlate are clustered together into a single system state.

11. A method as claimed in claim 10 comprising clustering feature vectors which are separated by a pre-determined distance in feature space and wherein clustering of two or more feature vectors is expressed by a clustered vector in the same feature space.

12. (canceled)

13. (canceled)

14. A method as claimed in claim 11 wherein the clustering process is repeated one or more times, the method comprising, for the second and subsequent runs of the clustering process, clustering: two or more feature vectors together; one or more feature vectors with one or more generated cluster vectors; or two or more cluster vectors.

15. A method as claimed in claim 1 wherein the at least one value of the or each variable parameter is acquired by monitoring equipment comprising one or more sensors.

16. (canceled)

17. A method as claimed in claim 1 comprising generating each vector component of the encoded vector as a binary number having a number of digits equal to a number of classes into which measured values of one or more variable parameters of the system may be classified, each digit of the binary number corresponding to a specific class; and wherein the generated binary number is repeated one or more times to form a repeated binary number.

18. (canceled)

19. (canceled)

20. A method as claimed in claim 17 wherein the repeated binary number is formed by repeating the whole of the generated binary number, or by repeating each of the digits of the generated binary number in turn.

21. A method as claimed in claim 1 used to monitor the operational state of a system to detect a fault within the system, wherein any discrepancies in identified spatial and temporal changes in the measured value or values of the or each variable parameter are taken to be an indication of a fault within the monitored system.

22. A method as claimed in claim 21 wherein a fault is identified when at least one component of the feature vector differs from a corresponding vector component of the predicted vector such that one or more of the components of the residual vector are not equal to zero, or when two or more components of the feature vector differ from a corresponding vector component of the predicted vector such that two or more of the components of the residual vector are not equal to zero, or when the magnitude of the residual vector exceeds a predetermined threshold value.

23. A method as claimed in claim 21 comprising isolating one or more components of a system in the event a fault has been identified in the operation of the system.

24. (canceled)

25. A residual vector formation system for performing for monitoring the state of an operational system, the residual vector formation system comprising: a vector encoder for encoding an input vector from measurements of one or more variable parameters of the operational system to form a feature vector; a prediction engine for generating a predicted vector characterising the predicted state of the operational system wherein the predicted vector is generated by using a Markov chain to statistically model the probability of various spatial or temporal transitions occurring on the basis of one or more previously generated feature vectors; and a residual vector generator for forming a residual vector from the generated feature and predicted vectors, so as to thereby identify a fault in the operation of the monitored system.

26. A system as claimed in claim 25 wherein the vector encoder is operable to encode an input vector from the measurements of the one or more variable parameters by generating an encoded vector wherein vector components of the encoded vector are represented as a binary number having a number of digits equal to a number of classes into which measured values of one or more variable parameters of the system may be classified; and wherein the vector encoder is operable to repeat the digits of the generated binary number one or more times to form a repeated binary number.

27. (canceled)

28. (canceled)

29. A system as claimed in claim 25 comprising one or more sensors operable to measure the one or more variable parameters of the operational system.

30. (canceled)

31. (canceled)

32. (canceled)

33. (canceled)

34. A system as claimed in claim 25 comprising an output device comprising a visual and/or audio output device for communicating information to a user.

35. (canceled)

36. (canceled)

37. (canceled)

38. (canceled)

39. A method as claimed in claim 17 comprising measuring at least one value of a variable parameter which may be in one of two categories and generating each vector component as a binary number having a single digit, the single digit being assigned a value of 0 is the measured parameter is in a first state, and a value of 1 if measured in the other of the two states.

40. A method as claimed in claim 17 comprising measuring at least one value of a variable parameter which may fall into one of two or more categories, and generating each vector component as a binary number having a number of digits corresponding to the number of categories into which the value of the variable parameter may fall.

41. A method as claimed in claim 17 comprising measuring at least one value of a variable parameter wherein the values which the variable parameter can take are continuous, and generating each vector component as a binary number having a number of digits corresponding to a number of intervals into which the measured variable parameter may fall.

42. (canceled)

Description

DETAILED DESCRIPTION OF THE INVENTION

[0061] In order that the invention may be more clearly understood an embodiment thereof will now be described, by way of example only, with reference to the accompanying drawings, of which:

[0062] FIG. 1 is a flow diagram illustrating embodiments of a method in accordance with the invention;

[0063] FIG. 2 is a table illustrating the generation of a residual vector in accordance with embodiments of the invention;

[0064] FIG. 3 is a graphical representation of an embodiments of an encoding process forming part of a method in accordance with embodiments of the invention;

[0065] FIG. 4 is a graphical representation of a further embodiment of an encoding process forming part of a method in accordance with embodiments of the invention; and

[0066] FIGS. 5A and 5B are a graphical representation of a further embodiment of an encoding process forming part of a method in accordance with embodiments of the invention.

[0067] A method in accordance with the present invention is now described with reference to FIG. 1, which is a flow diagram illustrating embodiments of the present invention.

[0068] Signals are obtained from a plurality of sensors S.sub.1, S.sub.2, S.sub.n and fed into an encoder 12. The encoder 12 is operable in use to receive input signals from each of the sensors S.sub.1, S.sub.2, S.sub.n and translate said signals into one or more vectors characterising the state of one or more of the operational parameters of the monitored system, hereinafter referred to as an encoded vector V.sub.E.

[0069] The signals from the sensors S.sub.1, S.sub.2, S.sub.n may relate to one or more different operational parameters of a connected system and may provide data to the encoder 12 in one or more different forms. For example, one or more of the sensors S.sub.1, S.sub.2, S.sub.n may comprise binary sensors which output a binary code to the encoder 12, or may comprise categorical sensors which output a signal relating to which of one or more categories the measured parameters falls within. In some embodiments one or more of the sensors S.sub.1, S.sub.2, S.sub.n may be a sensor operable to output data relating to a continuous parameter. The operation of the different sensor types are illustrated in FIGS. 3, 4, 5A and 5B and are discussed in detail below.

[0070] The encoded vector V.sub.E is fed into a translation engine 13 which translates the encoded vector V.sub.E into feature space to form a feature vector V.sub.F. The feature vector V.sub.F is subsequently fed into a residual vector generator 16 which compares the feature vector V.sub.F with a predicted vector V.sub.p generated by a prediction engine 14. The predicted vector V.sub.p preferably characterises any expected spatial and/or temporal variations in the values of the one or more parameters of the monitored system measured by sensors S.sub.1, S.sub.2, S.sub.n.

[0071] The residual vector generator 16 is operable to compare the feature vector V.sub.F with a predicted vector V.sub.p and output a residual vector V.sub.R which characterises any differences between the feature vector V.sub.F and the predicted vector V.sub.p, i.e. characterises any differences between the expected variations in the values of the parameters of the monitored system and the actual observed variations in measured values taken from the data obtained by the sensors. The operation of the residual vector generator 16 in generating the residual vector V.sub.R is illustrated in FIG. 2 and is discussed in detail below.

[0072] In addition, the feature vector V.sub.F is also fed directly into the prediction engine 14. In this way, the current operational state of each of the variable parameters of the monitored system can be input into the prediction engine 14 to update subsequent predictions made by the prediction engine 14. This is performed using statistical modelling of one or more feature vectors, which may include a number of previously obtained feature vectors along with the latest feature vector V.sub.F generated by the translation engine 13. The statistical modelling of the feature vectors V.sub.F to form the predicted vector V.sub.p by the prediction engine 14 is described in detail below.

[0073] The formed residual vector V.sub.R is then input into a computation unit 18 for analysis. Typically, this analysis may involve determining whether the differences identified between the predicted and feature vectors V.sub.p, V.sub.F indicate that there is a fault in the monitored system.

[0074] The method of the invention may be used (but is not limited to use) in the following applications:

[0075] The vehicle instrument clusters production line requires the assembly of a number of components and their testing to ensure correct functionality and product conformity meeting client needs. This process is highly complex and distributed, involving over 600 different operations, for example screw torque and angle measurements, electrical voltage tests, LCD brightness functional quality checks, CAN/LIN interface tests, gauge position angle measurements and many other functional and visual tests. Data pertaining to these production tests is collected from various workstations at different assembly points over a specific production line. The proposed approach is used for automatically learning the correlations and causative relationships between variables both in time and space, to be used to identify patterns in data to infer potential process abnormalities and fault occurrences. This is based on based on data pertaining to measured vectors and predictions of normal process operations from which a residual vector indicating a specific fault (based on the measured sensor parameters) can be detected. This information may be used to determine machine reliability (wear and tear, sensory or actuation failures) and suggest corrective actions and tolerances for adapting and configuring production machines based on conformity constrains, for example.

[0076] Fault Detection and predictive maintenance for transport fleet management based on categorical, real valued and binary data collected from vehicle maintenance records, Engine management system and CAN-Bus data on vehicle and driver interaction parameters, route, operational load previous faults. This data could be used to build models for modelling spatial/temporal causeeffect relationships predictive maintenance which could be used to determine what selected equipment/parts need to be maintained to avoid loss of revenue. The detection of these faults can be based on data pertaining to measured vectors and predictions of normal fleet operational conditions from which a residual vector indicating a specific behaviour change (based on the measured parameters) can be detected to indicate the occurrence of a specific type of fault, for example.

[0077] Detection of behaviour abnormalities in daily living activities of people suffering from dementia wherein a plethora of environmental sensors and actuators sensing aspects of user activities and behaviour in the environment and user actuations of various computational artefacts can be captured. The approach can be used to discover complex spatial and temporal correlations, which can be used to determine the occurrence of different categories of behaviour abnormalities, which can be a sign of cognitive impairments. The detection of these abnormalities can be based on monitored data pertaining to measured vectors and predictions of normal behaviours and activities from which a residual vector indicating a specific behaviour changes (based on the measured parameters) can be detected. This can be used to determine appropriate care intervention, for example.

[0078] FIG. 2 is a table illustrating how the residual vector V.sub.R may in some embodiments be generated by the residual vector generator 16 from the generated feature vector V.sub.F and the predicted vector V.sub.p. In the illustrated method, the residual vector V.sub.R is generated by subtracting the feature vector V.sub.F from the predicted vector V.sub.p. This is done by subtracting each separate component of the feature vector V.sub.F from the corresponding component in the predicted vector V.sub.p as shown in FIG. 2. As will be described below, each component of the vectors may relate to the value of one or more variable parameters of the monitored system. In this way, where the residual vector V.sub.R shows a difference between the measured and predicted values for a given component (see the 8.sup.th component in the illustrated embodiment) this can be used to directly identify which variable parameter of the monitored system has a measured value which is different to or has changed spatially or temporally by a greater or lesser amount to what was expected.

[0079] FIGS. 3, 4, 5A and 5B illustrate embodiments of methods for encoding an encoded vector V.sub.E. It should, however, be understood that the present invention is not limited to these methods but rather is intended to include any method for encoding an encoded vector V.sub.E for use in monitoring the operational state of a system as per the method of the invention.

[0080] FIG. 3 illustrates how a sensor S.sub.B may be used to measure at least one value of a variable parameter which may be in one of two classes, i.e. in one of two separate states. In the illustrated embodiment, the sensor S.sub.B comprises a binary sensor which may output one of two values, which in this case is shown to be True or False (relating to off or on/open or closed, etc. of a monitored parameter). As shown, in order to translate the measured values from the binary sensor S.sub.B to real-number variables for inclusion in the encoded vector V.sub.E, the method comprises simply assigning a binary value, i.e. 0 or 1, to a corresponding component within the encoded vector V.sub.E. In the illustrated method, this comprises assigning the binary value of 1 to a True output and the binary value 0 to a False output. The formed encoded vector V.sub.E comprises only a single component which takes the binary value of the corresponding output. However, in some variations the components of the encoded vector V.sub.E may be repeated n times. For example, a repeated encoded vector V.sub.E may read either [1, 1, 1, 1, . . . , 1] or [0, 0, 0, 0, . . . , 0], where the components are repeated n times.

[0081] FIG. 4 illustrates how a sensor S.sub.Cat may be used to measure at least one value of a variable parameter which may fall into one of two or more classes, i.e. where the measured variable parameter may have one of two or more categorical values. In the illustrated embodiment, the sensor S.sub.Cat may comprise a categorical sensor operable to output one or two or more values corresponding to two or more categories into which the measured value of the variable parameter falls. Variable parameters having two or more categorical values may include colours, or operational settings having two or more discrete values, for example. As shown, in order to translate the measured values from the categorical sensor S.sub.Cat to real-number variables for inclusion in the encoded vector V.sub.E, the method comprises assigning each category to a corresponding component within the encoded vector V.sub.E. Where a variable parameter is measured to fall within a given category, this is output from the sensor S.sub.Cat and a binary value of 1 is assigned to the corresponding component of the encoded vector V.sub.E, as shown. The formed encoded vector V.sub.E comprises a number of components equal to the number of categories into which the measured parameter of the system may fall. However, in some variations the components of the encoded vector V.sub.E may be repeated n times. For example, a formed encoded vector V.sub.E may read [0, 1, 0, 0, . . . , 0] where the measured parameters falls within the second category and the corresponding repeated encoded vector would read [0, 1, 0, 0, . . . , 0, 0, 1, 0, 0, . . . , 0] where the whole formed encoded vector V.sub.E is repeated twice. Alternatively, a formed encoded vector V.sub.E may read [0, 1, 0, 0, . . . , 0] where the measured parameter falls within the second category and the corresponding repeated encoded vector would read [0, 0, 1, 1, 0, 0, 0, 0, . . . , . . . , 0, 0] where each component of the formed encoded V.sub.E vector is repeated in turn.

[0082] FIGS. 5A and 5B illustrate how a sensor S.sub.Cont may be used to measure at least one value of a variable parameter wherein the values which the variable parameter can take are continuous. In the illustrated embodiment, the method comprises assigning components of the encoded vector V.sub.E to one or more classes in the form of intervals within which the value of the measured variable parameter may fall. This is shown graphically in FIG. 5B where the components of the encoded vector V.sub.E are each set to have a real-number variable of 0, which is increased to 1 in the event that a measured value of the variable parameter falls within a corresponding interval.

[0083] FIG. 5A illustrates how the intervals are generated from a normal distribution of previously measured values of the variable parameter. The normal distribution is then used to calculate the mean value of the variable parameter and also the standard deviation from this mean value , including multiples of standard deviation therefrom. Each component of the encoded vector V.sub.E is assigned to a class defined by an interval between two distinct multiples of standard deviation from the mean of the normal distribution of data points. In the illustrated embodiment the interval width is equal to the standard deviation . Therefore, the first component of the encoded vector V.sub.E is assigned to values falling in the range 4 to 3 , the second component of the encoded vector V.sub.E is assigned to values falling between 3 and 2 , and so on up to the eight component of the encoded vector V.sub.E assigned to values falling between 3 and 4 . In the event that a measured value of a given variable parameter falls within a given interval, the corresponding vector component assigned to that interval has its value set at 1, whereas each of the other components relating to the further intervals have their value set at 0. In this way, the method comprises a means to encode continuous data in a binary vector.

[0084] Again, in some variations the components of the encoded vector V.sub.E may be repeated n times. For example, a formed encoded vector V.sub.E may read [0, 1, 0, 0, 0, 0, 0, 0] where the measured parameter falls within the second category, corresponding to the interval between 3 and 2 and the corresponding repeated encoded vector would read [0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0] where the whole formed encoded vector is repeated twice. Alternatively, a formed encoded vector V.sub.E may read [0, 1, 0, 0, 0, 0, 0, 0] where the measured parameter falls within the second category, corresponding to the interval between 3 and 2 and the corresponding repeated encoded vector would read [0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] where each component of the formed encoded vector is repeated in turn.

[0085] The prediction engine 14 preferably uses a Markov chain to statistically model previously generated feature vectors V.sub.F in order to generate a predicted vector V.sub.p. The Markov chain is used to determine transition properties of the monitored system by forming a look up table of possible states of that system and statistically analysing the frequency of different states measured to determine said transition properties. The transition properties relate to spatial transitions (i.e. how the variable parameter/s change in parameter space) and to temporal transitions (i.e. how the variable parameter/s change over time). To generate the predicted vector, V.sub.p, the prediction engine 14 predicts the next state of a system by looking at the probability of various spatial or temporal transitions occurring (i.e. being seen in a subsequent measured vector) on the basis of the current state of the system (characterised by the most recently obtained feature vector V.sub.F). For example, in the event that the statistical model suggests that the probability of there being any alternative variation in one or more of the parameters to what is shown in the previously obtained feature vector V.sub.F is low (which may be below a specified threshold value) the predicted vector V.sub.p may be equal to the last feature vector V.sub.F. Alternatively, in the event that the statistical model suggests that the probability of there being any alternative variation in one or more of the parameters to what is shown in the previously obtained feature vector V.sub.F is high (which may be above a specified threshold value) the predicted vector V.sub.p may differ from the last feature vector V.sub.F in relation to the variables which are deemed to have a high probability of varying to a greater or lesser extent to that illustrated by the feature vector V.sub.F, for example. The prediction engine 14 is operable to store the spatial and temporal transition properties are for future reference in the look-up table and in this way, the look-up table is continually updated each time a feature vector V.sub.F is obtained to continually update the prediction process to better predict the next state of the monitored system.

[0086] The above embodiment is described by way of example only. Many variations are possible without departing from the scope of the invention as defined in the appended claims.