JP6223936B2 - Abnormal trend detection method and system - Google Patents

Description

  The present invention relates to an abnormal tendency detection method for detecting a gradual increase in the degree of abnormality of a sensor output signal based on multi-dimensional time series data output from various sensors attached to a plant or equipment, and the like About the system.

  Electric power companies use waste heat from gas turbines to supply hot water for district heating, and supply high-pressure steam and low-pressure steam to factories. Petrochemical companies operate gas turbines and other power sources. As described above, in various plants and facilities using a gas turbine or the like, abnormality detection for detecting a malfunction of the facility or its sign is extremely important for minimizing damage to society.

  Not only gas turbines and steam turbines, but also water turbines in hydroelectric power plants, nuclear reactors in nuclear power plants, wind turbines in wind power plants, aircraft and heavy machinery engines, railway vehicles and tracks, escalators, elevators, equipment / parts level, Equipment that requires preventive maintenance as described above, such as on-board battery deterioration and life, has no spare time.

  For this reason, attaching a some sensor to object equipment or a plant, and judging whether it is normal or abnormal according to the monitoring standard for every sensor is performed. Japanese Patent Application Laid-Open No. 2011-70635 (Patent Document 1) discloses a normal model in an abnormality detection method for detecting the presence or absence of an abnormality based on an abnormality measure calculated by comparison with a model created from past normal data. The creation by a local subspace method is disclosed. In this method, the presence or absence of an abnormality can be accurately determined based on the magnitude of the abnormality measure. However, there is a need to detect not only the presence / absence of abnormality but also the degree of abnormality gradually worsening, that is, the rising trend of abnormality measure. Furthermore, it is necessary to specify a sensor related to the detected increase in the abnormality measure.

  In relation to such needs, Japanese Patent Application Laid-Open No. 2009-115481 (Patent Document 2) describes a facility by comparing an inclination of an upward trend or a downward tendency from accumulated data of sensor signals with a preset reference value. A monitoring diagnosis system for determining a deterioration state is disclosed.

JP 2011-70635 A JP 2009-115481 A

  In the method described in Patent Document 1, it is possible to determine the presence or absence of an abnormality based on the magnitude of the abnormality measure, but there is no description about the detection of an abnormal measure increasing tendency. Further, in the method described in Patent Document 2, an upward or downward tendency can be detected for each sensor, but it is necessary to calculate an inclination for each sensor and compare it with a preset reference value. Although the slope is calculated by regression analysis, it varies depending on the calculation interval, so it is difficult to set an appropriate interval, and it is generally difficult to set an appropriate reference value for each sensor. is there. This is particularly difficult when dealing with a large number of sensors. In addition, it is possible to detect an upward or downward trend for each sensor by visual inspection of the time series graph, but there is a problem that the standard changes depending on the person or day, and it is performed when the number of sensors is large. It becomes difficult.

  Therefore, an object of the present invention is a method and system for solving the above-described problems and detecting an upward trend of an abnormal measure and identifying a related sensor in equipment state monitoring using a multidimensional time series sensor signal, It is an object of the present invention to provide a method and a system that do not require human intervention in the process and do not require complicated standard setting.

  In order to achieve the above object, in the present invention, in a method for detecting an abnormal tendency of equipment, a plurality of sensor signals are processed by processing a plurality of sensor signals output in time series from a plurality of sensors attached to the equipment. A feature vector is created for each time for each time, an abnormal measure is calculated based on the created feature vector, an abnormal measure increasing tendency is determined for each time based on the time series change of the calculated abnormal measure, and the determination result The abnormal measure rising section where the abnormal trend increasing trend is detected from the plurality of sensors, the sensor contributing to the abnormal measure increasing in the abnormal measure increasing section detected from the plurality of sensors is identified, and in the abnormal measure increasing section of the identified sensor Based on the inclination of the contribution amount and the length of the abnormal measure increasing section, the presence or absence of an abnormal measure increasing tendency in the abnormal measure increasing section is determined.

  Alternatively, in the above-described abnormality detection method, the abnormality measure calculation method is such that a normal model is created using learning data, and the feature vector at each time is calculated based on the distance from the normal model.

  In order to achieve the above object, according to the present invention, a system for detecting an abnormal tendency of equipment based on a plurality of sensor signals output in time series from a plurality of sensors mounted on the equipment is provided. Raw data accumulating means for accumulating the sensor signal output to, feature vector creating means for creating a feature vector for each time from the sensor signal output in time series accumulated in the raw data accumulating means, and An anomaly measure calculating means for calculating an anomaly measure based on the feature vector at each time created by the feature vector creating means, and an increasing trend of the anomaly measure at each time based on the time series change of the anomaly measure calculated by the anomaly measure calculating means An abnormal measure increasing tendency determining means for determining the anomaly measure, and an interval in which the abnormal measure increasing tendency continues in the abnormal measure increasing tendency determined by the abnormal measure increasing tendency determining means An abnormal measure increasing section detecting means for detecting, a related sensor specifying means for specifying a sensor contributing to the abnormal measure increasing in an abnormal measure increasing section detected from a plurality of sensors, and a sensor specified by the related sensor specifying means A trend presence / absence determining means for determining the presence / absence of an abnormal measure increasing tendency in the abnormal measure increasing section based on the slope of the contribution amount in the abnormal measure increasing section and the length of the section is provided.

  According to the present invention, in equipment condition monitoring using a plurality of sensor signals, in order to detect an upward trend of an abnormal measure calculated based on a plurality of sensor signals, a slope is set for each sensor by setting a section for each sensor. Need not be calculated, and it is not necessary to compare with the inclination calculated with reference setting for each sensor. That is, according to the present invention, it is possible to detect a continuous section having an abnormal trend of increasing trend, specify a related sensor for each section, and determine the presence or absence of the trend. Can be realized with less computer load.

  By applying the above method, not only equipment such as gas turbines and steam turbines, but also water turbines at hydroelectric power plants, nuclear reactors at nuclear power plants, wind turbines at wind power plants, engines for aircraft and heavy machinery, railway vehicles, At the orbit, escalator, elevator, factory production equipment, and equipment / parts level, not only the presence / absence of abnormality in sensing data targeting humans, such as the deterioration / life of the on-board battery, electroencephalogram, electrocardiogram, etc. It is possible to detect a tendency that the degree of the deterioration gradually deteriorates at an early stage.

It is a block diagram which shows the structure of the outline of the abnormal tendency detection system in one Example of this invention. It is a flowchart which shows the flow of the process using the abnormal tendency detection system in Example 1 of this invention. It is a table | surface of the signal list | wrist which shows the example of the sensor signal in Example 1 of this invention. It is a flowchart which shows the flow of the process of the feature vector creation in Example 1 of this invention. It is a flowchart which shows the flow of a process of abnormality measure calculation in Example 1 of this invention. It is a figure explaining a local subspace method. It is a figure explaining the method to determine the presence or absence of an abnormal measure rising tendency from the graph of the time series data of the abnormal measure in Example 1 of this invention. It is a flowchart which shows the flow of a related sensor specific process of the abnormal measure raise area in Example 1 of this invention. It is an example of the time series graph of one element of the feature vector and the starting point vector in Example 1 of this invention. It is a graph which shows the change of the difference data for every time in the area explaining the inclination of the least square approximate line of the time series data in Example 1 of the present invention. It is a front view of the display screen showing an example of GUI for the recipe setting in Example 1 of this invention. It is a front view of the screen which displayed the some time series data in the example of the result display screen in Example 1 and 2 of this invention. It is a front view of the screen which expanded and displayed several time series data in the example of the result display screen in Example 1 and 2 of this invention. It is a front view of the screen which displayed the detail of the information of the abnormal measure raise area in the example of the result display screen in Example 1 and 2 of this invention. It is a front view of the display screen showing an example of the GUI for the recipe setting in Example 2 of this invention. It is a flowchart which shows the flow of the process using the abnormal tendency detection system in Example 2 of this invention. It is a front view of the display screen showing an example of GUI for showing the evaluation result in Example 2 of this invention.

The present invention processes a plurality of sensor signals output in time series from a plurality of sensors attached to a facility, creates a feature vector for each of the plurality of sensor signals at each time, and based on the created feature vector An abnormal measure is calculated, and an abnormal measure increasing tendency is determined at each time based on the time series change of the calculated abnormal measure. The sensor that contributed to the abnormal measure increase in the abnormal measure increase section detected from the sensors is identified, and the abnormal measure increases based on the slope of the contribution amount of the specified sensor in the abnormal measure increase section and the length of the abnormal measure increase section The present invention relates to a method and a system for detecting an abnormal tendency of equipment that determines the presence or absence of an abnormal measure increasing tendency in a section.
Hereinafter, embodiments of the present invention will be described with reference to the drawings.

Hereinafter, the contents of the present invention will be described in detail with reference to the drawings.
FIG. 1A shows a configuration example of an abnormal tendency detection system 100 that realizes the abnormal tendency detection method of the present invention.

  The abnormal tendency detection system 100 includes a data analysis unit 120 that processes the sensor signal 102 output from the facility 101, an input / output unit 110 that includes a screen that displays a GUI (Graphic User Interface), and a storage unit 111 that stores data. It is configured with.

  The data analysis unit 120 includes a sensor signal storage unit 103 that stores the sensor signal 102, a feature vector generation unit 104 that generates a feature vector based on the sensor signal 102 stored in the sensor signal storage unit 103, and based on the feature vector An abnormal measure calculating unit 105 that calculates an abnormal measure, an abnormal measure increasing tendency determining unit 106 that determines an abnormal measure increasing tendency at each time based on a time series change of the abnormal measure, and an abnormality that detects a section in which the abnormal measure increasing tendency continues. The measure rising section detecting unit 107, the related sensor specifying unit 108 for specifying the related sensor in each section based on the time series change of the contribution amount to the abnormal measure of each sensor, the inclination of the contribution amount of the related sensor and the section length A tendency presence / absence determination unit 109 that determines the presence / absence of an abnormal trend increasing tendency in the section is provided.

In the present embodiment, using the system shown in FIG. 1A, an abnormal tendency of equipment is detected by a processing flow as shown in FIG. 1B. That is, first, the sensor signal 102 output from the equipment 101 stored in the sensor signal storage unit 103 is input to the feature vector creation unit 104 (S121), and the feature vector creation unit 104 creates a feature vector (S122). The abnormal measure calculating unit 105 calculates the abnormal measure (S123), the abnormal measure increasing tendency determining unit 106 determines the abnormal measure increasing tendency at each time (S124), and the abnormal measure increasing section detecting unit 107 shows the abnormal measure increasing tendency. The continuous section is detected (S125), the related sensor specifying unit 108 specifies the related sensor for each section (S126), and the tendency presence / absence determination unit 109 finally determines whether or not there is an abnormal measure increasing tendency (S127). ), The determination result is output from the input / output unit 110 (S128).
Details of each step will be described below.

  The equipment 101 subject to state monitoring is equipment or a plant such as a gas turbine, a steam turbine, a chemical plant, or a water plant. The facility 101 outputs a sensor signal 102 indicating the state. The sensor signal 102 is stored in the sensor signal storage unit 103. An example in which the sensor signals 102 are listed and displayed in a tabular form is shown in FIG. The sensor signal 102 is a multi-dimensional time series signal acquired at regular intervals, and a table in which the sensor signal 102 is listed includes a plurality of sensor values provided in the date / time column 201 and the equipment 101 as shown in FIG. It consists of a data column 202. The number of types of sensors may be several hundreds to thousands, for example, the temperature of cylinders, oil, cooling water, etc., the pressure of oil or cooling water, the rotational speed of the shaft, the room temperature, the operating time, etc. In addition to representing an output or state, there may be a control signal for controlling something to a certain value.

  In this system, it is assumed that processing is performed using the sensor signal stored in the sensor signal storage unit 103. Although not shown, a “learning period” for learning the normal state and an “evaluation period” for trend detection are set in advance, and the learning period and evaluation are performed from the sensor signal storage unit 103 in S121. A sensor signal included in the period is extracted.

  Using the sensor signal extracted from the sensor signal storage unit 103, the feature vector creation unit 104 creates a feature vector. A processing flow in the feature vector creation unit 104 (detailed processing flow in S122) will be described with reference to FIG.

  First, the sensor signal 102 extracted from the sensor signal storage unit 103 is input (S301), and canonicalization is performed for each sensor (S302). In the canonicalization, for example, the average μ of the learning data and the standard deviation σ are used, and conversion is performed so that the average becomes 0 and the variance becomes 1 by the calculation formula of (original signal value−μ) / σ. The evaluation data is also converted by the same calculation formula using the same average and standard deviation. Alternatively, the maximum value and the minimum value of each sensor signal are used for conversion so that the maximum is 1 and the minimum is 0. Alternatively, a preset upper limit value and lower limit value may be used instead of the maximum value and the minimum value. Sensor signal canonicalization is a process for simultaneously handling sensor signals of different units and scales.

  Next, smoothing is performed for each sensor (S303). Specifically, the data at a certain time is the average of the data for the window width including and including that time. The window width is a parameter and is the number of data for which the average is calculated. The smoothing of the sensor signal is performed for the purpose of making a judgment from a macro viewpoint without being constrained by the ups and downs of numerical values.

  Next, feature vectors are created for each time (S304). A feature vector at a certain time is obtained by extracting sensor signals of a predetermined order including and including the time and collecting them for all the sensors. At that time, not the continuous data but the data of the interval at which the smoothing windows contact without overlapping each other is extracted. Further, it may be performed not at every time but at a predetermined sampling interval. This is equivalent to thinning out data and is not an essential process, but an effect of shortening the calculation time can be obtained. Since smoothing is performed, there is little loss of information even if thinning out due to overlapping windows.

  Based on the feature vector created by the feature vector creation unit 104 in S122, the abnormality measure calculation unit 105 calculates an abnormality measure for each feature vector in S123. The flow of processing in the abnormality measure calculation unit 105 (details of processing in S123) will be described with reference to FIG.

  First, feature vectors created from sensor signals during the learning period are used as learning data. However, feature vectors that satisfy a predesignated condition may be excluded from the learning data (S401). The designation of the condition may be for date and time, for the magnitude or rate of change of a specific sensor signal, or for a calculation result between two or more specific sensor signals. In addition, although not shown in the drawing, the presence / absence of a specific event or a combination of two or more events using the event signal relating to the operation start or operation end of the facility output from the facility, the number of times in a predetermined section, interval, etc. And a condition for the calculation result may be designated.

  Next, the following steps S403 to S405 are performed for all feature vectors (S402). Paying attention to one feature vector, the normal model creation unit 106 creates a normal model using learning data excluding the same section as the feature vector (focused vector) of interest (S403). The section here may be divided by a predetermined length such as one day, for example, but any division may be used as long as it is defined in advance. If the attention vector is included in the evaluation period, there is no section to be excluded, and all learning data is used.

  Next, an abnormality measure is calculated based on the distance between the vector of interest and the normal model (S404). Here, the starting point vector used for the distance measurement is recorded (S405). If processing has been completed for all feature vectors, the loop is exited, and the calculated abnormality measure is smoothed (S407). Specifically, an average of a predetermined number of abnormality measures before and after the time is calculated as the abnormality measure at a certain time. By this processing, it is possible to easily detect an abnormal measure increasing tendency.

  As a normal model creation method in S403, a local subspace method (LSC) or a projection distance method (PDM) can be considered.

  The local subspace method is a method of creating a k−1 dimensional affine subspace using the k−neighbor vector of the vector of interest q. FIG. 5 shows an example in the case of k = 3. As shown in FIG. 5, since the abnormal measure is expressed by the projection distance shown in the figure, the point Xb on the affine subspace closest to the attention vector q may be obtained. This point is a leg of a perpendicular drawn from the attention vector q to the affine subspace, and is the starting point vector in step S405. From the evaluation data q and its k-neighbor vector xi (i = 1, i can be calculated from the step, from the matrix Q and the matrix X that arranges k q

により相関行列Cを求め、

To obtain the correlation matrix C,

によりbを計算する。ｂは、xiの重み付けを表す係数ベクトルである。

To calculate b. b is a coefficient vector representing the weighting of xi.

  Since the abnormal measure d is a distance between q and Xb, it is expressed by the following equation.

  Although FIG. 5 illustrates the case of k = 3, any number may be used as long as it is sufficiently smaller than the number of dimensions of the feature vector. When k = 1, the processing is equivalent to the nearest neighbor method.

  The projection distance method is a method of creating a subspace having an original origin for a selected feature vector, that is, an affine subspace (space of maximum dispersion). A plurality of feature vectors corresponding to the vector of interest are selected by some method, and the affine subspace is calculated by the following method.

  First, the average μ of the selected feature vectors and the covariance matrix Σ are obtained, then the eigenvalue problem of Σ is solved, and a matrix U in which eigenvectors corresponding to r eigenvalues designated in advance from the larger one are arranged. Let it be an orthonormal basis of the affine subspace. r is a number smaller than the dimension of the feature vector and smaller than the number of selected data. Alternatively, r may not be a fixed number, and may be a value when the contribution rate accumulated from the larger eigenvalue exceeds a predesignated ratio.

  The anomaly measure is the projection distance of the vector of interest onto the affine subspace. Similar to the local subspace method, it is possible to calculate the position of the foot of the perpendicular line dropped into the affine subspace, and this is recorded as the starting point vector in S405. Here, as a method for selecting a plurality of feature vectors, a method of selecting several tens to several hundreds of feature vectors specified in advance in order from the vector of interest is conceivable. Further, feature vectors to be learned may be clustered in advance, and feature vectors included in a cluster closest to the vector of interest may be selected.

  In addition, a local average distance method in which the distance of the vector of interest q to the average vector of k-neighbor vectors is an abnormal measure, a Gaussian process, a Mahalanobis distance method, a clustering method, or the like may be used.

  Next, based on the time series change of the calculated abnormal measure, the abnormal measure increasing tendency determining unit 106 determines the abnormal measure increasing tendency at each time in S124. An example of the method for determining the upward trend of the abnormal measure in S124 will be described with reference to FIG.

  FIG. 6 is a graph in which a part of the calculated time series data of the abnormality measure is cut out and displayed. The horizontal axis in FIG. 6 represents time, and the vertical axis represents an abnormal measure. Based on a predetermined period, the value obtained by subtracting the anomaly measure 1 cycle before and 2 cycles before the anomaly measure at the time of interest is the reference time anomaly at y0, y1, 1 cycle, and 2 cycles after the anomaly If the value obtained by subtracting the measure is y2 and y3, respectively, and (y0> 0 || y1> 0) && (y2> 0 || y3> 0) or y1 + y3> th, then the time is an abnormal measure Judge that there is an upward trend. Here, th is a preset threshold value. The period is a parameter set according to the operation cycle of the equipment. For example, in the case of a device that is started at a fixed time of the day and stopped at a fixed time, it is set to 1 day.

  Here, the determination logic may be another predetermined one. Further, only the abnormal measure before and after one cycle may be used, or the abnormal measure separated by three cycles or more may be used. The aim of this process is to determine the rising tendency of the abnormality measure by comparing the time expected to be in the same state, thereby canceling the change caused by the change of the state.

  In another embodiment of the method of determining an abnormal measure increasing tendency, a process of repeatedly checking one point at a time by taking out one point at a time from a preset section before and after the attention time and checking whether the abnormal measure is increasing is repeated. If the ratio is greater than or equal to a predetermined threshold, it is determined that the time has a tendency to increase the abnormal measure.

  Next, as processing in S125, the abnormal measure increasing section detecting unit 107 detects a section in which the abnormal measure increasing tendency continues based on the presence or absence of the abnormal measure increasing tendency at each time. At that time, if there is an allowable gap (interval that interrupts the continuity of the upward trend of abnormal measurement), the interruption time (period) that can be considered that the abnormal trend of upward trend is ignored ignoring the interrupted interval or The number of samplings to be interrupted) is given as a parameter, and when the abnormal measure upward trend is interrupted shorter (less) than the allowable gap, it is detected as a continuous interval.

  In another embodiment of the processing in the abnormal measure increasing section detecting unit 107, there is an abnormality and an abnormal measure increasing tendency based on the presence or absence of an abnormal measure increasing tendency at each time and the presence or absence of an abnormality determined by a method described later. Detect continuous intervals. Similarly to the above, when the interruption is shorter (less) than the allowable gap, it is detected as a continuous section.

  The determination of the presence or absence of an abnormality at each time is performed after the processing in the abnormality measure calculation unit 105. First, based on the abnormality measure of learning data, a threshold value that does not determine learning data as abnormal is set. That is, the maximum value of the abnormality measure of the learning data is set as the threshold value. Alternatively, the abnormal measures of the learning data are sorted in ascending order, and a value reaching a ratio close to 1 specified in advance is set as a threshold value. In this case, a threshold value that does not determine that most of the learning data is abnormal is set. Next, the abnormality measure is compared with the calculated threshold value every time, and when the abnormality measure exceeds the threshold value, it is determined that the time is abnormal.

  Next, as a process for specifying the related sensor in S126, the related sensor specifying unit 108 specifies the related sensor for the detected abnormal measure increasing section. A detailed flow of processing in the related sensor specifying unit 108 will be described with reference to FIG.

  The related sensor identification process is performed for all sections (S701). First, the index of intra-section data is acquired (S702). This is for making it possible to refer to all feature vectors and the starting point vector recorded in S405, and any format can be used as long as data at the same time can be associated.

  Next, the following processing is performed for all sensors (S703). First, difference data for each time in the section is calculated (S704).

  FIG. 8A shows an example of a time series graph of one element of the feature vector and the starting point vector. The horizontal axis represents time, and the vertical axis represents element values. A thick line 801 represents an element of a feature vector, and a thin line 802 represents an element of an origin vector.

  Here, as shown in FIG. 8B, the difference data for each time in the calculated interval is the absolute value (| (802) − (801) |) of the difference between the element 802 of the starting vector and the element 801 of the feature vector. Obtained as time-series data.

  From the time-series data of the difference data, the inclination is calculated with time on the horizontal axis and the difference on the vertical axis (S705). The slope is, for example, the slope of a least square approximation line of time series data. That is, as shown in FIG. 8B, Δx / Δt is defined as a slope when the amount of change of the least square approximation line in a certain period Δt is ΔΔx. This inclination can be considered to represent the degree of time series change of the anomaly measure of one element shown in FIG. 8A. In other words, the larger the inclination, the more the time series change of the abnormal measure is contributed. Hereinafter, this inclination is referred to as a contribution amount inclination.

  The difference data is calculated using the element extracted from the sensor to be processed in step S304 among the elements of the difference vector between the feature vector and the starting point vector at that time. For example, the sum of these absolute values, the sum of squares, etc. are defined as difference data. Or it is good also as the data of the newest time in them. The difference data has a larger value as the contribution to the abnormality measure is larger.

  When the processing is completed for all the sensors, the sensor having the maximum inclination (the sensor having the largest change rate) is searched for, and the sensor is set as the related sensor. The related sensor number and its inclination are recorded (S706). The section length is also recorded (S707).

  Another embodiment of the related sensor specifying process will be described. Steps S701 to S703 are the same as described above. In the step corresponding to step S704, instead of the difference data, the data at the latest time extracted from the sensor to be processed in step S304 is extracted from the feature vector in the section. What is obtained here is time-series data indicated by a black line 801 in FIG. 8A. In the step corresponding to step S705, the inclination of the data obtained in the step corresponding to step S704 is calculated.

  In a step corresponding to step S706, a sensor having the maximum inclination absolute value is searched for, and that sensor is set as a related sensor. In the step corresponding to step S707, the same processing as in step S707 described above is performed.

  By any of the above processes, it is possible to identify a sensor that contributes to an increase in the abnormality measure.

  Next, as a process for finally determining the presence or absence of an abnormal measure increasing tendency in S127, after the related sensor is specified for each section by the related sensor specifying unit 108, the tendency presence / absence determining unit 109 determines whether or not there is an abnormal measure increasing tendency in each section. Final decision is made. Specifically, rules for the slope recorded in step S706 and the section length recorded in step S707 are determined in advance, and determination is made according to the rules. The rule is, for example, a logical operation of a conditional expression represented by a threshold value and an inequality sign such as “the inclination is greater than TH1 and the section length is greater than TH2”, that is, a combination of “and” and “or”. is there. There may be a plurality of such rules.

  Finally, in S128, the determination result is output from the input / output unit 110. Details of how to output this will be described in the following description of the GUI.

  As shown above, in this example, for the need to detect that the degree of abnormality is gradually worsening, an abnormality measure calculated based on normal data learning is used. Since an upward tendency is detected, an increase in the abnormal measure can be detected without calculating the downward or upward inclination of the time series signal of each sensor.

  For related sensor identification, detect the section where the anomaly measure is rising, calculate the slope of each sensor's contribution to the anomaly measure for each section, and use the maximum slope as the related sensor, so the section where the slope is calculated It is possible to avoid the difficult procedure of adjusting the system appropriately. In addition, since related sensors are automatically identified, it is not necessary for a person to confirm a large number of unrelated sensor signals.

  Although this embodiment is intended for accumulated sensor signals, semi-real time analysis is possible if processing is performed periodically at appropriate intervals according to the target, and an abnormal measure can be obtained at an early stage. It is possible to detect a rising tendency.

  An embodiment of a GUI of a system that realizes the above method will be described.

  An example of a GUI for setting an evaluation target and processing parameters is shown in FIG. In the following description, this setting is simply referred to as recipe setting. The past sensor signal 102 is assumed to be stored as a database in the sensor signal storage unit 103 in association with the equipment ID and time. In the GUI recipe setting screen 901 displayed on the screen of the input / output unit 110, the target device, learning period, evaluation period, used sensor, feature vector creation parameter, normal model parameter, trend determination parameter, and final determination parameter are input. . In the equipment ID input window 902, the ID of the target equipment is input.

  Although a list of device IDs of data stored in the database is displayed by pressing the equipment list display button 903, the list is selected and input from the list. In the learning period input window 904, the start date and end date of the period for which learning data is to be extracted are input. In the evaluation period input window 905, a start date and an end date of a period to be detected as an abnormal measure increasing tendency are input. In the sensor selection window 906, a sensor to be used is input. By clicking the list display button 907, although not shown, a sensor list of the target equipment is displayed.

  In the feature vector creation parameter input window 908, parameters used for processing in the feature vector creation unit 104 are set. The parameters set here are the window width used in the smoothing in step S303, the order used in the feature vector creation in step S304, and the sampling interval.

  In a normal model parameter input window 909, parameters used in normal model creation are input. The figure shows an example in which a local subspace is adopted as a normal model, and the number of neighboring vectors used for model creation and regularization parameters are input. The regularization parameter is a small number added to the diagonal component in order to prevent the inverse matrix of the correlation matrix C from being obtained in Equation 2.

  In the trend determination parameter input window 910, a cycle used in the process in the abnormal measure increasing tendency determining unit 106 and an allowable gap used in the process in the abnormal measure increasing section detecting unit 107 are set. Although the unit is an example of time, any unit representing time such as second, minute, day, etc. may be used, and “unit” may be used in the sense of the number of data.

  In the final determination parameter input window 911, a rule used for processing in the tendency presence / absence determination unit 109 is set. In this example, a plurality of input units are prepared, and a rule whose check box is checked is used. A threshold value for the section length is input to the left side of each stage, and a threshold value for the slope is input to the right side. If any of the checked stages is satisfied, it is determined that there is an upward trend. In the recipe name input window 912, a unique name associated with the input information and the processing result is input. After the above information is input, when the execution button 913 is pressed, the above-described abnormal measure increasing tendency detection is executed.

  After the process is executed, the device ID, the learning period, the evaluation period, the used sensor information, each parameter, and the evaluation result input on the recipe setting screen 901 are stored in a temporary data folder. The evaluation result is stored as feature vector data for the learning period and the evaluation period, the starting point vector data recorded in step S405, the abnormal measure data, and the abnormal measure increasing tendency determination result data obtained by the abnormal measure increasing tendency determining unit 106. The series data and the abnormal measure increasing section information obtained after the processing in the abnormal measure increasing section detecting unit 107. The information on the abnormal measure increasing section includes the section start and end times, the section length, the gradient of the difference data of each sensor, its maximum value, the related sensor name, and the final determination result.

  Then, the processing result is shown to the user by the input / output unit 110. Examples of GUIs for this purpose are shown in FIGS. 10A, 10B, and 10C. By selecting a tab displayed at the top of each screen, the result display screen 1001, the result enlarged display screen 1002, and the abnormal measure increasing section information display screen 1003 can be switched.

  FIG. 10A shows a result display screen 1001 that displays the entire result. The result display screen 1001 displays the abnormal measure for all the specified periods, the abnormal measure increasing tendency determination result, the time series graph of the sensor signal, and the detected abnormal measure increasing section.

  In the period display window 1004, the designated learning period and evaluation period are displayed. In the related sensor name display window 1005, the related sensor name of the specified section is displayed on the related sensor name display unit 10051. Initially, the related sensor name of section number 1 is displayed, and switching is performed by clicking an arrow or directly inputting a section number into the section number display section 10052. The abnormal measure display window 1006 displays the abnormal measure and the abnormal measure increasing tendency determination result (trend) during the designated learning period and evaluation period. In addition, a circle is displayed on the day used for learning.

  In the abnormal measure increasing section display window 1007, the abnormal measure increasing section is displayed in a rectangle. A black rectangle 10071 represents a section in which the final determination result has an upward trend, and a white rectangle 10072 represents a section in which there is no upward trend.

  The sensor signal display window 1008 displays the signal value of the specified sensor for the specified period. A solid line 10081 represents a feature vector, and a broken line 10024 represents an element corresponding to a sensor for which a starting point vector is designated. The sensor to be displayed in the sensor signal display window 1008 is designated by input to the sensor name selection window 1009. However, before the user designates, the related sensor displayed in the related sensor name display portion 10051 of the related sensor name display window 1005 is selected in the sensor name selection window 1009.

  When the section number is switched in the section number display section 10052 in the related sensor name display window 1005, the selected sensor is changed together, and the sensor signal display window 1008 is also redrawn on the graph of the selected sensor. A cursor 1010 represents a starting point for enlarged display and can be moved by a mouse operation. In the display day designation window 1011, the number of days from the start point to the end point of the enlarged display on the result enlarged display screen 1002 is displayed although it is not used on this screen. You can also enter on this screen. The date display window 1012 displays the date at the cursor position. When the end button 1013 is pressed, the result display screen 1001, the result enlarged display screen 1002, and the abnormal measure increasing section information display screen 1003 are erased and the processing ends.

  FIG. 10B shows a result enlarged display screen 1002. In the result enlarged display screen 1002, the abnormal measure of the number of days specified in the display day specification window 1011 and the abnormal measure increase from the date indicated by the cursor 1010 and displayed in the date display window 1012 in the result display screen 1001. A trend determination result and a time series graph of sensor signals are displayed. In the period display window 1004, the same information as the result display screen 1001 is displayed. The related sensor name display window 1005 displays the same information as the result display screen 1001. In this screen, the section number may be switched by the section number display unit 10052.

  In the abnormal measure display window 1006, the abnormal measure rising section display window 1007, and the sensor signal display window 1008, the same information as the result display screen 1001 is displayed in an enlarged manner. In the display day specification window 1011, the number of days from the start point to the end point of the enlarged display is specified. The date display window 1012 displays the date of the starting point of the enlarged display.

  It is also possible to change the display start point with the scroll bar 1014, and this change is reflected in the position of the cursor 1010 and the display of the date display window 1012. The entire length of the scroll bar display area 1015 corresponds to the entire period displayed on the result display screen 1001. The length of the scroll bar 1014 corresponds to the number of days specified in the display day specification window 1011, and the left end portion of the scroll bar 1014 corresponds to the starting point of the enlarged display. The process ends when the end button 1013 is pressed.

  FIG. 10C shows an example of the abnormal measure increasing section information display screen 1003. The abnormal measure increasing section information display screen 1003 displays abnormal measure increasing section information. In the period display window 1004, the same information as the result display screen 1001 is displayed. The related sensor name display window 1005 displays the same information as the result display screen 1001. In this screen, the section number may be switched by the section number display unit 10052. The section information list display window 1016 displays a list of section start and end times, section length, maximum slope value, related sensor names, and final determination results. In the final judgment result, an upward trend is indicated by ◯, and none is indicated by no mark.

  The slope graph display window 1017 displays a bar graph of the slope of the difference data of each sensor for the section specified by the section number display section 10052 of the related sensor name display window 1005. The section information detail display window 1018 displays the start and end times, section length, maximum slope value, associated sensor name, and final determination result of the same section specified in the section number display unit 10052. The process ends when the end button 1013 is pressed.

  When confirmation of the abnormal measure increasing tendency detection result is completed by pressing the end button 1013 on any of the screens shown in FIGS. 10A to 10C, the display returns to the recipe setting screen 901 shown in FIG. By pressing the display button 914 on this screen, the data stored in the temporary data folder of the storage unit 111 can be read and the screens of FIGS. 10A to 10C can be displayed again. When the registration button 915 is pressed, the data stored in the temporary data folder is stored in the storage unit 111 in association with the recipe name input in the recipe name input window 912.

  When the end button 916 is pressed, the data stored in the temporary data folder is deleted, and the recipe setting screen is deleted. Although not shown, the recipe once saved in the storage unit 111 can be called again by designation and the recipe setting screen 901 can be opened. Data stored in the storage unit 111 is copied to a temporary data folder. By pressing the display button 914, the screens of FIGS. 10A to 10C can be displayed. It is also possible to change the conditions for execution. When the recipe name is changed and the registration button 915 is pressed, the data in the temporary data folder is stored in the storage unit 111 in association with the new recipe name.

As described above, in the first embodiment, according to the GUI shown in FIG. 9, one evaluation can be performed with one apparatus and one parameter set, and the result can be displayed and stored.
In the second embodiment, an example of a GUI for periodically performing evaluation with one apparatus and one set of parameters is shown in FIG.

  In the recipe setting screen 1101 in this embodiment, the target device, learning period, evaluation period, used sensor, feature vector creation parameter, normal model parameter, trend determination parameter, and final determination parameter are input. As in the case of the recipe setting screen 901 described with reference to FIG. 9 in the first embodiment, information on the target device and the learning period is input to the facility ID input window 1102 and the learning period input window 1104 as in the GUI.

  In the equipment ID input window 1102, a list of device IDs of data stored in a database (not shown) is displayed when the equipment list display button 1103 is pressed. The end date of the learning period set in the learning period input window 1104 is limited so as not to be later than the current time. In the evaluation period input window 1105, a start date and time 11051 of a period to be detected for an abnormal measure increasing tendency and an evaluation execution interval 11052 are input.

  The sensor selection window 1106, the feature vector creation parameter input window 1108, the normal model parameter input window 1109, the tendency determination parameter input window 1110, and the final determination parameter input window 1111 are the same as those in the recipe setting screen 901 described with reference to FIG. Enter information. In the sensor selection window 1106, the sensor list of the target equipment stored in the database (not shown) is displayed when the list display button 1107 is pressed, and a plurality of selections are input from the list.

  In the recipe name input window 1112, a unique name associated with the input information and the processing result is input. A test is performed by pressing the test button 1113. First, the abnormal measure increasing tendency detection as described in the first embodiment is performed using the period from the start date and time input to the evaluation period input window 1105 to the latest time when the sensor signal of the target device has been acquired as a temporary evaluation period.

  After the process is executed, the input device ID, learning period, evaluation period, used sensor information, each parameter, temporary evaluation period, and evaluation result are stored in the temporary data folder of the storage unit 111, and a result display screen shown in FIG. 10A 1001 is displayed. 10A to C, when the confirmation is completed, the process is terminated by pressing the end button 1013, and the process returns to the recipe setting screen 1101. By pressing the display button 1114, the data stored in the temporary data folder of the storage unit 111 can be read and the screens of FIGS. 10A to 10C can be displayed again.

  When the registration button 1115 is pressed, the data stored in the temporary data folder of the storage unit 111 is stored in the storage unit 111 in association with the recipe name input in the recipe name input window 1112. The next execution date is stored.

  The next execution date / time is the time that is an integral multiple of the execution interval entered in the execution interval 11052 from the start date / time entered in the start date / time 11051 of the evaluation period input window 1105, and is the first time after the period evaluated in the test. And When the end button 1116 is pressed, the data stored in the temporary data folder of the storage unit 111 is deleted, and the recipe setting screen 1101 is deleted.

  After setting the recipe, the system always monitors the latest time when the sensor signal of the target device has been acquired, and if it exceeds the next execution date that was stored in association with the recipe, As an evaluation period, abnormal measure increasing tendency detection is executed, and the processing result and the evaluation period are stored in association with the recipe name. Update the next execution date and time by adding the execution interval to the next execution date and time.

FIG. 12 shows a flow of processing for periodically performing evaluation in this embodiment.
First, the evaluation start date 11051 is set on the GUI screen 1101 shown in FIG. 11 (S1201), and the execution interval is input and set in the execution interval column 11052 (S1202). Next, the following processing is executed at each set execution interval (S1203). First, as in the case described with reference to FIG. 1B in the first embodiment, the sensor signal 102 output from the equipment 101 stored in the sensor signal storage unit 103 is input to the feature vector creation unit 104 (S1204), The vector creation unit 104 creates a feature vector (S1205), the abnormal measure calculation unit 105 calculates an abnormal measure (S1206), and the abnormal measure increase tendency determination unit 106 determines the abnormal measure increase tendency for each time (S1207). The abnormal measure increasing section detecting unit 107 detects a section in which the abnormal measure increasing tendency continues (S1208), the related sensor specifying unit 108 specifies the related sensor for each section (S1209), and the tendency presence / absence determining unit 109 determines each section. Finally, a final determination is made as to whether or not there is an upward trend in the abnormal measure (S1210). To output (S1211).

  As described above, the abnormal measure increasing tendency detection is periodically performed. An example of a GUI for showing the evaluation result to the user is shown in FIG.

  FIG. 13 is an example of a GUI that designates a display target of periodic evaluation results. A display target equipment and recipe are specified from a display target specifying screen 1301. First, a device ID is selected using a device ID selection window 1302. Next, a recipe to be displayed is selected from a list of recipes with the device ID as a target by using a recipe name selection window 1303.

  When the display button 1304 is pressed, the latest processing result and evaluation period stored in association with the recipe name in the storage unit 110 are copied to a temporary data folder, and a result display screen 1001 shown in FIG. 10A is displayed based on the information. . By switching the tabs, it is possible to display the result enlarged display screen 1002 shown in FIG. 10B or the abnormal measure increasing section information display screen 1003 shown in FIG. 10C. When the end button 1013 is pressed, the data in the temporary data folder is deleted and the screen is deleted. When the end button 1305 is pressed on the display target designation screen 1301, the screen is deleted.

  DESCRIPTION OF SYMBOLS 101 ... Equipment 102 ... Sensor signal 103 ... Sensor signal storage part 104 ... Feature vector creation part 105 ... Abnormality measure calculation part 106 ... Abnormality measure upward tendency determination part 107 ... Abnormality Increasing measure section detecting unit 108 ... related sensor specifying unit 109 ... presence / absence determining unit 110 ... input / output unit 111 ... storage unit 901 ... recipe setting screen 1001 ... result display screen 1002 .. Result enlarged display screen 1003... Abnormal measure increasing section information display screen 1101... Recipe setting screen 1201.

Claims (17)

A method of detecting an abnormal tendency of equipment, which processes a plurality of sensor signals output in time series from a plurality of sensors attached to the equipment and creates a feature vector for each of the plurality of sensor signals at each time. And calculating an anomaly measure based on the created feature vector,
Based on the time series change of the calculated abnormal measure, determine an abnormal measure increasing tendency at each time,
An abnormal measure increasing section in which an abnormal measure increasing tendency continues from the determined result, and the detected abnormal measure increasing section based on a time-series change of the contribution amount to the abnormal measure from the plurality of sensors. The sensor that contributed to the abnormal measure increase is identified, and the presence or absence of an abnormal measure increase tendency in the abnormal measure increase section based on the slope of the contribution amount of the identified sensor in the abnormal measure increase section and the length of the abnormal measure increase section An abnormal tendency detection method characterized by determining the above.   The abnormal measure is calculated based on a distance of the feature vector at each time from a normal model created using learning data in a learning period for learning a normal state designated in advance from the feature vector. The method of detecting an abnormal tendency according to claim 1.   2. The method for detecting an abnormal tendency according to claim 1, wherein when the feature vector is created, the sensor signal is smoothed and the feature vector is created using the smoothed sensor signal.   The abnormality trend detection method according to claim 2, wherein the abnormality measure calculated based on the distance from the normal model is smoothed when calculating the abnormality measure.   Based on a predetermined cycle, the abnormal measure increasing tendency for each time to be determined is an abnormal measure that is 1 to a few cycles before the attention time and the attention time and 1 to a few cycles after the attention time. The abnormal tendency detection method according to claim 1, wherein the determination is based on a difference between the two.   A rate of increase by repeating the process of checking whether or not the abnormal measure is increasing by randomly taking out the abnormal measure increasing tendency at each time to be determined from a preset section before and after the attention time and checking whether the abnormal measure is increasing or not. 2. The method according to claim 1, wherein the time is determined to have an increasing tendency of an abnormal measure when the value is equal to or greater than a predetermined threshold value.   The abnormal trend detection method according to claim 1, wherein when detecting the abnormal measure increasing section, an abnormal measure increasing tendency is detected while an interruption period shorter than a predetermined allowable period is regarded as continuous. . Identifying the sensor is an element of a vector difference vector that is a starting point for measuring the distance from the feature vector and the normal model at each sensor time for all sensors in the abnormal measure increasing section. Difference data defined as absolute value sum or square sum or latest time data of those extracted from each of the sensors,
The abnormal trend detection method according to claim 2, wherein an inclination of the difference data in the abnormal measure increasing section is calculated and a sensor having the maximum inclination is specified as the sensor.   In order to identify the sensor, the data of the latest time among the elements of the feature vector extracted from each sensor for each sensor at each time for each sensor in the abnormal measure increasing section. The abnormal trend detection method according to claim 1, further comprising: extracting, calculating a slope of the extracted data in the abnormal measure increasing section, and identifying a sensor having the maximum absolute value of the slope as the sensor.   2. The abnormal trend detection method according to claim 1, wherein the process from the step of creating the feature vector to the step of determining the presence or absence of the abnormal measure increasing tendency is repeatedly executed at predetermined time intervals.   An abnormal tendency detection system for detecting an abnormal tendency of the equipment or device based on sensor signals outputted in time series from a plurality of sensors attached to the equipment or device, wherein the sensors are outputted in time series Raw data storage means for storing signals, feature vector creation means for creating feature vectors at each time from the sensor signals output in time series accumulated in the raw data storage means, and the feature vector creation means An anomaly measure calculating means for calculating an anomaly measure based on the created feature vector at each time, and an anomaly measure for determining an upward trend of the anomaly measure at each time based on a time series change of the anomaly measure calculated by the anomaly measure calculating means In the upward trend determination means and the abnormal measure upward tendency determined by the abnormal measure upward tendency determination means, a section in which the abnormal measure upward trend continues is detected. An abnormal measure increasing section detecting means, and a related sensor specifying means for specifying a sensor contributing to the abnormal measure increasing in the detected abnormal measure increasing section based on a time-series change of the contribution amount to the abnormal measure of the plurality of sensors. And a tendency to determine the presence or absence of an abnormal measure increasing tendency in the abnormal measure increasing section based on the inclination of the contribution amount of the sensor specified by the related sensor specifying means in the abnormal measure increasing section and the length of the abnormal measure increasing section An abnormal tendency detection system comprising: presence / absence determination means.   The abnormal measure calculating means creates a normal model using learning data in a learning period for learning a normal state designated in advance from the feature vector, and the feature at each time from the created normal model The abnormal trend detection system according to claim 11, wherein the abnormal measure is calculated based on a vector distance.   13. The abnormal tendency detection system according to claim 12, wherein the feature vector creating means smoothes the sensor signal and creates the feature vector using the smoothed sensor signal.   13. The abnormal trend detection system according to claim 12, wherein the abnormal measure calculating means further includes an abnormal measure smoothing means for smoothing the abnormal measure calculated based on a distance from the normal model.   The anomaly measure increasing tendency determination means is based on a predetermined period, and the difference between the anomaly measures 1 to before the attention time and the attention time to several predetermined cycles and after one to several specified cycles of the attention time. The abnormal trend detection system according to claim 11, wherein the presence or absence of an upward tendency is determined based on the calculated difference.   The abnormal measure increasing tendency determination means counts a case where the abnormal measure is increasing a number of times by randomly picking out one point from a preset section before and after the attention time and checking whether the abnormal measure is increasing. The abnormal tendency detection system according to claim 11, wherein when the ratio of the number of times of increase is equal to or greater than a predetermined threshold value, the time is determined to have a tendency to increase the abnormal measure.   12. The abnormal trend detecting system according to claim 11, wherein the abnormal measure increasing section detecting means detects a section in which an abnormal measure increasing tendency is continued while considering an interruption period shorter than a predetermined allowable period as continuous. .

Images