JP6898607B2 - Abnormality sign detection system and abnormality sign detection method - Google Patents

Description

The present invention relates to an abnormality sign detection system and method for machinery and equipment, and particularly to a normal pattern learning method and an abnormality sign detection method in condition monitoring of machinery and equipment.

If infrastructure equipment is shut down due to a sudden failure, a great deal of human and economic loss will occur. Therefore, equipment diagnosis technology to prevent such failures is indispensable.

The conventional concept of anomaly detection has mainly been a method of setting a threshold value for a measured value in advance based on human experience, but individual differences in specifications of each equipment and measurement errors have caused deterioration in accuracy.

With the recent development of data collection infrastructure, it has become possible to easily collect data at all times from sensors installed in equipment. Based on the large amount of collected data, it is possible to directly construct more accurate judgment criteria for abnormal / normal.

However, it is usually difficult to collect all the anomalous data of machinery and equipment. Therefore, a method is conceivable in which a boundary surface in a normal range is constructed only from normal data, and the data deviating from the boundary surface or the ratio of the deviated data is regarded as an abnormality of the equipment.

For such a one-class identification problem, for example, application of a one-class support vector machine (one-class SVM; hereinafter may be referred to as OCSVM) can be considered. However, the judgment device using one one-class support vector machine cannot cope with the case where the state that can be regarded as normal changes with time (for example, the rotation speed of the rotating machine changes with time and the vibration pattern changes accordingly).

Patent Document 1 is a diagnostic method assuming a case where there are a plurality of normal patterns. In Patent Document 1, the training data is divided into clusters by clustering, each cluster is modeled by the subspace method, classified into those having the closest distance to each model, and the distance is defined as the degree of anomaly.

Japanese Patent No. 5048625

Patent Document 1 describes in advance that data in a normal state is divided into clusters over time, but the details of the clustering method are not mentioned, and the number of clusters is unknown, so accurate clustering is performed. Cannot be done. That is, since there are a plurality of normal patterns but the number is unknown, it is not possible to create a set of each normal pattern.

Moreover, since the method of updating the normal class is not mentioned, it is not possible to deal with the case where the normal state changes with time.

The present invention solves the above problems, and an object of the present invention is to provide an abnormality sign detection system and an abnormality sign detection method capable of creating an appropriate class set even when the number of normal classes and the normal range are unknown. It is in.

The abnormality sign detection system according to claim 1 for solving the above problems is
It is an abnormality sign detection system that detects an abnormality sign of machinery and equipment based on the state detection data that detects the state of machinery and equipment.
A data collection unit that collects and records status detection data that detects the status of machinery and equipment targeted for abnormality sign detection,
Of the state detection data collected by the data collection unit, the time-series segment data collected at a certain time is divided into frame units, the feature parameter vector is calculated for each frame unit, and the feature parameter vector is calculated for each frame unit. A feature amount extraction unit that creates a set of feature parameter vectors as segment feature vectors,
It is a segment feature vector created by the feature amount extraction unit, and it is determined whether or not each segment is normal based on a plurality of segment feature vectors that go back a certain time from the current time, and a plurality of segment feature vectors. A class set creation unit that creates a judgment device for identifying one normal / abnormal from each of the above , and creates a set of the created multiple judgment devices as a class set.
Using each judgment device created by the class set creation unit, a class feature vector which is a feature quantity of each judgment device is obtained from the segment feature vector used for creating the judgment device, and the class feature vector is used. A normal class set creation unit that makes the remaining set of judgments a normal class set by making a judgment of the judgment device by the 1-class support vector machine and excluding the judgment device judged to be out of order from the class set. Prepare,
The class set creation unit creates a judgment device for the first segment feature vector among a plurality of segment feature vectors that go back a certain time from the current time by using a method of detecting an outlier, and adds it to the class set. The second and subsequent segment feature vectors are evaluated using the judgment devices belonging to the class set, and a new judgment device is created for the segment feature vectors evaluated to be out of normal by all the judgment devices. It is characterized by adding to the class set.

Further, the abnormality sign detection method according to claim 7 is an abnormality sign detection method for detecting an abnormality sign of machinery / equipment based on state detection data for detecting the state of machinery / equipment.
A data collection step in which the data collection unit collects and records the state detection data that detects the state of the machinery and equipment for which the abnormality sign is detected.
The feature amount extraction unit divides the time-series segment data collected at a certain time from the state detection data collected by the data collection unit into frame units, calculates the feature parameter vector in each frame unit, and calculates the feature parameter vector. A feature amount extraction step for creating a set of feature parameter vectors calculated for each frame as a segment feature vector, and
The class set creation unit is a segment feature vector created by the feature amount extraction unit, and determines whether or not each segment is normal based on a plurality of segment feature vectors that go back a certain time from the current time. , A class set creation step that creates a judgment device for discriminating one normal / abnormal from each of a plurality of segment feature vectors, and creates a set of the created multiple judgment devices as a class set.
The normal class set creation unit obtains the class feature vector, which is the feature quantity of each judgment device, from the segment feature vector used for creating the judgment device, using each judgment device created by the class set creation unit. , The judgment device is judged by the one-class support vector machine using the class feature vector, and the judgment device judged to be out of order is excluded from the class set, so that the remaining set of judgment devices is set as the normal class set. With a class set creation step,
In the class set creation step, among a plurality of segment feature vectors that go back a certain time from the current time, for the first segment feature vector, a determination device is created by using a method of detecting an outlier and added to the class set. The second and subsequent segment feature vectors are evaluated using the judgment devices belonging to the class set, and a new judgment device is created for the segment feature vectors evaluated to be out of normal by all the judgment devices. It is characterized by adding to the class set.

Further, the abnormal sign detection system according to claim 3, in claim 1 or 2, wherein the normal class set creating unit, for each feature parameter vectors of the segment feature vectors were used to create the determiner, 1 class The distances to the identification planes of the support vector machines are obtained, the frequencies of these distances are set as the class feature vectors corresponding to the segment feature vectors, and the class feature vectors are used to determine whether or not the one-class support vector machine is normal. The feature is that the judgment is performed and the judgment device corresponding to the abnormal class feature vector is excluded from the class set as an out-of-order judgment device.

According to the above configuration, the class set creation unit determines whether or not each segment data is normal and creates the class set. Since the normal class set creation unit determines whether or not the class set (set of each judgment device created from multiple segments) that is higher than each segment is normal, the accuracy of normal class set creation is improved. To do. As a result, in the detection of abnormal signs, an appropriate class set can be created even when the number of normal classes and the normal range are unknown.

Further, in the abnormality sign detection system according to claim 2, in claim 1, the class set creation unit executes the creation of the class set in real time or periodically to update the class set sequentially.
The normal class set creation unit is characterized in that the normal class set is created in real time or periodically and the normal class set is sequentially updated.

According to the above configuration, when the state of the normal class changes, it is possible to follow the change.

Further, the abnormality sign detection system according to claim 4 is when the state detection data newly collected by the data collection unit in any one of claims 1 to 3 is collected at a certain time. A segment feature vector is created from the segment data of the series using the feature amount extraction unit, and each determination of the normal class set created by the normal class set creation unit is performed on the feature parameter vector of each frame of the segment feature vector. The feature is that it is further equipped with an in-operation judgment unit that evaluates using a device and judges whether the segment data is normal / abnormal based on the ratio of abnormal frames.

According to the above configuration, it is possible to detect an abnormality in segment units using a normal class set determining device for the newly collected state detection data.

Further, in the abnormality sign detection system according to claim 5 , in any one of claims 1 to 4 , the number of determination devices existing in the normal class set and the deviation excluded by the normal class set creation unit. It is characterized by further including a class diagnosis unit that diagnoses abnormalities in the class set based on the number of judgment devices.

According to the above configuration, it is possible to diagnose the abnormality of the created class set, and based on this, it is possible to grasp the tendency (variation, etc.) and abnormality of the data in the vicinity where the class set is created.

Further, in the abnormality sign detection system according to claim 6 , in any one of claims 1 to 5 , the class set creation unit uses a one-class support vector machine for the first segment feature vector. It is characterized by creating a judge and adding it to the class set.

(1) According to the inventions of claims 1 to 7 , an appropriate class set can be created even when the number of normal classes and the normal range are unknown in the abnormality sign detection.
(2) According to the second aspect of the present invention, when the state of the normal class changes, the change can be followed.
(3) According to the invention of claim 4 , it is possible to detect an abnormality in segment units by using a determination device of a normal class set for newly collected state detection data.
(4) According to the invention of claim 5 , it is possible to diagnose the abnormality of the created class set, and based on this, grasp the tendency (variation, etc.) and abnormality of the data in the vicinity where the class set is created. be able to.

The block diagram of the abnormality sign detection system according to one Embodiment of this invention. FIG. 6 is an image diagram of two-dimensional data showing a state of abnormality / normality determination of OCSVM used in an example of one embodiment of the present invention. An explanatory diagram showing an abnormality sign detection algorithm implemented by the abnormality sign detection device in FIG. 1. An explanatory diagram of a deviation rate calculation algorithm when performing segment determination in an example of one embodiment of the present invention. An explanatory diagram of a class determination algorithm when creating a normal class set from a class set in an embodiment of the present invention. FIG. 5 shows how the class feature vector is obtained from the segment feature vector in the class determination algorithm of FIG. 5, where FIG. 5A is an explanatory diagram showing a distance from the identification surface of the determination device, and FIG. 5B is an explanatory diagram showing the frequency of the distance.

Hereinafter, embodiments of the present invention will be described with reference to the drawings, but the present invention is not limited to the following examples of embodiments. In the present embodiment, the determination by the binary discriminator (judgment device) is repeated to create a set of judgment devices, and the set of judgment devices is evaluated to exclude the existence of the deviating judgment device. This makes it possible to create a more appropriate class set even when the number of normal classes is unknown. In addition, it is possible to follow the change in the normal state by sequentially creating a set of judgment devices for a section that goes back a certain time from the current time and performing an abnormality judgment while updating the normal class set.

FIG. 1 shows the overall configuration of the abnormality sign detection system according to the example of the present embodiment. In FIG. 1, the mechanical equipment 10 (one or more units are provided and two units are shown in FIG. 1) to be diagnosed (target for detecting an abnormality sign) is composed of, for example, electric power equipment and various devices in a power plant. The mechanical equipment 10 is provided with a sensor (not shown) for detecting, for example, voltage, current, motor rotation speed, pressure, temperature, and the like.

The operating state (operating or stopped state) of the mechanical equipment 10 is detected by a detecting means (not shown), and the operating data is transmitted to the abnormality sign detecting device 100 by the communication means 20 together with the data of the sensor.

The sensor data constitutes the state detection data of the present invention. Further, the data is sensed at a certain interval (for example, once an hour) for a certain period of time (for example, 10 seconds).

The sensor data and operation data transmitted via the communication means 20 are stored in the data recording unit 110 in the abnormality sign detection device 100 (the data collection function (data collection unit) of the abnormality sign detection device 100 records the data. Record in section 110).

The feature amount extraction unit 120 divides the sensor data stored in the data recording unit 110 into frames as necessary, calculates a feature parameter vector for each frame, and creates a segment feature vector as a set of feature parameter vectors. Is. The feature amount extraction unit 120 calculates a Fourier coefficient, cepstrum, or the like for data having periodicity such as vibration.

Reference numeral 130 denotes a determination device creating unit that creates a determination device for discriminating between normal and abnormal from the measurement signal (sensor data accumulated in the data recording unit 110) by, for example, a one-class support vector machine.

Reference numeral 140 denotes a segment determination unit for determining whether or not a segment belongs to the class set or the normal class set by a determination device extracted from the class set H or the normal class set H ^{* described later.}

Reference numeral 150 denotes a class set creation unit that creates a set (H) of determination devices created from a plurality of segments that are traced back from the current time by a certain time by processing of the determination device creation unit 130 and the segment determination unit 140.

Reference numeral 160 denotes a class determination unit that determines a class set by extracting a deviation determination device from the determination devices included in the class set H by, for example, a one-class support vector machine, and excluding it from the class set H. The class set H is updated in real time or periodically.

The 170 is a normal class set creation unit that creates a ^{set of the remaining judgment devices (normal class set H *} ) excluding the out-of-order judgment device by the class judgment unit 160 for the class set H created by the class set creation unit 150. Is. The normal class set H ^* is updated in real time or periodically.

Reference numeral 180 denotes a class diagnosis unit that determines an abnormality in each class set based on the number of classes in the normal class set H ^{* and the number of classes in the excluded class set excluded by the class determination unit 160.}

The 190 is an operation in which a segment feature vector is extracted from the newly measured segment using the feature amount extraction unit 120, and an abnormality determination is made by processing of the segment determination unit 140 using the determination device included ^{in the normal class set H *.} It is a medium judgment unit.

In the present embodiment, three hierarchical time widths including a frame, a segment, and a class set are defined for the time series data (sensor data) recorded in the data recording unit 110.

1. 1. Frame: The shortest analysis interval, which is the interval for which the feature parameter vector _CN is analyzed.

2. Segment: A section consisting of multiple consecutive frames. From the set of feature parameter vectors C _m = {c ₁ , c ₂ , ..., C _N } in one segment, one OCSVM determiner h (c) for determining normality / deviation is created.

3. 3. Class set: A set of each judge made from multiple segments that go back a certain period of time.

The abnormality sign detection device 100 of FIG. 1 is composed of, for example, a computer, and includes ordinary computer hardware resources such as ROM, RAM, CPU, input device, output device, communication interface, hard disk, recording medium, and a drive device thereof. ing.

As a result of the collaboration between the hardware resource and the software resource (OS, application, etc.), the abnormality sign detection device 100 includes a data recording unit 110 (data collection unit), a feature amount extraction unit 120, and as shown in FIG. The determination device creation unit 130, the segment determination unit 140, the class set creation unit 150, the class determination unit 160, the normal class set creation unit 170, the class diagnosis unit 180, and the operating determination unit 190 are implemented.

The data recording unit 110 is constructed in a storage means / storage means such as a hard disk or RAM.

As described above, in the present embodiment, the respective parts 110 to 190 in the abnormality sign detection device 100 are provided.

As an embodiment, the class set creation unit of the present invention creates the class set H by using the functions of the segment determination unit 140 and the determination device creation unit 130.

Further, the normal class set creation unit of the present invention, as an embodiment, creates a normal class set H ^* by using the function of the class determination unit 160.

Further, the in-service determination unit of the present invention, as an embodiment, makes a determination using each function of the feature amount extraction unit 120 and the segment determination unit 140.

Next, the operation of the abnormality sign detection system configured as described above will be described with reference to FIGS. 2 to 6. First, it is difficult to collect abnormal data because equipment designed for a long life, such as electrical equipment, rarely breaks down. Therefore, it is necessary to set the abnormality / normal judgment criteria only with the normal data.

As a method of detecting outliers (abnormalities) from data of only one class, for example, there is a one-class support vector machine (OCSVM). OCSVM can also construct a non-linear classification boundary surface (discriminator) by using the kernel method together. FIG. 2 is a diagram showing a state of abnormality / normality determination of a one-class support vector machine using two-dimensional data. The graph represents two-dimensional data as an example, and there is no axis unit. ○ The plot is the measurement data (collected data).

FIG. 3 is a diagram illustrating an abnormality sign detection algorithm implemented by the abnormality sign detection device 100 of FIG. In FIG. 3, the same parts as those in FIG. 1 are indicated by the same reference numerals.

Since the process executed when the class set H is created and the process executed when determining new measurement data are slightly different in the segment determination unit 140, the operation determination unit 190 in FIG. 3 is different from the segment determination unit 140a. It is described as a segment determination unit 140b in the class set creation unit 150.

Further, 210 in FIG. 3 is a storage unit ^{in which the normal class set H *} created by the normal class set creation unit 170 is stored. The class set H created by the class set creation unit 150 and the excluded class set, which is a set of out-of-class classes excluded by the class determination unit 160, are also stored in the same storage unit as the storage unit 210. It is not shown in the figure.

^{First, assuming that the normal class set H *} created by the normal class set creation unit 170 is already stored in the storage unit 210, the operation of the operating determination unit 190 is shown in FIG. 3 and the deviation rate calculation algorithm. This will be described with reference to FIG.

The feature amount extraction unit 120 divides the new measurement data recorded in the data recording unit 110 into frames as necessary, calculates a feature parameter vector for each frame, and segments feature vectors as a set of feature parameter vectors. Create C (step S101).

The segment determination unit 140a takes out the class determination device h (c) in order from the ^{normal class set H * stored in the storage unit 210, and performs the following determination.}

(A) The feature parameter vector c is taken out in order from the segment feature vector C, and a determination is made by h (c) (step S102).

(B) All the feature parameter vectors c are determined, and the deviation rate _Te = count (h (c) == FALSE) / N is calculated (step S103).

(C) a threshold value compared to the T _e as T _seg, when T _e <T _seg, segment assumes that corresponds to h, the process ends. If not, the process returns to step S102 and evaluation is performed using another determination device h (c).

(D) If the corresponding class is not found for all h (c), the segment alerts as an "abnormal" segment.

Next, the class set creation algorithm implemented by the class set creation unit 150 will be described. The class set creation unit 150 creates a set H of determination devices, which is created from a plurality of segments that are traced back from the current time by a certain time, by the following procedure using the processing of the segment determination unit 140b.

(A) obtaining a segment feature vector C from the respective (C ₁ -C _N) segments until a predetermined time past (processing similar to step S101 of FIG. 4).

(B) From the segment feature vector C ₁ of the first segment, the determination device creation unit 130 creates a determination device by OCSVM and adds it to the class set H.

(C) For the second segment feature vector C ₂ -C _N subsequent segments is evaluated using any determiner belonging to class set H ( "set of class set (determiner that now in Figure 3 ) ”) (Processing similar to steps S102 and S103 in FIG. 4). Then, if even one judgment device belongs, no judgment device is created from that segment, and when all the judgment devices are evaluated as out of order, a new judgment device is created and added to the class set H (FIG. 3). "If it does not exist in the class set, create a new judge and add it to the set").

Next, the normal class set creation algorithm implemented by the normal class set creation unit 170 will be described with reference to FIGS. 3, 5, and 6. The class set H covers the characteristics of the segment within a certain period of time in the past, and there is a possibility that the class of the abnormal segment is included in this. Therefore, the normal class set creation unit 170 further excludes the classes that are out of the class set H by the class determination unit 160, and sets the remaining class set as the normal class set H ^* .

The class determination unit 160 executes the processing of the class determination algorithm shown in FIG. 5 as follows.

1. 1. The segment feature vectors C used to create each determiner of the class set H are collected and set as C ^∀ (step S201).

2. The following processing is performed on all the judgment devices h ₁ (c), h ₂ (c), ... Of the class set H to obtain the class feature vector which is the feature quantity of each judgment device.

(A) Each judgment device h is used to evaluate the ^{corresponding C ∀} _{(steps S202 -1} , S202 _-2 , ...). At this time, for each feature parameter vector c, the distance from the identification surface (boundary surface) of the OCSVM is obtained as shown in FIG. 6A. If this sign is positive, it is judged to be normal, and if it is negative, it is judged to be off.

(B) A histogram of the distance (frequency characteristics for each distance) obtained in (a) above is created as shown in FIG. 6 (b) and used as a class feature vector d (d ₁ , d ₂ , ...) (Step S203). _-1 , S203 _-2 , ...). The method for determining the bin width of the histogram is determined as follows, for example, using Scott's selection method.

Here, σ is the standard deviation of the data, and n is the total number of data.

3. 3. The OCSVM determination is performed using the class feature vector d (d ₁ , d _{2, ...) Obtained through the above processing (step S204).} Then, the class considered to be normal is set as the normal class set H ^* , and the outlier class is excluded and used as the excluded class set.

Each process of the class set creation unit 150 and the normal class set creation unit 170 is executed in real time or periodically, and the class set H, the normal class set H ^*, and the excluded class set are sequentially updated.

Next, the operation of the class diagnosis unit 180 will be described.

In the class diagnosis unit 180, the ^{number of classes of the normal class set H *} (the number of judgment devices existing in the normal class set) and the number of classes of the excluded class set excluded from the class set H by the class judgment unit 160 (excluded). If the ratio of the number of outliers) or the rate of change of the class exceeds a certain threshold, an alert is issued.

The class diagnosis unit 180 is not limited to the above, and may perform the above processing based on the number of ^{classes of the class set H and the normal class set H *.}

By the processing of the class diagnosis unit 180, it is possible to diagnose the abnormality of the created class set, and based on this, it is possible to grasp the tendency (variation, etc.) and abnormality of the data in the vicinity where the class set is created.

The class set creation unit 150 and the normal class set creation unit 170 sequentially create a set of judgment devices for a section that goes back a certain period of time from the current time, ^{and update the normal class set H *} while operating the judgment unit. The abnormality may be determined by 190. As a result, it is possible to follow the change in the normal state.

10 ... Mechanical equipment 20 ... Communication means 100 ... Abnormal sign detection device 110 ... Data recording unit 120 ... Feature quantity extraction unit 130 ... Judgment device creation unit 140 ... Segment judgment unit 150 ... Class set creation unit 160 ... Class determination unit 170 ... Normal Class set creation unit 180 ... Class diagnosis unit 190 ... Operational judgment unit 210 ... Storage unit

Claims (7)

It is an abnormality sign detection system that detects an abnormality sign of machinery and equipment based on the state detection data that detects the state of machinery and equipment.
A data collection unit that collects and records status detection data that detects the status of machinery and equipment targeted for abnormality sign detection,
Of the state detection data collected by the data collection unit, the time-series segment data collected at a certain time is divided into frame units, the feature parameter vector is calculated for each frame unit, and the feature parameter vector is calculated for each frame unit. A feature amount extraction unit that creates a set of feature parameter vectors as segment feature vectors,
It is a segment feature vector created by the feature amount extraction unit, and it is determined whether or not each segment is normal based on a plurality of segment feature vectors that go back a certain time from the current time, and a plurality of segment feature vectors. A class set creation unit that creates a judgment device for identifying one normal / abnormal from each of the above , and creates a set of the created multiple judgment devices as a class set.
Using each judgment device created by the class set creation unit, a class feature vector which is a feature quantity of each judgment device is obtained from the segment feature vector used for creating the judgment device, and the class feature vector is used. A normal class set creation unit that makes the remaining set of judgments a normal class set by making a judgment of the judgment device by the 1-class support vector machine and excluding the judgment device judged to be out of order from the class set. Prepare,
The class set creation unit creates a judgment device for the first segment feature vector among a plurality of segment feature vectors that go back a certain time from the current time by using a method of detecting an outlier, and adds the determination device to the class set. The second and subsequent segment feature vectors are evaluated using the judgment devices belonging to the class set, and a new judgment device is created for the segment feature vectors evaluated to be out of normal by all the judgment devices. Anomalous sign detection system characterized by adding to the class set. The class set creation unit executes the creation of the class set in real time or periodically to update the class set sequentially.
The abnormality sign detection system according to claim 1, wherein the normal class set creation unit sequentially executes the creation of the normal class set in real time or periodically to update the normal class set. The normal class set creation unit obtains the distances from the identification surface of the one-class support vector machine for each feature parameter vector of the segment feature vector used to create the determination device, and determines the frequency of these distances. A class feature vector corresponding to the feature vector is used, and the class feature vector is used to judge whether the class is normal or not by a one-class support vector machine. The abnormality sign detection system according to claim 1 or 2, wherein the system is excluded from the above. Of the state detection data newly collected by the data collection unit, a segment feature vector is created from the time-series segment data collected at a certain time using the feature amount extraction unit, and the segment feature vector of the segment feature vector is created. The feature parameter vector of each frame is evaluated using each judgment device of the normal class set created by the normal class set creation unit, and the normality / abnormality of the segment data is judged based on the ratio of the abnormal frames. The abnormality sign detection system according to any one of claims 1 to 3, further comprising a middle determination unit. It is characterized by further including a class diagnosis unit that diagnoses an abnormality in the class set based on the number of judgment devices existing in the normal class set and the number of deviation judgment devices excluded by the normal class set creation unit. The abnormality sign detection system according to any one of claims 1 to 4. The class set creation unit according to any one of claims 1 to 5, wherein the first segment feature vector is added to the class set by creating a determination device using a one-class support vector machine. The described anomaly sign detection system. It is an abnormality sign detection method that detects an abnormality sign of machinery and equipment based on the state detection data that detects the state of machinery and equipment.
A data collection step in which the data collection unit collects and records the state detection data that detects the state of the machinery and equipment for which the abnormality sign is detected.
The feature amount extraction unit divides the time-series segment data collected at a certain time from the state detection data collected by the data collection unit into frame units, calculates the feature parameter vector in each frame unit, and calculates the feature parameter vector. A feature amount extraction step for creating a set of feature parameter vectors calculated for each frame as a segment feature vector, and
The class set creation unit is a segment feature vector created by the feature amount extraction unit, and determines whether or not each segment is normal based on a plurality of segment feature vectors that go back a certain time from the current time. , A class set creation step that creates a judgment device for discriminating one normal / abnormal from each of a plurality of segment feature vectors, and creates a set of the created multiple judgment devices as a class set.
The normal class set creation unit obtains the class feature vector, which is the feature quantity of each judgment device, from the segment feature vector used for creating the judgment device, using each judgment device created by the class set creation unit. , The judgment device is judged by the one-class support vector machine using the class feature vector, and the judgment device judged to be out of order is excluded from the class set, so that the remaining set of judgment devices is set as the normal class set. With a class set creation step,
In the class set creation step, among a plurality of segment feature vectors that go back a certain time from the current time, for the first segment feature vector, a determination device is created by using a method of detecting an outlier and added to the class set. The second and subsequent segment feature vectors are evaluated using the judgment devices belonging to the class set, and a new judgment device is created for the segment feature vectors evaluated to be out of normal by all the judgment devices. Anomalous sign detection method characterized by adding to the class set.

Images