JP2020038594A - Abnormality detecting apparatus, abnormality detecting method, and program - Google Patents

Abstract

To provide a technique which, under various conditions, even if the same conditions are not satisfied, can detect abnormality with the same inference model.SOLUTION: The present invention is directed to an abnormality detecting apparatus for establishing an abnormality detecting model for detecting abnormality on the basis of an algorithm to detect any abnormality in an optional device. The apparatus has a data acquiring unit for acquiring data relating to an optionally set item, a model establishing unit for reading data acquired by the data acquiring unit and for carrying out learning by a self-encoder to establish and keep a model and a weight, an inference unit for reading the model and the weight from the model establishing unit and reading a newly collected data from the data acquiring unit to calculate a loss value, and a determining unit for determining a state of an optional device by comparing the loss value calculated by the inference unit and the reference value calculated by the model establishing unit.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an abnormality detection device, an abnormality detection method, and a program. More specifically, an abnormality detection model is constructed using measurement data collected in advance, and an abnormality of a monitoring target is detected based on the abnormality detection model. Related to technology.

  Conventionally, when an abnormality of a monitoring target is detected, a threshold is set for data collected from a current sensor, a vibration sensor, an image sensor, and the like, and a rule-based original model is created and handled. Alternatively, a normal value and an abnormal value are learned by using machine learning (SVM, k-nearest neighbor method, k-average method, LOF method, etc.), and an abnormality is detected by an inference model.

  For example, Patent Literature 1 discloses a diagnostic unit that performs abnormality diagnosis from sensor data of a machine, a false report / unreport detection unit that detects a false report / unreport by comparing a diagnosis result of the diagnostic unit with a history of abnormality of the machine, There is described an abnormality diagnosis / analysis apparatus including a sensor difference detection unit that presents a sensor having different specifications between an individual having a large number of false reports / unreports and an individual having a small number of false reports to an analyst. Thereby, it is said that an analyst who prepares the process of the abnormality diagnosis can see a difference in the specification of the sensor that causes a false report / unreport.

JP-A-2006-133944

  However, in general, industrial equipment has a low frequency of failure, and there are countless failure points and failure causes, so it takes a huge number of samples and time to directly collect data when a failure occurs, that is, when an abnormality occurs in the equipment. .

  Furthermore, there is an individual difference between devices.For example, when anomaly detection is performed using machine learning, even if an inference model is constructed using one device, it cannot be adapted if the devices are different, and prediction greatly deviates. There is also a problem that often occurs. These problems arise from the fact that in supervised machine learning, only limited learning with the same outlier label can be performed under the same condition.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to provide a technique capable of detecting an abnormality using the same inference model even in a situation where the same condition cannot be satisfied under various conditions. To provide.

  In order to solve the above problems, the abnormality detection device according to the present invention is an abnormality detection device that detects an abnormality of an arbitrary device by constructing an abnormality detection model that detects an abnormality based on an algorithm, and is arbitrarily set. A data acquisition unit that acquires data for the item, a model construction unit that reads the data acquired by the data acquisition unit, performs learning by a self-encoder, and constructs and stores models and weights, and a model and weights from the model construction unit By reading the newly collected data from the data acquisition unit and calculating the loss value, and comparing the loss value calculated by the inference unit with the reference value calculated by the model construction unit. A determination unit that determines a state of an arbitrary device.

  Also, the abnormality detection method according to the present invention is an abnormality detection method for detecting an abnormality of an arbitrary device by constructing an abnormality detection model for detecting an abnormality based on an algorithm, and acquiring data for an arbitrarily set item. A data acquisition step for reading, reading the acquired data, performing a learning by a self-encoder, constructing and storing a model and a weight, and reading a constructed model and a weight, and re-reading the newly collected data. It includes a calculating step of reading and calculating a loss value, and a determining step of comparing the calculated loss value with the calculated reference value to determine the state of any device.

  A program according to the present invention is a program for causing a computer to function as the abnormality detection device.

  According to the abnormality detection device of the present invention, under various conditions such as individual differences of arbitrary devices, daily fluctuations, failure locations / failure causes, and differences in motion, the same condition can not be satisfied even in a situation where the same condition cannot be satisfied. An abnormality can be detected by the inference model. Note that the effects described here are not necessarily limited, and may be any of the effects described in the present technology.

It is a block diagram showing composition of an abnormal detection device concerning one embodiment of the present invention. 5 is a flowchart illustrating an abnormality detection method according to an embodiment of the present invention. 4 is a graph illustrating anomaly detection by a self-encoder according to an embodiment of the present invention.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The embodiment described below is an example of a typical embodiment of the present invention, and the scope of the present invention is not construed as being narrow.

  First, an abnormality detection device 100 according to an embodiment of the present invention will be described. FIG. 1 is a block diagram illustrating a configuration of the abnormality detection device 100 according to the present embodiment. FIG. 1 illustrates the configuration in each block in a simplified manner. In FIG. 1, only blocks relating to the present invention are shown, and other necessary components are omitted.

  As shown in FIG. 1, an abnormality detection apparatus 100 acquires input data in one or a plurality of computers, detects an abnormality based on an algorithm, and a data acquisition unit 101 that stores the acquired data. A model construction unit 102 for constructing an anomaly detection model to perform, an inference unit 103 for calculating a required value based on the acquired data, a determination unit 104 for comparing a plurality of values to determine the state of an arbitrary device, And a display unit 105 for displaying important information on a display screen.

  The data acquisition unit 101 includes a control unit 111 that acquires data such as time-series data for an item set arbitrarily from an arbitrary device 106 such as a robot, and a recording medium that stores the data (collected data) acquired by the control unit 111. 112.

  The model construction unit 102 reads the data stored in the data acquisition unit 101 and performs machine learning by a self-encoder (auto encoder) to construct a model and a weight. And a recording medium 114 for storing. Here, the self-encoder refers to an algorithm for dimension compression using a neural network in machine learning.

  The learning unit 113 can also set the level of the compression ratio of the encoding, and calculate the reference value (threshold) in the normal state and its variance based on the constructed model. In the method of calculating the reference value and its variance, the final loss value and variance when the machine learning of the data in the normal state is terminated by the learning unit 113 are used as the reference value and its variance. As an example, the learning unit 113 adjusts the data of the arbitrarily set items as a pre-process, as a pre-processing, with an over-learning avoiding unit that reduces the complexity of the model and avoids over-learning, and aligns the scale between the data items And a scale adjusting unit. The over-learning avoiding unit reduces the number of items by principal component analysis, and the scale adjusting unit performs data scaling by normalization or standardization.

  The inference unit 103 reads the learned model and the weight from the recording medium 114 of the model construction unit 102, reads the unknown data newly collected from the recording medium 112 of the data acquisition unit 101, and calculates a loss value.

  The determination unit 104 compares the loss value calculated by the inference unit 103 with the reference value calculated by the learning unit 113 of the model construction unit 102 to determine whether any device 106 is in a normal state or an abnormal state. Is determined.

  Next, an example of an abnormality detection method using the abnormality detection device 100 will be described with reference to FIG. FIG. 2 is a flowchart illustrating the abnormality detection method according to the present embodiment. In the present embodiment, as an example, a process of detecting an abnormality of a robot that is an arbitrary device 106 will be described. This flowchart starts by executing the inference unit (103), the determination unit (104), and the display unit (105).

  As shown in FIG. 2, first, in S201, the data acquisition unit 101 collects time-series data for an arbitrary item from the robot, which is an arbitrary device 106 in a normal state, via the control unit 111, and stores the data in the storage medium 112.

  Next, in S202, the model construction unit 102 reads the data collected by the data acquisition unit 101 from the storage medium 112 of the data acquisition unit 101, and performs machine learning using a self-encoder. Then, the model construction unit 102 stores the learned model and the weight in the storage medium 114.

  Next, in S203, the inference unit 103 reads the stored learned model and weight from the storage medium 114 of the model construction unit 102.

  Next, in S204, the inference unit 103 collects, through the control unit 111, time series data for the same item as the time series data collected in S201 from any device 106 in a state of unknown whether it is normal or abnormal. Then, the inference unit 103 calculates a loss value from the collected time-series data.

  Next, in S205, the determination unit 104 compares the loss value calculated by the inference unit 103 with a normal state reference value (threshold) calculated by the learned model stored in the model construction unit 102. Specifically, the determination unit 104 determines whether the loss value is equal to or greater than a reference value.

  When the loss value is less than the reference value (when the determination result is No), the process proceeds to S206, and the determination unit 104 determines that the robot as the arbitrary device 106 is in a normal state.

  When the loss value is equal to or larger than the reference value (when the determination result is Yes), the process proceeds to S207, and the determining unit 104 determines that the robot as the arbitrary device 106 is in an abnormal state. The flowchart of FIG. 2 ends by forcibly terminating the inference unit (103), the determination unit (104), and the display unit (105).

  As described above, according to the abnormality detection device 100 and the abnormality detection method of the present embodiment, the same condition cannot be satisfied in various conditions such as individual differences of robots, daily fluctuations, failure locations / failure causes, and differences in motion. Even in a situation, an abnormality can be detected by the same inference model.

  Further, according to the abnormality detection device 100 and the abnormality detection method of the present embodiment, it is found that the loss value of the self-encoder using the numerical value and the time-series data forms separate clusters in the normal state and the abnormal state. Was. Thus, by detecting an "outlier" different from the cluster in the normal state, it is possible to detect an abnormality of an arbitrary device 106. The configuration of the abnormality detection device 100 has an effect of absorbing differences in various conditions, and can be expected to be applicable at a practical level.

<Example>
Next, an embodiment of abnormality detection by the self-encoder using the abnormality detection device 100 will be described with reference to FIG. FIG. 3 is a graph illustrating abnormality detection by the self-encoder according to the present embodiment. In the present embodiment, as an example, a description will be given of the result of abnormality detection processing of two-axis and three-axis belt looseness by machine learning of a six-axis articulated robot that is an arbitrary device 106.

  In the present embodiment, two robots of the six-axis articulated robot with the two-axis and three-axis belts loosened were driven, and a failure sign model was constructed from data analysis of current, speed, and the like. At this time, the degree of the abnormal value increases as the belt slackness increases, and the degree of the abnormal value decreases as the belt slackness decreases.

  As shown in FIG. 3, the horizontal axis represents a value obtained by subtracting the compression ratio from 1 (1-compression ratio), and the vertical axis represents a loss value. Here, a region 301 surrounded by a dotted line below the graph represents a region of a normal value. An area 302 surrounded by a dotted line above the graph represents an abnormal value area. The graph in FIG. 3 shows the degree of looseness of the belt. In the graph, 〇 indicates a reference value of the looseness of the belt, △ indicates that the looseness is small, and × indicates that the looseness is large.

  Then, as shown in the area 301 of FIG. 3, all the reference values の of the slack of the belt could be detected as normal values. Further, as shown in the area 302 of FIG. 3, in the case where the slack of the belt is small Δ and the case where the belt slack is large X, all were detected as abnormal values.

  Here, as shown in FIG. 3, the respective intervals between two reference values 〇 which are the same on the horizontal axis and are arranged vertically, two が which have small looseness △, and × which have two large looseness 同 士Represents the magnitude of the variance due to individual differences between the robots. The variance is large when the interval is large, and small when the interval is small. Then, as shown in an area 301 in FIG. 3, when the compression ratio of the coding is high (in FIG. 3, 1-compression ratio value is around 0.5), the variance between the reference values 〇 increases, and the compression ratio decreases. When the value of 1-compression ratio is about 0.9 in FIG. 3, the variance between the reference values 〇 is small.

  As described above, since the variance between the reference values is reduced by setting the compression ratio to be low, individual differences are reduced, and the reference values can be converged to a predetermined value. Therefore, if the compression ratio is set low, the loss value can be compared with a predetermined reference value to detect an abnormality regardless of the individual difference of the robot, so that the accuracy of the abnormality detection can be improved.

  In the present embodiment, the model is created only from the looseness of the belt, but the created model can also detect abnormal values of other types (such as bearings). In addition, the accuracy of the model can be improved by increasing the collected data of the normal values.

  The present invention relates to an abnormality detection device that detects an abnormality of an arbitrary device, and has industrial applicability.

REFERENCE SIGNS LIST 100 abnormality detection device 101 data acquisition unit 102 model construction unit 103 inference unit 104 determination unit 105 display unit 106 arbitrary device 111 control unit 112, 114 recording medium 113 learning unit 301, 302 area

Claims (7)

An abnormality detection device that detects an abnormality of an arbitrary device by constructing an abnormality detection model that detects an abnormality based on an algorithm,
A data acquisition unit for acquiring data for an arbitrarily set item;
A model construction unit that reads data acquired by the data acquisition unit, performs learning by a self-encoder, constructs a model and weights, and saves the model.
An inference unit that reads a model and a weight from the model construction unit, reads newly collected data from the data acquisition unit, and calculates a loss value.
A determination unit that determines the state of any device by comparing the loss value calculated by the inference unit and the reference value calculated by the model construction unit,
An abnormality detection device comprising: The model construction unit has a learning unit that reads data acquired by the data acquisition unit, and performs machine learning by the self-encoder to construct a model and weights.
The abnormality detection device according to claim 1, wherein the learning unit sets a level of a compression ratio of the encoding, and calculates a reference value in a normal state and a variance thereof based on the constructed model. The learning unit performs pre-processing on the data of the arbitrarily set items, as an over-learning avoiding unit for reducing the complexity of the model and avoiding over-learning, and adjusts the scale between the data items to be the same. A scale adjustment unit,
The abnormality detection device according to claim 1, wherein the overlearning avoiding unit reduces the number of items by principal component analysis, and the scale adjusting unit performs scaling of data by normalization or standardization. An abnormality detection method for detecting an abnormality of an arbitrary device by constructing an abnormality detection model for detecting an abnormality based on an algorithm,
A data acquisition step for acquiring data for an arbitrarily set item;
A model construction step of reading the acquired data, performing learning by a self-encoder, constructing and storing a model and weights,
Reading the constructed model and weights, reading the newly collected data, and calculating a loss value;
A determination step of determining the state of any device by comparing the calculated loss value with the calculated reference value,
Abnormality detection method including. The model building step includes a learning step of reading the obtained data, performing machine learning by the self-encoder to build a model and weights,
5. The abnormality detection method according to claim 4, wherein the learning step includes a calculation step of setting a level of an encoding compression ratio and calculating a reference value in a normal state and a variance thereof based on a constructed model. The learning step includes, as pre-processing, data of arbitrarily set items, an over-learning avoiding step for reducing the complexity of the model and avoiding over-learning, and adjusting a scale between data items. A scaling step;
The abnormality detection method according to claim 4, wherein the overlearning avoiding step reduces the number of items by principal component analysis, and the scale adjusting step performs data scaling by normalization or standardization.   A program for causing a computer to function as the abnormality detection device according to claim 1.

Images