WO2022091248A1

WO2022091248A1 - Degradation detection device

Info

Publication number: WO2022091248A1
Application number: PCT/JP2020/040417
Authority: WO
Inventors: 裕貴太中; 浩司脇本
Original assignee: 三菱電機株式会社
Priority date: 2020-10-28
Filing date: 2020-10-28
Publication date: 2022-05-05
Also published as: JPWO2022091248A1; JP6961126B1

Abstract

A degradation detection device according to the present disclosure is characterized by comprising: a learning unit that learns and constructs a decision tree for inferring a category of degradation events to which data belongs from the feature of the data on the basis of learning data; a feature amount extraction unit that calculates a feature amount from input time-series sensor data; and a degradation candidate inference unit that selects a feature amount having high evaluation from the calculated feature amount and infers a candidate of plausible degradation events from the learned decision tree, wherein the learning unit groups the degradation events to learn and construct the decision tree.

Description

Deterioration detection device

This disclosed technology relates to a deterioration detection device.

At various sites, there is a need to improve the equipment utilization rate by predictive maintenance of equipment. To meet this need, a prediction model is created, the values of multiple types of features calculated from the operating state data are acquired, and the abnormalities that occur in the device are obtained from the values of each type of features acquired during normal and abnormal conditions. The technique of associating with each kind of feature amount is disclosed. Further, it is disclosed that a decision tree is used as an algorithm for specifying the degree of association between an abnormality occurring in an apparatus and each type of feature (for example, Patent Document 1). A machine learning method for creating a decision tree from data is called decision tree learning, or simply decision tree for short.

Japanese Unexamined Patent Publication No. 2018-116545

In the prior art exemplified in Patent Document 1, when the deterioration event of the device is defined, a decision tree is created to try to classify all the defined deterioration events. However, sometimes, even when looking at the distribution of features of each type, there may be deterioration events that are difficult to classify because they mix with each other. At this time, if a plurality of deterioration events are selected and grouped by trial and error, the correct answer rate of the classification is improved, and the nature of the grouped deterioration events may be clarified by the decision tree. However, the prior art does not disclose what degradation events should be grouped and how grouping can be incorporated into decision tree learning.

The deterioration detection device according to the present disclosure includes a learning unit that learns and constructs a decision tree that infers the category of deterioration events to which the data belongs from the feature amount of the data based on the learning data, and an input time-series sensor. A feature amount extraction unit that calculates a feature amount from data, and a deterioration candidate inference unit that selects a feature amount with a high evaluation from the calculated feature amount and estimates a plausible deterioration event candidate from the learned decision tree. , The learning unit is characterized in that the deterioration event is grouped and the decision tree is learned and constructed.

Since the deterioration detection device according to the present disclosure technique has the above configuration, deterioration events are grouped by decision tree learning, the correct answer rate of classification is improved, and the nature of the grouped deterioration events is clarified by the decision tree. Can be done.

FIG. 1 is a block diagram showing a configuration of a deterioration detection factor analysis system according to the first embodiment. FIG. 2 is a schematic diagram showing a process performed by a deterioration candidate inference unit of the deterioration detection device according to the first embodiment. FIG. 3 is a flowchart showing a processing flow of the inference phase of the deterioration detection device according to the first embodiment. FIG. 4 is a block diagram showing a configuration of a learning unit of the deterioration detection device according to the first embodiment. FIG. 5 is a block diagram showing a configuration of a model generation unit of the deterioration detection device according to the first embodiment. FIG. 6 is a diagram showing an example of an adjacency matrix generated by the adjacency matrix calculation unit of the feature amount evaluation unit according to the first embodiment. FIG. 7 is a flowchart showing a processing flow of the learning phase of the deterioration detection device according to the first embodiment.

The form for implementing the disclosed technique will be clarified by the following description along with the drawings. Further, the description of each embodiment includes the description of the inference phase and the learning phase in decision tree learning. The learning related to the disclosed technique is a kind of "supervised learning" because the learning data is labeled. Labeled data refers to data that is labeled to indicate which category the data belongs to.

Embodiment 1.
The deterioration detection device 100 according to the first embodiment constitutes the deterioration detection factor analysis system 1000 in the inference phase in the decision tree learning. FIG. 1 is a block diagram showing a configuration of a deterioration detection factor analysis system according to the first embodiment.

As shown in FIG. 1, the deterioration detection factor analysis system 1000 includes a deterioration detection device 100 and an external device 200. More specifically, the deterioration detection factor analysis system 1000 includes sensor ESs (ES1, ES2, ..., ESn) attached to n units (n is an integer of 1 or more) and these sensor ESs (ES1, ES1). The deterioration detection device 100 that receives vibration data or current data acquired from each of ES2, ..., ESn) via the communication network NW, and the user inputs various settings and displays the output result of the deterioration detection device 100. It is composed of an external device 200.

The deterioration detection device 100 performs deterioration diagnosis and factor analysis on vibration data or current data (hereinafter, generally referred to as "sensor data") received from the sensor ES (ES1, ES2, ..., ESn). The deterioration detection device 100 stores a state descriptor, a geographical descriptor, or a temporal descriptor indicating the result of deterioration detection in association with the sensor data D2, and deteriorates in response to an input from the external device 200. Display the possibility and deterioration factors. Here, the state descriptor is an integer that distinguishes the state of the electric device to which the sensor is attached. The geographical descriptor is an integer that distinguishes the positions of the sensors ES (ES1, ES2, ..., ESn) and the like. The temporal descriptor is an integer that distinguishes the sensor data acquisition time and the like.

The deterioration detection device 100 includes a labeled data acquisition unit 1, a feature amount extraction unit 2, a deterioration candidate inference unit 3, a storage unit 6, an interface unit 7, and a learning unit 300. The operation of the deterioration detection device 100 according to the first embodiment will be clarified by the following description of the operation of each part.

The labeled data acquisition unit 1 of the deterioration detection device 100 receives the distribution data D1 from the sensors ES (ES1, ES2, ..., ESn). The distribution data D1 includes sensor data D2 from each sensor and data consisting of information concomitant to the sensor data D2. The labeled data acquisition unit 1 refers to the information regarding the data structure stored in the storage unit 6 and extracts the sensor data D2 from the received distribution data D1. The labeled data acquisition unit 1 outputs the extracted sensor data D2 to the feature amount extraction unit 2.

The feature amount extraction unit 2 of the deterioration detection device 100 calculates the feature amount such as the mean value, the amplitude, and the spectrum peak from the time-series sensor data D2 input from the labeled data acquisition unit 1. The calculated feature amount is output to the deterioration candidate inference unit 3 as the feature amount data D3.

The deterioration candidate inference unit 3 of the deterioration detection device 100 selects a feature amount that is highly evaluated by the feature amount evaluation descriptor described later with respect to the feature amount data D3 output by the feature amount extraction unit 2. The deterioration candidate inference unit 3 measures the distance from the feature quantity distribution that has been learned and is created by each of the deterioration events, which will be described later, and estimates a plausible deterioration event candidate from the measured distance. The deterioration candidate inference unit 3 outputs the deterioration event candidate to which the data is presumed and the estimation accuracy thereof to the interface unit 7. The estimation accuracy may be a calculation of the probability using Bayes' theorem. In the present disclosed technique, the learned deterioration events include a "deterioration event group" in which a plurality of deterioration events are grouped.

Specifically, the inference process performed by the deterioration candidate inference unit 3 is performed using a decision tree. The decision tree consists of nodes and branches. In the decision tree in the present disclosure technique, a node has a type of feature amount used for classification, a classification surface of the feature amount for classification, and a deterioration event as a result of classification.

The storage unit 6 of the deterioration detection device 100 stores various information, and is realized by a storage device such as a hard disk. In the first embodiment, the storage unit 6 stores the feature amount data D3 and the feature amount evaluation descriptor D5 described later in association with each other.

The interface unit 7 of the deterioration detection device 100 connects the external device 200 and each part of the deterioration detection device 100 to enable communication and various controls. The interface unit 7 sets the accuracy to be obtained in the deterioration factor estimation, and is used to confirm the factor estimation result.

The processing performed by the feature amount extraction unit 2 will be clarified by the following specific example. The target device in the specific example is a platform door for the purpose of preventing a fall from a platform and a contact accident with a train at a railway station. When this device is targeted, the feature quantities include a door opening / closing start position, torque average, torque standard deviation, maximum torque value, and the like. Further, the feature amount may include an ID indicating an electric device to which the sensor is attached and a data acquisition time.

The processing performed by the deterioration candidate inference unit 3 is clarified by the following specific example. FIG. 2 is a schematic diagram showing a process performed by a deterioration candidate inference unit of the deterioration detection device according to the first embodiment. FIG. 2A shows the process A performed by the deterioration candidate inference unit 3 in the learning phase. As a specific example of the platform door, four categories of belt tension, belt loosening, twist installation a, and twist installation b are considered as deterioration events. The graph on the left of FIG. 2A is an example of a feature space in which the horizontal axis is the feature X5 and the vertical axis is the feature X6. This graph shows the distribution in which the features in each deterioration event are plotted. Specifically, the process A measures the sensor data D2 for a plurality of home doors in an abnormal state such as "belt tension", and calculates the feature amount X5 and the feature amount X6 from the measured sensor data D2. , Plot in the feature space. Process A plots platform doors in other categories of degradation events as well. Although FIG. 2A can express only a two-dimensional space, the actual feature space is a p-dimensional space if the number of types of features is p.

Looking at the feature space represented by the graph on the left of FIG. 2A, the plot for belt tension and belt loosening is where the feature X5 is small, and the plot for twist installation a and twist installation b is the feature X5. Are distributed in large areas. Since these two distributions do not overlap, they can be classified by determining the boundaries of the regions. The deterioration candidate inference unit 3 learns the boundary of the region for the classification of this plot. In a one-dimensional feature space, the boundaries of the regions are represented by thresholds. In the two-dimensional feature space as shown in the graph on the left of FIG. 2A, the boundary of the region is represented by a straight line. Generally, in an N-dimensional feature space, the boundary of this region is called a classification plane. For the classification surface, it is conceivable to use a support vector machine algorithm that maximizes the margin, which is the distance between the sample closest to the classification surface and the classification surface. The method of determining the classification surface is not limited to this, and may be obtained by learning by another algorithm.

The deterioration candidate inference unit 3 of the deterioration detection device 100 that has learned the classification surface then constructs a corresponding decision tree. The figure to the right of FIG. 2A shows the decision tree in this embodiment. The feature amount X5 and the feature amount X6 corresponding to the horizontal axis and the vertical axis of the feature amount space correspond to the "type of feature amount used for classification" of the node of the decision tree. In addition, the learned classification surface corresponds to the "classification surface of the feature quantity for classification" of the node of the decision tree. Further, the two categories classified by the classification plane correspond to the "deterioration event as a result of classification" of the node of the decision tree. In the decision tree in the figure on the right of FIG. 2A, the deterioration candidate at the tip of the two branches corresponds to the "deterioration event as a classification result". In this specific example, on the left side of the classification surface, belt tension and belt loosening can be mentioned as candidates for deterioration events. On the right side of the classification surface, twist installation a and twist installation b are listed as candidates for deterioration events.

FIG. 2B shows the process B performed by the deterioration candidate inference unit 3 in the inference phase. The graph on the left side of FIG. 2B is a feature space in which the feature X5 and the feature X6 are plotted for the platform door to be investigated. The graph on the right side of FIG. 2B is the same as the graph on the left side of FIG. 2A. The image of the process B performed by the deterioration candidate inference unit 3 is to infer which category the plot for the platform door to be investigated belongs to by comparing it with the learned past plot. For example, when the plot for the platform door to be investigated is on the left side of the classification plane, the deterioration candidate inference unit 3 infers that the deterioration event is belt tension or belt loosening. That is, the deterioration candidate inference unit 3 infers the category of deterioration events from the plot of the feature amount to be investigated by using the decision tree learned in the learning phase.

The deterioration event candidate inferred by the deterioration candidate inference unit 3 may be output to the interface unit 7 as a predetermined deterioration candidate descriptor. That is, in the deterioration detection device according to the present disclosure technique, deterioration events are numbered and defined in advance. In a specific example of a platform door, the deterioration candidate descriptor is 1, if the belt is meandering, 2, if the belt is loose, 3, if the belt is tight, 9 if it is normal, and so on. The deterioration candidate descriptor is stored in the storage unit 6 together with the information of the above-learned decision tree.

The operation of the deterioration detection device 100 in the inference phase will be clarified by the following explanation along the flowchart. FIG. 3 is a flowchart showing a processing flow of the inference phase of the deterioration detection device 100 according to the first embodiment.

As shown in FIG. 3, in the processing operation in the inference phase of the deterioration detection device 100, a step of extracting a feature amount (ST1), a step of inferring a deterioration event (ST2), and a step of outputting the inferred result (ST3). ) And.

The learning unit 300 of the deterioration detection device 100 is used in the learning phase. FIG. 4 is a block diagram showing the configuration of the learning unit 300 of the deterioration detection device 100 according to the first embodiment. As shown in FIG. 4, the learning unit 300 has a learning data acquisition unit 301 and a model generation unit 302. The model generation unit 302 is connected to the storage unit 6.

The model generation unit 302 is further subdivided into constituent members. FIG. 5 is a block diagram showing a configuration of a model generation unit 302 of the deterioration detection device 100 according to the first embodiment. As shown in FIG. 5, the model generation unit 302 has an inter-distribution distance calculation unit 303, a feature amount evaluation unit 304, and an aggregation unit 305. The model generation unit 302 and the storage unit 6 are connected to each other via the aggregation unit 305 of the model generation unit 302. The feature amount evaluation unit 304 further includes an adjacency matrix calculation unit 311, a separation degree evaluation unit 312, and a similarity evaluation unit 313.

The learning data acquisition unit 301 of the deterioration detection device 100 acquires data as a set of various feature quantities of the investigation target in the state of the deterioration event and the category of the deterioration event as the learning data D11.

The model generation unit 302 of the deterioration detection device 100 learns and constructs a decision tree based on the learning data D11 input via the learning data acquisition unit 301. The decision tree is an algorithm that distinguishes from which of many variables the most useful classification conditions can be obtained. In determining the superiority or inferiority of this classification condition, the decision tree uses the impureness of category identification as an index. In addition, information gain is used to score which variable is useful as a branching condition. The decision tree constructed by the model generation unit 302 is constructed by comparing the information gains of the learning data D11 by changing the type of the feature amount to be used and the classification surface to be used. By increasing the training data D11, the model generation unit 302 learns the classification surface and the decision tree.

The model generation unit 302 of the deterioration detection device 100 outputs the learned decision tree to the storage unit 6 as the learning model D12.

The storage unit 6 of the deterioration detection device 100 stores the trained learning model D12 output by the model generation unit 302.

The details of the operation of the model generation unit 302 will be clarified by the explanation of the operation of the inter-distribution distance calculation unit 303, the feature amount evaluation unit 304, and the aggregation unit 305, which are the components of the model generation unit 302.

The inter-distribution distance calculation unit 303 of the model generation unit 302 calculates the distance between the distributions of the two types of plots in the feature space for the data with two types of labels in the learning data D11. Since the present disclosure technique considers the impureness of category identification to be important, the distance between distributions is defined in terms of impureness. Purity is a concept that is paired with impureness. The purity is a value obtained by dividing the number of data that can be correctly classified by the total number of data when the learning data D11 is classified in terms of classification. Being able to classify with high purity can be interpreted as having a sufficient distance from each other. Therefore, the present disclosure technique defines this purity as the distance between two distributions classified by the classification plane.
When classifying by the threshold value of one feature amount, it is conceivable to define the distance between distributions by Area Under The Curve (hereinafter referred to as “AUC”). AUC is an area under the Receiver Operating Characteristic Curve (hereinafter referred to as “ROC curve”) and takes a value from 0 to 1. An AUC of 1 indicates that there is no impureness and the two distributions are sufficiently separated.

The feature amount evaluation unit 304 of the model generation unit 302 is further divided into an adjacency matrix calculation unit 311, a separation degree evaluation unit 312, and a similarity evaluation unit 313. The details of the operation of the feature amount evaluation unit 304 will be clarified by the explanation of each operation of the adjacency matrix calculation unit 311 which is a component of the feature amount evaluation unit 304, the separation degree evaluation unit 312, and the similarity evaluation unit 313. ..

The adjacency matrix calculation unit 311 of the feature quantity evaluation unit 304 uses the adjacency matrix F as an element of the distance between the distributions of all the labels based on the information of the distance between the distributions of the two labels output by the distance calculation unit 303. Generate and output the adjacency matrix F to the separation evaluation unit 312 and the similarity evaluation unit 313. Here, the adjacency matrix F is a square matrix representing a finite graph used in graph theory and computer science. FIG. 6 is a diagram showing an example of an adjacency matrix F generated by the adjacency matrix calculation unit 311 of the feature quantity evaluation unit 304 according to the first embodiment. As shown in FIG. 6, since the adjacency matrix F in the disclosed technique is an undirected graph, it is a triangular matrix in the upper half. Also, since the diagonal component represents the distance to itself, all the elements are 0.

The separation evaluation unit 312 of the feature amount evaluation unit 304 determines whether or not the separation distance condition is satisfied for each element _{Fi, j} of the adjacency matrix F output by the adjacency matrix calculation unit 311, and determines whether or not the separation distance condition is satisfied, and the adjacency matrix F Binarize each element _{Fi and j} . The separation distance condition is given by the following equation.
IF F _{i, j} > 1-ε ₁ THEN G _{i, j} = 1, ELSE G _{i, j} = 0 ... (1)
However, ε ₁ is a parameter of the separation distance condition. Further, the matrix obtained by binarizing the elements _{Fi and j} of the adjacency matrix F is called the separation evaluation matrix G. The separation evaluation unit 312 outputs this separation evaluation matrix G to the aggregation unit 305.

The similarity evaluation unit 313 of the feature quantity evaluation unit 304 determines whether or not the similarity distance condition is satisfied for each element _{Fi, j} of the adjacency matrix F output by the adjacency matrix calculation unit 311, and determines whether or not the similarity distance condition is satisfied, and the adjacency matrix F Binarize each element _{Fi and j} . The similar distance condition is given by the following equation.
IF F _{i, j} <ε ₂ THEN Hi _{i, j} = 0, ELSE Hi _{, j} = 1 ... (2)
However, ε ₂ is a parameter of the similar distance condition. Further, the matrix obtained by binarizing the elements _{Fi and j} of the adjacency matrix F is called the similarity evaluation matrix H. The similarity evaluation unit 313 outputs this similarity evaluation matrix H to the aggregation unit 305.

The aggregation unit 305 of the model generation unit 302 selects the category of deterioration events to be grouped from the separation evaluation matrix G and the similarity evaluation matrix H output by the feature quantity evaluation unit 304. Whether to use the separation evaluation matrix G or the similarity evaluation matrix H is a design matter, but after all, the deterioration events to be grouped are difficult to classify in the classification by the classification surface indicated by the decision tree, and the classification is performed. Even so, select deterioration events with high impurities.
After selecting the deterioration events to be grouped, the model generation unit 302 reflects the information of the grouped "deterioration event group" in the learning of the classification surface and the decision tree. The grouping here may change as the learning progresses.

The operation of the deterioration detection device 100 in the learning phase will be clarified by the following explanation along the flowchart. FIG. 7 is a flowchart showing a processing flow of the learning phase of the deterioration detection device 100 according to the first embodiment.

As shown in FIG. 7, the processing operation in the learning phase of the deterioration detection device 100 includes a step of acquiring learning data D11 (ST21), a step of calculating the distance between distributions (ST22), and a step of calculating an adjacency matrix. (ST23), a step of calculating the separation evaluation matrix G (ST24), a step of calculating the adjacency matrix H (ST25), a step of selecting a category of deterioration events to be grouped (ST26), and grouping. It has a step of reflecting and advancing the learning (ST27) and a step of confirming whether to end the learning (ST28).

The step (ST28) for confirming whether to finish learning requires explanation, so it is performed here. In general, learning methods are divided into batch learning and online learning. In the case of batch learning, it is considered that this step (ST28) is not necessary because all the training data D11 is used at once. However, when learning data D11 having different properties such as a difference in the type of the target device is used, there may be a case where it is desired to perform learning step by step and observe the difference in the learning result. In such cases, this step (ST28) is included.

With the above configuration, the deterioration detection device 100 according to the first embodiment improves the correct answer rate of classification and can clarify the nature of grouped deterioration events by a decision tree.

1 Labeled data acquisition unit; 2 Feature amount extraction unit; 3 Deterioration candidate inference unit; 6 Storage unit; 7 Interface unit; 100 Deterioration detection device; 200 External device; 300 Learning unit; 301 Learning data acquisition unit; 302 Model generation Part; 303 Distance between distribution calculation part; 304 Feature amount evaluation part; 305 Aggregation part; 311 Adjacent matrix calculation part; 312 Separation degree evaluation part; 313 Similarity evaluation part; 1000 Deterioration detection factor analysis system.

Claims

A learning unit that learns and builds a decision tree that infers the category of deterioration events to which the data belongs from the features of the data based on the learning data.
A feature amount extractor that calculates a feature amount from the input time-series sensor data,
A deterioration candidate inference unit that selects a feature with a high evaluation from the calculated features and infers a plausible deterioration event candidate from the learned decision tree.
It is a deterioration detection device equipped with
The learning unit is a deterioration detection device characterized by grouping deterioration events and learning and constructing the decision tree.
The learning unit
The first aspect of claim 1, wherein when the learning data is classified according to the decision tree, the deterioration events to be grouped are determined based on the value obtained by dividing the number of data that can be correctly classified by the total number of data. Deterioration detection device.
The learning unit
Among the training data, the data with two types of labels is provided with an inter-distribution distance calculation unit for calculating the inter-distribution distance of the two types of plots in the feature space.
The deterioration detection device according to claim 1, wherein the deterioration event to be grouped is determined based on the distance between distributions calculated by the distance calculation unit between distributions.
The deterioration detection device according to claim 3, wherein the distance between distributions is defined by Area Under The Curve.