JP7330754B2 - Abnormality diagnosis device and method - Google Patents

Description

The present invention relates to an abnormality diagnosis apparatus and method for diagnosing an abnormality by monitoring the state of a machine, parts, and systems (hereinafter referred to as "machine") using measurement data or sensor data.

In the status monitoring of machines such as wind turbines, sensors installed in the machines are used to collect physical quantities such as temperature and acceleration of the machine at a certain frequency, and the collected data is analyzed and processed to determine the state of the machine. is monitored and judged as normal/abnormal. In recent years, a method has been developed that utilizes machine learning technology, a field of artificial intelligence technology, to learn (also called “training”) machine operation data collected in the past, and to estimate and judge the current state. It is

However, in the majority of machines, the occurrence of abnormalities is rare or non-existent. Therefore, even if past operation data is learned, only a normal model can be constructed, making it difficult to distinguish between abnormal and normal conditions. In order to avoid this problem, the system described in Patent Document 1 proposes an abnormality diagnosis method using a plurality of similar machines and using a meta-judgment model common to similar machines.

JP 2010-287011 A

In the technology disclosed in Patent Document 1, in a plurality of similar devices, when operating data is classified into two types, an explanatory variable and an objective variable, regression analysis is performed on the explanatory variable and the objective variable, and multiple Create a regression model for It also proposes creating a meta-prediction model for predicting coefficients of the regression model and diagnosing the state of the target device using the meta-prediction model. This technique can be expected to have a certain effect if the operation data of the target device can be expressed as a simple regression model.

However, even if a statistical method is used, it may not be possible to classify the operating data into explanatory variables and objective variables. Moreover, even if classification is possible, the relationship between explanatory variables and objective variables is complicated, and construction of a regression model may be practically difficult. For example, in a wind power generator, if wind speed, generator bearing temperature, generator bearing acceleration, gearbox generator temperature, gearbox generator acceleration are selected as operating data, they can be classified into explanatory variables and objective variables. difficult. Patent Literature 1 does not describe how to deal with such a case.

The present invention has been made in view of such problems, and its purpose is to provide a method for analyzing a machine that has rare or no abnormal data, and even if the operating data of the target machine cannot be expressed as a simple regression model. An object of the present invention is to provide an abnormality diagnosis device and method capable of diagnosing the state of a machine.

In view of the above, in the present invention, an abnormality diagnosis apparatus for diagnosing an abnormality in a machine acquires and stores sensor data of a machine to be diagnosed and a similar machine similar to the machine to be diagnosed. a learning model creation unit that creates a learning model using the first feature created from the sensor data acquired for the similar machine, and the second feature created from the sensor data acquired for the machine to be diagnosed. characterized by comprising a feature amount correction unit that corrects the first feature amount and a machine state determination unit that determines the state of the diagnosis target machine by applying the corrected second feature amount to the learning model. do.

Further, in the present invention, there is provided an abnormality diagnosis method for diagnosing an abnormality in a machine, in which sensor data of a machine including a diagnosis target machine to be diagnosed and a similar machine similar to the diagnosis target machine are acquired, and the similar machine is acquired. A learning model is created using the first feature amount created from the sensor data obtained from the machine to be diagnosed, the second feature amount created from the sensor data acquired from the machine to be diagnosed is corrected by the first feature amount, and the corrected second feature amount is corrected. 2 is applied to the learning model to determine the state of the machine to be diagnosed.

According to the present invention, it is possible to diagnose the condition of a target machine even for a machine with rare or no abnormal data, and even when the operating data of the target machine cannot be expressed as a simple regression model.

The figure which shows the structural example of the abnormality-diagnosis apparatus which concerns on this invention. The figure which shows an example of the data structure of sensor data. A diagram showing a wind speed-power generation curve, which is an example of the system curve of the machine 10A. The figure which shows the image of the data structure of feature-value FA. FIG. 4 is a diagram showing an image of the data structure of learning data consisting of feature values and normal/abnormal labels. The figure which shows an example of the processing flow based on this invention. FIG. 4 is a diagram showing correlations between extracted feature amounts in comparison with each other on the vertical axis and the horizontal axis. The figure which shows an example of abnormality degree distribution.

Hereinafter, embodiments of the present invention will be described in detail based on the drawings.

In this embodiment, an example will be described in which a wind power generator that converts wind energy into electric energy is used as a machine to be diagnosed.

First, with reference to FIG. 1, an example configuration of an abnormality diagnosis apparatus according to an embodiment of the present invention will be described. The abnormality diagnosis apparatus of this embodiment operates on a computer 100, receives information from a machine to be diagnosed 10A and a similar machine 10B, and uses a database section DB, a sensor data acquisition section 101, and a similar machine selection section 102. , a similar machine data acquisition unit 103 , a feature amount creation unit 104 , a feature amount correction unit 105 , a learning model creation unit 106 , a machine state determination unit 107 and a visualization unit 108 .

The machine 10A to be diagnosed and the similar machine 10B are wind turbine generators in this embodiment, and are equipped with measuring devices such as sensors, which are means for collecting physical quantity data such as temperature, pressure, and acceleration at regular time intervals. Equipped with Also provided is communication means for transmitting the collected data to the abnormality diagnosis apparatus of this embodiment. As will be described later, the similar machine 10B is a machine selected from a plurality of machines 10 (wind turbine generators) from the viewpoint that it matches the diagnosis target machine 10A in terms of characteristics and the like.

In the database DB, all data such as sensor data related to the abnormality diagnosis device of this embodiment are stored in one or more storage devices in electronic file format.

In the abnormality diagnosis apparatus of this embodiment, first, the sensor data acquisition unit 101 acquires the sensor data from the diagnosis target machine 10A through communication means, and stores it in the database DB.

FIG. 2 shows an example of the data structure of sensor data. Sensor data is a set of time-series data collected from each sensor and can be described as a matrix as shown in FIG . A column of the sensor data SA corresponds to each physical quantity, and each row corresponds to each time ts. Here, all sensor data of the diagnosis target machine 10A is described as SA, and individual sensor data SA is given a branch number i (i=1 to N). If there are N sensor items (physical quantities), SA can be mathematically described as in equation (1). Here, SAi (i=1 to N) is the time-series data of the i-th sensor item.

Among all the sensor data SA of the diagnosis target machine 10A, there are normally sensor items that do not contribute to the calculation of the feature amount. Let SAA be a subset of the data actually used for calculating the feature quantity in all the sensor data SA.

Returning to FIG. 1, the similar machine selection unit 102 selects a machine that satisfies all of the following selection conditions from among a plurality of machines 10 to be selected as a similar machine 10B.

Selection condition 1: The inputs and outputs of the machines 10B and 10A are the same, and the system curves representing the system characteristics of the machines also have the same characteristics. If the machine 10A is a windmill, the input is wind energy and the output is electrical energy, so the machine 10B must have similar inputs and outputs.

FIG. 3 shows an example of a wind speed-power generation curve, which is a system curve in a wind turbine. As shown in FIG. 3, in the wind speed-power generation curve, which is the system curve of the machine 10A, the power generation is zero before the wind speed reaches the cut-in wind speed, and then the power generation gradually increases as the wind speed increases. However, when the wind speed reaches the cutout wind speed, the power generation amount is maintained as it is. The system curve for machine 10B should also have such characteristics.

In the case of wind turbines, in this example of the wind speed - power generation curve, which is the system curve, the state of the diagnosis target machine can be determined using sensor data when the wind speed is equal to or higher than the cut-in wind speed and equal to or lower than the cut-out wind speed. become.

Selection condition 2: The past sensor data of the machine 10B must include both normal data and abnormal data.

Selection condition 3: Any sensor data in the data subset SAA that is actually used for feature calculation must have the same or a strong correlation in the sensor data SB of the machine 10B. For example, if there is sensor data of generator bearing front temperature in the subset SAA, and if there is sensor data of generator bearing front temperature in the sensor data SB of the machine 10B, the conditions are satisfied. Or, in the sensor data SB of the machine 10B, there is no sensor data of the generator bearing front temperature, but the sensor data of another sensor, for example, "generator bearing rear temperature", is always the generator bearing front temperature of the subset SAA. If there is a strong correlation with , the condition is satisfied by using it as an alternative sensor data.

If there are multiple similar machine candidates that satisfy all of the above selection conditions, one or more of the multiple can be selected. In this embodiment, the number of similar machines is one.

In FIG. 1, the similar machine data acquisition unit 103 acquires the past sensor data SB of the similar machine 10B selected by the similar machine selection unit 102 from the database DB. The sensor data SB includes all sensor items of the subset SAA out of all the sensor data SA in the diagnosis target machine 10A. The sensor data SB preferably has the same data structure as the sensor data SA.

The feature quantity creation unit 104 creates input data for machine learning from the sensor data for the target machine 10A and the similar machine 10B. In the target machine 10A, the subset SAA of the sensor data SA is used to calculate feature values for abnormality diagnosis. If FA is the feature amount data of the machine 10A, FA is expressed as in Equation (2).

Here, f is a function when calculating the feature amount. For example, a new feature amount may be created by taking the difference between two pieces of sensor data. In addition, there are cases in which feature quantities are created by using transformations such as Fourier transforms and wavelet transforms on sensor data values. In addition, a dimension reduction algorithm such as principal component analysis may be used to create features. As a special example, sensor data values may be directly used as feature amounts. The feature quantity created in this manner can be expressed mathematically as N vectors. That is, FA=[FA1, FA2, . . . , FAN]. Here, FAi (i=1 to N) is the time-series data of the i-th feature amount.

Furthermore, since processing such as resampling may be performed on time-series data of sensor data values when calculating feature values, time stamps tf1, tf2, . The stamps ts1, ts2, . . . , tsn may differ.

FIG. 4 shows an image of the data structure of the feature amount FA. In FIG. 4, the upper part shows the subset SAA of the data actually used for calculating the feature value in all the sensor data SA in FIG. 2 arranged by time. The feature amount data FA of the machine 10A to be diagnosed is organized by time. For example, a plurality of feature quantities derived from the subset SAA at time ts1 are described for each time as feature quantities at time tf1.

For the similar machine 10B as well, using the same conversion function f, the feature amount FB is calculated from the subset SBB of the sensor data SB of the similar machine 10B. That is, FB=f(SBB). Since the number of data of feature amount FB is the same as that of feature amount FA, it can be described as in equation (3). Here, FBi (i=1 to N) is the i-th feature amount data of the machine 10B.

In FIG. 1, the feature quantity correction unit 105 maps the feature quantity FA to the data space of the feature quantity FB by matching the feature quantity FA of the machine 10A with the value of the feature quantity FB of the machine 10B. For example, using the following equation (4), each vector FAi of the feature amount FA is mapped to the space of the feature amount FB. Note that the expression (4) means that the feature amount FA and the feature amount FB are normalized to values in the range of 1 to 0, instead of being evaluated as absolute values.

Here, FA˜i is the i-th feature quantity vector after mapping to the feature quantity space of the machine 10B. FAimin and FAimax are the minimum and maximum values of FAi, and FBimin and FBimax are the minimum and maximum values of FBi. The feature amount FA~ after mapping can be described as in the following equation (5).

The learning model creation unit 106 creates a learning model by applying a machine learning algorithm using the feature amount FB. In this embodiment, a supervised algorithm represented by a neural network will be described as an example. In the space of the feature amount FB, learning data is created by assigning a normal/abnormal label to the state of the machine at each time based on the past operation performance. For example, 0 is given as a label when normal, and 1 is given as a label when abnormal.

FIG. 5 shows an image of the data structure of learning data consisting of feature values and normal/abnormal labels. The learning data shown in FIG. 5 is obtained by assigning normal/abnormal labels to the data group of the feature amount data FB for each time.

For such learning data, association between anomaly labels and feature data can be represented by a learning model such as Equation (6). where Y is the value of the anomaly label, L() is the function of the learning model, and wi (i=1 to K) are the coefficients of the learning model.

The machine state determining unit 107 in FIG. 1 uses a learning model to calculate the abnormal label value of the machine 10A by inputting the feature amount FA˜ of the machine 10A corrected by the feature amount correcting unit 105. FIG. Thereby, the normal/abnormal state of the machine 10A can be determined. For example, if the fault label is 0, the state of the machine is normal, and if the fault label is 1, the state of the machine is abnormal.

The visualization unit 108 visualizes the results determined by the abnormality determination unit 107 and the feature values created by the feature value creation unit 104 on a display device. The sensor data, feature values, machine states, etc. are displayed side by side on the same screen so that the machine 10A and the machine 10B can be compared.

FIG. 6 illustrates the processing flow of the abnormality diagnosis device in this embodiment. The processing procedure of this embodiment consists of the following steps. The following procedure includes processing for handling sensor data SB from the similar machine 10B and processing for handling sensor data SA from the diagnostic target machine 10A.

First, the processing procedure in the similar machine 10B will be explained. In step S1B, the similar machine selection unit 102 is used to select a similar machine 10B having similar characteristics from a plurality of machines. In step S2B, the similar machine data acquisition unit 103 is used to acquire the sensor data SB of the similar machine 10B from the database DB. In step S3B, the feature amount FB of the similar machine 10B is created using the feature amount creation unit 104. FIG. In step S4B, a learning model of similar machine 10B is created using learning model creation unit 106, and stored in database DB.

Next, a processing procedure in the diagnosis target machine 10A will be described. In step S1A, the sensor data acquisition unit 101 is used to acquire the sensor data SA of the machine 10A to be diagnosed. In step S2A, the feature amount FA of the target machine 10A is created using the feature amount creating unit 104. FIG. In step S3A, the feature amount correction unit 105 is used to correct the feature amount FA to create FA˜. In step S4A, the learning model is read, and the state of the target machine 10A is determined using the machine state determination unit 107. FIG. In step S5A, the visualization unit 108 is used to visualize the diagnosis result.

When collecting and storing a group of data to be accumulated in the database DB, the system curve representing the system characteristics of the machine, for example, if the machine 10A is a wind turbine, is represented by a wind speed-power generation curve, which is the system curve of the wind turbine. , the data collected within the range of the cut-in wind speed and the cut-out wind speed in FIG. 3 are accumulated, and the characteristic amount is calculated, corrected, and anomaly determination is performed from the data of this section. The excluded range is not within the active control range of the control device, so it is not suitable for determining abnormality as a system.

Although the embodiment of the abnormality diagnosis apparatus of the present invention has been described above using the wind power generator as an example, it goes without saying that the present invention can also be applied to machines other than the wind power generator.

In this embodiment, an unsupervised algorithm is used for creating a learning model in the abnormality diagnosis apparatus in the first embodiment. That is, a learning model is created by using a method such as distance between data or clustering without assigning normal/abnormal labels to the learning data.

An example of the unsupervised learning model of this embodiment will be described. FIG. 7 is a diagram showing the correlation between the extracted feature amounts in comparison with the vertical axis and the horizontal axis. In this case, since the sensor data SB of the similar machine 10B includes both failure data and normal data, two clusters of failure data and normal data are formed in the feature space of FIG. can be calculated, for example, by an algorithm such as a support vector machine.

If this boundary line is described as a function g(X), the learning model can be described in pseudo code such as the following equation (7). Here, X is a matrix representing the feature amount of the observation point, and p(X) is the function value of the observation point in the feature amount space.

The state of the machine 10A can be determined by applying the corrected feature amount FA˜ of the machine 10A to this learning model.

In the present embodiment, among the abnormality diagnosis apparatuses in the first embodiment, a statistical algorithm such as the Mahalanobis-Taguchi method is used to create a learning model, and an index of the degree of abnormality is calculated from the feature amount.

The degree of abnormality referred to here is a numerical value indicating the degree of abnormality of the machine, and can be calculated, for example, by the following equation (8). Note that a is the degree of abnormality, x is the feature amount vector, μ is the average value of the feature amount, and σ is the variance of the feature amount.

FIG. 8 shows an example of anomaly distribution calculated by such a method. The threshold in the figure is the boundary line between abnormal and normal calculated from the degree of abnormality of the machine 10B. A learning model can be described in pseudo code such as Equation (9) using the value of the degree of anomaly and the threshold.

The state of the machine 10A can be determined by applying the corrected feature amount FA˜ of the machine 10A to this learning model.

In this embodiment, a plurality of similar machines among the abnormality diagnosis apparatuses in the first to third embodiments are used. In such a case, it is possible to perform the processing as in the first to third embodiments for each machine, and then determine the final normality/abnormality judgment by a method such as majority voting. Alternatively, by performing processing such as totaling or averaging the operation data of multiple machines, it is possible to treat them as one machine.

Although the database is described as being stored in one device in the present invention, it is also possible to store data in different computers or storages by using a network environment. Also in the computer, it is possible to distribute the calculation processing to a plurality of computers.

Although the preferred embodiments of the present invention have been described above, it goes without saying that the present invention is not limited to the above embodiments.

100: Computer 101: Sensor data acquisition unit 102: Similar machine selection unit 103: Similar machine data acquisition unit 104: Feature amount creation unit 105: Feature amount correction unit 106: Learning model creation unit 107: Machine state determination unit 108: Visualization unit 10A :: Target machine 10B :: Similar machine DB: Database SA: Sensor data of target machine SB: Sensor data of similar machine SAA: Sensor data of target machine used for feature value calculation SBB: Similar machine Among the sensor data, sensor data used for feature amount calculation FA: feature amount data of target machine FB: feature amount data of similar machine

Claims (7)

An abnormality diagnosis device for diagnosing an abnormality in a machine,
a database that acquires and stores sensor data of a machine to be diagnosed and a machine that includes similar machines similar to the machine to be diagnosed;
a learning model creation unit that creates a learning model using a first feature amount created from sensor data acquired for the similar machine;
a feature amount correcting unit that corrects a second feature amount created from sensor data acquired for the diagnosis target machine with the first feature amount;
An abnormality diagnosis apparatus, comprising: a machine state determination unit that determines the state of the diagnosis target machine by applying the corrected second feature amount to the learning model. The abnormality diagnosis device according to claim 1,
When selecting the similar machine similar to the diagnosis target machine from among a plurality of machines,
The input/output is the same as that of the machine to be diagnosed, the system characteristics of the machine are the same, the sensor data of the similar machine includes both normal data and abnormal data, and the sensor data of the similar machine and selecting the similar machine on the basis that it contains data necessary for creating a feature amount or contains data that can be substituted. The abnormality diagnosis device according to claim 1 or claim 2,
The abnormality diagnosis apparatus, wherein the learning model creation unit learns normal/abnormal labels from past operating states of the similar machine as teacher data. The abnormality diagnosis device according to claim 1 or claim 2,
The abnormality diagnosis apparatus, wherein the learning model creation unit performs learning using an unsupervised algorithm such as clustering in a data space of feature amounts of the similar machines. The abnormality diagnosis device according to claim 1 or claim 2,
The abnormality diagnosis apparatus, wherein the learning model creation unit calculates the degree of anomaly at each time using a statistical method from the feature quantity of the similar machine, and creates the learning model from the degree of anomaly and its threshold value. The abnormality diagnosis device according to any one of claims 1 to 5,
The machine is a wind power generator, and the machine to be diagnosed uses the sensor data at a wind speed equal to or higher than the cut-in wind speed and equal to or lower than the cut-out wind speed in a wind speed-power generation curve that is a system curve of the wind turbine. An abnormality diagnosis device characterized by judging the state of An abnormality diagnosis method for diagnosing an abnormality in a machine, comprising:
Obtaining sensor data of a machine including a machine to be diagnosed and a similar machine similar to the machine to be diagnosed, and generating a learning model using a first feature created from the sensor data obtained for the similar machine. and correcting a second feature amount created from sensor data acquired for the diagnosis target machine with the first feature amount, and applying the corrected second feature amount to the learning model, An abnormality diagnosis method characterized by determining the state of a machine to be diagnosed.

Images