JP7127477B2 - LEARNING METHOD, APPARATUS AND PROGRAM, AND EQUIPMENT FAILURE DIAGNOSIS METHOD - Google Patents

Description

TECHNICAL FIELD The present invention relates to a learning method, an apparatus, a program, and a method for diagnosing facility abnormality.

2. Description of the Related Art There is an anomaly diagnosis technology for detecting an anomaly in an equipment used in a process such as a manufacturing process, using measurement data measured by the equipment. For example, various statistics (average, standard deviation, etc.) can be calculated from measurement data to analyze trends. Acceleration, etc.) for evaluation, or FFT (Fast Fourier Transform) or envelope processing (envelope processing) of the measured data to capture abnormal signs in the frequency domain.

In equipment abnormality diagnosis through such analysis, evaluation, and processing, the know-how of skilled engineers is needed to appropriately interpret the obtained results (statistics, feature values, FFT, etc.) in conjunction with the actual operation and equipment status. It depends a lot. This is because it is necessary to change the determination criteria each time according to the state of the process and appropriately determine the obtained results.
In order to overcome such a situation, in recent years, pattern recognition technology such as machine learning is used to diagnose facility abnormalities, thereby reducing the workload of engineers (see Patent Documents 1 and 2). ).

In the case of diagnosing an abnormality of equipment using pattern recognition technology, it is necessary to register pattern information of normal conditions and pattern information of abnormal conditions in advance. At that time, since the pattern information of the normal state and the pattern information of the abnormal state differ depending on the type of the component equipment, the pattern information of the normal state and the pattern information of the abnormal state for the wide variety of equipment to be diagnosed must be collected. However, since the occurrence of an abnormal event is extremely rare, it is difficult to register in advance the pattern information of abnormal states for all types of constituent devices.
Therefore, by detecting vibration etc. according to the operation of equipment that operates periodically, converting the detected time-series data into frequency data, and analyzing the frequency spectrum, It is known to detect equipment anomalies (see Patent Literatures 3 and 4).

However, due to physical changes due to the surrounding environment (operational conditions, temperature, etc.) of the equipment, wear, etc., the signal intensity and the generated frequency of the spectrum signal indicating the normal operation of the equipment may change over time. Therefore, unlike an ideal spectrum signal, a spectrum signal indicating normal operation has a noise component, so that it is superimposed on an abnormal component indicating abnormal operation of the equipment. was difficult.
Therefore, a technique for improving detection accuracy by removing noise components has been proposed (see Patent Document 5).

JP 2008-262375 A JP 2010-191556 A JP 2013-30015 A JP 2011-59790 A JP 2016-62258 A

In Patent Document 5, a regularization model is used to remove noise components from measurement data. In the regularization model, the degree of removal of noise components differs depending on the regularization parameter. An excessively large regularization parameter may remove even normal partial correlations, while an excessively small regularization parameter fails to remove noise components.
However, in Patent Document 5, only an empirical criterion (regularization parameter ≤ 0.01) is shown as a regularization parameter, and it was difficult to set it appropriately.

SUMMARY OF THE INVENTION The present invention has been made in view of the above points, and an object of the present invention is to appropriately remove noise components from measurement data so as to be able to perform abnormality diagnosis of equipment.

A learning method of the present invention is a learning method for performing learning for abnormality diagnosis of the equipment using measurement data measured by the equipment, wherein the learning data acquires measurement data of the normal state of the equipment as learning data. In the acquisition step , sample data of the same size as the learning data is created using normal random numbers, and a regularization parameter is set to 0 in order to calculate the partial correlation coefficient between variables of the sample data using a regularization model. A regularization parameter calculation step that gradually increases from , sequentially calculates the partial correlation coefficient using the regularization parameter, and calculates the minimum regularization parameter at which the partial correlation coefficient between all variables converges to zero and a learning data processing step of calculating a partial correlation coefficient between feature quantities in the learning data using a regularization model using the calculated regularization parameter.
An abnormality diagnosis method for equipment according to the present invention is a method for diagnosing abnormality of equipment using measurement data measured by the equipment, wherein the measurement data of the normal state of the equipment is acquired as learning data. a step of acquiring learning data, creating sample data of the same size as the learning data using normal random numbers, and calculating a partial correlation coefficient between variables of the sample data using a regularization model, regularizing Regularization that gradually increases the parameter from 0, sequentially calculates the partial correlation coefficient using the regularization parameter, and calculates the minimum regularization parameter that causes the partial correlation coefficient between all variables to converge to zero . a parameter calculation step; a learning data processing step of calculating a partial correlation coefficient between feature quantities in the learning data using a regularization model using the calculated regularization parameter; and the detection data and the regularization parameter calculated in the regularization parameter calculation step, the partial correlation coefficient between the feature quantities in the detection data is calculated using a regularization model and a detection data processing step for diagnosing equipment abnormality based on the partial correlation coefficient calculated in the learning data processing step and the partial correlation coefficient calculated in the detection data processing step. Characterized by
A learning device of the present invention is a learning device that performs learning for abnormality diagnosis of the equipment using measurement data measured by the equipment, and acquires measurement data of the normal state of the equipment as learning data. and an acquisition means for creating sample data of the same size as the learning data using normal random numbers, and setting a regularization parameter to 0 in order to calculate a partial correlation coefficient between variables of the sample data using a regularization model. A regularization parameter calculation means for gradually increasing from the and learning data processing means for calculating, using the calculated regularization parameter, a partial correlation coefficient between feature quantities in the learning data using a regularization model.
A program of the present invention is a program for performing learning for abnormality diagnosis of the equipment using measurement data measured by the equipment, and learning data for acquiring measurement data of the normal state of the equipment as learning data. and an acquisition means for creating sample data of the same size as the learning data using normal random numbers, and setting a regularization parameter to 0 in order to calculate a partial correlation coefficient between variables of the sample data using a regularization model. A regularization parameter calculation means for gradually increasing from the Then, using the calculated regularization parameter, the computer functions as learning data processing means for calculating a partial correlation coefficient between feature quantities in the learning data using a regularization model.

ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to remove a noise component appropriately from measurement data, and to perform an abnormality diagnosis of equipment.

It is a figure which shows the functional structure of the abnormality-diagnosis apparatus of the installation which concerns on embodiment. It is a flowchart which shows the learning method by the abnormality-diagnosis apparatus of the installation which concerns on embodiment. 4 is a flow chart showing a method for diagnosing equipment abnormality by the equipment abnormality diagnosis device according to the embodiment. FIG. 5 is a characteristic diagram showing time series changes in the degree of abnormality of each hydraulic cylinder by applying the present invention. FIG. 5 is a characteristic diagram showing time-series changes in the degree of abnormality of each hydraulic cylinder when the regularization parameter is not appropriate;

Preferred embodiments of the present invention will now be described with reference to the accompanying drawings.
FIG. 1 shows the configuration of an equipment abnormality diagnosis device 100 according to an embodiment.
The facility abnormality diagnosis device 100 according to the embodiment uses measurement data measured by the facility to perform abnormality diagnosis for detecting abnormal operation due to deterioration or damage of the facility, and is a learning device to which the present invention is applied. And it functions as an abnormality diagnosis device for equipment.
The facility abnormality diagnosis device 100 includes a learning data acquisition unit 101, a learning data processing unit 102, a detection data acquisition unit 103, a detection data processing unit 104, a storage unit 105, a difference calculation unit 106, and an abnormality determination unit. 107 and an output unit 108 .

The learning data acquisition unit 101 acquires measurement data of the normal state of the equipment as learning data from the detection device. In this case, the measurement data used as the learning data may be subjected to a predetermined process, for example, the measurement data may be appropriately converted into a statistic amount or a feature amount. The detection devices are sensors, such as acceleration sensors and vibration sensors, that are arranged around or near the equipment and detect the operation of the equipment. Note that, in the present embodiment, the learning data acquisition unit 101 functions as the learning data acquisition means of the present invention, and executes the learning data acquisition step.

The learning data processing unit 102 calculates a regularization parameter λ for removing noise components from the learning data acquired by the learning data acquisition unit 101 . Then, using the calculated regularization parameter λ, a partial correlation coefficient (correlation coefficient excluding pseudo-correlation) between feature quantities in the learning data is calculated using a regularization model. In this embodiment, the learning data processing unit 102 functions as regularization parameter calculation means and learning data processing means of the present invention, and executes a regularization parameter calculation step and a learning data processing step.
Here, since the noise component has a characteristic that it changes depending on the size of the measurement data (the number of data n×the number of feature amounts m), this is utilized for the calculation of the regularization parameter. Specifically, the sample data (number of data n × variable m) of the same size as the learning data (number of data n × number of features m) is normal random numbers (average: 0, standard deviation: 1 random numbers generated from a normal distribution according to ), and the partial correlation coefficients between the variables of the created sample data are calculated using the regularization model. At that time, the regularization parameter λ is gradually increased from 0 (for example, by 0.00001), and the partial correlation coefficient is calculated sequentially using the regularization parameter λ, and the partial correlation coefficient between all variables is Compute the smallest regularization parameter λ that converges to zero. Using the regularization parameter λ calculated in this way, the partial correlation coefficient between the feature quantities in the learning data is calculated using the regularization model.

The detection data acquisition unit 103 acquires, as detection data, measurement data for diagnosing an abnormality of equipment from the detection device. When the learning data acquisition unit 101 converts the measurement data into statistics and feature values to obtain learning data, the detection data acquisition unit 103 similarly converts the measurement data to obtain detection data. Also, the size of the detection data is assumed to be the same as the size of the learning data (number of data n×number of variables m).

The detection data processing unit 104 uses the detection data acquired by the detection data acquisition unit 103 and the regularization parameter λ calculated by the learning data processing unit 102, which is shared by the storage unit 105, to obtain the feature amount in the detection data. The partial correlation coefficient between is calculated using a regularized model.

Storage unit 105 is connected to learning data processing unit 102 , detection data processing unit 104 , and difference calculation unit 106 . The storage unit 105 stores the regularization parameter λ calculated by the learning data processing unit 102 and the partial correlation coefficient between feature amounts in the learning data.

The difference calculation unit 106 calculates the difference between the partial correlation coefficient between the feature amounts in the learning data stored in the storage unit 105 and the partial correlation coefficient between the feature amounts in the detection data calculated by the detection data processing unit 104 as the degree of abnormality. Calculate as

The abnormality determination unit 107 determines whether the facility is in a normal state or an abnormal state by comparing the degree of abnormality calculated by the difference calculation unit 106 with a preset abnormality determination threshold value.

The output unit 108 outputs the calculation result of each unit and the determination result of the abnormality determination unit 107 . The output here means, for example, displaying on a display (not shown) or transmitting to an external device of the apparatus 100 .

The equipment abnormality diagnosis apparatus 100 thus constructed is implemented by a computer device including, for example, a CPU, a ROM, a RAM, and the like. In FIG. 1, the facility abnormality diagnosis device 100 is illustrated as one device, but it may be configured by a plurality of devices. For example, a device responsible for the learning function by the learning data acquisition unit 101 and the learning data processing unit 102, and a device responsible for the abnormality diagnosis function by the detection data acquisition unit 103, the detection data processing unit 104, the difference calculation unit 106, and the abnormality determination unit 107 may be configured separately.

Here, the regularization model used for calculating the partial correlation coefficient will be described.
Consider a general linear model as in equation (1).

where φ(x) represents basis functions and w represents weights. Suppose that (y, φ(x)) is obtained as in Equation (2) for n pieces of observation data (y, x). Replacing this with a regularized model results in Equation (3).

When q=2, it is L2 regularization, and when q=1, it is L1 regularization. Here, argminf(x) indicates x when f(x) is minimized, and means to calculate vector w when equation (3) is minimized. The first term in equation (3) represents the squared loss function. Also, |w _j | ¹ in the second term of Equation (3) indicates an L1 regularization term, which is expressed by Equation (4).

The L1 regularization term of the second term is a regularization term added to make the matrix sparse so that most of the components of the squared loss function of the first term are zero. That is, the learning data processing unit 102, which is a model learning unit, calculates the vector w for modeling the learning data with fewer variables by introducing the L1 regularization term.
Here, the regularization parameter λ, which sets the magnitude of the regularization term, is the regularization parameter included in the estimated squared loss function. By appropriately defining the regularization parameter λ, it is possible to learn a model that matches the target facility.

In a general linear regression model, the regularization parameter λ that minimizes the squared error of f(x) can be determined as an appropriate parameter by varying the regularization parameter λ.
However, in the abnormality diagnosis of the facility targeted this time, it is important to appropriately calculate the partial correlation coefficient between the feature values of the learning data in the normal state, not to predict the objective function. Therefore, even if the regularization parameter λ is changed, it cannot be determined uniquely. That is, since there is no actual value of the objective variable, it is not possible to determine the regularization parameter λ that minimizes the squared error between the predicted value and the actual value.
Therefore, the feature of the L1 regularization model is utilized. In the L1 regularization model, by appropriately defining the regularization parameter λ, it is possible to make a sparse matrix in which the noise component between feature quantities converges to zero. Taking advantage of this property, the minimum regularization parameter λ that appropriately removes the noise component that varies depending on the size (number of data n × number of features m) in the sample data is the appropriate regularization for training data of the same size. parameter λ.

A learning method and an equipment abnormality diagnosis method by the equipment abnormality diagnosis device 100 according to the embodiment will be described below. FIG. 2A is a flowchart showing a learning method by the facility abnormality diagnosis device 100 according to the embodiment. FIG. 2B is a flow chart showing a method for diagnosing an abnormality of equipment by the equipment abnormality diagnosis device 100 according to the embodiment. In the present embodiment, an example of diagnosing an abnormality of a hydraulic shearing machine will be described. The facility abnormality diagnosis device 100 acquires measurement data acquired from the rise and fall times of a plurality of hydraulic cylinders and a plurality of lubricating pumps, and detects an abnormal operation of the facility.

As shown in FIG. 2A, in step S1, the learning data acquisition unit 101 acquires normal state measurement data as learning data from the detection device. The learning data acquisition unit 101 acquires measurement data detected by the detection device at predetermined intervals as learning data. In the present embodiment, the learning data acquisition unit 101 uses the feature amount obtained by converting the measured data detected each time the equipment is started up into an hourly average as the learning data, and sets the number of data n to 100 and the number of feature amounts m to 24. .

In step S2, the learning data processing unit 102 creates sample data (number of data: 100 x variable: 24) of the same size as the learning data (number of data: 100 x number of feature values: 24) acquired in step S1 using normal random numbers. .

In step S3, the learning data processing unit 102 gradually sets the regularization parameter λ from 0 (for example, 0.00001), and sequentially calculate the partial correlation coefficient using the regularization parameter λ, and calculate the minimum regularization parameter λ at which the partial correlation coefficient between all variables converges to zero. , to the storage unit 105 . In this embodiment, the regularization parameter λ was 0.00705.

In step S4, the learning data processing unit 102 uses the regularization parameter λ calculated in step S3 to calculate the partial correlation coefficient between the feature values in the learning data using the regularization model, and stores it in the storage unit 105. supply.

As shown in FIG. 2B, in step S5, the detection data acquisition unit 103 acquires measurement data for diagnosing an abnormality of equipment from the detection device as detection data. The detected data acquisition unit 103 sets the feature amount obtained by converting the measurement data detected each time the equipment is started up to an hourly average as the learning data, similarly to the learning data, and sets the data number n to 100 and the feature amount number m to 24. do.

In step S6, the detection data processing unit 104 uses the detection data acquired in step S5 and the regularization parameter λ calculated in step S3, which is shared by the storage unit 105. A partial correlation coefficient is calculated using a regularized model.

In step S7, the difference calculation unit 106 stores the partial correlation coefficient between the feature amounts in the learning data calculated in step S4 and the partial correlation between the feature amounts in the detection data calculated in step S6, which are stored in the storage unit 105. The difference from the number is calculated as the degree of anomaly.
For example, the Kullback-Leibler information amount (KL information amount) may be used as the difference between the partial correlation coefficient between the feature amounts in the learning data and the partial correlation coefficient between the feature amounts in the detection data. The KL information amount is a measure for calculating the information amount difference between probability distributions, and is obtained by Equation (5). P(i), Q(i) refer to probability distributions, and _DKL is a measure for calculating the difference between two probability distributions.

In step S8, the abnormality determination unit 107 determines whether the equipment is in a normal state or an abnormal state by comparing the degree of abnormality calculated in step S7 with a preset abnormality determination threshold value. do.
The abnormality determination threshold is obtained by, for example, cross-validation. Specifically, the measurement data, which is normal data acquired by the learning data acquisition unit 101, is divided into k (five in this example) data sets, and one data set is divided into learning data (n: 20 in this example). × m: 24), create sample data of the same size (n: 20 × m: 24) using normal random numbers, and calculate the partial correlation coefficient between variables of the created sample data using a regularization model do. At that time, the regularization parameter λ is gradually increased from 0 (for example, by 0.00001), and the partial correlation coefficient is sequentially calculated using the regularization parameter λ, and the partial correlation coefficient between all variables is Compute the smallest regularization parameter λ that converges to zero. Using the calculated regularization parameter λ, a partial correlation coefficient between feature amounts in the learning data (n: 20×m: 24) is calculated using a regularization model.
Next, in the remaining k-1 data sets, create (n x (k-1) + 1) pieces of detection data (n: 20 x m: 24) in chronological order, and for each detection data Using the regularization parameter λ calculated from the learning data, the partial correlation coefficient between feature quantities is calculated using a regularization model. A difference between the calculated partial correlation coefficient between feature quantities in each detection data and the partial correlation coefficient between feature quantities in learning data is calculated as the degree of abnormality.
In the above-described calculation procedure, the degree of abnormality is calculated by changing k pieces of learning data. The probability density is q% (99% in this example) by estimating the kernel density (a method of estimating the distribution of the population from finite sample data) from the calculated p (105 in this example) data set of the degree of anomaly A threshold value that becomes an abnormality determination threshold value.

In step S9, the output unit 108 outputs the result determined in step S8.

FIG. 3 shows an example of the result of facility abnormality diagnosis by applying the present invention. The regularization parameter λ was 0.00705. FIGS. 3(a) to 3(c) show chronological changes in the KL information amount, which is the degree of abnormality of each hydraulic cylinder (clamp CYL, main CYL, sizing CYL) of the hydraulic shearing machine. As shown in FIG. 3(a), in clamp CLY, the degree of anomaly increases around September 16th. After actually diagnosing the hydraulic shearing machine and actually inspecting and inspecting damage and wear, it was found that oil was seeping from the hose of the clamp CYL around September 16th, and then the hose was replaced. As a result, it did not lead to a major equipment failure. In this way, an abnormality of the clamp CYL actually occurring could be detected.

On the other hand, FIG. 4 shows an example of the result of facility abnormality diagnosis when the regularization parameter λ is not appropriate. The regularization parameter λ was 0.001. 4(a) to (c), similarly to FIGS. 3(a) to (c), each hydraulic cylinder (clamp CYL, main CYL, sizing CYL) of the hydraulic shearing machine is abnormal KL information amount shows the time-series change of As shown in FIG. 4(a), clamp CLY had a large degree of anomaly around July 17, but it was found that there was no anomaly when it was actually inspected and inspected. The degree of anomaly that increased around July 17th was affected by the noise component, which could not be removed from the measurement data.

As described above, the measurement data acquired when the equipment operates normally is used as learning data to calculate the partial correlation coefficient using a regularization model, and the measurement data is used as detection data to regularize the partial correlation coefficient. Abnormal operation of the facility is detected by calculating the difference between the partial correlation coefficients calculated using the model and calculated between the learning data and the detection data. Thereby, it is possible to calculate the partial correlation coefficient from which the noise component is appropriately removed in accordance with the equipment, and to set the regularization parameter λ that appropriately removes the noise component. Therefore, it is possible to appropriately remove the noise component from the measurement data and perform the abnormality diagnosis of the equipment, thereby accurately performing the abnormality diagnosis of the equipment.

As described above, the present invention has been described together with the embodiments, but the above-described embodiments merely show specific examples for carrying out the present invention, and the technical scope of the present invention is not construed in a limited manner. It should not be. That is, the present invention can be embodied in various forms without departing from its technical concept or main features.
According to the present invention, the software (program) that realizes the learning function and equipment abnormality diagnosis function of the present invention is supplied to a system or device via a network or various storage media, and the computer of the system or device reads the program. It can also be realized by executing

100: equipment abnormality diagnosis device, 101: learning data acquisition unit, 102: learning data processing unit, 103: detection data acquisition unit, 104: detection data processing unit, 105: storage unit, 106: difference calculation unit, 107: abnormality determination unit 108: output unit

Claims (9)

A learning method for learning for abnormality diagnosis of the equipment using measurement data measured by the equipment,
a learning data acquisition step of acquiring measurement data of the normal state of the equipment as learning data;
Create sample data of the same size as the learning data using normal random numbers, and gradually increase the regularization parameter from 0 in order to calculate the partial correlation coefficient between variables of the sample data using a regularization model. a regularization parameter calculation step of sequentially calculating the partial correlation coefficient using the regularization parameter and calculating the minimum regularization parameter at which the partial correlation coefficient between all variables converges to zero ;
and a learning data processing step of calculating a partial correlation coefficient between feature quantities in the learning data using a regularization model using the calculated regularization parameter. An abnormality diagnosis method for equipment for diagnosing an abnormality of the equipment using measurement data measured by the equipment,
a learning data acquisition step of acquiring measurement data of the normal state of the equipment as learning data;
Create sample data of the same size as the learning data using normal random numbers, and gradually increase the regularization parameter from 0 in order to calculate the partial correlation coefficient between variables of the sample data using a regularization model. a regularization parameter calculation step of sequentially calculating the partial correlation coefficient using the regularization parameter and calculating the minimum regularization parameter at which the partial correlation coefficient between all variables converges to zero ;
A learning data processing step of calculating a partial correlation coefficient between feature quantities in the learning data using a regularization model using the calculated regularization parameter;
a detection data acquisition step of acquiring measurement data of the equipment as detection data;
and a detection data processing step of calculating a partial correlation coefficient between feature quantities in the detection data using a regularization model, using the detection data and the regularization parameter calculated in the regularization parameter calculation step. death,
An abnormality diagnosis method for equipment, wherein an abnormality diagnosis for equipment is performed based on the partial correlation coefficient calculated in the learning data processing step and the partial correlation coefficient calculated in the detected data processing step. 3. The method according to claim 2 , further comprising a difference calculation step of calculating a difference between the partial correlation coefficient calculated in the learning data processing step and the partial correlation coefficient calculated in the detection data processing step as a degree of abnormality. Abnormality diagnosis method for the described equipment. 4. The equipment abnormality diagnosis method according to claim 3 , further comprising an abnormality determination step for determining whether the equipment is in a normal state or an abnormal state based on the degree of abnormality. 5. The apparatus according to claim 4 , wherein in the abnormality determination step, it is determined whether the equipment is in a normal state or an abnormal state by comparing the degree of abnormality with a preset abnormality determination threshold value. equipment abnormality diagnosis method. 6. The equipment abnormality diagnosis method according to claim 2 , wherein, in said detection data acquisition step, the size of said detection data is made the same as the size of said learning data. In the detection data acquisition step, when the measurement data used as the learning data is subjected to a predetermined process in the learning data acquisition step, the measurement data used as the detection data is also subjected to the predetermined process. The equipment abnormality diagnosis method according to any one of claims 2 to 6 . A learning device that performs learning for abnormality diagnosis of the equipment using measurement data measured by the equipment,
learning data acquisition means for acquiring measurement data of the normal state of the equipment as learning data;
Create sample data of the same size as the learning data using normal random numbers, and gradually increase the regularization parameter from 0 in order to calculate the partial correlation coefficient between variables of the sample data using a regularization model. regularization parameter calculation means for calculating the minimum regularization parameter at which the partial correlation coefficient between all variables converges to zero by sequentially calculating the partial correlation coefficient using the regularization parameter;
and learning data processing means for calculating a partial correlation coefficient between feature quantities in the learning data using the regularization parameter calculated using the regularization model. A program for learning for abnormality diagnosis of the equipment using measurement data measured by the equipment,
learning data acquisition means for acquiring measurement data of the normal state of the equipment as learning data;
Create sample data of the same size as the learning data using normal random numbers, and gradually increase the regularization parameter from 0 in order to calculate the partial correlation coefficient between variables of the sample data using a regularization model. regularization parameter calculation means for calculating the minimum regularization parameter at which the partial correlation coefficient between all variables converges to zero by sequentially calculating the partial correlation coefficient using the regularization parameter;
A program for causing a computer to function as learning data processing means for calculating a partial correlation coefficient between feature quantities in the learning data using the regularization model calculated using the regularization parameter.

Images