A Fault Diagnosis Method Based on Density Induced Support Vector Data Description

CN118551328BActive Publication Date: 2024-11-15CHANGSHU INSTITUTE OF TECHNOLOGY
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411017400.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15
Estimated Expiration
2044-07-29

AI Technical Summary

Technical Problem

通过密度诱导支持向量数据描述方法得到的故障检测模型可以实现异常数据的监测,但是进一步分析异常数据发生故障的原因还没有实现,需要进一步地完善诊断流程,建立完整的故障诊断模型

Benefits of technology

[0048] The present invention is a fault diagnosis method based on density-induced support vector data description recursive feature elimination, and builds a fault diagnosis model based on the density-induced support vector data description method. First, use normal data training to build a detection model described by density-induced support vector data. The density-induced support vector data description method takes into account the density distribution of the training data and can achieve the purpose of anomaly detection more accurately than the existing support vector data description method. . Secondly, the collected fault data is trained using the density-induced support vector data description recursive feature elimination method to obtain a fault feature ranking set. Then the test data is detected according to the constructed detection model, and when abnormal data is detected, it is diagnosed. Therefore, the feature sorting of fault data can be obtained through the density-induced support vector data description recursive feature elimination method proposed by the present invention, and the corresponding instrument part is searched for inspection and the fault type is determined based on the obtained feature sorting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118551328B_ABST
    Figure CN118551328B_ABST
Patent Text Reader

Abstract

The present invention discloses a fault diagnosis method based on density induced support vector data description and recursive feature elimination, including: collecting data and performing standardized preprocessing on the data to obtain normal data and fault data; using normal data to train to obtain a detection model; using fault data to train to obtain a diagnostic model, and obtaining a fault feature sorting set through the diagnostic model; performing standardized preprocessing on the test data and inputting it into the detection model for detection, if the test data is predicted to be abnormal data, then entering the next step, otherwise outputting it as normal data: obtaining feature sorting of the abnormal data, and comparing it with the fault feature sorting set to confirm the fault type. The present invention uses a feature selection method of recursive feature elimination based on density induced support vector data description to obtain feature sorting, thereby confirming the fault type and realizing a complete fault diagnosis process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of fault diagnosis, relates to support vector data description technology, and specifically relates to a fault diagnosis method based on density-induced support vector data description recursive feature elimination. Background Art

[0002] Due to the rapid development of modern technology, digital monitoring has occupied an indispensable position in industrial processes. Collecting data such as the status, changes and processes of machines during operation through sensors and other equipment can help operators better monitor machines. Timely and accurate detection of faults and the causes of faults are the main tasks of monitoring industrial data. To achieve this goal, data-driven fault detection and fault diagnosis methods are widely used in industrial processes.

[0003] Support Vector Data Description (SVDD) is a learning method for anomaly detection and is widely used in fault diagnosis of industrial data. However, in addition to detecting faults, it is more important to find out the cause of the fault and analyze it. Therefore, the collected fault data is classified using the feature selection method. According to the feature subset obtained from a certain fault data, it can be inferred that an abnormality has occurred at a certain position or state in the process, which requires investigation and maintenance. Therefore, combining the support vector data description method and the feature selection method to build a learning model can achieve the purpose of fault diagnosis. However, the support vector data description only constructs a hypersphere based on normal data, only considers the influence of the support vector on the hypersphere, and ignores most of the data, which is susceptible to noise interference. In order to fully consider the distribution of data, researchers proposed to use density-induced support vector data description (DSVDD) to improve the performance of traditional support vector data description in anomaly detection. The density-induced support vector data description method introduces the density of data when constructing the hypersphere to balance the influence of data on the learning model. The fault detection model obtained by the density induced support vector data description method can realize the monitoring of abnormal data, but further analysis of the causes of abnormal data failure has not been achieved. It is necessary to further improve the diagnosis process and establish a complete fault diagnosis model. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, a fault diagnosis method based on density induced support vector data description and recursive feature elimination is provided. The innovation of this method is to obtain feature ranking by using a feature selection method of recursive feature elimination based on density induced support vector data description, thereby confirming the fault type and realizing a complete fault diagnosis process.

[0005] Technical solution: To achieve the above object, the present invention provides a fault diagnosis method based on density-induced support vector data description recursive feature elimination, comprising the following steps:

[0006] S1: Collect data and perform standardized preprocessing on the data to obtain normal data and fault data;

[0007] S2: Use normal data to train the detection model;

[0008] S3: training the fault data to obtain a fault feature sorting set, and obtaining a diagnosis model through the training of the fault feature sorting set;

[0009] S4: Standardize and preprocess the test data and input it into the detection model for detection. If the test data is predicted to be abnormal data, go to step S5, otherwise it is output as normal data:

[0010] S5: Obtain feature ranking of abnormal data and compare it with the fault feature ranking set to confirm the fault type.

[0011] Furthermore, the specific process of the standardization preprocessing in step S1 is as follows:

[0012] A1: For the normal data collected , , perform standardization preprocessing, where is the total number of normal data; the standardization formula is

[0013] ;

[0014] in for The mean of for The standard deviation of

[0015] A2: Fault data Perform standardization preprocessing, the standardization formula is ,in for The mean of for The standard deviation of .

[0016] Furthermore, the training method of the detection model in step S2 is:

[0017] B1: Calculate the relative density between samples , calculated as follows:

[0018] ,

[0019] in For sample The K nearest neighbor samples, For sample and The Euclidean distance between Represents the average distance of all samples to their K nearest neighbor samples; the closer the distance between sample points, the higher the relative density between them;

[0020] B2: The preprocessed normal training data is learned to obtain the density-induced support vector data description model. The objective function of the model is as follows:

[0021] ;

[0022] ;

[0023] in is the relative density, is the penalty term coefficient, is the radius of the optimal hypersphere, is the center of the optimal hypersphere, is the slack variable;

[0024] B3: Solve the objective function by Lagrange multiplier method and obtain its dual form:

[0025] ;

[0026] in is the Lagrange coefficient;

[0027] B4: Get the coefficient After calculating the radius of the model , calculated as follows:

[0028] ;

[0029] in is the mean relative density.

[0030] Furthermore, in step S3, the preprocessed fault training data is learned using a density induced support vector data description recursive feature elimination method to obtain a fault feature ranking set.

[0031] Furthermore, the acquisition of the fault feature ranking set in step S3 includes the following steps:

[0032] C1: Initialize the feature set to the entire feature set of the data, that is , is the number of features, and the feature subset after initialization sorting is ;

[0033] C2: Using the feature set The fault data is used to train the density-induced support vector data description model and obtain the radius of the model. ;

[0034] C3: Calculate the elimination features , then the feature set is , that is, the current feature set is ;

[0035] C4: Extract the characteristics of fault data For training data, a density-induced support vector data description model is constructed, and the radius of the model is ;

[0036] C5: Calculation criteria function ;

[0037] C6: Find the feature with the largest criterion value , remove this feature from the feature subset , update the feature ranking subset ;

[0038] C7: Return to step C2 until The last remaining feature in ;

[0039] C8: Update sorted subset , and obtain the final feature sorting set ,According to the sorted set of fault features, the fault type is determined, and the features ranked in the front indicate the greater correlation with the fault.

[0040] Furthermore, the method for detecting the model in step S4 is:

[0041] D1: Standardize and preprocess the test data;

[0042] D2: Predict the test sample points, calculate the distance from the sample points to the center of the hypersphere, and determine abnormal data and normal data based on the obtained distance.

[0043] Furthermore, the normalization formula in step D1 is: ,in For test data The mean of For test data The standard deviation of .

[0044] Furthermore, the distance from the sample point to the center of the hypersphere in step D2 is expressed as:

[0045] ;

[0046] When the distance between the sample points in the hypersphere , it indicates that the sample is normal data, otherwise , that is, the sample is abnormal data.

[0047] Furthermore, in step S5, the density induced support vector data is used to describe the recursive feature elimination method to obtain the feature ranking of the abnormal data.

[0048] The present invention is a fault diagnosis method based on density induced support vector data description recursive feature elimination, and a fault diagnosis model is constructed on the basis of the density induced support vector data description method. First, a detection model described by density induced support vector data is constructed by training with normal data. The density induced support vector data description method takes into account the density distribution of the training data and can achieve the purpose of abnormality detection more accurately than the existing support vector data description method. Secondly, the collected fault data is trained using the density induced support vector data description recursive feature elimination method to obtain a fault feature sorting set. Then, the test data is detected according to the constructed detection model, and when the abnormal data is detected, it is diagnosed. Therefore, the feature sorting of the fault data can be obtained by the density induced support vector data description recursive feature elimination method proposed by the present invention, and the corresponding instrument part is searched according to the obtained feature sorting to check and determine the fault type.

[0049] Beneficial effects: Compared with the prior art, the present invention can not only detect abnormal data but also diagnose features related to faults, which is convenient for subsequent troubleshooting and maintenance. The present invention proposes a fault diagnosis method based on density-induced support vector data description recursive feature elimination, which has higher monitoring performance and can give fault feature ranking, thereby obtaining the fault type and determining the cause of the fault, which is convenient for subsequent maintenance detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of a training module of the method of the present invention;

[0051] Figure 2 Schematic diagram of the diagnostic module of the method of the present invention;

[0052] Figure 3 It is the control logic diagram of the dry integrated system;

[0053] Figure 4 This is the detection monitoring diagram of the test data using SVDD;

[0054] Figure 5 The present invention is used to detect and monitor test data. DETAILED DESCRIPTION

[0055] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0056] The present invention provides a fault diagnosis method based on density-induced support vector data description recursive feature elimination, comprising the following steps:

[0057] 1. Collect data and input it into the training module

[0058] like Figure 1 As shown in the figure, the operation process of the training module is:

[0059] S1: Collect data and perform standardized preprocessing on the data to obtain normal data and fault data:

[0060] The specific process of standardized preprocessing is:

[0061] A1: For the normal data collected , , perform standardization preprocessing, where is the total number of normal data; the standardization formula is

[0062] ;

[0063] in for The mean of for The standard deviation of

[0064] A2: Fault data Perform standardization preprocessing, the standardization formula is ,in for The mean of for The standard deviation of .

[0065] S2: Use normal data to train the detection model. The specific training method is:

[0066] B1: Calculate the relative density between samples , calculated as follows:

[0067] ,

[0068] in For sample The K nearest neighbor samples, For sample and The Euclidean distance between Represents the average distance of all samples to their K nearest neighbor samples; the closer the distance between sample points, the higher the relative density between them;

[0069] B2: The preprocessed normal training data is learned to obtain the density-induced support vector data description model. The objective function of the model is as follows:

[0070] ;

[0071] ;

[0072] in is the relative density, is the penalty term coefficient, is the radius of the optimal hypersphere, is the center of the optimal hypersphere, is the slack variable;

[0073] B3: Solve the objective function by Lagrange multiplier method and obtain its dual form:

[0074] ;

[0075] in is the Lagrange coefficient;

[0076] B4: Get the coefficient After calculating the radius of the model , calculated as follows:

[0077] ;

[0078] in is the mean relative density.

[0079] S3: Using the density-induced support vector data description recursive feature elimination method to learn the preprocessed fault training data, obtain the fault feature sorting set, and obtain the diagnosis model through the fault feature sorting set training;

[0080] The acquisition of the fault feature sorting set includes the following steps:

[0081] C1: Initialize the feature set to the entire feature set of the data, that is , is the number of features, and the feature subset after initialization sorting is ;

[0082] C2: Using the feature set The fault data is used to train the density-induced support vector data description model and obtain the radius of the model. ;

[0083] C3: Calculate the elimination features , then the feature set is , that is, the current feature set is ;

[0084] C4: Extract the characteristics of fault data For training data, a density-induced support vector data description model is constructed, and the radius of the model is ;

[0085] C5: Calculation criteria function ;

[0086] C6: Find the feature with the largest criterion value , remove this feature from the feature subset , update the feature ranking subset ;

[0087] C7: Return to step C2 until The last remaining feature in ;

[0088] C8: Update sorted subset , and obtain the final feature sorting set ,According to the sorted set of fault features, the fault type is determined, and the features ranked in the front indicate the greater correlation with the fault.

[0089] 2. Input test data into the diagnostic module

[0090] like Figure 2 As shown in the figure, the operation process of the diagnostic module is as follows:

[0091] S4: Standardize and preprocess the test data and input it into the detection model for detection. If the test data is predicted to be abnormal data, it goes to step S5, otherwise it is output as normal data;

[0092] The detection model is tested by:

[0093] D1: Perform standardization preprocessing on the test data. The standardization formula is: ,in For test data The mean of For test data The standard deviation of

[0094] D2: Predict the test sample points and calculate the distance from the sample points to the center of the hypersphere , the calculation formula is:

[0095] ;

[0096] When the distance between the sample points in the hypersphere , it indicates that the sample is normal data, otherwise , that is, the sample is abnormal data.

[0097] S5: The test data predicted as abnormal data is input into the diagnosis model, and the density-induced support vector data description recursive feature elimination method is used in the diagnosis model to obtain the feature ranking of the abnormal data, and compared with the fault feature ranking set to confirm the fault type. The density-induced support vector data description recursive feature elimination method used here is consistent with that in step S3, that is, steps C1 to C8.

[0098] The method of the present invention uses the collected normal data and fault data during model training, wherein the normal data is used to build a detection model and the fault data is used to build a diagnosis model. When the method of the present invention is tested, in the detection stage, the prediction result of each sample is obtained by the density induced support vector data description method. Whether it is fault data or abnormal data, their feature sorting sets are obtained by the density induced support vector data description recursive feature elimination method.

[0099] In order to verify the effectiveness and practical effect of the method of the present invention, a demonstration experiment was carried out on the present invention in this embodiment, as follows:

[0100] The data set used in the demonstration experiment is: a dry data set collected by the dry integrated system control instrument. The dry data set is a real industrial data set collected by the dry integrated system control. It includes 5289 samples, each of which contains 7 features, namely, the total pipe wind speed (m / s), cyclone dust collector temperature, dust collector temperature, system total negative pressure (Kpa), dust collector pressure difference (Pa), total pipe water pressure (kg / cm2) and air pressure (kg / cm2). The experiment contains two types of data, one is the normal operation data, that is, normal data, and the other is the abnormal data generated by system fluctuations. Figure 3 The control logic diagram of the dry integrated system is shown, and the process of collecting data is shown. In order to verify the effectiveness of the fault diagnosis method proposed in this invention, different amounts of data are used for training in the experiment, and the average value of 10 accuracy rates on the test set is taken as the evaluation index.

[0101] The experimental hardware environment in this embodiment is: MacbookPro with Intel Core i5 (2.7GHz) processor and 8GB memory. Code running environment: Matlab (R2015b).

[0102] The experimental results are as follows:

[0103] In order to verify the effectiveness and superiority of the present invention, the method of the present invention is compared with the traditional method of support vector data description in this embodiment. Table 1 records the accuracy of the traditional method of support vector data description and the present invention under different training data amounts. It can be seen from Table 1 that the detection accuracy of the present invention is higher than that of the traditional method of support vector data description, and abnormal data can be detected more effectively. In addition, by comparing the fault feature ranking set obtained by the present invention with the actual situation, the present invention can put the features related to the fault in the front, which is convenient for subsequent maintenance management.

[0104] Table 1 Average classification accuracy of SVDD and the present invention under different training data amounts

[0105]

[0106] Figure 4 and Figure 5 The monitoring results detected by using SVDD and the method of the present invention are given, wherein Figure 4 is the detection result of SVDD, Figure 5 This is the detection result of the method of the present invention, where red indicates abnormal data and blue indicates normal data. The test data contains 245 faulty data, 118 of which are detected by SVDD and 304 by the present invention, indicating that the present invention is more sensitive to abnormal data and has better detection effect than the traditional support vector data description method.

[0107] In addition, by using the method of the present invention to study fault data, the ranking of fault-related features is obtained as {1, 4, 2, 3, 5, 6, 7}, which is consistent with the faults recorded in reality, that is, the fault caused by system fluctuation is most related to feature 1 (total pipe wind speed), and in fault diagnosis, the first feature of the feature ranking set of abnormal data, that is, the most relevant feature, is feature 1 (total pipe wind speed). This shows the effectiveness of the present invention in fault diagnosis, which can well obtain the features related to the fault and determine the fault type, thereby facilitating the identification of the cause of the fault and solving the fault.

Claims

1. A fault diagnosis method based on density-induced support vector data description recursive feature elimination, characterized in that: The steps include: S1: Collect data and perform standardized preprocessing on the data to obtain normal data and fault data; S2: Use normal data to train the detection model; S3: training the fault data to obtain a fault feature sorting set, and obtaining a diagnosis model through the training of the fault feature sorting set; S4: Standardize and preprocess the test data and input it into the detection model for detection. If the test data is predicted to be abnormal data, go to step S5, otherwise it is output as normal data: S5: Obtain feature ranking of abnormal data and compare it with the fault feature ranking set to confirm the fault type; The training method of the detection model in step S2 is: B1: Calculate the relative density ρ between samples i , calculated as follows: in For sample x i The K nearest neighbor samples, For sample x i and The Euclidean distance between Represents the average distance of all samples to their K nearest neighbor samples; B2: The preprocessed normal training data is learned to obtain the density-induced support vector data description model. The objective function of the model is as follows: where ρ i is the relative density, C≥0 is the penalty coefficient, R is the radius of the optimal hypersphere, a is the center of the optimal hypersphere, ξ i is the slack variable; B3: Solve the objective function by Lagrange multiplier method and obtain its dual form: where α i is the Lagrange coefficient; B4: Get coefficient α i Then calculate the radius R of the model 2 , calculated as follows: in is the mean value of relative density; In step S3, the preprocessed fault training data is learned using a density-induced support vector data description recursive feature elimination method to obtain a fault feature ranking set; The acquisition of the fault feature ranking set in step S3 includes the following steps: C1: Initialize the feature set to the entire feature set of the data, that is, s = {1, 2, ..., m}, where m is the number of features, and the feature subset after initialization sorting is C2: Use the fault data with feature set s to train the density induced support vector data description model and obtain the radius of the model C3: Calculate and remove feature p. At this time, the feature set is sp, that is, the current feature set is s = {1, ..., p-1, p+1, ..., m}; C4: Take the feature s of the fault data as training data, build a density-induced support vector data description model, and get the radius of the model as R 2 (X s ); C5: Calculation criteria function C6: Find the feature with the largest criterion value Remove this feature from the feature subset, that is, s←ss * , update the feature sorting subset r←[s * ;r]; C7: Return to step C2 until the last feature remaining in s is s * ; C8: Update the sorted subset r←[s * ; r], and obtain the final feature sorting set r; In step S5, the density induced support vector data is used to describe the recursive feature elimination method to obtain the feature ranking of the abnormal data.

2. A fault diagnosis method based on density induced support vector data description recursive feature elimination according to claim 1, characterized in that: The specific process of the standardization preprocessing in step S1 is as follows: A1: For the collected normal data x i , i = 1, 2, ..., n, perform standardization preprocessing, where n is the total number of normal data; The standardized formula is in For x i The mean of i The standard deviation of A2: For fault data x' i Perform standardization preprocessing, the standardization formula is in For x' i The mean of i The standard deviation of .

3. A fault diagnosis method based on density induced support vector data description recursive feature elimination according to claim 1, characterized in that: The method for detecting the detection model in step S4 is: D1: Standardize and preprocess the test data; D2: Predict the test sample points, calculate the distance from the sample points to the center of the hypersphere, and determine abnormal data and normal data based on the obtained distance.

4. A fault diagnosis method based on density induced support vector data description recursive feature elimination according to claim 3, characterized in that: The normalization formula in step D1 is: in For the test data x i ^mean, σ^ is the test data x i ^Standard deviation.

5. A fault diagnosis method based on density induced support vector data description recursive feature elimination according to claim 4, characterized in that: The distance from the sample point to the center of the hypersphere in step D2 is expressed as: When the distance d(x) of the sample point in the hypersphere is ≤ R 2 When , it indicates that the sample is normal data, otherwise d(x)>R 2 , that is, the sample is abnormal data.

Citation Information

Patent Citations

  • Fault detection method and system

    CN106295712A

  • Subway short-time passenger flow prediction method based on machine learning

    CN107291668A

  • Key feature recognition method for anomaly detection of bulk grain container logistics transportation environment

    CN113592379A

  • Complex transmission device fault diagnosis traceability method and system based on time-frequency domain characteristics

    CN115096581A

  • Failure prediction diagnosis system and method through pattern analysis according to failure type

    KR102618023B1