Fault feature extraction method based on sensitivity analysis
Through the fault feature extraction method based on sensitivity analysis, the autoencoder and LASSO regression model are used to screen out the characteristics with high sensitivity in mechanical equipment failures, solving the problem of inaccurate fault feature extraction in the existing technology and improving fault diagnosis accuracy and interpretability.
Patent Information
- Application Number
- CN202411585111.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-07
- Publication Date
- 2025-02-07
AI Technical Summary
The prior art is difficult to effectively extract fault features in mechanical equipment fault diagnosis, resulting in poor accuracy and robustness of fault feature extraction.
The fault feature extraction method based on sensitivity analysis is adopted, and effective features with high sensitivity to faults are screened through the offline training stage and the online detection stage using the autoencoder model and the LASSO regression model.
It realizes automatic screening of features with high fault sensitivity, improves the performance and interpretability of the fault diagnosis model, and enhances the accuracy and robustness of fault diagnosis.
Smart Images
Figure CN119397244A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of mechanical equipment status monitoring, and in particular to technical methods such as feature extraction, linear regression, and anomaly detection. Background Art
[0002] Feature selection plays a vital role in the field of mechanical equipment fault diagnosis. It can select the most informative features from the original data, improve model performance, reduce overfitting, and speed up model training and prediction. Feature selection is particularly important in large-scale data sets and high-dimensional data, because unnecessary features will increase computational complexity and introduce redundant information.
[0003] The common feature selection methods currently available mainly include filtering methods and packaging methods. The filtering method is a common feature selection method that selects features with high correlation with the target variable by screening the features before training the model. Typical methods include variance threshold, mutual information, and correlation coefficient. However, the filtering method ignores the interactions and nonlinear relationships between features and may fail to capture some important features. The packaging method is a more complex and time-consuming feature selection method that uses a learning model to evaluate the importance of features and selects features based on the importance. Since the packaging method requires multiple model training, the computational complexity is high and it may not be applicable to large-scale data sets.
[0004] In addition, these methods do not take into account the interaction between features and cannot automatically select features with strong correlation with the target variable, resulting in poor accuracy and robustness of fault feature extraction. Therefore, an efficient fault feature extraction method is needed to help users accurately locate fault features. Summary of the invention
[0005] The present invention aims to provide a fault feature extraction method based on sensitivity analysis, which helps engineers to screen effective features with high fault sensitivity from a large number of data features, and then optimize the fault diagnosis model, while making the fault diagnosis conclusion more interpretable.
[0006] To achieve the above objectives, the technical solution of the present invention is: a fault feature extraction method based on sensitivity analysis, which is divided into an offline training stage and an online detection stage. The steps of the offline training stage are:
[0007] Step 1: Normal data preparation
[0008] Collect time series data samples of all characteristics of mechanical equipment in normal state, the number of characteristic values is N, and the number of normal data samples is M1;
[0009] Step 2: Anomaly detection model construction
[0010] The autoencoder model is trained based on the semi-supervised learning method. A new sample x is input into the trained model, and the result reconstructed by the autoencoder is x A , then the mean square error MSE is used as the evaluation index of reconstruction or abnormality:
[0011]
[0012] The steps of the online detection phase are:
[0013] Step 3: Anomaly Detection Data Accumulation
[0014] The anomaly detection model is used to monitor the operating status of mechanical equipment in real time until a certain number of abnormal data samples are obtained, and the number of samples is M2;
[0015] Step 4: Normalization
[0016] In order to unify the differences in the magnitude of each feature value, we first i , i = 1, 2, ..., M1 + M2, including normal data and abnormal data, marked as y i =f(X i ), where y i ∈{0,1};
[0017] Using Z-score normalization, we can calculate the data set X i The mean of all features in is μ j , j = 1, 2, ..., N, the variance is σ j , j = 1, 2, ..., N, normalized eigenvalues
[0018] Step 5: Fault feature selection
[0019] Based on the LASSO regression model, the data set X i Perform fitting to obtain the fitting coefficient β of all features i , i = 1, 2, ... N, and then sort the feature sensitivity according to the size of the fitting coefficient, determine the initial threshold k, and screen out r fault features X with high sensitivity k =(x1, x2, ..., x r )′;
[0020] Step 6: Fault Characteristic Evaluation
[0021] Based on the extracted fault feature β∈R r×1 , reconstruct the regression model y = β T x+b, using the test data set X t The model is evaluated according to the following criteria:
[0022] sst=Σ(y t -E(y t )) 2
[0023] sse=Σ(yy t ) 2 rsq=(1-sse / sst)∈[0,1]
[0024] When rsq>0.7, it means that the model can respond to the variability of the variable well and fit the data; if rsq<0.7, repeat step 5 and reselect the threshold k=k-α, where α is the model learning rate, a decimal greater than zero.
[0025] The beneficial effects of the present invention are:
[0026] In the present invention, a fault feature extraction method based on sensitivity analysis is proposed to help engineers automatically screen effective features with high fault sensitivity from a large number of data features, thereby facilitating manual data analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a fault feature extraction flow chart of the fault feature extraction method based on sensitivity analysis of the present invention. DETAILED DESCRIPTION
[0028] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0029] refer to Figure 1 , a fault feature extraction method based on sensitivity analysis proposed in an embodiment of the present invention is implemented as follows:
[0030] This method is divided into two stages: offline training and online detection. The main process in the offline training stage is:
[0031] Step 1: Normal data preparation. Collect time series data samples of all characteristics of mechanical equipment under normal conditions, the number of characteristic values is N, and the number of normal data samples is M1;
[0032] Step 2: Anomaly detection model construction. Train the autoencoder model based on the semi-supervised learning method. Input a new sample x into the trained model, and the result reconstructed by the autoencoder is x. A , then the mean square error MSE is used as the evaluation index of reconstruction or abnormality:
[0033]
[0034] The main process of the online detection stage is:
[0035] Step 3: Anomaly detection data accumulation. Use the anomaly detection model to monitor the operating status of mechanical equipment in real time until a certain amount of abnormal data samples are obtained, and the sample size is M2;
[0036] Step 4: Normalization. i (i=1, 2, ..., M1+M2) (including normal data and abnormal data) is normalized by Z-score and the data set X is calculated. i The mean of all features in is μ j (j=1,2,…,N), the variance is σ j (j=1,2,…,N), normalized eigenvalue
[0037] Step 5: Fault feature selection. Based on the LASSO regression model, the dataset X i Perform fitting to obtain the fitting coefficient β of all features i (i=1, 2, ...N), and then sort the feature sensitivity according to the size of the fitting coefficient, determine the initial threshold k, and screen out r fault features X with high sensitivity k =(x1, x2, ..., x r )′;
[0038] Step 6: Fault feature evaluation. Based on the extracted fault feature β∈R r×1 , reconstruct the regression model y = β T x+b, using the test data set X t The model is evaluated. When rsq≥0.7, it means that the model can respond to the variability of the variable well and fit the data. If rsq<0.7, repeat step 5 and reselect the threshold k=k-α, where α is the model learning rate (a decimal greater than zero).
Claims
1. A fault feature extraction method based on sensitivity analysis, characterized in that: This method is divided into an offline training stage and an online detection stage. The steps of the offline training stage are as follows: Step 1: Normal data preparation Collect time series data samples of all features in the normal state of mechanical equipment. The number of feature values is N, and the number of normal data samples is M1. Step 2: Abnormal detection model construction The autoencoder model is trained based on the semi-supervised learning method. A new sample x is input into the trained model, and the result reconstructed by the autoencoder is x A , then the mean square error MSE is used as the evaluation index of reconstruction or abnormality: The steps of the online detection stage are as follows: Step 3: Abnormal detection data accumulation Use the abnormal detection model to monitor the running state of mechanical equipment in real time until a certain number of abnormal data samples are obtained. The number of samples is M2. Step 4: Normalization processing In order to unify the differences in the magnitude of each feature value, we first i , i=1,2,...,M1+M2, including normal data and abnormal data, marked as y i =f(X i ), where y i ∈{0,1}; Using Z-score normalization, we can calculate the data set X i The mean of all features in is μ j , j = 1, 2, ..., N, with a variance of σ j , j = 1, 2, ..., N, normalized eigenvalues Step 5: Fault feature selection Based on the LASSO regression model, the data set X i Perform fitting to obtain the fitting coefficient β of all features i , i = 1, 2, ... N, and then sort the feature sensitivity according to the size of the fitting coefficient, determine the initial threshold k, and screen out r fault features X with high sensitivity k =(x1, x2, ..., x r )′; Step 6: Fault feature evaluation Based on the extracted fault feature β∈R r×1 , reconstruct the regression model y = β T x+b, using the test data set X t The model is evaluated according to the following criteria: sst=∑(y t -E(y t )) 2 sse=∑(yy t ) 2 rsq = (1 - sse / sst) ∈ [0, 1] When rsq > 0.7, it indicates that the model can better respond to the variability of variables and fit the data. If rsq < 0.7, repeat Step 5 and reselect the threshold k = k - □, where □ is the model learning rate, a decimal greater than zero.