Rotary machinery fault diagnosis method based on adaptive feature optimization and integrated model

Abstract

The invention provides a rotating machinery fault diagnosis method based on adaptive feature optimization and an integrated model, and belongs to the field of rotating machinery equipment fault diagnosis. The system comprises a data processing module, a feature screening and combination module, a fault category evaluation module, a feature enhancement generation module, a performance monitoring and model protection module and a fault diagnosis module. Basic input of a diagnosis model is initially obtained through data processing and feature screening, after the basic input is evaluated by the fault category evaluation module, if performance does not meet requirements, the feature set is adjusted by the enhanced feature generation module and the performance monitoring and model protection module, and the adjusted features enter the fault category evaluation module again to be evaluated again. The process can be cyclically performed and optimized in real time until the diagnostic performance reaches a preset threshold or an iteration termination condition. The method can dynamically optimize features and models according to composite faults and complex working conditions, improves the accuracy and robustness of fault diagnosis, and has a wide application prospect.

Description

Technical Field

[0001] The present invention belongs to the field of rotating machinery fault diagnosis, and in particular relates to a rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model. Background Art

[0002] With the increasing complexity and automation of industrial equipment, equipment fault diagnosis plays a critical role in ensuring system safety and reliability. Vibration signal analysis is a key method for fault detection in rotating machinery (such as bearings and gearboxes). It typically extracts characteristic parameters from the vibration signal and performs pattern recognition to determine the fault type and severity. Traditional methods often determine a feature set and a single classification model (such as RBF or SVM) during the system design phase, which is then applied permanently. However, in complex operating conditions or complex fault scenarios, these pre-defined features often fail to fully capture key information, resulting in reduced diagnostic accuracy. Existing feature selection methods, such as filtering, wrapping, and embedding, are used to select important features from a large number of features to improve model performance. However, these methods are typically static and lack the ability to dynamically adjust during the diagnostic process. When new fault modes or multiple faults occur simultaneously, existing features may not be sufficient to distinguish different fault categories, and there is no mechanism to automatically add new feature dimensions to address this. Furthermore, blindly adding features can lead to redundancy, overfitting, and even diagnostic model degradation. Existing methods lack performance protection and fallback mechanisms to avoid this problem. In summary, existing systems lack performance protection and fallback mechanisms to avoid performance degradation. There is an urgent need for a diagnostic framework that takes adaptive feature optimization as the core and combines ensemble learning to improve robustness. Summary of the Invention

[0003] To solve the above problems, the present invention proposes a rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model. The feature set is dynamically optimized through a closed-loop feedback mechanism, and an integrated diagnostic model is used for fault judgment to improve the diagnostic accuracy and stability.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated models includes the following six functional modules: data processing module, feature screening and combination module, fault category assessment module, feature enhancement generation module, performance monitoring and model protection module, and fault diagnosis module. These modules are interconnected through data channels to form a closed-loop iterative process, enabling the system to adaptively optimize the diagnostic process based on diagnostic results and performance feedback. Specifically:

[0006] Data processing module: This module pre-processes the collected raw vibration signal, dividing the continuous vibration data into fixed-length time windows for feature extraction. Preferably, a sliding time window of 1024 points is used to segment the vibration signal and extract both time-domain and frequency-domain features.

[0007] Feature Screening and Combination Module: This module selects a subset of features highly relevant to fault diagnosis from the large number of initial features extracted by the data processing module. This module uses a combination of three feature selection strategies: filtering, wrapping, and embedding. It also employs a voting fusion mechanism to integrate the results of each algorithm to select the feature subset with the greatest diagnostic value.

[0008] Fault Category Assessment Module: This module performs preliminary fault type determination and performance evaluation based on the currently selected feature set. This module monitors the performance metrics (such as accuracy, recall, and F1 score) output by the fault diagnosis module and compares them against preset thresholds to determine the reliability of the diagnosis. It also provides error analysis based on feature importance, which serves as a basis for feature enhancement. For example, if the F1 score falls below the preset threshold, the current feature set may not be sufficient for accurate fault identification, and new features may be needed.

[0009] The Feature Enhancement Generation Module introduces new features to enrich the feature space when existing features are deemed insufficient for fault identification. This module determines a feature generation strategy based on information provided by the Fault Classification Assessment Module. It generates feature enhancement factors based on feature importance analysis using SHAP values. The generated new features are then fused with the original features to enhance the subsequent diagnostic model's ability to discriminate against difficult samples. Once the new features are generated, this module submits them to the Fault Classification Assessment Module for evaluation.

[0010] Performance monitoring and model protection module: used to monitor the impact of the introduction of new features on the diagnostic performance, and perform a rollback operation when it is found that the performance has not improved or has decreased. Specifically, after the supplementary feature generation module adds new features and retrains the fault diagnosis model, the rollback mechanism module compares the performance indicators of the new and old models. If it is found that the F1 score indicator of the model does not meet the improvement requirements or even decreases after the addition of the new feature, a rollback is triggered to restore the feature set and model to the state before the introduction of the feature to avoid interference caused by invalid features. At the same time, the rollback mechanism module can also limit the number of iterations or the size of the feature set, thereby preventing the unlimited increase of features from causing the model to be overly complex or overfitting. This module is equivalent to providing performance protection logic for the entire optimization process to ensure that feature optimization is carried out in the direction of improving diagnostic performance.

[0011] Fault diagnosis module: It is composed of parallel diagnostic classification models (such as support vector machines, random forests, gradient boosting trees, and shallow neural networks). Each diagnostic classification model independently judges the same input features and generates its own category probability distribution. The fusion unit adopts a weighted soft voting strategy to integrate the outputs of each diagnostic classification model according to the weights obtained in advance or online learning to calculate the final fault category.

[0012] The integrated approach of these modules forms a closed-loop fault diagnosis process. Initially, data processing and feature screening generate the basic inputs for the diagnostic model. After evaluation by the fault category assessment module, if performance falls short of expectations, the enhanced feature generation module and the performance monitoring and model protection module adjust the feature set. The adjusted features are then re-evaluated by the fault category assessment module. This process can be repeated, with real-time optimization, until diagnostic performance reaches a preset threshold or an iteration termination condition is reached.

[0013] The advantages and positive effects of the present invention are as follows: Compared with existing technologies, the framework has a highly modular design, with each functional module being independent and decoupled, making it easy to adjust and expand for different applications; a feedback mechanism for diagnostic results is introduced, which can dynamically optimize the feature set and select the optimal model based on model performance, significantly improving the diagnostic accuracy in complex fault situations. At the same time, this module also provides a rollback mechanism for performance protection to prevent invalid features or improper adjustments from negatively affecting the diagnostic results, thereby ensuring the reliability and stability of the diagnostic process; the entire framework implements automated feature optimization and model selection, reducing manual intervention and improving the intelligent level of equipment health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Flowchart of the rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1 As shown in Figure 1, the rotating machinery fault diagnosis method based on adaptive feature optimization and integrated modeling includes the following six functional modules: data processing module, feature screening and combination module, fault category assessment module, feature enhancement generation module, performance monitoring and model protection module, and fault diagnosis module. These modules are interconnected via data channels to form a closed-loop iterative process, enabling the system to adaptively optimize the diagnostic process based on diagnostic results and performance feedback. Specifically:

[0017] Data processing module: This module pre-processes the collected raw vibration signal, dividing the continuous vibration data into fixed-length time windows for feature extraction. Preferably, a sliding time window of 1024 points is used to segment the vibration signal and extract both time-domain and frequency-domain features.

[0018] Feature Screening and Combination Module: This module selects a subset of features highly relevant to fault diagnosis from the large number of initial features extracted by the data processing module. This module utilizes a combination of feature selection strategies, including filtering, wrapping, and embedding, and employs a voting fusion mechanism to integrate the results of each algorithm to select the feature subset with the greatest diagnostic value.

[0019] Fault Category Assessment Module: This module performs preliminary fault type determination and performance evaluation based on the currently selected feature set. This module monitors the performance metrics (such as accuracy, recall, and F1 score) output by the fault diagnosis module and compares these metrics against preset thresholds to determine the reliability of the diagnosis. It also provides error analysis based on feature importance, which serves as a basis for feature enhancement. If key metrics such as the F1 score fall below the preset threshold, the current feature set may not be sufficient for accurate fault identification, and new features may be needed.

[0020] The Feature Enhancement Generation Module introduces new features to enrich the feature space when existing features are deemed insufficient for fault identification. This module determines a feature generation strategy based on information provided by the Fault Classification Assessment Module. It generates feature enhancement factors based on feature importance analysis using SHAP values. The generated new features are then fused with the original features to enhance the subsequent diagnostic model's ability to discriminate against difficult samples. Once the new features are generated, this module submits them to the Fault Classification Assessment Module for evaluation.

[0021] Performance monitoring and model protection module: used to monitor the impact of the introduction of new features on the diagnostic performance, and perform a rollback operation when it is found that the performance has not improved or has decreased. Specifically, after the supplementary feature generation module adds new features and retrains the fault diagnosis model, the rollback mechanism module compares the performance indicators of the new and old models. If it is found that the F1 score indicator of the model does not meet the improvement requirements or even decreases after the addition of the new feature, a rollback is triggered to restore the feature set and model to the state before the introduction of the feature to avoid interference caused by invalid features. At the same time, the rollback mechanism module can also limit the number of iterations or the size of the feature set, thereby preventing the unlimited increase of features from causing the model to be overly complex or overfitting. This module is equivalent to providing performance protection logic for the entire optimization process to ensure that feature optimization is carried out in the direction of improving diagnostic performance.

[0022] Fault diagnosis module: It is composed of parallel diagnostic classification models (such as support vector machines, random forests, gradient boosting trees, shallow neural networks, etc.). Each diagnostic classification model independently judges the same input features and generates its own category probability distribution. The fusion unit adopts a weighted soft voting strategy to integrate the outputs of each diagnostic classification model according to the weights obtained in advance or online learning to calculate the final fault category.

[0023] The integrated approach of these modules forms a closed-loop fault diagnosis process. Initially, data processing and feature screening generate the basic inputs for the diagnostic model. After evaluation by the fault category assessment module, if performance falls short of expectations, the enhanced feature generation module and the performance monitoring and model protection module adjust the feature set. The adjusted features are then re-evaluated by the fault category assessment module. This process can be repeated, with real-time optimization, until diagnostic performance reaches a preset threshold or an iteration termination condition is reached.

[0024] Data processing module: This module is responsible for preprocessing and feature extraction of vibration signals collected online from rotating machinery to generate the initial time-frequency domain feature set required for subsequent diagnosis. Its specific implementation is as follows:

[0025] Step 1: Take a The sliding window segmentation is preferably 50% overlap (i.e. 512 points), and the discrete signal segments are obtained. .

[0026] Step 2: For each signal segment Calculate the time domain features listed in formulas (1)-(12) to form the time domain feature vector .

[0027] Step 3: Transmit the same signal Performing discrete Fourier transform yields , and calculate the power spectrum , its total energy Then, the frequency domain feature vector is extracted according to formulas (13)-(16): .

[0028] Step 4: Transform the time domain feature vector and frequency domain eigenvector Cascade to form the initial feature vector of the signal segment Repeat the above steps for all signal segments to construct the initial feature pool , this module will then Output to the "Feature Screening and Combination Module" for further processing.

[0029] (1)

[0030] (2)

[0031] (3)

[0032] (4)

[0033] (5)

[0034] (6)

[0035] (7)

[0036] (8)

[0037] (9)

[0038] (10)

[0039] (11)

[0040] (12)

[0041] (13)

[0042] (14)

[0043] (15)

[0044] , (16)

[0045] Feature screening and combination module: This module uses three strategies: filtering, wrapping and embedding. The features are screened and a preliminary feature subset is formed by voting fusion , specifically including:

[0046] Step 1: Input the initial feature set , target fault label set .

[0047] Step 2: Filtering (1) Pearson correlation coefficient: , according to the absolute value Select first in descending order Features (2) Mutual Information Method: ,according to Descending order retains the first Features.

[0048] Step 3: Request for Feedback (RFE) Taking random forest as the base model, the iterative steps are: (1) Train a random forest using the current feature set; (2) Calculate feature importance based on the reduction in impurity (Gini) when the node is split; (3) Eliminate the least important features in each round until only indivual.

[0049] Step 4: Embedded Screening: according to Before the election item.

[0050] Step 5: Voting Integration (1) Each method outputs its screening result list; (2) Count the votes for each feature (1 point is added for each occurrence); (3) The first one with the highest number of votes Features (preferred )constitute .

[0051] Step 6: Passed to the "Fault Category Evaluation Module" for preliminary diagnosis.

[0052] Fault category assessment module: Based on the above preliminary screening features, the fault diagnosis module is used to conduct an assessment and provide the assessment results. The specific steps are as follows: Step 1: Calculate the macro-average based on the validation set ; Step 2: If , directly output the diagnosis results; Step 3: Otherwise, filter out the The category list is used as the category to be enhanced and the feature enhancement generation module is triggered.

[0053] Feature enhancement generation module: for fault categories with low F1 scores ( ), SHAP value is used to perform feature importance analysis, the specific operations are as follows.

[0054] Step 1: Enter the set of fault categories that need to be enhanced ,in Representation category F1 score of the current preliminary feature subset ; SHAP value matrix of the integrated model trained on the validation set , , .

[0055] Step 2: Alternative Feature Pool Preparation: Maintain a candidate feature library in the system, including but not limited to: energy of each sub-band of wavelet packet; main peak of envelope spectrum; energy of short-time Fourier transform (STFT) frequency band; energy of each mode of variational mode decomposition (VMD) and other statistical features.

[0056] Step 3: SHAP directional selection, for each fault category : (1) Set the current feature set All features in according to Sort descending, select first The most critical feature set ; (2) In the spare feature pool , calculate each candidate feature and The correlation coefficient (such as Pearson correlation coefficient or mutual information) of the two is taken as the highest correlation. (Preferred ) features, and get the incremental feature set .

[0057] Step 4: Incremental feature fusion: Merge the incremental features of all weak categories: ,right Each feature in is normalized to zero-mean and unit variance to obtain , forming an enhanced feature set: .

[0058] Step 5: Output will enhance the feature set Return to the "Feature Screening and Combination Module" or directly transfer to the "Fault Category Assessment Module" for retraining and evaluation.

[0059] Performance monitoring and model protection module: This module uses enhanced feature set in the "fault category assessment module" After training and evaluating the integrated model, it is responsible for comparing the performance of the new and old models and deciding whether to roll back or continue iterating based on the results. The specific monitoring steps are as follows.

[0060] Step 1: Input (1) Initial optimal state: feature set: ; Corresponding integrated model: ; Macro average F1: ; (2) Current enhancement status: Feature set: ;based on Trained ensemble model: .

[0061] Step 2: Performance Evaluation (1) Calculate the macro-average F1 of the two models on the same validation set: , ; (2) Computational performance gain: .

[0062] Step 3: Rollback and update logic (1) If (The new model does not bring any improvement) 1) Rollback operation: , , ; 2) Set the end flag, terminate feature addition and model iteration, and enter the output module; (2) Otherwise (New model performance improvement) 1) Update the best status: , , ; 2) Threshold judgment: If , then set the end flag and enter the output module; otherwise, if the number of additions or feature dimensions have not reached the upper limit, return to the "Feature Enhancement Generation Module" to continue the next round of additions; otherwise, the maximum number of iterations or maximum dimension is reached, set the end flag and enter the output module.

[0063] Step 4: Iteration Cap and Protection: Maximum Number of Iterations Used to avoid infinite loops; maximum feature dimension :Prevent overfitting and computational bottlenecks caused by too many features. When any upper limit is triggered and the performance is still not up to standard, the current 、 Output the optimal result.

[0064] The fault diagnosis module includes the following steps:

[0065] Step 1: Diagnostic classification model training: For diagnostic classification models such as support vector machine, random forest, gradient boosting tree, shallow neural network, etc., train them on the validation set to obtain classification probability and F1 score .

[0066] Step 2: Weight Normalization: Calculate weights .

[0067] Step 3: Soft Voting Fusion: Fusion Probability , and Determine the fault category and output the diagnostic accuracy.

[0068] In this implementation, the fault diagnosis of rotating machinery (rolling bearings) is used as an application example, and the Case Western Reserve University (CWRU) bearing fault public dataset is selected. The specific data selection is shown in Table 1.

[0069] Table 1 Data Table

[0070] Category number Category Name Fault diameter bearing components Corresponding data file (.mat) illustrate 1 Normal — — 99.mat Baseline data, 2HP, 1750rpm 2 Inner race fault 0.007″ (IR007) 0.007″ Inner ring 107.mat 2HP, 1750rpm 3 Rolling element failure 0.007″ (BF007) 0.007″ rolling elements 120.mat 2HP, 1750rpm 4 Outer ring fault 0.007″ (OR007) 0.007″ Outer Circle @ 6:00 132.mat 2HP, 1750rpm 5 Inner race fault 0.014″ (IR014) 0.014″ Inner ring 171.mat 2HP, 1750rpm 6 Rolling element failure 0.014″ (BF014) 0.014″ rolling elements 187.mat 2HP, 1750rpm 7 Outer ring fault 0.014″ (OR014) 0.014″ Outer Circle @ 6:00 199.mat 2HP, 1750rpm 8 Inner race fault 0.021″ (IR021) 0.021″ Inner ring 211.mat 2HP, 1750rpm 9 Rolling element failure 0.021″ (BF021) 0.021″ rolling elements 224.mat 2HP, 1750rpm 10 Outer ring fault 0.021″ (OR021) 0.021″ Outer Circle @ 6:00 236.mat 2HP, 1750rpm

[0071] Data processing module: All sample segments are segmented using a sliding window of 1024 points. The window shift is preferably 50% overlap (i.e., 512 points), resulting in 3074 sample segments. According to formulas (1)-(16), 12 time domain features and 4 frequency domain features are calculated; the output initial feature pool is constructed. .

[0072] Feature screening and combination module: for : Filtering method (mutual information), wrapping method (random forest), embedding method (Lasso regression) voting fusion, take the first 10 features to form the preliminary feature subset: .

[0073] Fault category assessment module: Input [S6 fault diagnosis module], the overall accuracy (Accuracy): 92.84%, the diagnostic performance indicators of each category are shown in Table 2, the category set in the table , transfer to "Feature Enhancement Generation Module".

[0074] Table 2 Diagnostic performance indicators for each category

[0075] Class Precision Recall <![CDATA[ F 1 macro ]]> 1 0.98233 0.96864 0.97544 2 0.98611 0.91026 0.94667 3 0.81429 0.89062 0.85075 4 1 1 1 5 0.95775 0.95775 0.95775 6 0.81944 0.89394 0.85507 7 0.83099 0.80822 0.81944 8 1 1 1 9 0.77465 0.74324 0.75862 10 0.95775 1 0.97842

[0076] Feature enhancement generation module: Each category: select by SHAP value The first 6 features are selected, and the first 4 features with the highest correlation are selected from the spare feature pool (energy of each sub-band of wavelet packet, main peak of envelope); Each feature in is normalized to zero-mean and unit variance to obtain : {'WP3_E_band1'}, {'WP3_E_band2'}, {'EnvE_1_2kHz'}, {'EnvE_2_3kHz'}, and with Fusion , the feature dimension increases from 10 to 14.

[0077] Performance monitoring and model protection module: based on Train the integrated model and calculate the new , as shown in Table 3, it can be seen that , and all categories , enhancement is completed. Using the best feature set ; Best ensemble model Output fault category, overall diagnosis accuracy (Accuracy): 98.2%.

[0078] Table 3 Categories F 1 table

[0079] <![CDATA[

F 1 best

F 1 cur

[0080] This example fully demonstrates the feasibility and effectiveness of the present invention. In practical applications, the feature set and base learner combination can be adjusted according to specific working conditions and requirements. The adaptive architecture of the present invention has good flexibility and robustness, and is suitable for the intelligent health management of various equipment such as rotating machinery.

[0081] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. A rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model, characterized in that: The system consists of six functional modules: data processing module, feature screening and combination module, fault category assessment module, feature enhancement generation module, performance monitoring and model protection module, and fault diagnosis module. These modules are interconnected through data channels to form a closed-loop iterative process, enabling the system to adaptively optimize the diagnostic process based on diagnostic results and performance feedback. Specifically: Data processing module: used to pre-process the collected raw vibration signal and divide the continuous vibration data into time windows of fixed length to extract features; preferably, a sliding time window of 1024 points is used to segment the vibration signal and extract time domain features and frequency domain features; Feature screening and combination module: This module is used to select a subset of features with high relevance to fault diagnosis from the large number of initial features extracted by the data processing module. This module uses a combination of feature selection strategies such as filtering, wrapping, and embedding, and integrates the screening results of each algorithm using a voting fusion mechanism to select the feature subset with the greatest diagnostic value. Fault Category Assessment Module: This module is used to make preliminary fault type determinations and performance evaluations based on the currently selected feature set. It monitors the performance indicators (accuracy, recall, F1 score, etc.) output by the fault diagnosis module and compares the performance with preset thresholds to determine whether the diagnosis is reliable. It also provides error analysis based on feature importance as a basis for feature enhancement. For example, when key indicators such as the F1 score fall below the preset threshold, it indicates that the current feature set may not be sufficient to support accurate fault identification and that new features need to be introduced. Feature enhancement generation module: used to introduce new features to enrich the feature space when it is determined that the existing features cannot fully identify the fault; This module can determine the feature generation scheme based on the information provided by the fault category assessment module; Generate feature enhancement factors based on the feature importance analysis results of SHAP values; The generated new features are fused with the original features to enhance the discrimination ability of the subsequent diagnosis model for difficult samples. Once the new features are generated, this module sends them to the fault category assessment module for evaluation; Performance Monitoring and Model Protection Module: This module is used to monitor the impact of new feature introductions on diagnostic performance and to roll back performance if it detects no improvement or a decrease in performance. Specifically, after the Supplementary Feature Generation Module adds new features and retrains the fault diagnosis model, the Rollback Mechanism Module compares the performance indicators of the new and old models. If it is found that the F1 score and other indicators of the model do not meet the improvement requirements or even decrease after the addition of the new feature, a rollback is triggered to restore the feature set and model to the state before the introduction of the feature to avoid interference caused by invalid features. At the same time, the Rollback Mechanism Module can also limit the number of iterations or the size of the feature set to prevent the unlimited addition of features from leading to excessive complexity or overfitting of the model. This module is equivalent to providing performance protection logic for the entire optimization process, ensuring that feature optimization proceeds in the direction of improving diagnostic performance. Fault diagnosis module: It is composed of several diagnostic classification models (such as support vector machines, random forests, gradient boosting trees, shallow neural networks, etc.) in parallel. Each diagnostic classification model independently judges the same input features and generates its own category probability distribution. The fusion unit adopts a weighted soft voting strategy to integrate the outputs of each diagnostic classification model according to the weights obtained in advance or online learning to calculate the final fault category.

2. The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model according to claim 1 is characterized in that: The data processing module includes the following steps: Step 1: Take a The sliding window segmentation is preferably 50% overlap (i.e. 512 points), and the discrete signal segments are obtained. ; Step 2: For each signal segment Calculate time domain features (mean, variance, standard deviation, RMS value, peak-to-peak value, absolute peak value, kurtosis, skewness, kurtosis factor, margin factor, impulse factor, and crest factor) to form a time domain feature vector ; Step 3: Transmit the same signal Performing discrete Fourier transform yields , and calculate the power spectrum , its total energy , and then extract the frequency domain feature vector (spectral energy, average spectral energy, spectral entropy, spectral variance); Step 4: Transform the time domain feature vector and frequency domain eigenvector Cascade to form the initial feature vector of the signal segment , repeat the above operation for all signal segments to construct the initial feature pool, ,in is the total number of segments.

3. The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model according to claim 1 is characterized in that: Feature screening and combination module: using filtering, wrapping and embedding strategies to The features are screened and a preliminary feature subset is formed by voting fusion , specifically including the following steps: Step 1: Input the initial feature set , target fault label set ; Step 2: Filtering (1) Pearson correlation coefficient: , according to the absolute value Select first in descending order Features (2) Mutual Information Method: ,according to Descending order retains the first Features Step 3: Request for Feedback (RFE) Taking random forest as the base model, the iterative steps are: (1) Train a random forest using the current feature set; (2) Calculate feature importance based on the reduction in impurity (Gini) when the node is split; (3) Eliminate the least important features in each round until only indivual; Step 4: Embedded Screening: according to Before the election item; Step 5: Voting Integration (1) Each method outputs its screening result list; (2) Count the votes for each feature (1 point is added for each occurrence); (3) The first one with the highest number of votes Features (preferred )constitute ; Step 6: Passed to the "Fault Category Evaluation Module" for preliminary diagnosis.

4. The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model according to claim 1 is characterized in that: Fault category assessment module: Based on the above preliminary screening features, the fault diagnosis module is used to evaluate and give the evaluation results, which specifically includes the following steps: Step 1: Calculate the macro-average based on the validation set ; Step 2: If , directly output the diagnosis results; Step 3: Otherwise, filter out the The category list is used as the category to be enhanced and the feature enhancement generation module is triggered.

5. The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model according to claim 1 is characterized in that: Feature enhancement generation module: For fault categories with lower F1 scores ( ), SHAP value is used to perform feature importance analysis, which includes the following steps: Step 1: Enter the set of fault categories that need to be enhanced ,in Representation category F1 score; Current preliminary feature subset ; SHAP value matrix of the integrated model trained on the validation set , , ; Step 2: Alternative Feature Pool Preparation: Maintain a candidate feature library in the system, including but not limited to: energy of each sub-band of wavelet packet; main peak of envelope spectrum; energy of frequency band of short-time Fourier transform (STFT); energy of each mode of variational mode decomposition (VMD), and other statistical features; Step 3: SHAP directional selection, for each fault category : (1) Set the current feature set All features in according to Sort descending, select first The most critical feature set ; (2) In the spare feature pool , calculate each candidate feature and The correlation coefficient (such as Pearson correlation coefficient or mutual information) of the two is taken as the highest correlation. (Preferred ) features, and get the incremental feature set ; Step 4: Merge the incremental features of all weak categories: ;right Each feature in is normalized to zero-mean and unit variance to obtain , forming an enhanced feature set: ; Step 5: Output enhanced feature set Return to the "Feature Screening and Combination Module" or directly transfer to the "Fault Category Assessment Module" for retraining and evaluation.

6. The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model according to claim 1 is characterized in that: The performance monitoring and model protection module includes: Using enhanced feature sets in the Fault Class Assessment Module After training and evaluating the ensemble model, you are responsible for comparing the performance of the old and new models and deciding whether to roll back or continue iterating based on the results. This involves the following steps: Step 1: Input (1) Initial optimal state: feature set: ; Corresponding integrated model: ; Macro average F1: ; (2) Current enhancement status: Feature set: ;based on Trained ensemble model: ; Step 2: Performance Evaluation (1) Calculate the macro-average F1 of the two models on the same validation set: ， ； (2) Computational performance gain: ; Step 3: Rollback and update logic (1) If (The new model did not bring any improvement) 1) Rollback operation: , , ; 2) Set the end flag, terminate feature addition and model iteration, and enter the output module; (2) Otherwise (New model performance improvement) 1) Update the best status: , , ; 2) Threshold judgment: If , then set the end flag and enter the output module; otherwise, if the number of additions or feature dimensions have not reached the upper limit, return to the "Feature Enhancement Generation Module" to continue the next round of additions; otherwise, the maximum number of iterations or maximum dimension has been reached, set the end flag and enter the output module; Step 4: Iteration Cap and Protection: Maximum Number of Iterations Used to avoid infinite loops; maximum feature dimension :To prevent overfitting and computational bottlenecks caused by too many features, when any upper limit is triggered and the performance is still not up to standard, the current 、 Output the optimal result.

7. The rotating machinery fault diagnosis method based on adaptive feature optimization and integrated model according to claim 1 is characterized in that: The fault diagnosis module includes the following steps: Step 1: Diagnostic classification model training: For diagnostic classification models such as support vector machine, random forest, gradient boosting tree, shallow neural network, etc., train them on the validation set to obtain classification probability and F1 score ; Step 2: Weight Normalization: Calculate weights ; Step 3: Soft Voting Fusion: Fusion Probability , and Determine the fault category and output the diagnostic accuracy.

Images