Multi-working-condition-oriented equipment fault diagnosis method based on unbalanced learning

Abstract

The invention discloses a multi-working-condition-oriented equipment fault diagnosis method based on unbalanced learning, and the method comprises the following steps: collecting the operation data of equipment under different working conditions through a plurality of sensors, carrying out the preprocessing, feature extraction and feature selection of the collected original operation data, and obtaining a feature set; processing the feature set by adopting an unbalanced learning technology and a cost-sensitive learning algorithm to obtain a balanced data set; constructing a multi-working-condition fault diagnosis model based on learning, and optimizing model parameters by adopting a method of combining cross validation and grid search according to a balanced data set in a model training process; deploying a multi-working-condition fault diagnosis model obtained through training, carrying out the online diagnosis of equipment faults, and continuously updating and optimizing the model through a continuous learning mechanism. According to the method, the unbalanced learning technology is applied to multi-working-condition fault diagnosis, the accuracy and robustness of diagnosis can be remarkably improved, and real-time monitoring and predictive maintenance of the equipment health state are achieved.

Description

Technical Field

[0001] The present invention relates to the field of combining mechanical equipment fault diagnosis with artificial intelligence, and in particular to a multi-operating condition-oriented equipment fault diagnosis method based on non-equilibrium learning. Background Art

[0002] With the rapid development of intelligent manufacturing, equipment fault diagnosis technology is gradually evolving towards intelligence and automation. Traditional fault diagnosis methods often rely on expert experience and data analysis under a single working condition, and have problems such as low diagnostic efficiency, poor generalization ability, and difficulty in coping with complex and changing working conditions. The multi-condition equipment fault diagnosis method based on unbalanced learning can quickly and accurately identify equipment faults under various working conditions by processing unbalanced fault sample data and using advanced machine learning algorithms, providing equipment maintenance personnel with more reliable diagnostic results.

[0003] In the modern industrial field, equipment fault diagnosis is a key link to ensure production stability and equipment reliability. Traditional fault diagnosis methods mainly rely on the experience and judgment of maintenance personnel, the analysis of expert systems, and data monitoring based on a single working condition. However, these methods have many limitations, such as low diagnostic efficiency, poor generalization ability, and difficulty in adapting to complex and changing working environments.

[0004] With the rapid development of industrial Internet and artificial intelligence technology, especially the progress of machine learning algorithms, new ideas and methods have been provided for equipment fault diagnosis. As an important branch of machine learning, imbalanced learning is specifically used to process data sets with uneven category distribution, and can effectively solve the common problem of imbalanced fault samples in actual industrial environments. In recent years, imbalanced learning has been increasingly used in various fields, such as fault detection, anomaly recognition, predictive maintenance, etc.

[0005] However, the existing equipment fault diagnosis system still has shortcomings when facing multiple working conditions. On the one hand, most systems lack the ability to effectively extract and analyze fault characteristics under different working conditions, making it difficult to accurately identify fault types in complex and changing working environments; on the other hand, the system's ability to diagnose rare faults is weak, making it easy to miss or misdiagnose. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a multi-condition equipment fault diagnosis method based on unbalanced learning in view of the defects in the prior art. The method combines advanced data preprocessing technology, feature engineering methods and integrated learning algorithms, and can effectively process unbalanced fault data under multiple conditions, improve the generalization ability of the diagnosis model and the recognition rate of rare faults. By applying unbalanced learning technology to multi-condition fault diagnosis, the accuracy and robustness of diagnosis can be significantly improved, and real-time monitoring and predictive maintenance of equipment health status can be achieved.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] The present invention provides a multi-operating condition-oriented equipment fault diagnosis method based on non-equilibrium learning, comprising the following steps:

[0009] Multiple sensors are set up to collect the operation data of the equipment under different working conditions, and the collected raw operation data is preprocessed, feature extracted and selected to obtain a feature set;

[0010] Use unbalanced learning technology and cost-sensitive learning algorithm to process feature sets and obtain balanced data sets;

[0011] Construct a multi-condition fault diagnosis model based on learning. During the model training process, a combination of cross-validation and grid search is used to optimize model parameters according to the balanced data set.

[0012] Deploy the trained multi-condition fault diagnosis model to perform online diagnosis of equipment failures, and continuously update the optimization model through a continuous learning mechanism.

[0013] Furthermore, in the method, the sensors include vibration sensors, temperature sensors, pressure sensors, and current sensors; the operating data under different working conditions include different loads, speeds, temperatures, vibration signals, pressures, and currents.

[0014] Furthermore, the method of preprocessing the collected original operation data in the method includes:

[0015] Clean the collected raw operating data to remove outliers, missing values, and duplicate data; use statistical methods or machine learning algorithms to detect and correct outliers; for missing values, select interpolation methods or delete them directly according to data characteristics; after data cleaning, perform standardization to unify data of different dimensions to the same scale, including minimum-maximum normalization and Z-score normalization;

[0016] Time series synchronization is performed on the processed operating data through interpolation, resampling or time window segmentation.

[0017] The running data is split and the processed data set is divided into training set, validation set and test set.

[0018] Furthermore, the feature extraction methods in the method include time domain, frequency domain and time-frequency domain analysis methods to extract features simultaneously; time domain analysis focuses on the statistical characteristics of the signal, including mean, standard deviation, peak, skewness, kurtosis, period, and waveform factor; frequency domain analysis converts the time domain signal to the frequency domain, uses fast Fourier transform and power spectral density analysis to extract the main frequency, harmonic amplitude, and band energy characteristics from the spectrum; time-frequency analysis methods include short-time Fourier transform, wavelet transform and Hilbert-Huang transform to extract the instantaneous frequency and energy distribution characteristics of the signal;

[0019] Feature selection methods include filtering, wrapping and embedding. The filtering method selects based on the statistical properties of the features; the wrapping method combines feature selection with the learning algorithm to select the optimal feature combination by repeatedly training and evaluating the performance of different feature subsets; the embedding method integrates feature selection into the model training process, using L1 regularized Lasso regression or feature importance evaluation in decision trees;

[0020] In the multi-operating condition fault diagnosis scenario, feature selection takes into account the influence of operating conditions and uses conditional mutual information or multi-task learning techniques to select features that have good discrimination capabilities under different operating conditions.

[0021] Furthermore, the method of processing the unbalanced learning technology and the cost-sensitive learning method in the method includes:

[0022] Perform sample distribution analysis on the processed feature set, including counting the number of fault samples of each category, calculating the imbalance ratio between categories, and analyzing the changes in sample distribution under different working conditions;

[0023] Based on the results of sample distribution analysis, unbalanced data is processed, including oversampling, undersampling, cost-sensitive learning methods, and data enhancement technology; oversampling, through synthetic minority class oversampling technology or adaptive synthetic sampling, interpolation between minority class samples to generate new samples, increasing the representation of minority classes while avoiding overfitting; undersampling, using EasyEnsemble or BalanceCascade methods, through multiple undersampling and combined with ensemble learning; cost-sensitive learning methods, by assigning different costs to different categories of misclassification, guide the model to pay more attention to minority class samples; data enhancement technology, generate high-quality synthetic samples through adversarial generative networks.

[0024] Furthermore, the method for constructing a multi-operating condition fault diagnosis model in the method includes:

[0025] The ensemble learning method is used as the basic architecture of the model, and random forest and gradient boosting decision tree are selected as ensemble learning algorithms. In a multi-operating environment, the random feature selection mechanism of random forest is used to capture the key features under different operating conditions. The gradient boosting decision tree corrects the errors of the previous learner by iteratively training the weak learner, and focuses on the difficult-to-classify samples by adjusting the sample weights when processing unbalanced data sets.

[0026] By adopting the working condition perception mechanism, the working condition information is input into the model as an additional feature, or a dedicated sub-model is trained for each major working condition, so that the model can dynamically adjust its diagnostic strategy according to the current working condition.

[0027] Furthermore, the model training method in the method includes:

[0028] During the model training process, a combination of cross-validation and grid search is used to optimize model parameters. Cross-validation ensures the reliability of model performance evaluation, and grid search is used to find the optimal hyperparameter combination. For multi-operating scenarios, stratified cross-validation is used.

[0029] Focal loss is introduced in the training process. Focal loss dynamically adjusts sample weights to make the model pay more attention to difficult-to-classify samples. Sample importance weighting technology is also used to assign different weights to different samples according to the severity and frequency of the fault.

[0030] Furthermore, the multi-condition fault diagnosis model in this method adopts a multi-level classification strategy; the first-level classifier is used to distinguish between normal and fault states, and the second-level classifier further identifies the specific fault type; and feature importance analysis and local interpretability technology are introduced to obtain the key factors for fault diagnosis by analyzing the degree of dependence of the model on different features; the SHAP value is used to explain the single prediction result of the model, providing maintenance personnel with an intuitive diagnostic basis.

[0031] Furthermore, the method of online diagnosis and model optimization model in the method includes:

[0032] During the online diagnosis process, multi-source sensor data from the equipment is received and processed in real time, and then input into the trained multi-condition fault diagnosis model after preprocessing and feature extraction. The model outputs the current health status assessment of the equipment and potential fault warning.

[0033] The system of the multi-condition fault diagnosis model also integrates a confidence assessment mechanism. When the model's prediction confidence for certain samples is lower than the threshold, manual intervention or more in-depth analysis is triggered;

[0034] In the model optimization model, an incremental learning strategy is adopted to regularly use newly collected data to update the model, including data screening, labeling and model fine-tuning. During the update process, rare fault samples are continuously monitored to prevent the model performance from improving on common faults and degrading on rare faults.

[0035] In the model optimization model, a model drift detection mechanism is adopted to monitor the performance changes of the model on new data, timely discover the deviation of the model from the actual situation, and trigger necessary retraining or adjustment.

[0036] The present invention provides a multi-operating condition-oriented equipment fault diagnosis system based on non-equilibrium learning, comprising:

[0037] a memory for storing executable computer programs;

[0038] The processor is used to implement the above-mentioned equipment fault diagnosis method for multiple working conditions based on non-equilibrium learning when executing the executable computer program stored in the memory.

[0039] The beneficial effects produced by the present invention are:

[0040] (1) Improve the accuracy and efficiency of fault diagnosis

[0041] The present invention realizes intelligent identification and prediction of equipment faults through a multi-operating condition fault diagnosis method based on unbalanced learning. Multi-operating condition data processing and feature engineering can fully capture the characteristics of the equipment under different working conditions. Unbalanced learning technology effectively solves the problem of sample imbalance and significantly improves the accuracy of diagnosis, especially the ability to identify rare faults. At the same time, the automated diagnosis process greatly reduces manual participation and improves diagnostic efficiency.

[0042] (2) Enhance the generalization ability of the diagnostic model

[0043] The present invention can more comprehensively simulate actual industrial scenarios by considering multiple operating conditions, including different operating environments, load conditions, and working modes. The diagnostic model based on non-equilibrium learning and ensemble learning has stronger generalization ability and can adapt to various complex working environments and equipment states. This helps maintenance personnel to more accurately assess the health status of equipment and formulate reasonable maintenance plans.

[0044] (3) Achieving adaptive optimization of diagnostic models

[0045] Through online diagnosis and continuous learning mechanisms, the diagnostic model is automatically optimized. The system can automatically adjust and optimize the model based on newly collected data and diagnostic results, reducing deviations and errors caused by human factors. This adaptive optimization process can be carried out continuously, ensuring that the diagnostic model is always consistent with the actual status of the equipment, improving the long-term reliability of the system.

[0046] (4) Improve the ability to adapt to complex working conditions

[0047] Through multi-condition data processing and feature engineering, as well as a diagnostic model based on non-equilibrium learning, the system's adaptability to complex working conditions has been improved. The system can process equipment operating data under different working conditions in real time, including various sensor signals and environmental parameters, to provide maintenance personnel with an accurate description of the equipment status. At the same time, the diagnostic model can adjust the diagnostic strategy in real time according to different working conditions, providing maintenance personnel with reliable fault diagnosis results and predictive maintenance suggestions.

[0048] (5) Promote the implementation of predictive maintenance

[0049] The method of the present invention provides strong technical support for predictive maintenance. By accurately identifying potential equipment failures and predicting failure development trends, maintenance personnel can take preventive measures before failures cause serious losses. This can not only reduce unplanned equipment downtime, but also optimize maintenance plans, reduce maintenance costs, and improve the overall operating efficiency and life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0051] Figure 1 It is a multi-operating-condition equipment fault diagnosis method based on non-equilibrium learning in an embodiment of the present invention;

[0052] Figure 2 It is a specific process of an embodiment of the present invention;

[0053] Figure 3 This is the unbalanced learning data processing flow of an embodiment of the present invention;

[0054] Figure 4 This is the construction of a multi-operating condition fault diagnosis model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] Example 1

[0057] The multi-operating condition equipment fault diagnosis method based on non-equilibrium learning in the embodiment of the present invention makes full use of the advanced nature of non-equilibrium learning technology and the richness of multi-operating condition data, and aims to realize the automatic and intelligent diagnosis and prediction of equipment faults in complex industrial environments.

[0058] Specifically, the main contents of the present invention include the following aspects:

[0059] S01: Multi-condition data processing and feature engineering

[0060] Set up multiple sensors to collect the operating data of the equipment under different working conditions, including vibration signals, temperature, pressure, current and other characteristics. Preprocess the collected raw data, including denoising, normalization, time synchronization and other operations. For different working conditions and fault types, use time domain, frequency domain and time-frequency domain analysis methods to extract features simultaneously, and use feature selection algorithms to select the most representative feature set.

[0061] S02: Data processing based on imbalanced learning

[0062] To address the problem of imbalanced fault samples, techniques such as oversampling, undersampling, or mixed sampling are used to process training data. At the same time, a classification algorithm based on cost-sensitive learning is designed to improve the ability to identify minority class faults. Through this step, the common problem of imbalanced fault samples in actual industrial environments is effectively solved.

[0063] S03: Construction of multi-condition fault diagnosis model

[0064] Based on the processed balanced data set, a fault diagnosis model that adapts to multiple working conditions is constructed. Ensemble learning methods such as random forest or gradient boosting decision tree are used to improve the generalization ability of the model. Model parameters are optimized through methods such as cross-validation and grid search to obtain the best diagnostic performance. The model can effectively handle equipment fault diagnosis problems under different working conditions.

[0065] The embodiment of the present invention uses advanced non-equilibrium learning technology to build an intelligent diagnosis model that adapts to various working conditions, thereby realizing the automatic identification and diagnosis of equipment faults in various complex industrial environments. This technology not only improves the accuracy of fault diagnosis, but also enhances the robustness and adaptability of the diagnosis model, providing strong support for predictive maintenance and health management of industrial equipment.

[0066] S04: Online diagnosis and model optimization

[0067] Deploy the trained model to the actual industrial environment for online diagnosis. Design a continuous learning mechanism to continuously update and optimize the model based on newly collected data and diagnostic results to adapt to changes in equipment status and working environment. In this way, adaptive optimization of the diagnostic model is achieved to improve the long-term reliability of the system.

[0068] In summary, the present invention realizes the automatic and intelligent diagnosis and prediction of equipment failures in complex industrial environments by constructing a multi-condition fault diagnosis system based on non-equilibrium learning and combining advanced feature engineering and machine learning technologies. Through this step, the accuracy and efficiency of fault diagnosis are improved, and a more scientific, reliable and efficient solution is provided for the predictive maintenance and health management of industrial equipment.

[0069] Example 2

[0070] The flowchart of the multi-operating-condition equipment fault diagnosis method based on non-equilibrium learning according to an embodiment of the present invention is as follows: Figure 2 As shown, the specific implementation of the present invention includes the following steps:

[0071] Step 1: Multi-condition data collection and preprocessing

[0072] Multi-operating condition data collection and preprocessing are the key foundations for realizing multi-operating condition equipment fault diagnosis methods based on non-equilibrium learning. This step aims to obtain comprehensive and accurate equipment operation data, and clean, standardize and synchronize the raw data, laying a solid foundation for subsequent feature extraction and model construction.

[0073] First, in the multi-operating condition data collection stage, it is necessary to reasonably arrange various types of sensors according to the equipment characteristics and working environment. These sensors may include but are not limited to vibration sensors, temperature sensors, pressure sensors, current sensors, etc. The selection and arrangement of sensors should take into account the key components of the equipment, common fault types, and characteristic changes under different operating conditions. The data acquisition system should have high sampling rate, high accuracy and anti-interference ability to ensure that subtle changes in the equipment under various operating conditions are captured. At the same time, the data acquisition process should cover the operating status of the equipment under different load, speed, temperature and other working conditions to fully reflect the dynamic characteristics of the equipment.

[0074] The data preprocessing stage is a key step in converting the raw collected data into a high-quality data set that can be used for subsequent analysis. First, the raw data is cleaned to remove outliers, missing values, and duplicate data. Outliers can be detected and corrected using statistical methods or machine learning algorithms. For missing values, interpolation methods or direct deletion can be selected based on data characteristics. After data cleaning, standardization is performed to unify data of different dimensions to the same scale. Commonly used methods include minimum-maximum normalization and Z-score normalization. Standardization helps to eliminate dimensional differences between different features and improve the training effect of subsequent models.

[0075] Time series synchronization is an important part of multi-condition data processing. Since different sensors may have different sampling frequencies or time delays, all data need to be aligned to the same timeline. This is achieved through methods such as interpolation, resampling, or time window segmentation. Synchronous processing ensures the temporal consistency between different sensor data and provides a reliable data foundation for subsequent feature extraction and model training.

[0076] In addition, in the data preprocessing stage, the data segmentation strategy also needs to be considered. The processed data set is usually divided into training set, validation set and test set. The training set is used for model learning and parameter optimization, the validation set is used to adjust model hyperparameters and prevent overfitting, and the test set is used to evaluate the generalization ability of the model. In multi-operating scenarios, data segmentation should ensure that each subset contains samples under various operating conditions to ensure that the model can learn the characteristics of different operating conditions.

[0077] Through the above multi-condition data collection and preprocessing steps, a high-quality, multi-dimensional data set was obtained, laying a solid foundation for subsequent feature extraction and unbalanced learning. This data set not only contains the operating status information of the equipment under various working conditions, but also has undergone strict cleaning, standardization and synchronization processing, which can truly reflect the health status and potential fault characteristics of the equipment.

[0078] Step 2: Feature extraction and selection

[0079] Feature extraction and selection are the core links in the multi-condition equipment fault diagnosis method based on imbalanced learning, which directly affects the performance and accuracy of the diagnosis model. This step aims to extract information that can effectively characterize the equipment status and fault characteristics from the preprocessed multi-condition data, and select the most representative and discriminative feature subset.

[0080] In the feature extraction stage, a combination of time domain, frequency domain and time-frequency domain analysis is used to fully capture the operating characteristics of the equipment under different working conditions. Time domain analysis focuses on the statistical characteristics of the signal, including mean, standard deviation, peak, skewness and kurtosis. These features can intuitively reflect the overall distribution and fluctuation of the signal. For signals with obvious periodicity, features such as period and waveform factor can also be extracted. Time domain features are simple to calculate and easy to process in real time, but may not fully reflect the frequency components of the signal.

[0081] Frequency domain analysis reveals the frequency composition of a signal by converting the time domain signal into the frequency domain. Common methods include fast Fourier transform (FFT) and power spectral density (PSD) analysis. Features such as main frequency, harmonic amplitude, and frequency band energy can be extracted from the spectrum. Frequency domain analysis is particularly suitable for detecting periodic faults in equipment, including bearing faults, gear faults, etc. In a multi-operating environment, different operating conditions may cause changes in the spectrum characteristics, so it is necessary to consider the impact of operating conditions on frequency domain characteristics.

[0082] Time-frequency domain analysis combines the advantages of time domain and frequency domain, and can simultaneously reflect the changes of signals in time and frequency dimensions. Commonly used time-frequency analysis methods include short-time Fourier transform (STFT), wavelet transform, and Hilbert-Huang transform (HHT). These methods can extract the instantaneous frequency, energy distribution and other characteristics of the signal, and are particularly suitable for analyzing non-stationary signals and capturing transient fault characteristics. In a multi-operating environment, time-frequency domain analysis can effectively identify the dynamic changes of fault characteristics under different operating conditions.

[0083] After feature extraction, we are faced with a high-dimensional feature space. However, not all extracted features contribute significantly to fault diagnosis, and some features may be redundant or noisy. Therefore, feature selection becomes a key step to improve the performance of the diagnostic model. The goal of feature selection is to reduce feature dimensions while retaining key information, thereby improving the generalization ability and computational efficiency of the model.

[0084] Feature selection can be done by filtering, wrapping, and embedding. The filtering method selects based on the statistical properties of the features, such as variance analysis and mutual information. This method is computationally efficient, but may ignore the interactions between features. The wrapping method combines feature selection with the learning algorithm to select the optimal feature combination by repeatedly training and evaluating the performance of different feature subsets. Although the computational cost is high, it is possible to find the feature subset that best matches the specific learning algorithm. The embedding method integrates feature selection into the model training process, using L1 regularized Lasso regression or feature importance evaluation in decision trees.

[0085] In the multi-operating condition fault diagnosis scenario, feature selection also needs to consider the influence of operating conditions. Technologies such as conditional mutual information or multi-task learning can be used to select features that have good discrimination capabilities under different operating conditions. At the same time, in order to cope with the challenges of unbalanced learning, the feature selection process should focus on the feature performance of minority class samples to ensure that the selected features also have good recognition capabilities for rare faults.

[0086] The ultimate goal of feature extraction and selection is to construct a feature set that can fully characterize the equipment status and has good generalization ability. This feature set should be able to adapt to multiple working conditions, have good recognition capabilities for different types of faults (including rare faults), and have moderate dimensions to ensure the training efficiency of subsequent models. By comprehensively using time domain, frequency domain and time-frequency domain analysis methods, combined with advanced feature selection algorithms, it is possible to provide high-quality input features for the multi-working condition fault diagnosis model of non-equilibrium learning, thereby improving the accuracy and reliability of diagnosis.

[0087] Step 3: Imbalanced Learning Data Processing

[0088] Unbalanced learning data processing is a key step to solve the sample imbalance problem in multi-condition equipment fault diagnosis. In actual industrial environments, normal operation data is often much more than fault data, especially the number of samples of some rare faults may be very small. This imbalance in data distribution will cause traditional machine learning algorithms to tend to ignore minority classes, thereby affecting the diagnostic model's ability to identify rare faults. Therefore, unbalanced learning data processing aims to adjust the data distribution and improve the model's sensitivity to minority class samples.

[0089] First, a detailed sample distribution analysis of the data set is required. This includes counting the number of samples of each type of fault, calculating the imbalance ratio between categories, and analyzing the changes in sample distribution under different working conditions. Through this analysis, it is possible to clarify which fault types belong to the minority class and the severity of the imbalance. At the same time, the impact of multiple working condition factors on sample distribution also needs to be considered, because some faults may be more likely to occur under specific working conditions.

[0090] Based on the results of sample distribution analysis, a variety of techniques are used to handle unbalanced data. Oversampling is a common method that balances data distribution by increasing the number of minority class samples. Simple oversampling methods such as random oversampling may lead to overfitting, so more advanced techniques such as synthetic minority oversampling technology (SMOTE) or adaptive synthetic sampling (ADASYN) are used. These methods generate new samples by interpolating between minority class samples, which increases the representation of the minority class while avoiding overfitting. In a multi-operating environment, the operating condition information needs to be considered when oversampling to ensure that the generated samples meet the characteristic distribution of each operating condition.

[0091] Undersampling is another method to balance data by reducing the number of majority class samples. Simple random undersampling may lead to information loss, so smarter undersampling strategies such as EasyEnsemble or BalanceCascade are used. These methods balance the data distribution and retain important information of the majority class through multiple undersampling and ensemble learning. In multi-operating scenarios, undersampling needs to ensure that samples under each operating condition are fully retained.

[0092] In addition to adjusting the number of samples, a cost-sensitive learning method is also used to deal with the imbalance problem. This method guides the model to pay more attention to minority class samples by assigning different costs to different categories of misclassification. In multi-condition fault diagnosis, the cost matrix can be dynamically adjusted according to the severity and frequency of the fault to ensure that the model remains highly sensitive to critical faults.

[0093] In addition, some special data processing techniques are introduced to enhance the effect of unbalanced learning. Data augmentation techniques are used to generate new minority class samples while maintaining the original features, and high-quality synthetic samples are generated through adversarial generative networks (GANs). These methods not only increase the number of minority class samples, but also improve the diversity of samples, which helps the model learn more robust feature representations.

[0094] When implementing unbalanced learning data processing, cross-validation is used to evaluate the effects of different processing methods. By comparing the recognition performance of the model before and after processing on various types of faults, especially the recall rate and precision rate of minority faults, the unbalanced learning strategy that best suits the current multi-condition fault diagnosis task can be selected.

[0095] Through the above unbalanced learning data processing steps, the negative impact of sample imbalance on the fault diagnosis model can be effectively alleviated, and the model's ability to identify various types of faults (especially rare faults) can be improved, thereby laying a solid data foundation for the subsequent construction of a multi-condition fault diagnosis model.

[0096] Step 4: Construction of multi-condition fault diagnosis model

[0097] The construction of a multi-condition fault diagnosis model is the core of the method of the present invention, which aims to use data processed by unbalanced learning to build an intelligent diagnosis system that can adapt to various working conditions and accurately identify various types of faults. The model needs to have strong generalization capabilities to cope with complex and changing industrial environments while maintaining high sensitivity to rare faults.

[0098] First, an ensemble learning method is used as the basic framework of the model. Ensemble learning can significantly improve the performance and robustness of the model by combining the prediction results of multiple base learners. In the present invention, random forest (RandomForest) and gradient boosting decision tree (XGBoost) are selected as ensemble learning algorithms. This algorithm has the ability to process high-dimensional features, has a strong ability to capture nonlinear relationships, and is relatively not prone to overfitting.

[0099] Random forests can effectively reduce the variance of a single tree and improve the generalization ability of the model by constructing multiple decision trees and integrating the results by voting or averaging. In a multi-operating environment, the random feature selection mechanism of random forests helps capture key features under different operating conditions. Gradient boosting decision trees correct the errors of previous learners by iteratively training weak learners. They are particularly suitable for processing unbalanced data sets because they can focus on difficult-to-classify samples by adjusting sample weights.

[0100] In order to further improve the model's adaptability to multiple working conditions, a working condition perception mechanism is introduced. Specifically, working condition information is input into the model as an additional feature, or a dedicated sub-model is trained for each major working condition. This approach enables the model to dynamically adjust its diagnostic strategy according to the current working condition, improving the accuracy and reliability of the diagnosis.

[0101] During the model training process, a combination of cross-validation and grid search is used to optimize model parameters. Cross-validation ensures the reliability of model performance evaluation, while grid search helps find the optimal hyperparameter combination. For multi-operating scenarios, stratified cross-validation is used to ensure that each compromise contains samples under various operating conditions, thereby evaluating the performance of the model under different operating conditions.

[0102] In order to further improve the model's ability to identify rare faults, focal loss is introduced during the training process. Focal loss dynamically adjusts sample weights to make the model pay more attention to samples that are difficult to classify, which is particularly effective for identifying rare faults in a dataset. At the same time, sample importance weighting technology is also used to assign different weights to different samples according to the severity and frequency of the fault, ensuring that the model can accurately identify critical faults.

[0103] In the model structure design, a multi-level classification strategy is adopted. The first-level classifier is used to distinguish between normal state and fault state, and the second-level classifier further identifies the specific fault type. This hierarchical structure helps to improve the overall accuracy of the model, especially when dealing with multi-class imbalance problems.

[0104] In order to enhance the interpretability of the model, feature importance analysis and local interpretability techniques are introduced. By analyzing the model's dependence on different features, we can understand which factors play a key role in fault diagnosis, which is of great significance for understanding the fault mechanism and guiding preventive maintenance. At the same time, SHAP (SHapley AdditiveexPlanations) values ​​are used to explain the model's single prediction results, providing maintenance personnel with an intuitive basis for diagnosis.

[0105] In the model evaluation phase, we not only focus on the global accuracy, but also pay special attention to the recall and precision of various faults, especially the ability to identify rare faults. This step uses multiple evaluation indicators such as F1 score, confusion matrix, ROC curve, etc. to comprehensively evaluate the performance of the model.

[0106] Step 5: Online diagnosis and model update

[0107] Online diagnosis and model updating are key links to ensure the long-term effective operation of the multi-condition equipment fault diagnosis system based on non-equilibrium learning. This step realizes the transformation of the diagnosis model from static to dynamic, enabling the system to adapt to the continuous changes in equipment status and working environment.

[0108] During the online diagnosis process, the system receives and processes multi-source sensor data from the device in real time. The data first undergoes preprocessing and feature extraction steps consistent with the training phase, and then is input into the trained diagnostic model. The model outputs the current health status assessment of the device and potential fault warnings. In order to improve the reliability of diagnosis, the system also integrates a confidence assessment mechanism. When the model's prediction confidence for certain samples is lower than the threshold, it will trigger manual intervention or more in-depth analysis.

[0109] Model updating is the key to maintaining the performance of the diagnostic system. Adopt an incremental learning strategy and regularly update the model with newly collected data. This process includes data screening, labeling, and model fine-tuning. In particular, it is necessary to pay attention to rare fault samples during the update process to prevent the model performance from improving on common faults and degrading on rare faults.

[0110] In addition, the system also includes a model drift detection mechanism. By monitoring the performance changes of the model on new data, the deviation of the model from the actual situation can be discovered in time, and necessary retraining or adjustment can be triggered. This adaptive mechanism ensures that the diagnostic system can continuously adapt to the dynamic changes of the equipment status and working environment, and maintain long-term diagnostic accuracy and reliability.

[0111] Through the above specific implementation methods, the equipment fault diagnosis method for multiple working conditions based on unbalanced learning proposed in the present invention successfully solves the challenges faced by traditional diagnostic technologies in complex industrial environments. The method comprehensively uses innovative technologies such as multi-condition data processing, advanced feature engineering, unbalanced learning technology, and intelligent diagnostic model construction to achieve high-precision identification and prediction of various types of faults (especially rare faults). The method of the present invention is not only suitable for equipment fault diagnosis in traditional manufacturing industries, but can also be extended to equipment health management in high-tech fields such as smart factories, energy industries, and aerospace. By improving the accuracy and reliability of fault diagnosis, the method helps to reduce unplanned downtime of equipment, optimize maintenance strategies, and improve overall production efficiency. At the same time, its adaptive learning and online update mechanisms ensure that the system maintains high performance in long-term operation, provides strong technical support for predictive maintenance and intelligent management of industrial equipment, has broad application prospects and significant economic benefits, and is of great significance to promoting industrial intelligence and digital transformation.

[0112] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0113] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A multi-condition equipment fault diagnosis method based on non-equilibrium learning, characterized in that: The following steps are involved: Multiple sensors are set up to collect the operation data of the equipment under different working conditions, and the collected raw operation data are preprocessed, feature extracted and selected to obtain a feature set; Use unbalanced learning technology and cost-sensitive learning algorithm to process feature sets and obtain balanced data sets; Construct a multi-condition fault diagnosis model based on learning. During the model training process, a combination of cross-validation and grid search is used to optimize model parameters according to the balanced data set. Deploy the trained multi-condition fault diagnosis model to perform online diagnosis of equipment failures, and continuously update the optimization model through a continuous learning mechanism.

2. The equipment fault diagnosis method for multiple working conditions based on non-equilibrium learning according to claim 1 is characterized in that: In the method, the sensors include vibration sensors, temperature sensors, pressure sensors, and current sensors; the operating data under different working conditions include different loads, speeds, temperatures, vibration signals, pressures, and currents.

3. The equipment fault diagnosis method for multiple working conditions based on non-equilibrium learning according to claim 1 is characterized in that: The method for preprocessing the collected original operation data in the method includes: Clean the collected raw operating data to remove outliers, missing values, and duplicate data; use statistical methods or machine learning algorithms to detect and correct outliers; for missing values, select interpolation methods or delete them directly according to data characteristics; after data cleaning, perform standardization to unify data of different dimensions to the same scale, including minimum-maximum normalization and Z-score normalization; Time series synchronization is performed on the processed operating data through interpolation, resampling or time window segmentation. The running data is split and the processed data set is divided into training set, validation set and test set.

4. The method for equipment fault diagnosis based on non-equilibrium learning and facing multiple working conditions according to claim 1, characterized in that: The feature extraction methods in this method include time domain, frequency domain and time-frequency domain analysis methods to extract features simultaneously; time domain analysis focuses on the statistical characteristics of the signal, including mean, standard deviation, peak, skewness, kurtosis, period, and waveform factor; frequency domain analysis converts the time domain signal to the frequency domain, uses fast Fourier transform and power spectral density analysis to extract the main frequency, harmonic amplitude, and band energy characteristics from the spectrum; time-frequency analysis methods include short-time Fourier transform, wavelet transform and Hilbert-Huang transform to extract the instantaneous frequency and energy distribution characteristics of the signal; Feature selection methods include filtering, packaging and embedding. The filtering method selects based on the statistical properties of features; the packaging method combines feature selection with learning algorithms to select the optimal feature combination by repeatedly training and evaluating the performance of different feature subsets. Embed The method integrates feature selection into the model training process, using L1 regularized Lasso regression or feature importance evaluation in decision trees; In the multi-operating condition fault diagnosis scenario, feature selection takes into account the influence of operating conditions and uses conditional mutual information or multi-task learning techniques to select features that have good discrimination capabilities under different operating conditions.

5. The method for equipment fault diagnosis based on non-equilibrium learning and facing multiple working conditions according to claim 1, characterized in that: The method of using the unbalanced learning technique and the cost-sensitive learning method in this method includes: Perform sample distribution analysis on the processed feature set, including counting the number of fault samples of each category, calculating the imbalance ratio between categories, and analyzing the changes in sample distribution under different working conditions; Based on the results of sample distribution analysis, unbalanced data is processed, including oversampling, undersampling, cost-sensitive learning methods, and data enhancement technology; oversampling, through synthetic minority class oversampling technology or adaptive synthetic sampling, interpolation between minority class samples to generate new samples, increasing the representation of minority classes while avoiding overfitting; undersampling, using EasyEnsemble or BalanceCascade methods, through multiple undersampling and combined with ensemble learning; cost-sensitive learning methods, by assigning different costs to different categories of misclassification, guide the model to pay more attention to minority class samples; data enhancement technology, generate high-quality synthetic samples through adversarial generative networks.

6. The method for equipment fault diagnosis based on non-equilibrium learning and facing multiple working conditions according to claim 1, characterized in that: The method for constructing a multi-condition fault diagnosis model in this method includes: The ensemble learning method is used as the basic architecture of the model, and random forest and gradient boosting decision tree are selected as ensemble learning algorithms. In a multi-operating environment, the random feature selection mechanism of random forest is used to capture the key features under different operating conditions. The gradient boosting decision tree corrects the errors of the previous learner by iteratively training the weak learner, and focuses on the difficult-to-classify samples by adjusting the sample weights when processing unbalanced data sets. By adopting the working condition perception mechanism, the working condition information is input into the model as an additional feature, or a dedicated sub-model is trained for each major working condition, so that the model can dynamically adjust its diagnostic strategy according to the current working condition.

7. The method for equipment fault diagnosis based on non-equilibrium learning and facing multiple working conditions according to claim 6 is characterized in that: The model training method in this method includes: During the model training process, a combination of cross-validation and grid search is used to optimize model parameters. Cross-validation ensures the reliability of model performance evaluation, and grid search is used to find the optimal hyperparameter combination. For multi-operating scenarios, stratified cross-validation is used. Focal loss is introduced in the training process. Focal loss dynamically adjusts sample weights to make the model pay more attention to difficult-to-classify samples. Sample importance weighting technology is also used to assign different weights to different samples according to the severity and frequency of the fault.

8. The method for equipment fault diagnosis based on non-equilibrium learning and facing multiple working conditions according to claim 6, characterized in that: The multi-condition fault diagnosis model in this method adopts a multi-level classification strategy; the first-level classifier is used to distinguish between normal and fault states, and the second-level classifier is used to further identify the specific fault type; feature importance analysis and local interpretability technology are introduced to obtain the key factors for fault diagnosis by analyzing the degree of model dependence on different features; SHAP values ​​are used to explain the single prediction results of the model, providing maintenance personnel with an intuitive diagnostic basis.

9. The method for equipment fault diagnosis based on non-equilibrium learning and facing multiple working conditions according to claim 1, characterized in that: The online diagnosis and model optimization methods in the method include: During the online diagnosis process, multi-source sensor data from the equipment is received and processed in real time, and then input into the trained multi-condition fault diagnosis model after preprocessing and feature extraction. The model outputs the current health status assessment of the equipment and potential fault warning. The system of the multi-condition fault diagnosis model also integrates a confidence assessment mechanism. When the model's prediction confidence for certain samples is lower than the threshold, manual intervention or more in-depth analysis is triggered; In the model optimization model, an incremental learning strategy is adopted to regularly use newly collected data to update the model, including data screening, labeling and model fine-tuning. During the update process, rare fault samples are continuously monitored to prevent the model performance from improving on common faults and degrading on rare faults. In the model optimization model, a model drift detection mechanism is adopted to monitor the performance changes of the model on new data, timely discover the deviation of the model from the actual situation, and trigger necessary retraining or adjustment.

10. A multi-condition equipment fault diagnosis system based on non-equilibrium learning, characterized in that: include: a memory for storing executable computer programs; A processor is used to implement the equipment fault diagnosis method for multiple working conditions based on non-equilibrium learning as described in any one of claims 1 to 9 when executing an executable computer program stored in a memory.

Images