Variable working condition rolling bearing fault diagnosis method and system

Abstract

The invention discloses a variable working condition rolling bearing fault diagnosis method and system, and the method comprises the steps: firstly obtaining bearing vibration signal data, temperature data and current spectrum data of a motor under different working conditions, carrying out the adaptive decomposition through employing an EEMD algorithm, and extracting a stable feature subset; then, constructing a cross-working-condition diagnosis model by utilizing transfer learning, and introducing an HMM (Hidden Markov Model) time sequence model to capture signal dynamic characteristics; during diagnosis, the working condition of the signal is firstly judged, and the corresponding model is selected for fault diagnosis. And if the diagnosis confidence is low, triggering an online updating mechanism, and performing incremental learning by using new data. Through continuous data collection and model updating, a dynamic evolution system capable of automatically adjusting the diagnosis strategy according to the working condition change is constructed, the problem that the diagnosis performance is reduced due to the working condition change in bearing fault diagnosis is effectively solved, and high-accuracy self-adaptive diagnosis is achieved.

Description

Technical Field

[0001] The invention belongs to the field of bearing fault diagnosis, and in particular relates to a rolling bearing fault diagnosis method and system under variable working conditions. Background Art

[0002] During the actual operation of rolling bearings, their operating conditions often change dynamically, such as fluctuations in speed, load, temperature and other parameters. These changes will directly affect the characteristic performance of the bearing vibration signal and increase the difficulty of fault diagnosis. Traditional machine learning models are usually trained for specific working conditions. Once the working conditions change, the diagnostic performance of the model will drop significantly. How to build a transfer learning diagnostic model that can adapt to different working conditions is a key technical problem that needs to be solved urgently.

[0003] To address this issue, it is necessary to conduct a comprehensive analysis and reasonable settings of the various working conditions that the bearing may encounter before training the transfer learning model, and to collect vibration signal data under different working conditions in a targeted manner. In the data preprocessing stage, it is necessary to study how to effectively extract fault features under different working conditions through signal processing technology (such as the EEMD algorithm), and try to eliminate the impact of changes in working conditions on the features. In addition, it is also necessary to explore methods that combine time series models such as HMM with transfer learning, in order to establish a new bearing fault diagnosis model that can adapt to changes in working conditions and dynamically update. These are all technical difficulties that must be overcome to build a transfer learning fault diagnosis model for variable working conditions. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a rolling bearing fault diagnosis method and system for variable working conditions. Among them, a rolling bearing fault diagnosis method for variable working conditions includes:

[0005] Obtain vibration signal data, temperature data and motor current spectrum data of rolling bearings under different working conditions, use EEMD algorithm to adaptively decompose the vibration signal data, and extract fault characteristics under different working conditions; extract temperature gradient and hot spot position characteristics from the temperature data; extract current harmonic components and frequency offset characteristics from the motor current spectrum data;

[0006] Eliminate the influence of working condition changes on features through feature selection method and obtain feature subsets;

[0007] Based on the feature subset, a transfer learning method is used to build an adaptive diagnosis model, a basic model is trained using source domain data, and fine-tuned on target domain data to obtain an adaptive fault diagnosis model across working conditions;

[0008] Based on the cross-operating condition adaptive fault diagnosis model, comprehensive fault diagnosis is performed on the newly collected vibration signal, temperature signal and electrical frequency signal to obtain a fault diagnosis result.

[0009] Preferably, the process of obtaining vibration signal data of the rolling bearing under different working conditions includes:

[0010] Obtain vibration signal data of rolling bearings under different speed, load and temperature conditions, and collect comprehensive data samples according to the changes of different parameters;

[0011] Constructing a data set covering various working conditions according to the data samples;

[0012] Then, the vibration signal data is preprocessed by denoising and normalization, and the signal processing methods of time domain analysis and frequency domain analysis are used to extract the time domain statistical characteristics and frequency domain characteristics of the vibration signal as the input features of the transfer learning model.

[0013] Preferably, the vibration signal data is adaptively decomposed by using an EEMD algorithm to extract fault features under different working conditions, and the influence of working condition changes on the features is eliminated by a feature selection method. The process of obtaining a feature subset includes:

[0014] According to the vibration signal data under different working conditions, the EEMD algorithm is used to adaptively decompose the vibration signal to obtain several intrinsic mode function components;

[0015] Extracting fault characteristic parameters from the intrinsic mode function components to construct an original high-dimensional feature vector;

[0016] The high-dimensional feature vector is subjected to dimensionality reduction processing by a feature selection algorithm, fault features are screened according to the feature selection algorithm, and a low-dimensional feature subset is constructed to serve as an input for subsequent fault diagnosis.

[0017] Preferably, according to the feature subset, a transfer learning method is used to construct an adaptive diagnosis model, a basic model is trained using source domain data, and fine-tuned on target domain data to obtain an adaptive fault diagnosis model across working conditions, the process includes:

[0018] According to the acquired source domain data and target domain data, the optimal feature subset is extracted through the feature selection algorithm to build the initial diagnosis model;

[0019] Using a transfer learning method, the initial diagnosis model is trained using source domain data to obtain a basic diagnosis model;

[0020] The distribution difference between the target domain data and the source domain data is determined. If the difference is greater than a preset threshold, the basic diagnosis model is fine-tuned to finally obtain an adaptive fault diagnosis model across working conditions.

[0021] Preferably, the process of adopting a transfer learning method and using source domain data to train the initial diagnosis model to obtain a basic diagnosis model includes:

[0022] The HMM time series model is introduced into the transfer learning model, and the feature representation learning of the time domain statistical features and frequency domain features of the extracted vibration signal is performed using the transfer learning model. Then, the feature representation obtained by the transfer learning is input into the HMM time series model, and the time series dynamic characteristics of the vibration signal are captured through the HMM time series model.

[0023] By fusing the transfer learning model and the HMM time series model, a basic diagnostic model of vibration signals is established, and on the basis of the basic diagnostic model, an online learning mechanism is introduced to dynamically update the basic diagnostic model according to the newly collected vibration data.

[0024] Preferably, the process of performing comprehensive fault diagnosis on the newly collected vibration signal, temperature signal and electrical frequency signal based on the cross-operating condition adaptive fault diagnosis model includes:

[0025] The newly collected vibration signal data, temperature signal, and electrical frequency signal are used to determine the working condition, and the corresponding diagnostic model is selected according to the working condition category. The signal data is input into the model for fault diagnosis to obtain the diagnostic result. If the diagnostic result indicates that the equipment is abnormal, the early warning mechanism is triggered to notify relevant personnel to perform maintenance.

[0026] If the confidence level of the diagnosis result is lower than the preset threshold, the online update mechanism of the model is triggered, and the model is incrementally learned using the newly collected vibration signal data. Through continuous data collection and model updating, an adaptive and dynamically evolving bearing fault diagnosis model is constructed.

[0027] Preferably, the newly collected vibration signal data, temperature signal, and electric frequency signal are subjected to working condition judgment, a corresponding diagnostic model is selected according to the working condition category, and the signal data is input into the model for fault diagnosis. The process of obtaining the diagnostic result includes:

[0028] The extracted feature vector is input into the pre-trained working condition classification model, and the working condition category of the current equipment is determined through model reasoning;

[0029] According to the determined working condition category, a diagnostic model corresponding to the working condition is selected from the fault diagnosis model library, and model parameters are loaded;

[0030] Input the previously extracted feature vector into the selected fault diagnosis model, and obtain the fault probability distribution of the equipment through forward calculation of the model;

[0031] Analyze the fault probability distribution to find the fault type with the highest probability, and make a judgment based on the preset fault threshold to determine the final fault diagnosis result;

[0032] The fault diagnosis result is compared and analyzed with the historical fault cases of the equipment, the credibility of the diagnosis result is evaluated, and fault alarm information is generated when necessary to notify relevant personnel to perform equipment maintenance and status inspection.

[0033] Preferably, if the confidence level of the diagnosis result is lower than a preset threshold, the online update mechanism of the model is triggered, and the process of incrementally learning the model using the newly collected vibration signal data includes:

[0034] Preprocess and extract features of the newly collected vibration signals to obtain a data set that can be used for incremental learning;

[0035] According to a preset incremental learning algorithm, the incremental learning data set is used to perform incremental training on the existing model, update model parameters, and obtain a new model with improved performance;

[0036] Using a cross-validation method and based on performance evaluation indicators, the new model is evaluated to determine whether the model performance meets expectations;

[0037] If the model performance does not meet expectations, return to continue incremental learning and model optimization;

[0038] If the model performance has reached expectations, the new model is deployed to the production environment to replace the original model;

[0039] During the model application process, the confidence of the diagnosis results is continuously monitored. When the confidence falls below the threshold again, a new round of online model update and optimization is repeatedly triggered.

[0040] Preferably, the preset incremental learning algorithm includes an online gradient descent method and an incremental support vector machine;

[0041] The performance evaluation indicators include calculation accuracy, recall rate, and F1 value.

[0042] The present invention also provides a rolling bearing fault diagnosis system for variable working conditions, characterized in that the system comprises:

[0043] A data acquisition module is used to obtain vibration signal data of rolling bearings under different working conditions;

[0044] A feature extraction module is used to adaptively decompose the vibration signal data using an EEMD algorithm, extract fault features under different working conditions, and eliminate the influence of working condition changes on the features through a feature selection method to obtain a feature subset;

[0045] A diagnosis model construction module, used to construct an adaptive diagnosis model based on the feature subset by using a transfer learning method, train a basic model using source domain data, and perform fine-tuning on target domain data to obtain an adaptive fault diagnosis model across working conditions;

[0046] A fault diagnosis module, used to perform fault diagnosis on the newly collected vibration signal based on the cross-operating condition adaptive fault diagnosis model to obtain a fault diagnosis result;

[0047] The model updating module is used for online updating and optimizing the cross-operating condition adaptive fault diagnosis model.

[0048] Compared with the prior art, the present invention has the following advantages and technical effects:

[0049] The present invention first obtains the bearing vibration signal data under different working conditions, uses the EEMD algorithm to perform adaptive decomposition and extract stable feature subsets. Then, a cross-working condition diagnosis model is constructed using transfer learning, and the HMM timing model is introduced to capture the dynamic characteristics of the signal. During diagnosis, the present invention first determines the working condition to which the signal belongs, and selects the corresponding model for fault diagnosis. If the diagnostic confidence is low, the online update mechanism is triggered, and incremental learning is performed using new data. Through continuous data collection and model updating, the present invention constructs a dynamically evolving system that can automatically adjust the diagnostic strategy according to changes in working conditions, effectively solving the problem of decreased diagnostic performance caused by changes in working conditions in bearing fault diagnosis, and achieving highly accurate adaptive diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0051] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;

[0052] Figure 2 Schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0055] Embodiment 1

[0056] like Figure 1 As shown, this embodiment provides a rolling bearing fault diagnosis method for variable working conditions, including the following steps:

[0057] Obtain vibration signal data, temperature data and motor current spectrum data of rolling bearings under different working conditions, use EEMD algorithm to adaptively decompose the vibration signal data, and extract fault characteristics under different working conditions; extract temperature gradient and hot spot position characteristics from the temperature data; extract current harmonic components and frequency offset characteristics from the motor current spectrum data;

[0058] Eliminate the influence of working condition changes on features through feature selection method and obtain feature subsets;

[0059] Based on the feature subset, the adaptive diagnosis model is constructed by using the transfer learning method. The basic model is trained with the source domain data and fine-tuned on the target domain data to obtain an adaptive fault diagnosis model across working conditions.

[0060] Based on the cross-operating condition adaptive fault diagnosis model, comprehensive fault diagnosis is performed on the newly collected vibration signal, temperature signal and electrical frequency signal to obtain a fault diagnosis result.

[0061] Furthermore, the process of obtaining vibration signal data of the rolling bearing under different working conditions includes:

[0062] Obtain vibration signal data of rolling bearings under different speed, load and temperature conditions, and collect comprehensive data samples according to the changes of different parameters;

[0063] Based on the data samples, a data set covering various working conditions is constructed to ensure the comprehensiveness and representativeness of the data;

[0064] Then, the vibration signal data is preprocessed by denoising and normalization, and the signal processing methods of time domain analysis and frequency domain analysis are used to extract the time domain statistical characteristics and frequency domain characteristics of the vibration signal as the input features of the transfer learning model.

[0065] Specifically, the vibration signal data of the rolling bearing under different operating conditions such as speed, load, temperature, etc. are obtained, and comprehensive data samples are collected according to the changes in each parameter. Rolling bearings are common components in mechanical equipment, and their operating status directly affects the performance and life of the equipment. The vibration signal is an important indicator reflecting the state of the bearing. In order to fully understand the vibration characteristics of the bearing under different operating conditions, it is necessary to collect a large amount of vibration signal data under different speeds, loads and temperature conditions. For example, in this embodiment, the vibration signal data are collected when the bearing speed is 1,000 revolutions per minute, 2,000 revolutions per minute and 3,000 revolutions per minute, the load is 50 kilograms, 100 kilograms and 150 kilograms, and the temperature is 30 degrees Celsius, 50 degrees Celsius and 70 degrees Celsius. Through this method, data samples covering a variety of operating conditions can be obtained to ensure the comprehensiveness and representativeness of the data.

[0066] Since the vibration characteristics of the bearing will change with the changes in speed, load and temperature. Only by collecting data covering various working conditions can an accurate bearing condition monitoring model be established. Construct a data set covering various working conditions to ensure the comprehensiveness and representativeness of the data. After collecting the vibration signal data under different working conditions, the data needs to be organized into a data set. This data set should contain data under various working conditions, and there should be enough data under each working condition to ensure the training effect of the model. For example, the previously collected data with a speed of 1,000 rpm, a load of 50 kg, and a temperature of 30 degrees Celsius can be used as a subset, and the data with a speed of 2,000 rpm, a load of 100 kg, and a temperature of 50 degrees Celsius can be used as another subset, and so on. Combining these subsets into a large data set constitutes a data set covering various working conditions. The purpose of this is to ensure that the data set is representative and can reflect the actual vibration conditions of the bearing under various working conditions. The quality and availability of the data can be improved by preprocessing the vibration signal data, including denoising, normalization and other operations.

[0067] The collected vibration signal data may be affected by factors such as environmental noise, and preprocessing is required to improve data quality. The purpose of denoising is to remove the noise components in the signal. The filtering algorithm used in this embodiment is low-pass filtering, which can remove high-frequency noise and retain the main components in the signal. The purpose of normalization is to scale the data to a certain range, such as between zero and one, to eliminate the dimensional differences between different features and make model training more stable. For example, assuming that there is high-frequency noise in the collected vibration signal, a low-pass filter can be used to filter out the high-frequency noise and retain the effective low-frequency signal. Then, the filtered signal is normalized to scale the amplitude of the signal to between zero and one. In this way, the preprocessed data has higher quality and is more suitable for model training.

[0068] This embodiment takes into account the possible existence of noise and dimensional differences in the original data, which will affect the training effect of the model. Through preprocessing, the signal-to-noise ratio of the data can be improved, making it easier for the model to learn the inherent laws of the data. Signal processing methods such as time domain analysis and frequency domain analysis are used to extract the time domain statistical features and frequency domain features of the vibration signal as input features of the machine learning model. Time domain analysis and frequency domain analysis are commonly used signal processing methods. Time domain analysis is to analyze the characteristics of the signal in the time domain, such as statistical features such as mean, variance, and peak value. Frequency domain analysis is to convert the signal to the frequency domain for analysis, such as obtaining the spectrum of the signal through Fourier transform, and extracting features in the spectrum, including frequency amplitude, frequency peak value, etc. These features can be used as input features of the machine learning model. For another example, for the collected vibration signal, its time domain statistical features such as mean, variance, and peak value can be calculated, and Fourier transform can be performed at the same time to obtain the spectrum of the signal, and the main frequency components and their amplitudes in the spectrum can be extracted as frequency domain features. These time domain features and frequency domain features together constitute the input feature vector of the machine learning model.

[0069] This embodiment also takes into account that the original vibration signal data has a high dimension, and directly inputting it into the model will result in a large amount of calculation and difficulty in model training. Through feature extraction, the original data can be converted into a low-dimensional feature vector to improve the training efficiency and accuracy of the model. According to the extracted features, machine learning algorithms such as support vector machine and random forest are used to establish a vibration signal classification model of rolling bearings under different working conditions. Support vector machine and random forest are commonly used machine learning algorithms. Support vector machine divides data of different categories by finding an optimal hyperplane, which is suitable for the classification of high-dimensional data. Random forest classifies by constructing multiple decision trees, with high robustness and accuracy. Selecting a suitable machine learning algorithm and using the extracted features can establish a vibration signal classification model of the bearing under different working conditions. For example, the extracted time domain features and frequency domain features can be used as the input of the support vector machine to train a classification model, which can classify the bearing vibration signals under different working conditions. At the same time, another classification model can also be trained using the random forest algorithm for comparison.

[0070] Parameter tuning refers to improving the performance of a model by adjusting the parameters of the model. For example, you can use 10-fold cross validation to evaluate the classification performance of a support vector machine model. Divide the data set into 10 parts, take one part as the validation set each time, and the remaining nine parts as the training set. Repeat 10 times and calculate the average accuracy and recall rate of the model. Then, you can improve the generalization ability of the model by adjusting the parameters of the support vector machine, such as the kernel function type and penalty coefficient.

[0071] Through methods such as cross-validation, the classification performance of the model is evaluated, and parameters are tuned to improve the generalization ability of the model. Cross-validation divides the data set into multiple parts, one for training and one for validation. By repeating this process multiple times, the average performance index of the model can be obtained.

[0072] This embodiment can evaluate the real performance of the model through cross-validation to avoid overfitting. The optimal model parameters can be found through parameter tuning, and the generalization ability of the model can be improved, so that it performs better in practical applications. The trained model is applied to the actual rolling bearing vibration signal monitoring, and the working conditions of the bearing are judged according to the vibration data collected in real time, so as to realize online monitoring and diagnosis of the bearing state. The trained model can be used to monitor the state of the bearing in real time. The vibration signal collected in real time is preprocessed and feature extracted, and input into the trained model, and the model will output the working condition category of the bearing. According to the working condition category, the operating state of the bearing can be judged, such as whether it is normal or whether there is a fault. For example, the trained support vector machine model is deployed to the bearing monitoring system, and the system collects the vibration signal of the bearing in real time, and inputs it into the model after preprocessing and feature extraction, and the model outputs the working condition category of the bearing. If the model outputs an abnormal working condition, the system will issue an alarm to prompt the operator to check and maintain. The automatic monitoring and diagnosis of the bearing state is realized through the diagnostic model, potential faults are discovered in time, equipment damage is avoided, and production efficiency and safety are improved.

[0073] Furthermore, the EEMD algorithm is used to adaptively decompose the vibration signal data, extract the fault features under different working conditions, and eliminate the influence of working condition changes on the features through feature selection method. The process of obtaining feature subsets includes:

[0074] According to the vibration signal data under different working conditions, the EEMD algorithm is used to adaptively decompose the vibration signal to obtain several intrinsic mode function components;

[0075] Extract fault characteristic parameters from intrinsic mode function components and construct original high-dimensional feature vectors;

[0076] The high-dimensional feature vector is reduced in dimension through feature selection algorithm to remove redundant features and eliminate the influence of working condition changes on the features.

[0077] Fault features are screened according to the feature selection algorithm to construct a low-dimensional feature subset as input for subsequent fault diagnosis.

[0078] Specifically, obtaining vibration signal data under different working conditions is the basis for rolling bearing fault diagnosis. For example, during the operation of mechanical equipment in a factory, the bearings may experience different working conditions such as speed (such as 500rpm, 1000rpm, 1500rpm), load (such as light load, medium load, heavy load) and temperature (such as normal temperature, high temperature). Through the acceleration sensor installed on the bearing, the vibration signal data under these working conditions can be collected in real time. The EEMD (Ensemble Empirical Mode Decomposition) algorithm is used to adaptively decompose the vibration signal. The EEMD algorithm can decompose complex non-stationary signals into several intrinsic mode function (IMF) components, each of which represents different frequency components in the signal. For example, for a collected vibration signal, the EEMD algorithm may decompose it into 5 IMF components and a residual component, each of which reflects the characteristics of the signal in different frequency bands. For the decomposed IMF components, the fault characteristic parameters such as energy, frequency center, bandwidth, etc. are extracted. Assume that the energy of the first IMF component is 0.8, the frequency center is 50Hz, and the bandwidth is 20Hz; the energy of the second IMF component is 0.5, the frequency center is 100Hz, and the bandwidth is 30Hz. Combine these feature parameters to construct an original high-dimensional feature vector, such as \[0.8,50,20,0.5,100,30,\ldots\].

[0079] The original high-dimensional feature vector is reduced in dimension by a feature selection algorithm. The feature selection algorithms that can be used in this embodiment include ReliefF, principal component analysis (PCA), etc. For example, the feature vector is processed using the ReliefF algorithm, and it is found that the energy and frequency center contribute more to the fault classification, while the bandwidth contributes less. Therefore, the redundant feature of the bandwidth is removed to eliminate the impact of the working condition change on the feature. Based on the features selected by the feature selection algorithm, a low-dimensional feature subset with strong stability is constructed, such as \[0.8,50,0.5,100\]. This low-dimensional feature subset not only reduces the computational complexity, but also improves the robustness of the feature, which serves as the input for subsequent fault diagnosis.

[0080] As an additional embodiment, the present embodiment can also use a support vector machine (SVM) algorithm to train feature subsets and establish a fault diagnosis model. SVM is an efficient classification algorithm that can find the optimal hyperplane to separate data of different categories. For example, the training data is divided into four categories: normal state, inner ring fault, outer ring fault and rolling element fault, and a classification model is obtained by training the SVM algorithm. When new vibration signal data is input, it is firstly decomposed by EEMD, the characteristic parameters of the IMF component are extracted, the characteristic vector is constructed, and then input into the trained fault diagnosis model. Assuming that the characteristic vector of the new data is \[0.7,55,0.6,105\], the model determines that the fault type to which it belongs is inner ring fault. According to the results of the fault diagnosis, combined with the operating parameters of the equipment, a fault report is automatically generated. For example, the fault report may contain the following content: equipment number, fault occurrence time, current speed 1000rpm, load is medium load, temperature is normal temperature, fault type is inner ring fault, recommended maintenance measures, etc. This report provides a decision-making basis for equipment maintenance personnel, helping them to quickly locate problems and perform repairs. In this way, not only the accuracy and efficiency of fault diagnosis are improved, but also downtime losses caused by equipment failure can be effectively prevented. The adaptive decomposition capability of the EEMD algorithm makes signal feature extraction more accurate, the dimensionality reduction processing of the feature selection algorithm simplifies the model complexity, and the classification capability of the SVM algorithm ensures the accuracy of fault diagnosis. All links cooperate with each other to form a complete rolling bearing fault diagnosis system, which provides a strong guarantee for the stable operation of the equipment.

[0081] Furthermore, based on the feature subset, the transfer learning method is used to build an adaptive diagnosis model, the basic model is trained using the source domain data, and fine-tuned on the target domain data. The process of obtaining an adaptive fault diagnosis model across working conditions includes:

[0082] According to the acquired source domain data and target domain data, the optimal feature subset is extracted through the feature selection algorithm to build the initial diagnosis model;

[0083] Using the transfer learning method, the source domain data is used to train the initial diagnosis model to obtain the basic diagnosis model;

[0084] The distribution difference between the target domain data and the source domain data is determined. If the difference is greater than the preset threshold, the basic diagnosis model is fine-tuned to finally obtain an adaptive fault diagnosis model across working conditions.

[0085] Specifically, based on the acquired source domain data and target domain data, the optimal feature subset is first extracted through the feature selection algorithm. Assume that the source domain data comes from the vibration signal of a certain model of motor under normal working conditions, and the target domain data comes from the vibration signal of the same model of motor under different working conditions. The feature selection algorithm uses the ReliefF algorithm to select the optimal feature by evaluating the importance of each feature in distinguishing different categories. For example, from the original 100 features, the ReliefF algorithm selects the 10 most discriminative features, such as frequency components, amplitudes, etc., to construct an initial diagnostic model. The transfer learning method is used to train the initial diagnostic model using the source domain data to obtain the basic diagnostic model.

[0086] More specifically, this embodiment uses the TrAdaBoost algorithm in transfer learning, which iteratively reweights the samples of the source domain and the target domain to enhance the adaptability of the source domain data to the target domain data. Assuming that there are 1000 samples in the source domain and 200 samples in the target domain, the weight of the samples in the source domain that are similar to the target domain distribution is increased through the TrAdaBoost algorithm, thereby improving the generalization ability of the model in the target domain. The distribution difference between the target domain data and the source domain data is judged. If the difference is greater than the preset threshold, the basic diagnostic model is fine-tuned. The distribution difference is evaluated by calculating the maximum mean difference (MMD) of the two domain data. Assuming that the preset threshold is 0.1, if the calculated MMD value is 0.15, it indicates that the distribution difference is large and the model needs to be fine-tuned. Fine-tuning can be achieved by further training the model in the target domain and adjusting the model parameters to make it better adapt to the target domain data. The target domain data is predicted by the fine-tuned diagnostic model to obtain the fault diagnosis results under cross-operating conditions. For example, the fine-tuned model predicts 200 samples in the target domain and identifies that 50 samples have bearing faults, 30 samples have gear faults, and the rest are normal. These diagnostic results provide an important basis for equipment maintenance. Determine the fault type based on the diagnostic results and adopt the corresponding maintenance strategy to achieve adaptive maintenance of the equipment. Assuming that the diagnostic results show a bearing fault, the maintenance strategy can be to replace the bearing or perform lubrication. Through this adaptive maintenance, the downtime of the equipment can be significantly reduced and the service life of the equipment can be extended. Continuously collect operating data of the target domain and regularly update the diagnostic model to adapt to changes in equipment conditions. For example, collect new vibration signal data every quarter, re-select features and retrain the model to ensure the timeliness and accuracy of the diagnostic model. This continuous update mechanism enables the model to continuously adapt to various changes in the operation of the equipment. Comprehensively analyze historical diagnostic results and maintenance records, optimize diagnostic models and maintenance strategies, and improve the efficiency of fault diagnosis and processing. For example, by analyzing historical data, it is found that a certain type of fault frequently occurs under specific conditions. The maintenance strategy can be adjusted in a targeted manner to increase preventive maintenance measures under this condition. At the same time, according to the feedback in the maintenance records, the feature weights in the model are adjusted to further improve the accuracy of diagnosis. Through the above steps, not only fault diagnosis under different working conditions is achieved, but also the effectiveness of the diagnosis model and maintenance strategy is improved through continuous optimization and updating. This comprehensive approach not only improves the operating reliability of the equipment, but also reduces maintenance costs, bringing significant economic benefits to the enterprise.

[0087] Furthermore, the transfer learning method is used to train the initial diagnosis model using the source domain data. The process of obtaining the basic diagnosis model includes:

[0088] The HMM time series model is introduced into the transfer learning model, and the feature representation learning of the time domain statistical features and frequency domain features of the extracted vibration signal is performed using the transfer learning model. Then, the feature representation obtained by the transfer learning is input into the HMM time series model, and the time series dynamic characteristics of the vibration signal are captured through the HMM time series model.

[0089] By fusing the transfer learning model and the HMM time series model, a basic diagnostic model of vibration signals is established. On the basis of the basic diagnostic model, an online learning mechanism is introduced to dynamically update the basic diagnostic model according to the newly collected vibration data.

[0090] Specifically, a transfer learning model is used to learn the feature representation of feature vectors. The basic idea of ​​transfer learning is to use a large amount of existing source domain data (such as motor vibration data under normal operating conditions) to train the model, and then transfer the learned knowledge to the target domain (such as motor vibration data under different working conditions). For example, a convolutional neural network (CNN) can be used to extract features and learn representations on source domain data, and then the learned feature representations can be applied to target domain data to improve the generalization and robustness of the model. The feature representation obtained by transfer learning is input into the hidden Markov model (HMM) time series model. As a powerful time series data analysis tool, HMM can capture the time series dynamic characteristics of vibration signals. Assuming that several states are defined, corresponding to different health states of the motor (such as normal, slight wear, severe wear, etc.), HMM can model the time series dynamic characteristics of vibration signals by learning the transition probabilities between states and the observation probabilities under each state.

[0091] Using the transfer learning method, we can fine-tune the existing diagnostic models in related fields. Suppose there is a fault diagnosis model trained based on a large amount of gearbox data. When applied to wind turbines, the direct use of the model may not work well due to differences in working conditions. Through transfer learning, the model can be fine-tuned using a small amount of vibration data from wind turbines to adapt it to new application scenarios. The specific operation includes freezing some layers of the model, adjusting only the last classification layer, and retraining with new data to improve the generalization ability of the model. The diagnostic model is fused with the HMM time series model to consider the time series dynamic characteristics of the vibration signal. The HMM model can capture the transition probability between states and reflect the changing trend of the equipment state. For example, the process from normal state to slight wear and then to serious fault of the bearing is a gradual process. The HMM model can identify this state transition by analyzing the vibration signals at consecutive time points, thereby improving the accuracy of diagnosis. Using the fused diagnostic model for real-time diagnosis can detect equipment abnormalities in a timely manner.

[0092] For example, in a real-time monitoring system, vibration signals are continuously input into the model, and the model outputs the current health status probability of the equipment. If the probability of the fault state output by the model exceeds the preset threshold within a certain period of time, the system triggers the early warning mechanism and notifies maintenance personnel to conduct inspections. This real-time diagnosis mechanism can effectively avoid downtime losses caused by sudden equipment failures. Continuously collecting vibration signal data of the equipment and regularly updating and optimizing the diagnosis model are to adapt to new operating conditions and failure modes that may occur during equipment operation.

[0093] For example, as equipment ages, new fault characteristics may gradually emerge. By regularly updating the model, these new characteristics can be incorporated into the diagnosis scope, improving the long-term effectiveness of the model. This includes regularly collecting new data, re-extracting features and retraining the model to ensure that the model's diagnostic performance is always optimal.

[0094] This embodiment also introduces an online learning mechanism based on the HMM model. During the operation of the equipment, the working conditions and environment will continue to change, and the newly collected vibration data may contain new information. The online learning mechanism can dynamically update the HMM model based on these new data to improve the model's adaptability. For example, after the motor has been running for a period of time, the bearing may gradually wear out, and the newly collected vibration data will reflect this change. The online learning mechanism can adjust the state transition probability and observation probability of the HMM in real time, so that the model always maintains adaptability to the current working conditions. Through the fusion of transfer learning and the HMM time series model, a diagnostic model for vibration signals is established. The model can not only utilize the prior knowledge of the source domain data, but also continuously adapt to the changes in the target domain data through online learning, thereby achieving accurate diagnosis and prediction of the health status of the equipment. For example, the model may predict that the motor has a high probability of bearing wear in the next month, prompting maintenance personnel to perform preventive maintenance in advance. According to the diagnosis results, if the equipment is found to be abnormal or has signs of failure, the alarm mechanism is triggered. The alarm mechanism can be an audible and visual alarm, or an automatically sent email or SMS notification to prompt relevant personnel to perform equipment maintenance or repair. For example, when the model diagnoses abnormal high-frequency components in the motor vibration signal, it may indicate that the bearing is about to fail. The system will immediately trigger an alarm to notify maintenance personnel to conduct timely inspections and replacements to avoid equipment damage or accidents. This vibration signal diagnosis method based on transfer learning and HMM time series model not only improves the accuracy and real-time performance of fault diagnosis, but also enhances the model's adaptive ability through online learning mechanisms, providing strong support for intelligent maintenance of equipment. By continuously collecting and analyzing vibration data and continuously optimizing diagnostic models and maintenance strategies, the operating reliability and maintenance efficiency of equipment can be significantly improved.

[0095] The advantages of this comprehensive diagnosis method are that it improves the generalization and robustness of the model through multi-domain feature extraction and transfer learning; enhances the model's ability to capture dynamic characteristics of time series by introducing the HMM time series model; and realizes dynamic updating and adaptive learning of the model through the online learning mechanism. Ultimately, it not only improves the accuracy of equipment health status diagnosis, but also significantly improves the timeliness and effectiveness of equipment maintenance, reduces operation and maintenance costs, and extends the service life of equipment.

[0096] In the specific implementation process, attention should also be paid to issues such as data collection quality, rationality of feature selection, and optimization of model parameters. For example, ensure that the installation position and angle of the vibration sensor are reasonable to obtain high-quality raw data; when selecting features, combine domain knowledge and data analysis to screen out the most representative features; in the model training process, optimize model parameters through cross-validation and other methods to avoid overfitting or underfitting. Controlling these details is the key to ensuring the efficient operation of the diagnostic model. Through the close cooperation and synergy of the above-mentioned links, an efficient and reliable equipment health status diagnosis system is finally built to provide strong guarantees for the stable operation of equipment and the safe production of enterprises.

[0097] Furthermore, the process of performing comprehensive fault diagnosis on the newly collected vibration signal, temperature signal, and electrical frequency signal based on the cross-operating condition adaptive fault diagnosis model includes:

[0098] The newly collected vibration signal data, temperature signal, and electrical frequency signal are used to determine the working condition. The corresponding diagnostic model is selected according to the working condition category, and the signal data is input into the model for fault diagnosis to obtain the diagnostic result. If the diagnostic result indicates that the equipment is abnormal, the early warning mechanism is triggered to notify relevant personnel to perform maintenance.

[0099] If the confidence level of the diagnosis result is lower than the preset threshold, the online update mechanism of the model is triggered, and the model is incrementally learned using the newly collected vibration signal data. Through continuous data collection and model updating, an adaptive and dynamically evolving bearing fault diagnosis model is constructed.

[0100] Furthermore, the newly collected vibration signal data, temperature signal, and electric frequency signal are subjected to working condition judgment, a corresponding diagnosis model is selected according to the working condition category, and the signal data is input into the model for fault diagnosis. The process of obtaining the diagnosis result includes:

[0101] The extracted feature vector is input into the pre-trained working condition classification model, and the working condition category of the current equipment is determined through model reasoning;

[0102] According to the determined working condition category, a diagnostic model corresponding to the working condition is selected from the fault diagnosis model library, and model parameters are loaded;

[0103] Input the previously extracted feature vector into the selected fault diagnosis model, and obtain the fault probability distribution of the equipment through forward calculation of the model;

[0104] Analyze the fault probability distribution to find the fault type with the highest probability, and make a judgment based on the preset fault threshold to determine the final fault diagnosis result;

[0105] Compare and analyze the fault diagnosis results with the historical fault cases of the equipment, evaluate the credibility of the diagnosis results, and generate fault alarm information when necessary to notify relevant personnel to perform equipment maintenance and status inspection.

[0106] Specifically, the extracted feature vector is input into a pre-trained working condition classification model, and the working condition category of the current equipment is determined by model reasoning. Assuming that a support vector machine (SVM) is used as a working condition classification model, the model has learned a large number of feature vectors under different working conditions during the training phase. For example, the model establishes classification boundaries by learning feature vectors under working conditions such as normal operation, slight wear, and severe wear. When a new feature vector is input, the model can determine the working condition category of the current motor based on these boundaries, such as judging it as "slight wear". According to the determined working condition category, a diagnostic model corresponding to the working condition is selected from the fault diagnosis model library, and the model parameters are loaded. Assume that the model library contains multiple fault diagnosis models for different working conditions, such as a model for the "slight wear" working condition. Each model has learned the fault characteristics and corresponding fault types under the working condition during the training phase. After loading the model parameters, the previously extracted feature vector is input into the selected fault diagnosis model, and the fault probability distribution of the equipment is obtained through model forward calculation. The fault probability distribution is analyzed to find the fault type with the highest probability, and the fault type with the preset fault threshold is combined for judgment to determine the final fault diagnosis result. For example, if the model output shows that the probability of "bearing outer ring wear" is 0.8, and the preset fault threshold is 0.7, it can be determined that the motor currently has a bearing outer ring wear fault. This method can accurately locate and diagnose the fault. Compare and analyze the diagnostic results with the historical fault cases of the equipment to evaluate the credibility of the diagnostic results. For example, if the historical records show that similar feature vectors have been diagnosed as "bearing outer ring wear" many times, and the actual maintenance results also verify this diagnosis, the credibility of the current diagnostic results is high. If necessary, generate fault alarm information to notify relevant personnel to perform equipment maintenance and status inspection. For example, when the diagnostic result is highly reliable and the fault is serious, the system will automatically trigger an audible and visual alarm, and send an email or SMS to notify the maintenance personnel, prompting them to check and replace the bearing in time to avoid equipment damage or accidents. Through the above steps, not only the accuracy and real-time nature of fault diagnosis are improved, but also the credibility of the diagnostic results is enhanced by comparing historical cases. The implementation of this method makes equipment maintenance more intelligent and efficient, and can significantly improve the operating reliability and maintenance efficiency of the equipment. For example, by continuously collecting and analyzing vibration data and constantly optimizing diagnostic models and maintenance strategies, preventive maintenance can be performed before serious equipment failures occur, thereby extending equipment life and reducing maintenance costs.

[0107] Furthermore, if the confidence level of the diagnosis result is lower than a preset threshold, the online update mechanism of the model is triggered, and the process of incremental learning of the model using the newly collected vibration signal data includes:

[0108] Preprocess and extract features of the newly collected vibration signals to obtain a data set that can be used for incremental learning;

[0109] According to the preset incremental learning algorithm, the existing model is incrementally trained using the incremental learning data set, the model parameters are updated, and a new model with improved performance is obtained;

[0110] Use the cross-validation method and performance evaluation indicators to evaluate the performance of the new model to determine whether the model performance meets expectations;

[0111] If the model performance does not meet expectations, return to continue incremental learning and model optimization;

[0112] If the model performance has reached expectations, the new model is deployed to the production environment to replace the original model;

[0113] During the model application process, the confidence of the diagnosis results is continuously monitored. When the confidence falls below the threshold again, a new round of online model update and optimization is repeatedly triggered.

[0114] Furthermore, the preset incremental learning algorithms include online gradient descent method and incremental support vector machine;

[0115] Performance evaluation indicators include calculating accuracy, recall rate, and F1 value.

[0116] Specifically, in the process of vibration signal data collection and fault diagnosis, if the confidence of the diagnosis result is lower than the preset threshold, it means that the prediction result of the model is not reliable enough and there may be a risk of misjudgment. At this time, it is necessary to trigger the online update mechanism of the model. For example, assuming that the preset confidence threshold is 0.8, and the confidence of the current diagnosis result is only 0.7, it indicates that the model needs to be further optimized. First, the newly collected vibration signal data is obtained, and preprocessed and feature extracted. Preprocessing includes denoising and filtering. Assuming that the original signal contains a lot of noise interference, the wavelet transform denoising method can effectively remove high-frequency noise and retain the low-frequency components reflecting the equipment status. Feature extraction extracts time domain and frequency domain features from the preprocessed signal, such as mean, variance, spectrum center, etc., to form a feature vector. Next, according to the preset incremental learning algorithm, such as the online gradient descent method, the newly acquired data set is used to perform incremental training on the existing model. Assuming that the existing model is a fault diagnosis model based on a deep neural network, the model can gradually adjust the weights and biases on the new data through the online gradient descent method to update the model parameters. For example, the new data set contains 10 samples, each with 20 features. The model gradually optimizes its parameters through each iteration update to improve the ability to identify faults. After the incremental training is completed, the new model is evaluated for performance using methods such as cross-validation. Assume that a five-fold cross-validation is used, and the data set is divided into five parts. Four parts are used for training and one part is used for validation each time. Indicators such as accuracy, recall, and F1 value are calculated. Assume that after evaluation, the accuracy of the new model is increased from the original 85% to 90%, the recall rate is increased from 80% to 88%, and the F1 value is increased from 82.5% to 89%, indicating that the model performance has improved. If the model performance does not meet expectations, for example, the F1 value is still lower than the preset threshold of 0.9, return to the second step and continue incremental learning and model optimization. It may be necessary to adjust the parameters of the incremental learning algorithm, such as the learning rate, number of iterations, etc., or introduce more feature engineering methods to improve the generalization ability of the model. If the model performance has met expectations, the new model is deployed to the production environment to replace the original model. For example, through the model deployment platform, the parameter files and inference code of the new model are updated to the production server to ensure that the new diagnostic system can apply the latest model in real time for fault diagnosis. During the model application process, the confidence of the diagnostic results is continuously monitored. Suppose a monitoring system is set up to record the confidence of each diagnosis in real time. When the confidence falls below the 0.8 threshold again, the above steps are repeated to trigger a new round of online model update and optimization. For example, when the monitoring system finds that the confidence of a certain diagnostic result is 0.75, it immediately triggers the update mechanism to obtain new data for incremental training and evaluation. By recording the data features and model performance improvement of each incremental learning, the incremental learning process is tracked and analyzed.For example, record the number of samples, feature distribution, model parameter changes, etc. for each incremental learning, analyze which features contribute more to the model performance improvement, and which samples have a significant impact on model training. This helps to continuously improve the incremental learning algorithm and strategy, and improve the adaptability and robustness of the model. Regularly perform full training on the accumulated diagnostic data to generate a new baseline model. Assume that full training is performed every six months, and a new model is retrained using all the diagnostic data accumulated within six months. This helps to adapt to long-term changes in data distribution and ensure sustainable optimization and evolution of the model. For example, through full training, it is found that the feature distribution of certain fault types has changed significantly. The new model can better capture these changes and improve the accuracy of diagnosis. Through the above steps, the online update mechanism of the model can not only respond to the changes in the confidence of the diagnostic results in a timely manner, but also continuously improve the performance of the model through a combination of incremental learning and full training, ensuring the reliability and accuracy of the fault diagnosis system. Doing so can not only reduce misjudgments and improve the efficiency of equipment maintenance, but also extend the service life of the equipment and reduce maintenance costs.

[0117] Embodiment 2

[0118] like Figure 2 As shown, this embodiment also provides a rolling bearing fault diagnosis system for variable working conditions, including:

[0119] A data acquisition module is used to obtain vibration signal data of rolling bearings under different working conditions;

[0120] A feature extraction module is used to adaptively decompose the vibration signal data using an EEMD algorithm, extract fault features under different working conditions, and eliminate the influence of working condition changes on the features through a feature selection method to obtain a feature subset;

[0121] A diagnosis model construction module, used to construct an adaptive diagnosis model based on the feature subset by using a transfer learning method, train a basic model using source domain data, and perform fine-tuning on target domain data to obtain an adaptive fault diagnosis model across working conditions;

[0122] A fault diagnosis module, used to perform fault diagnosis on the newly collected vibration signal based on the cross-operating condition adaptive fault diagnosis model to obtain a fault diagnosis result;

[0123] The model updating module is used for online updating and optimizing the cross-operating condition adaptive fault diagnosis model.

[0124] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A rolling bearing fault diagnosis method for variable working conditions, characterized in that: include: Acquire vibration signal data, temperature data and motor current spectrum data of rolling bearings under different working conditions, use EEMD algorithm to adaptively decompose the vibration signal data, and extract fault characteristics under different working conditions; Extracting temperature gradient and hot spot position characteristics from the temperature data; Extracting current harmonic components and frequency offset characteristics from the motor current spectrum data; Eliminate the influence of working condition changes on features through feature selection method and obtain feature subsets; Based on the feature subset, a transfer learning method is used to build an adaptive diagnosis model, a basic model is trained using source domain data, and fine-tuned on target domain data to obtain an adaptive fault diagnosis model across working conditions; Based on the cross-operating condition adaptive fault diagnosis model, comprehensive fault diagnosis is performed on the newly collected vibration signal, temperature signal and electrical frequency signal to obtain a fault diagnosis result.

2. The method according to claim 1, characterized in that: The process of obtaining vibration signal data of rolling bearings under different working conditions includes: Obtain vibration signal data of rolling bearings under different speed, load and temperature conditions, and collect comprehensive data samples according to the changes of different parameters; Constructing a data set covering various working conditions according to the data samples; Then, the vibration signal data is preprocessed by denoising and normalization, and the signal processing methods of time domain analysis and frequency domain analysis are used to extract the time domain statistical characteristics and frequency domain characteristics of the vibration signal as the input features of the transfer learning model.

3. The method according to claim 1, characterized in that: The EEMD algorithm is used to adaptively decompose the vibration signal data, extract fault features under different working conditions, and eliminate the influence of working condition changes on the features through feature selection methods. The process of obtaining feature subsets includes: According to the vibration signal data under different working conditions, the EEMD algorithm is used to adaptively decompose the vibration signal to obtain several intrinsic mode function components; Extracting fault characteristic parameters from the intrinsic mode function components to construct an original high-dimensional feature vector; The high-dimensional feature vector is subjected to dimensionality reduction processing by a feature selection algorithm, fault features are screened according to the feature selection algorithm, and a low-dimensional feature subset is constructed to serve as an input for subsequent fault diagnosis.

4. The method according to claim 1, characterized in that According to the feature subset, a transfer learning method is used to construct an adaptive diagnosis model, a basic model is trained using source domain data, and fine-tuned on target domain data to obtain an adaptive fault diagnosis model across working conditions, including: According to the acquired source domain data and target domain data, the optimal feature subset is extracted through the feature selection algorithm to build the initial diagnosis model; Using a transfer learning method, the initial diagnosis model is trained using source domain data to obtain a basic diagnosis model; The distribution difference between the target domain data and the source domain data is determined. If the difference is greater than a preset threshold, the basic diagnosis model is fine-tuned to finally obtain an adaptive fault diagnosis model across working conditions.

5. The method according to claim 4, characterized in that The process of using the transfer learning method to train the initial diagnosis model using the source domain data to obtain the basic diagnosis model includes: The HMM time series model is introduced into the transfer learning model, and the feature representation learning of the time domain statistical features and frequency domain features of the extracted vibration signal is performed using the transfer learning model. Then, the feature representation obtained by the transfer learning is input into the HMM time series model, and the time series dynamic characteristics of the vibration signal are captured through the HMM time series model. By fusing the transfer learning model and the HMM time series model, a basic diagnostic model of vibration signals is established, and on the basis of the basic diagnostic model, an online learning mechanism is introduced to dynamically update the basic diagnostic model according to the newly collected vibration data.

6. The method according to claim 1, characterized in that The process of performing comprehensive fault diagnosis on the newly collected vibration signal, temperature signal, and electrical frequency signal based on the cross-operating condition adaptive fault diagnosis model includes: The newly collected vibration signal data, temperature signal, and electrical frequency signal are used to determine the working condition, and the corresponding diagnostic model is selected according to the working condition category. The signal data is input into the model for fault diagnosis to obtain the diagnostic result. If the diagnostic result indicates that the equipment is abnormal, the early warning mechanism is triggered to notify relevant personnel to perform maintenance. If the confidence level of the diagnosis result is lower than the preset threshold, the online update mechanism of the model is triggered, and the model is incrementally learned using the newly collected vibration signal data. Through continuous data collection and model updating, an adaptive and dynamically evolving bearing fault diagnosis model is constructed.

7. The method according to claim 6, characterized in that The newly collected vibration signal, temperature signal, and electrical frequency signal data are used to determine the working condition, and the corresponding diagnostic model is selected according to the working condition category. The signal data is input into the model for fault diagnosis. The process of obtaining the diagnostic result includes: The extracted feature vector is input into the pre-trained working condition classification model, and the working condition category of the current equipment is determined through model reasoning; According to the determined working condition category, a diagnostic model corresponding to the working condition is selected from the fault diagnosis model library, and model parameters are loaded; Input the previously extracted feature vector into the selected fault diagnosis model, and obtain the fault probability distribution of the equipment through forward calculation of the model; Analyze the fault probability distribution to find the fault type with the highest probability, and make a judgment based on the preset fault threshold to determine the final fault diagnosis result; The fault diagnosis result is compared and analyzed with the historical fault cases of the equipment, the credibility of the diagnosis result is evaluated, and fault alarm information is generated when necessary to notify relevant personnel to perform equipment maintenance and status inspection.

8. The method according to claim 6, characterized in that If the confidence level of the diagnosis result is lower than the preset threshold, the online update mechanism of the model is triggered. The process of incremental learning of the model using the newly collected vibration signal data includes: Preprocess and extract features of the newly collected vibration signals to obtain a data set that can be used for incremental learning; According to a preset incremental learning algorithm, the incremental learning data set is used to perform incremental training on the existing model, update model parameters, and obtain a new model with improved performance; Using a cross-validation method and based on performance evaluation indicators, the new model is evaluated to determine whether the model performance meets expectations; If the model performance does not meet expectations, return to continue incremental learning and model optimization; If the model performance has reached expectations, the new model is deployed to the production environment to replace the original model; During the model application process, the confidence of the diagnosis results is continuously monitored. When the confidence falls below the threshold again, a new round of online model update and optimization is repeatedly triggered.

9. The method according to claim 8, characterized in that The preset incremental learning algorithm includes online gradient descent method and incremental support vector machine; The performance evaluation indicators include calculation accuracy, recall rate, and F1 value.

10. A rolling bearing fault diagnosis system for variable working conditions, characterized in that: include: A data acquisition module is used to obtain vibration signal data of rolling bearings under different working conditions; A feature extraction module is used to adaptively decompose the vibration signal data using an EEMD algorithm, extract fault features under different working conditions, and eliminate the influence of working condition changes on the features through a feature selection method to obtain a feature subset; A diagnosis model construction module, used to construct an adaptive diagnosis model based on the feature subset by using a transfer learning method, train a basic model using source domain data, and perform fine-tuning on target domain data to obtain an adaptive fault diagnosis model across working conditions; A fault diagnosis module, used to perform fault diagnosis on the newly collected vibration signal based on the cross-operating condition adaptive fault diagnosis model to obtain a fault diagnosis result; The model updating module is used for online updating and optimizing the cross-operating condition adaptive fault diagnosis model.

Images