Intelligent real-time diagnosis method of bearing unbalance fault based on wavelet phase space reconstruction

Abstract

The application discloses an intelligent real-time diagnosis method for bearing unbalance impact fault based on wavelet phase space oversampling reconstruction, constructs the oversampling method index positioning part of the concentrated fault signal energy set in the deep positioning data point of the phase space reconstruction continuous wavelet transform, and reduces the influence of noise and edge samples by controlling the distribution of the synthesized new samples; a two-stage feature learning random vector function link network fusing random projection and adaptive refinement is proposed, and a dynamic threshold perception focus balance loss function is designed, which reduces the loss contribution of easily classified samples while balancing the class balance weight, and adaptively adjusts the threshold to balance local noise and global statistics, so that the model pays more attention to difficult samples, and finally, small batch incremental learning is applied to online diagnosis to realize real-time fault diagnosis. The application has good diagnosis rate and stability, and is suitable for solving the problems of high misdiagnosis and missed diagnosis rate of fault samples and low overall diagnosis precision of the model under an unbalanced data set.

Description

Technical Field

[0001] This invention relates to intelligent fault detection and deep learning technology, specifically to an intelligent real-time diagnostic method for bearing imbalance impact faults reconstructed by wavelet phase space oversampling. Background Technology

[0002] Real-time bearing fault diagnosis is of great significance in ensuring equipment safety, improving efficiency, reducing costs, and promoting intelligent development. It is an indispensable technical means in modern industrial equipment management.

[0003] Bearings are among the most critical and fault-prone components in mechanical equipment, and their operating status directly affects the stability and safety of the equipment. Real-time diagnostic technology, combined with the Internet of Things (IoT) and big data analytics, enables remote acquisition and analysis of bearing operating data. On one hand, real-time fault diagnosis can promptly detect potential bearing failures, quickly locate the cause of the failure, shorten repair time, and reduce equipment downtime losses, avoiding equipment shutdowns or even catastrophic accidents caused by bearing failure, thereby ensuring safe equipment operation and improving overall equipment efficiency. On the other hand, bearings may face complex operating conditions in actual operation, such as high load, high temperature, and high humidity. Real-time diagnostic technology can adapt to these complex environments, promptly detect faults caused by changes in operating conditions, and ensure stable operation of equipment under harsh conditions.

[0004] With technological advancements, bearing fault diagnosis is evolving towards greater intelligence, precision, and efficiency. The application of real-time diagnostic technology not only improves diagnostic accuracy but also reduces the possibility of human intervention and misjudgment, providing strong support for the intelligent management of industrial equipment.

[0005] Among existing technical solutions, there are few methods for constructing pseudo-fault samples and applying them for real-time fault diagnosis. For example, CN202310774560.1 discloses a rolling bearing fault diagnosis method and terminal based on a temporal memory enhancement network, which mainly couples the forget gate and input gate of the LSTM unit and connects them to the memory unit respectively, but does not handle the problem of unbalanced faults and requires a large number of samples. CN202410570406.7 discloses a bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network, which increases the number and diversity of training samples through the VSE method and constructs TFIF-Net training, but does not perform real-time diagnosis and cannot evaluate its performance in a real-time diagnosis system. CN202210684731.7 discloses a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network, which uses a one-dimensional convolutional subdomain adaptive adversarial network model to extract domain-invariant features, but has poor generalization ability.

[0006] In summary, existing technical solutions are insufficient for real-time diagnosis of unbalanced bearing faults. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to overcome the shortcomings of the existing technology and provide an intelligent real-time diagnostic method for bearing imbalance impact faults by wavelet phase space oversampling reconstruction.

[0008] Technical solution: The present invention provides an intelligent real-time diagnostic method for bearing imbalance impact faults using wavelet phase space oversampling reconstruction, comprising the following steps:

[0009] Step 1: Obtain vibration signal data of the health and fault states of different parts of the bearing under different working conditions. Set the vibration signal of the health state as the majority class and set the signals of inner ring fault, outer ring fault and rolling element fault under different working conditions as the minority class. Divide the unbalanced original majority class signal dataset and the original minority class signal dataset.

[0010] Step 2: Construct the oversampling method ODLC-SMOTE for depth localization data points in phase space reconstruction continuous wavelet transform to generate pseudo-fault samples for the minority class of fault state signals;

[0011] Step 2.1: Use a mean-maximizing sliding window to index the deep fault features after continuous wavelet transform on the original minority class signal samples. Locate and extract signals containing original deep fault features ;

[0012] Step 2.2: Reconstruct the original one-dimensional majority class signal and the original minority class signal into a two-dimensional space through phase space, expand the original minority class fault state signal to generate pseudo fault samples, which become reconstructed minority class samples, and check the sample distance and sample generation quality.

[0013] Step 2.3: Compress the obtained reconstructed two-dimensional minority class samples to form a one-dimensional fault signal containing pseudo-fault samples, and combine it with the signal containing the original deep fault features to form a new minority class fault dataset;

[0014] The original majority class samples and the new minority class fault dataset are divided into an imbalanced training set, a balanced test set, and a validation set.

[0015] Step 3: Construct a two-stage feature learning stochastic vector functional link network (TSFL-RVFL) that integrates random projection and adaptive refinement. Obtain the fault diagnosis results for the offline stage through the two-stage feature learning stochastic vector functional link network (TSFL-RVFL).

[0016] The two-stage feature learning stochastic vector functional linker network TSFL-RVFL was trained using the training dataset. During the training process, the dynamic threshold-aware focus balance loss function DTAFBLoss was used for optimization. The class balance weights reduced the loss contribution of easily classified samples, while the adaptive threshold adjustment balanced local noise and global statistics, making the model pay more attention to difficult samples.

[0017] Adjusting appropriate model parameters improves diagnostic results, resulting in a fixed-parameter model. The validation and test sets are then input into the fixed-parameter model for offline diagnostic validation and real-time diagnostic testing.

[0018] Furthermore, in step 1, the original majority class signal dataset contains more samples than the original minority class signal dataset.

[0019] Furthermore, the formula for maximizing the sliding window index in step 2.1 is as follows:

[0020] ;

[0021] Where i refers to the i-th original minority class signal. The signal containing the original deep fault characteristics is located, where t is the time step. For the original minority class dataset, The original minority class dataset after continuous wavelet transform. Here, j represents the size of the sliding window, and j represents the starting position of the sliding window. Where is the length of a single sample in the original minority class dataset during the partitioning, and k is the index of the deep fault feature. This refers to the portion where the fault signal energy is concentrated.

[0022] Step 2.2 The specific method for obtaining the reconstructed minority class samples is as follows:

[0023] Calculate the Euclidean distance between the two-dimensional original majority class signal and the signal containing the original deep fault features, and find the K+1 nearest neighbors of the signal containing the original deep fault features on the original majority class signal. The calculation formula is as follows:

[0024]

[0025]

[0026] in, The set of data points representing the original two-dimensional majority class signal. Data points representing the original minority class signal in two dimensions The set, yes exist The A neighbor, Represents Euclidean distance. It is the (K+1)th smallest nearest neighbor value.

[0027] Furthermore, the two-stage feature learning stochastic vector functional linker network (TSFL-RVFL) includes a stochastic projection layer, a deep feature enhancement layer, a regularization mechanism, and a classification output layer.

[0028] In the random projection stage, given the input dimension d and the number of hidden layer units h, the input data is projected into a high-dimensional space through a fixed random mapping, inheriting the fast computation advantage of RVFL;

[0029] In the adaptive refinement stage, feature refinement is achieved by cascading trainable fully connected layers, activation functions are used to enhance nonlinear representation capabilities, and regularization constraints are introduced.

[0030] Furthermore, the dynamic threshold-sensing focus balance loss function DTAFBLoss in step 3 is based on the predicted probability. and adaptive threshold The loss function is divided into two branches: the hard sample branch loss function. and simple sample branch loss function The definition is as follows:

[0031] ;

[0032] ;

[0033] in, It is the loss amplification factor for hard samples, which controls the degree of loss amplification for hard samples and adjusts the influence of important local features of the time series captured by the ODLC-SMOTE algorithm on classification; It is the focus loss adjustment factor, which controls the adjustment strength of focus loss; yes Time step category The predicted probability, It is the predicted probability distribution of the model. It is a minimal constant that prevents the derivative from being zero;

[0034] Combining class-balanced weights and the focus loss branch, the final loss function is defined as:

[0035]

[0036] Among them, 1 / It is the true class of sample i (i.e., the original minority class signal). Category balance weights, It is the predicted probability of sample i. It is the adaptive threshold for sample i. C is the frequency of deep fault features, and C is the number of samples.

[0037] Furthermore, the real-time diagnostic testing process in step 3 is as follows:

[0038] First, obtain the initial weights. ;

[0039] Construct an augmented matrix by vertically stacking the historical output matrix and the current output matrix. The corresponding tag augmentation is ;

[0040] Calculate the gain matrix by combining historical and current data. At the same time, a regularization term is introduced. :

[0041] ;

[0042] Through forgetting factors Control the decay rate of historical information, subtract Uncertainty in the estimation of the compression parameters; the covariance matrix is ​​updated as follows:

[0043] ;

[0044] weight matrix Updated to:

[0045] ;

[0046] in, For historical output matrix, This is the current output matrix. An augmented matrix that is the transpose;

[0047] Through the above process, incremental learning in the real-time diagnosis process can effectively combine historical data with current data while maintaining numerical stability and computational efficiency.

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

[0049] (1) Compared with traditional signal processing methods, the present invention focuses directly on the impact part of bearing failure and uses healthy samples as the majority class to reconstruct minority class failure samples, which reduces the use of computing resources and is more effective for fault signal processing.

[0050] (2) This invention focuses on both the imbalance problem and the classification problem of easy and difficult samples during the model training process, which can balance the local noise and global statistics of model training, and make the model pay more attention to difficult samples.

[0051] (3) The present invention uses a lightweight training model, which shortens the computation time while ensuring training accuracy, thus providing a foundation for rapid real-time diagnosis.

[0052] (4) The present invention applies small-batch incremental learning. Incremental learning in the real-time diagnosis process can effectively combine historical data with current data while maintaining numerical stability and computational efficiency. Attached Figure Description

[0053] Figure 1 This is the overall flowchart of the present invention.

[0054] Figure 2a This is a two-dimensional rendering of the reconstructed bearing inner ring impact fault in the embodiment. Figure 2b This is a two-dimensional rendering of the reconstructed bearing outer ring impact fault in the embodiment.

[0055] Figure 3 This is a schematic diagram of the overall network framework of the present invention.

[0056] Figure 4 This is a schematic diagram illustrating the effects of the hard sample loss amplification factor of 5, the adaptive threshold weight of 0.5, and the minimum constant of 0.000001 in the embodiment.

[0057] Figure 5 This is the decision boundary diagram of the real-time diagnostic process of the model in the embodiment.

[0058] Figure 6 This is a G-means result graph of the real-time diagnostic process of the model in the embodiment. Detailed Implementation

[0059] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0060] Considering the data imbalance problem in practical engineering, this invention constructs an oversampling indexing method for depth localization data points in phase space reconstruction continuous wavelet transform to locate the energy concentration of fault signals, and reduces the influence of noise and edge samples by controlling the distribution of synthesized new samples. To address the feature coupling problem of fault signals, a two-stage feature learning stochastic vector functional link network integrating random projection and adaptive refinement is proposed. To address the insufficient learning capacity of the network in imbalanced tasks, a dynamic threshold-aware focus balancing loss function is designed. While class balancing weights reduce the loss contribution of easily classified samples, adaptive threshold adjustment balances local noise and global statistics, making the model focus more on difficult samples. Finally, mini-batch incremental learning is applied to online diagnosis to achieve real-time fault diagnosis. Online diagnosis results show that the proposed method has good diagnostic rate and stability, and is suitable for handling the problems of high misdiagnosis and false negative rates of fault samples and low overall diagnostic accuracy of the model under imbalanced datasets.

[0061] like Figure 1 and Figure 3 As shown in this embodiment, the intelligent real-time diagnosis method for bearing imbalance faults using wavelet phase space reconstruction includes the following steps:

[0062] Step 1: Obtain vibration signal data of the health and fault states of different parts of the bearing under different working conditions. Set the vibration signal of the health state as the majority class and set the signals of inner ring fault, outer ring fault and rolling element fault under different working conditions as the minority class. Divide the unbalanced original majority class signal dataset and the original minority class signal dataset.

[0063] Step 2: Construct the oversampling method ODLC-SMOTE for depth localization data points in phase space reconstruction continuous wavelet transform to generate pseudo-fault samples for the minority class of fault state signals;

[0064] Step 2.1: Use a mean-maximizing sliding window to index the deep fault features after continuous wavelet transform on the original minority class signal samples, and locate and extract signals containing the original deep fault features. ;

[0065] Step 2.2: Reconstruct the original one-dimensional majority class signal and the original minority class signal into a two-dimensional space through phase space, expand the original minority class fault state signal to generate pseudo fault samples, which become reconstructed minority class samples, and check the sample distance and sample generation quality.

[0066] Steps 2 and 3: Compress the obtained reconstructed two-dimensional minority class samples to form a one-dimensional fault signal containing pseudo-fault samples, and combine it with the signal containing the original deep fault features to form a new minority class fault dataset.

[0067] The original majority class samples and the new minority class fault dataset are divided into an imbalanced training set, a balanced test set, and a validation set.

[0068] Step 3: Construct a two-stage feature learning stochastic vector functional link network (TSFL-RVFL) that integrates random projection and adaptive refinement. Obtain the fault diagnosis results for the offline stage through the two-stage feature learning stochastic vector functional link network (TSFL-RVFL).

[0069] The two-stage feature learning stochastic vector functional linker network TSFL-RVFL was trained using the training dataset. During the training process, the dynamic threshold-aware focus balance loss function DTAFBLoss was used for optimization. The class balance weights reduced the loss contribution of easily classified samples, while the adaptive threshold adjustment balanced local noise and global statistics, making the model pay more attention to difficult samples.

[0070] Adjusting appropriate model parameters improves diagnostic results, resulting in a fixed-parameter model. The validation and test sets are then input into the fixed-parameter model for offline diagnostic validation and real-time diagnostic testing.

[0071] In this embodiment, the number of samples in the original majority class signal dataset in step 1 is greater than the number of samples in the original minority class signal dataset.

[0072] The formula for maximizing the sliding window index in step 2.1 of this embodiment is as follows:

[0073] ;

[0074] Where i refers to the i-th original minority class signal. The signal containing the original deep fault characteristics is located, where t is the time step. For the original minority class dataset, The original minority class dataset after continuous wavelet transform. Here, j represents the size of the sliding window, and j represents the starting position of the sliding window. Where is the length of a single sample in the original minority class dataset during the partitioning, and k is the index of the deep fault feature. This refers to the portion where the fault signal energy is concentrated.

[0075] Step 2.2 The specific method for obtaining the reconstructed minority class samples is as follows:

[0076] Calculate the Euclidean distance between the two-dimensional original majority class signal and the signal containing the original deep fault features, and find the K+1 nearest neighbors of the signal containing the original deep fault features on the original majority class signal. The calculation formula is as follows:

[0077] ;

[0078] ;

[0079] in, The set of data points representing the original two-dimensional majority class signal. Data points representing the original minority class signal in two dimensions The set, yes exist The A neighbor, Represents Euclidean distance. It is the (K+1)th smallest nearest neighbor value.

[0080] This embodiment of the two-stage feature learning stochastic vector functional linker network (TSFL-RVFL) includes a stochastic projection layer, a deep feature enhancement layer, a regularization mechanism, and a classification output layer.

[0081] In the random projection stage, given the input dimension d and the number of hidden layer units h, the input data is projected into a high-dimensional space through a fixed random mapping, inheriting the fast computation advantage of RVFL;

[0082] In the adaptive refinement stage, feature refinement is achieved by cascading trainable fully connected layers, activation functions are used to enhance nonlinear representation capabilities, and regularization constraints are introduced.

[0083] In this embodiment, step 3, the dynamic threshold sensing focus balance loss function DTAFBLoss, is based on prediction probability. and adaptive threshold The loss function is divided into two branches: the hard sample branch loss function. and simple sample branch loss function The definition is as follows:

[0084] ;

[0085] ;

[0086] in, It is the loss amplification factor for hard samples, which controls the degree of loss amplification for hard samples and adjusts the influence of important local features of the time series captured by the ODLC-SMOTE algorithm on classification; It is the focus loss adjustment factor, which controls the adjustment strength of focus loss; yes Time step category The predicted probability, It is the predicted probability distribution of the model. It is a minimal constant that prevents the derivative from being zero;

[0087] Combining class-balanced weights and the focus loss branch, the final loss function is defined as:

[0088] ;

[0089] Among them, 1 / Is it the true category of sample i? Category balance weights, It is the predicted probability of sample i. It is the adaptive threshold for sample i. C is the frequency of deep fault features, and C is the number of samples.

[0090] The real-time diagnostic test process in step 3 of this embodiment is as follows:

[0091] First, obtain the initial weights. ;

[0092] Construct an augmented matrix by vertically stacking the historical output matrix and the current output matrix. The corresponding tag augmentation is ;

[0093] Calculate the gain matrix by combining historical and current data. At the same time, a regularization term is introduced. :

[0094] ;

[0095] Through forgetting factors Control the decay rate of historical information, subtract Uncertainty in the estimation of the compression parameters; the covariance matrix is ​​updated as follows:

[0096] ;

[0097] weight matrix Updated to: ;

[0098] in, For historical output matrix, This is the current output matrix. An augmented matrix that is the transpose;

[0099] Through the above process, incremental learning in the real-time diagnosis process can effectively combine historical data with current data while maintaining numerical stability and computational efficiency.

[0100] To further verify the feasibility and technical effect of the present invention, this embodiment uses the BJTU-RAO bogie dataset for verification. This publicly available dataset is the world's first publicly available dataset for simulating faults in the transmission system of rail transit train bogies.

[0101] The experimental platform's drivetrain is powered by a three-phase asynchronous AC motor, with the motor speed controlled by a frequency converter. The motor bearings are SKF 6205-2RSH; the drive gear support bearings are HRB 32305; and the axle box bearings are HRB352213. The fault characteristic frequencies (rotational frequency f) of the inner ring, outer ring, and rolling elements are shown in Table 1. The sensors on the experimental platform highly replicate the actual measurement points of the train.

[0102] Table 1. Bearing Failure Characteristic Frequency Table

[0103] bearing model Inner circle Outer ring Rolling body cage SKF 6205-2RSH 5.415*f 3.585*f 2.357*f 0.398*f HRB 32305 7.889*f 5.111*f 2.203*f 0.393*f HRB 352213 10.929*f 8.071*f 3.164*f 0.425*f

[0104] This embodiment performs fault diagnosis on gearbox bearing inner ring faults, gearbox bearing rolling element faults, axle box bearing inner ring faults, axle box bearing outer ring faults, axle box bearing rolling element faults, and health status under the condition of motor speed of 1200 r / min and lateral load of 10 kN.

[0105] The two-stage feature learning stochastic vector functional linked network model TSFL-RVFL of this invention randomly initializes the weight matrix and bias vector to quickly construct hidden layers, performs linear transformation on the input signal, and applies ReLU activation. After grid search, the model contains three fully connected layers with 64, 32, and 16 neurons respectively. Each fully connected layer uses the ReLU activation function, and L2 regularization and random dropping of 10% of neurons are applied to prevent overfitting. To illustrate the effectiveness of the diagnostic method, the ratio of diagnostic training set, validation set, and test set is set to 0.6:0.2:0.2. The training set uses an unbalanced ratio of 1:1:1:1:1:10 for gearbox bearing inner race fault, gearbox bearing rolling element fault, axle box bearing inner race fault, axle box bearing outer race fault, axle box bearing rolling element fault, and healthy state. Algorithm sampling is performed on 300, 600, and 900 algorithm resampling data points respectively to verify the generalization performance of the method. The diagnostic process was trained for 500 rounds and the experiment was repeated. The validation set accuracy, balanced accuracy, F1 score, G-mean value, and Matthews correlation coefficient (MMC) of the offline diagnosis are shown in Table 2.

[0106] Table 2 Offline diagnostic results

[0107] Sampling points accuracy Balanced accuracy F1 score G-mean MMC 300 94.14±0.05% 94.14±0.04% 94.14±0.03% 93.99±0.08% 0.929±0.02 600 93.85±0.01% 93.85±0.02% 93.77±0.10% 93.58±0.04% 0.926±0.01 900 93.21±0.02% 93.21±0.12% 93.12±0.02% 92.89±0.02% 0.918±0.03

[0108] Accuracy represents the proportion of correctly classified samples out of all samples. Balanced accuracy avoids overestimating the performance of an imbalanced dataset by a single metric, thus mitigating the impact of imbalance in multi-class classification tasks.

[0109] As shown in Table 2, the offline diagnostic results demonstrate that the network model of this invention achieves a high accuracy level of 94.14±0.05% when processing datasets with an imbalance of 1:1:1:1:1:10, both in terms of accuracy and balance. In multi-class classification problems, G-mean reflects the geometric mean of classification accuracy for each class of samples. The proposed method achieves a G-mean of 93.99±0.08%, only 0.15% lower than the accuracy, indicating that the network model of this invention can handle imbalanced classification problems. MCC is a comprehensive evaluation metric that considers the performance of various aspects of the classification model more comprehensively, providing a more holistic perspective for evaluating the quality of the model. The proposed method achieves an MCC value of no less than 0.918±0.03, indicating good overall model performance.

[0110] Figure 2aThis is a diagram illustrating the inner circle fault effects in the embodiment, where blue dots represent healthy samples (majority class), red dots represent faulty samples (minority class), and green dots represent newly generated faulty samples. Figure 2b The image shows the fault detection effect on the outer ring, where blue dots represent healthy samples of the majority class, red dots represent fault samples of the minority class, and green dots represent newly generated fault samples. It can be seen that the fault impact data points located by the oversampling method ODLC-SMOTE of this invention are significantly more numerous than those of healthy samples, and the distribution of generated fault samples is consistent with that of the minority class fault samples, without losing fault features.

[0111] from Figure 4 As can be seen, the gray vertical dashed lines mark the segmented thresholds of the loss function. The hard sample loss amplification coefficient and adaptive threshold weight of the DTAFFBLoss in this invention will affect the model's judgment of easy and difficult samples. The loss is significantly higher than the blue baseline in the region where the predicted probability is less than 0.5, which reflects the model's stronger attention to difficult classification samples and ensures that the model assigns higher weights to faulty samples with low confidence. The loss decreases rapidly in the region where the predicted probability is greater than 0.5, which reflects the model's stronger suppression of simple samples, suppressing the weight of normal samples and preventing the model from being dominated by a large number of simple samples.

[0112] Figure 5 In the decision boundary diagram of the model's real-time diagnostic process, different colors represent faults with different labels; red crosses indicate incorrectly predicted data, while black dots represent the corresponding correct labels. From Figure 3 Firstly, the decision boundaries are clear rather than confusing, indicating that this invention can effectively distinguish the distribution of different fault categories in the feature space. Secondly, the predicted error labels of the model's real-time diagnosis are not concentrated, and there are no frequent fluctuations in the predicted error labels, indicating that this invention is not sensitive to noise or transient anomalies and has robustness.

[0113] from Figure 6 The G-mean results of the real-time diagnosis process show that as the number of data points in the online data stream increases, the real-time diagnosis of the model tends to stabilize. Although the G-mean value of 900 sampling points in the offline stage is 1.17% lower than the G-mean value of 300 sampling points, the performance of real-time diagnosis is significantly better. This indicates that the present invention can adapt to the data distribution of new samples in real-time diagnosis, and the higher the number of sampling points, the stronger this adaptability.

[0114] In summary, this invention directly focuses on the impact component of bearing failures, extracts bearing impact faults, and reconstructs minority class fault samples based on majority class healthy samples, resulting in more effective fault signal processing. The method of this invention simultaneously addresses the imbalance problem and the classification problem of easy and difficult samples during model training, balancing local noise and global statistics in model training, allowing the model to focus more on difficult samples. The mini-batch incremental learning in real-time diagnosis effectively combines historical and current data while maintaining numerical stability and computational efficiency.

Claims

1. A smart real-time diagnostic method for bearing imbalance impact faults using wavelet phase space oversampling reconstruction, characterized in that, Includes the following steps: Step 1: Obtain vibration signal data of the health and fault states of different parts of the bearing under different working conditions. Set the vibration signal of the health state as the majority class and set the signals of inner ring fault, outer ring fault and rolling element fault under different working conditions as the minority class. Divide the unbalanced original majority class signal dataset and the original minority class signal dataset. Step 2: The ODLC-SMOTE method, which uses oversampling of depth localization data points in phase space reconstruction continuous wavelet transform, is used to process the fault state signals of the minority class to generate pseudo-fault samples. Step 2.1: Use a mean-maximizing sliding window to index the deep fault features after continuous wavelet transform on the original minority class signal samples. Locate and extract signals containing original deep fault features ; Step 2.2: Reconstruct the original one-dimensional majority class signal and the original minority class signal into a two-dimensional space through phase space, expand the original minority class fault state signal to generate pseudo fault samples, which become reconstructed minority class samples, and check the sample distance and sample generation quality. Steps 2 and 3: Compress the obtained reconstructed two-dimensional minority class samples to form a one-dimensional fault signal containing pseudo-fault samples, and combine it with the signal containing the original deep fault features to form a new minority class fault dataset. The original majority class samples and the new minority class fault dataset are divided into an imbalanced training set, a balanced test set, and a validation set. Step 3: Construct a two-stage feature learning stochastic vector functional link network (TSFL-RVFL) that integrates random projection and adaptive refinement. Obtain fault diagnosis results for the offline stage through the TSFL-RVFL. Train the TSFL-RVFL using the training dataset. During training, optimize and adjust the loss function DTAFBLoss using dynamic threshold-aware focus balance. While reducing the loss contribution of easily classified samples by balancing the class weights, adaptive threshold adjustment balances local noise and global statistics, making the model focus more on difficult samples. Adjust appropriate model parameters to improve diagnostic results and obtain a fixed-parameter model. Input the validation set and test set into the fixed-parameter model for offline diagnostic verification and real-time diagnostic testing. The two-stage feature learning stochastic vector functional linker network TSFL-RVFL includes a stochastic projection layer, a deep feature enhancement layer, a regularization mechanism, and a classification output layer. In the stochastic projection stage, given the input dimension d and the number of hidden layer units h, the input data is projected into a high-dimensional space through a fixed random mapping, inheriting the fast computation advantage of RVFL. In the adaptive refinement stage, feature refinement is achieved by cascading trainable fully connected layers, activation functions are used to enhance nonlinear representation capabilities, and regularization constraints are introduced. The dynamic threshold-sensing focus balance loss function DTAFBLoss is based on prediction probability. and adaptive threshold The loss function is divided into a hard sample branch loss function and a simple sample branch loss function, and then the final loss function is obtained by combining the class balance weight and the focus loss branch.

2. The intelligent real-time diagnosis method for bearing imbalance impact faults by wavelet phase space oversampling reconstruction according to claim 1, characterized in that, In step 1, the original majority class signal dataset contains more samples than the original minority class signal dataset.

3. The intelligent real-time diagnosis method for bearing imbalance impact faults based on wavelet phase space oversampling reconstruction according to claim 1, characterized in that, The formula for maximizing the sliding window index in step 2.1 is as follows: ； Where i refers to the i-th original minority class signal. The signal containing the original deep fault characteristics is located, where t is the time step. For the original minority class dataset, The original minority class dataset after continuous wavelet transform. Here, j represents the size of the sliding window, and j represents the starting position of the sliding window. Where is the length of a single sample in the original minority class dataset during the partitioning, and k is the index of the deep fault feature. This refers to the portion where the fault signal energy is concentrated, i.e., the characteristics of a deep fault. Step 2.2 The specific method for obtaining the reconstructed minority class samples is as follows: Calculate the Euclidean distance between the two-dimensional original majority class signal and the signal containing the original deep fault features, and find the K+1 nearest neighbors of the signal containing the original deep fault features on the original majority class signal. The calculation formula is as follows: ； ； in, The set of data points representing the original two-dimensional majority class signal. Signals containing original deep fault characteristics The set, yes exist The A neighbor, Represents Euclidean distance. The (K+1)th smallest nearest neighbor value.

4. The intelligent real-time diagnosis method for bearing imbalance impact faults by wavelet phase space oversampling reconstruction according to claim 1, characterized in that, Step 3, the dynamic threshold-sensing focus balance loss function DTAFBLoss, is based on prediction probability. and adaptive threshold Divide the loss function into a hard sample branch loss function. and simple sample branch loss function The expression is as follows: ； ； in, It is the loss amplification factor for hard samples, which controls the degree of loss amplification for hard samples and adjusts the influence of important local features of the time series captured by the ODLC-SMOTE algorithm on classification; It is the focus loss adjustment factor, which controls the adjustment strength of focus loss; yes Time step category The predicted probability, It is the predicted probability distribution of the model. It is a minimal constant that prevents the derivative from being zero; Combining class-balanced weights and the focus loss branch, the final loss function is defined as: ； Among them, 1 / Is it the true category of sample i? Category balance weights, It is the predicted probability of sample i. It is the adaptive threshold for sample i. C is the frequency of deep fault features, and C is the number of samples.

5. The intelligent real-time diagnosis method for bearing imbalance impact faults by wavelet phase space oversampling reconstruction according to claim 1, characterized in that, The real-time diagnostic test process in step 3 is as follows: First, obtain the initial weights. Covariance Matrix ; Construct an augmented matrix by vertically stacking the historical output matrix and the current output matrix. The corresponding tag augmentation is ; Calculate the gain matrix by combining historical and current data. At the same time, a regularization term is introduced. : ； Through forgetting factors Control the decay rate of historical information, subtract Uncertainty in the estimation of the compression parameters; the covariance matrix is ​​updated as follows: ； weight matrix Updated to: ; in, For historical output matrix, This is the current output matrix. It is an augmented matrix that is transposed.

Images