Intelligent real-time diagnosis method for unbalanced impact fault of wavelet phase space oversampling reconstruction bearing

Abstract

The invention discloses an intelligent real-time diagnosis method for an unbalanced impact fault of a wavelet phase space oversampling reconstruction bearing, and the method comprises the steps: constructing an oversampling method index positioning fault signal energy concentration part of a depth positioning data point in phase space reconstruction continuous wavelet transformation; the influence of noise and edge samples is reduced by controlling the distribution of the synthesized new samples; a two-stage feature learning random vector function link network fusing random projection and adaptive refining is provided, a dynamic threshold perception focus balance loss function and category balance weight are designed to reduce loss contribution of easy-to-classify samples, and meanwhile, adaptive threshold adjustment is performed to balance local noise and global statistics, so that the model pays more attention to difficult samples, and the classification accuracy is improved. And finally, applying small-batch incremental learning to online diagnosis to realize real-time fault diagnosis. The method has good diagnosis rate and stability, and is suitable for processing the problems of fault sample misdiagnosis, high missed diagnosis rate and low overall diagnosis precision of the model under the unbalanced data set.

Description

Technical Field

[0001] The present invention relates to the technologies of intelligent fault detection and deep learning, and particularly to an intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults through wavelet phase space oversampling. Background Art

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

[0003] Bearing is one of the most critical and fault-prone components in mechanical equipment, and its operating state directly affects the stability and safety of the equipment. The real-time diagnosis technology combines the Internet of Things and big data analysis to realize the remote collection and analysis of bearing operation data. On the one hand, real-time fault diagnosis can timely detect potential faults of bearings, quickly locate the cause of the faults, shorten the maintenance time and reduce the equipment downtime loss, avoid equipment downtime or even catastrophic accidents caused by bearing failure, so as to ensure the safe operation of the equipment and improve the overall operation efficiency of the equipment. On the other hand, bearings may face complex working conditions in actual operation, such as high load, high temperature, high humidity, etc. The real-time diagnosis technology can adapt to these complex environments and timely detect faults caused by changes in working conditions to ensure the stable operation of the equipment under harsh conditions.

[0004] With the progress of technology, bearing fault diagnosis is developing towards the direction of intelligence, high precision, and high efficiency. The application of real-time diagnosis technology not only improves the accuracy of diagnosis but also reduces the possibility of manual intervention and misjudgment, providing strong support for the intelligent management of industrial equipment.

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

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

[0007] Object of the Invention: The object of the present invention is to solve the deficiencies existing in the prior art and provide an intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling.

[0008] Technical Solution: An intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling according to the present invention includes 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 signals in the healthy state as the majority class, and set the signals of inner ring faults, outer ring faults, and rolling element faults 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 an oversampling method ODLC-SMOTE for deep-locating data points in the phase space reconstruction continuous wavelet transform to generate pseudo-fault samples for the fault state signals of the minority class.

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

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

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

[0014] Divide the original majority class samples and the new minority class fault dataset into an unbalanced training set, a balanced test set, and a validation set.

[0015] Step 3: Construct a two-stage feature learning random vector functional link network TSFL-RVFL that combines random projection and adaptive refinement, and obtain the fault diagnosis results in the offline stage through the two-stage feature learning random vector functional link network TSFL-RVFL.

[0016] The two-stage feature learning random vector functional link network TSFL-RVFL is trained using the training set dataset. During the training process, the dynamic threshold-aware focus balance loss function DTAFBLoss is used for optimization and adjustment. While the class balance weight reduces the loss contribution of easy-to-classify samples, the adaptive threshold is adjusted to balance local noise and global statistics, making the model pay more attention to difficult samples;

[0017] Adjust appropriate model parameters to improve the diagnostic results and obtain a fixed-parameter model. Input the validation set and the test set into the fixed-parameter model for diagnostic verification and real-time diagnostic testing in the offline stage.

[0018] Further, the number of samples in the original majority-class signal dataset in step 1 is more than that in the original minority-class signal dataset.

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

[0020] ;

[0021] where i refers to the i-th original minority-class signal, is the signal containing the original deep fault feature located, t is the time step, is the original minority-class dataset, is the original minority-class dataset after continuous wavelet transform, is the sliding window size, j is the starting position of the sliding window, is the length of a single sample in the original minority-class dataset during partitioning, k is the index of the deep fault feature, is the part where the energy of the fault signal is concentrated;

[0022] The specific method for obtaining the reconstructed minority-class samples in step 2.2 is:

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

[0024]

[0025]

[0026] where, represents the set of data points of the two-dimensional original majority-class signal, represents the set of data points of the two-dimensional original minority-class signal 's set, is at 's K nearest neighbors, represents the Euclidean distance, is the (K+1)-th smallest nearest neighbor value.

[0027] Furthermore, the two-stage feature learning random vector functional link network TSFL-RVFL includes a random 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 units h, the input data is projected into a high-dimensional space through a fixed random mapping, inheriting the fast calculation advantage of RVFL;

[0029] In the adaptive refinement stage, feature refinement is achieved by cascading trainable fully connected layers, the activation function is used to enhance the non-linear representation ability, and a regularization constraint is introduced.

[0030] Furthermore, the step 3 dynamic threshold-aware focal balance loss function DTAFBLoss is based on the prediction probability and the adaptive threshold , and the loss function is divided into two branches: the hard sample branch loss function and the easy sample branch loss function , which are defined as follows:

[0031] ;

[0032] ;

[0033] where, is the hard sample loss amplification coefficient, which controls the degree of hard sample loss amplification and adjusts the influence degree of important local features of the time series captured by the ODLC-SMOTE algorithm on classification; is the focal loss adjustment factor, which controls the adjustment intensity of the focal loss; is the prediction probability of the time step category , is the prediction probability distribution of the model, is a very small constant to prevent the derivative from being 0;

[0034] Combining the class balance weight and the focal loss branch, the final loss function is defined as:

[0035]

[0036] where, 1 / is the class balance weight of the true class of sample i (i.e., the original minority class signal), is the prediction probability of sample i, is the adaptive threshold of sample i, is the frequency at which the deep fault feature appears, and C is the number of samples.

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

[0038] First, obtain the initial weights and the covariance matrix ;

[0039] Vertically stack the historical output matrix and the current output matrix to construct an augmented matrix , and the corresponding label is augmented to ;

[0040] Combine historical data with current data to calculate the gain matrix , and at the same time introduce a regularization term :

[0041] ;

[0042] Control the attenuation rate of historical information through the forgetting factor , and subtract to compress the uncertainty of parameter estimation; the covariance matrix is updated as:

[0043] ;

[0044] The weight matrix is updated as:

[0045] ;

[0046] where, is the historical output matrix, is the current output matrix, is the transposed augmented matrix;

[0047] Through the above process, the incremental learning in the real-time diagnostic 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 the traditional signal processing method, the present invention directly focuses on the impact part of the bearing fault, and uses healthy samples as the majority class to reconstruct the minority class fault samples, reducing the use of computing resources and being more effective in processing fault signals.

[0050] (2), The present invention pays attention to both the imbalance problem and the classification problem of easy and difficult samples during the model training process, can balance the local noise and global statistics of the model training, and makes the model pay more attention to difficult samples.

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

[0052] (4) The present invention applies mini-batch incremental learning. The incremental learning in the real-time diagnosis process can effectively combine historical data and current data, while maintaining numerical stability and computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0054] FIG. 2 is a two-dimensional effect diagram of reconstructing the impact fault of the inner and outer rings of the bearing in the embodiment.

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

[0056] Figure 4 is a schematic diagram of the function of the amplification factor of the hard sample loss being 5, the adaptive threshold weight being 0.5, and the extremely small constant being 0.000001 in the embodiment.

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

[0058] Figure 6 is a G-mean result diagram of the real-time diagnosis process of the model in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

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

[0060] Considering the data imbalance problem existing in actual engineering, the present invention constructs an oversampling method for indexing and locating the part with concentrated fault signal energy in the depth positioning of data points in the phase space reconstruction continuous wavelet transform, and reduces the influence of noise and edge samples by controlling the distribution of synthesized new samples. Aiming at the problem of fault signal feature coupling, a two-stage feature learning random vector functional link network integrating random projection and adaptive refinement is proposed. A dynamic threshold-aware focus balance loss function is designed for the problem of insufficient network learning ability in unbalanced tasks. While the class balance weight reduces the loss contribution of easy-to-classify samples, the adaptive threshold is adjusted to balance local noise and global statistics, making the model pay more attention to hard samples. Finally, mini-batch incremental learning is applied to online diagnosis to achieve real-time fault diagnosis. The online diagnosis results show that the proposed method has good diagnostic rate and stability, and is suitable for dealing with the problems of high misdiagnosis and missed diagnosis rates of fault samples and low overall diagnostic accuracy of the model under unbalanced data sets.

[0061] Such asFigure 1 and Figure 3 As shown in Figure 3 , the intelligent real-time diagnosis method for bearing unbalance impact faults based on wavelet phase space oversampling reconstruction in this embodiment includes the following steps:

[0062] Step 1: Obtain the vibration signal data of the health and fault states of different parts of the bearing under different working conditions. Set the vibration signals in the healthy state as the majority class, and set the signals of inner ring faults, outer ring faults, and rolling element faults 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 an oversampling method ODLC-SMOTE for deep positioning data points in the continuous wavelet transform of phase space reconstruction to generate pseudo-fault samples for the fault state signals of the minority class;

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

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

[0066] Step 2.3: Compress the obtained two-dimensional reconstructed minority class samples to form one-dimensional fault signals containing pseudo-fault samples, and combine them with the signals containing the original deep fault features to form a new minority class fault dataset;

[0067] Divide the original majority class samples and the new minority class fault dataset into an unbalanced training set, a balanced test set, and a validation set;

[0068] Step 3: Construct a two-stage feature learning random vector functional link network TSFL-RVFL that fuses random projection and adaptive refinement, and obtain the fault diagnosis results in the offline stage through the two-stage feature learning random vector functional link network TSFL-RVFL;

[0069] Use the training set dataset to train the two-stage feature learning random vector functional link network TSFL-RVFL. During the training process, use the dynamic threshold-aware focal balance loss function DTAFBLoss for optimization and adjustment. While reducing the loss contribution of easy-to-classify samples with the class balance weight, adaptively adjust the threshold to balance local noise and global statistics, so that the model pays more attention to difficult samples;

[0070] Adjust the appropriate model parameters to improve the diagnostic results to obtain a fixed-parameter model, and input the validation set and the test set into the fixed-parameter model for diagnostic verification and real-time diagnostic testing in the offline stage.

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

[0072] In step 2.1 of this embodiment, the formula for maximizing the mean sliding window index is as follows:

[0073] ;

[0074] Among them, i refers to the i-th original minority-class signal, is the signal containing the original depth fault feature located, t is the time step, is the original minority-class dataset, is the original minority-class dataset after continuous wavelet transform, is the sliding window size, j is the starting position of the sliding window, is the length of a single sample in the original minority-class dataset during partitioning, k is the index of the depth fault feature, is the part where the fault signal energy is concentrated;

[0075] The specific method for obtaining the reconstructed minority-class samples in step 2.2 is:

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

[0077] ;

[0078] ;

[0079] Among them, represents the set of data points of the two-dimensional original majority-class signal, represents the data point of the two-dimensional original minority-class signal is in the th nearest neighbor of, represents the Euclidean distance, is the (K + 1)-th smallest nearest neighbor value.

[0080] In this embodiment, the two-stage feature learning random vector functional link network TSFL-RVFL includes a random 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 calculation advantage of RVFL;

[0082] In the adaptive refinement stage, feature refinement is achieved by cascading trainable fully connected layers, the activation function is used to enhance the non-linear representation ability, and regularization constraints are introduced.

[0083] In step 3 of this embodiment, the dynamic threshold-aware focus balance loss function DTAFBLoss is based on the prediction probability and the adaptive threshold , and the loss function is divided into two branches: the loss function of the hard sample branch and the loss function of the easy sample branch , which are defined as follows:

[0084] ;

[0085] ;

[0086] where is the hard sample loss amplification coefficient, which controls the degree of hard sample loss amplification and adjusts the influence degree of important local features of the time series captured by the ODLC-SMOTE algorithm on classification; is the focus loss adjustment factor, which controls the adjustment intensity of the focus loss; is the prediction probability of the time step category , is the prediction probability distribution of the model, is a very small constant to prevent the derivative from being 0;

[0087] Combining the class balance weight and the focus loss branch, the final loss function is defined as:

[0088] ;

[0089] where, 1 / is the class balance weight of the true class of sample i, is the prediction probability of sample i, is the adaptive threshold of sample i, is the frequency of the occurrence of the deep fault feature, and C is the number of samples.

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

[0091] First, obtain the initial weights and the covariance matrix ;

[0092] Vertically stack the historical output matrix and the current output matrix to construct an augmented matrix , and the corresponding label is augmented as ;

[0093] Combine historical data and current data to calculate the gain matrix , and at the same time introduce a regularization term :

[0094] ;

[0095] Through the forgetting factor control the attenuation rate of historical information, and subtract the uncertainty of the item compression parameter estimation; the covariance matrix is updated as:

[0096] ;

[0097] The weight matrix is updated as: ;

[0098] where is the historical output matrix, is the current output matrix, is the transposed augmented matrix;

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

[0100] To further verify the feasibility and technical effects of the technical solution 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 drive system of rail transit train bogies.

[0101] The drive chain of the test bench is driven by a three-phase asynchronous AC motor, and the motor speed can be controlled by a frequency converter. Among them, the motor bearing model is SKF 6205-2RSH; the drive gear support bearing model is HRB 32305; the axle box bearing model is HRB352213, and the fault characteristic frequencies (rotation frequency f) of the inner ring, outer ring and rolling elements are shown in Table 1. The sensors on the test bench highly reproduce the real measurement points of the train.

[0102] Table 1 Bearing fault characteristic frequency table

[0103] In this embodiment, fault diagnosis is carried out on the inner ring fault of the gearbox bearing, the rolling element fault of the gearbox bearing, the inner ring fault of the axle box bearing, the outer ring fault of the axle box bearing, the rolling element fault of the axle box bearing, and the healthy state under the condition that the motor speed is 1200 r / min and the lateral load is 10 KN.

[0104] The two-stage feature learning random vector functional link network model TSFL-RVFL of the present invention randomly initializes the weight matrix and bias vector, which is used to quickly construct the hidden layer, linearly transform the input signal and apply the ReLU activation. After grid search, the model contains 3 fully connected layers, with 64, 32, and 16 neurons respectively. Each fully connected layer uses the ReLU activation function, and L2 regularization and randomly discarding 10% of the neurons are applied to prevent overfitting. To illustrate the effectiveness of the diagnosis method, the proportions of the diagnosis training set, validation set, and test set are set to 0.6:0.2:0.2. The training set adopts the unbalanced proportion of the inner ring fault of the gearbox bearing, the rolling element fault of the gearbox bearing, the inner ring fault of the axle box bearing, the outer ring fault of the axle box bearing, the rolling element fault of the axle box bearing, and the healthy state as 1:1:1:1:1:10, and 300, 600, and 900 algorithm resampling data points are sampled respectively to verify the generalization performance of the method. The diagnosis process is trained for 500 rounds and the experiment is repeated. The accuracy, balanced accuracy, F1 score, G-mean value, and Matthews correlation coefficient (MMC) of the validation set in the offline diagnosis are shown in Table 2.

[0105] Table 2 Offline diagnosis results

[0106] The accuracy rate represents the proportion of the number of correctly classified samples in all samples. The balanced accuracy rate avoids the overestimated performance of a single index for unbalanced data sets and alleviates the impact of the unbalanced situation in multi-classification tasks.

[0107] It can be seen from the offline diagnosis results in Table 2 that when the network model of the present invention processes the data set with an unbalanced degree of 1:1:1:1:1:10, the highest accuracy rate and balanced accuracy rate are 94.14±0.05%, reaching a relatively high accuracy level. In multi-classification problems, G-mean can reflect the geometric mean of the classification accuracy of each type of sample. The highest G-mean of the proposed method is 93.99±0.08%, which is only 0.15% lower than the accuracy rate, indicating that the network model of the present invention can handle unbalanced classification problems. MCC is a comprehensive evaluation index, which more comprehensively considers the performance of all aspects of the classification model and provides a more comprehensive perspective for comprehensively evaluating the quality of the model. The MCC value of the proposed method is not less than 0.918±0.03, indicating that the comprehensive performance of the model is good.

[0108] The blue dots in Figure 2 represent healthy samples of the majority class, the red dots represent faulty samples of the minority class, and the green dots represent newly generated faulty samples. It can be seen from Figure 2 that the fault impact data points located by the oversampling method ODLC-SMOTE of the present invention are more obvious than the healthy sample data points, and the generated faulty samples are distributed consistently with the faulty samples of the minority class, without losing the fault characteristics. (a) in Figure 2 is the inner ring fault effect diagram in the embodiment, and (b) in Figure 2 is the outer ring fault effect diagram.

[0109] From Figure 4 it can be seen that the gray vertical dotted line marks the segmentation threshold of the loss function. The difficult sample loss amplification factor and the adaptive threshold weight of the DTAFBLoss of the present 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 prediction probability is less than 0.5, reflecting that the model pays more attention to difficult classification samples, ensuring that the model assigns higher weights to faulty samples with low confidence. The loss drops rapidly in the region where the prediction probability is greater than 0.5, reflecting that the model has stronger suppression of easy samples, suppressing the weights of normal samples and preventing the model from being dominated by a large number of easy samples.

[0110] Figure 5 In the decision boundary diagram of the real-time diagnosis process of the model of, different colors represent faults with different labels, the red crosses represent the data predicted incorrectly, and the black dots represent the corresponding correct labels. From Figure 3 it can be seen that, firstly, the decision boundary is clear rather than confused, indicating that the present invention can effectively distinguish the distributions of different fault categories in the feature space. Secondly, the predicted incorrect labels of the model's real-time diagnosis are not concentrated, and there is no frequent fluctuation of the predicted incorrect labels, indicating that the present invention is not sensitive to the interference of noise or instantaneous anomalies and has robustness.

[0111] From Figure 6 the G-mean result diagram of the real-time diagnosis process of, it can be seen that as the number of data points in the online data stream increases, the real-time diagnosis of the model tends to be stable. 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 the real-time diagnosis is significantly better, indicating that the present invention can adapt to the data distribution of new samples in real-time diagnosis, and the higher the sampling points, the stronger this adaptability.

[0112] In summary, the present invention can directly focus on the impact part of bearing faults, extract bearing impact faults and reconstruct minority class fault samples according to majority class healthy samples, and is more effective in processing fault signals. The method of the present invention pays attention to both the imbalance problem and the classification problem of easy and difficult samples during the model training process, can balance the local noise and global statistics of model training, and makes the model pay more attention to difficult samples. The mini-batch incremental learning in the real-time diagnosis process can effectively combine historical data and current data, while maintaining numerical stability and computational efficiency.

Claims

1. An intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling, characterized in that, It includes the following steps: Step 1: Obtain the vibration signal data of the health and fault states of different parts of the bearing under different working conditions. Set the vibration signals in the healthy state as the majority class, and set the signals of inner ring faults, outer ring faults, and rolling element faults 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: Process the fault state signals of the minority class based on the oversampling method ODLC-SMOTE for deep positioning data points in the phase space reconstruction continuous wavelet transform to generate pseudo-fault samples; Step 2.1: Use the mean-maximized sliding window for the original minority-class signal samples to index the depth fault features after continuous wavelet transform, and locate and extract the signals containing the original depth fault features , and locate and extract the signals containing the original depth fault features ; Step 2.2: Reconstruct the one-dimensional original majority class signals and the original minority class signals into a two-dimensional space through the phase space, expand the fault state signals of the original minority class, and then generate pseudo-fault samples, which become the reconstructed minority class samples, and check the sample distance and sample generation quality; Step 2.3: Compress the obtained two-dimensional minority class samples to form one-dimensional fault signals containing pseudo-fault samples and combine them with the signals containing the original deep fault features to form a new minority class fault dataset; Divide the original majority class samples and the new minority class fault dataset into an unbalanced training set, a balanced test set, and a validation set; Step 3: Construct a two-stage feature learning random vector functional link network TSFL-RVFL that combines random projection and adaptive refinement, and obtain the fault diagnosis results in the offline stage through the two-stage feature learning random vector functional link network TSFL-RVFL; Use the training set dataset to train the two-stage feature learning random vector functional link network TSFL-RVFL. During the training process, use the dynamic threshold-aware focus balance loss function DTAFBLoss for optimization and adjustment. While reducing the loss contribution of easy-to-classify samples by the class balance weight, adaptively adjust the threshold to balance local noise and global statistics, so that the model pays more attention to difficult samples; Adjust appropriate model parameters to improve the diagnosis results to obtain a fixed-parameter model, and input the validation set and the test set into the fixed-parameter model for offline stage diagnosis verification and real-time diagnosis test.

2. The intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling according to claim 1, characterized in that In the above Step 1, the number of samples in the original majority class signal dataset is more than that in the original minority class signal dataset.

3. The intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling according to claim 1, wherein In the above Step 2.1, the formula for the mean-maximized sliding window index is as follows: ； Among them, i refers to the i th original minority class signal, is the signal with the original deep fault feature located, t is the time step, is the original minority class data set, is the original minority class data set after continuous wavelet transform, is the sliding window size, j is the starting position of the sliding window, is the length of a single sample in the original minority class data set during partitioning, k is the index of the deep fault feature, is the part where the energy of the fault signal is concentrated, that is, the deep fault feature; The specific method for obtaining the reconstructed minority class samples in Step 2.2 is: Calculate the Euclidean distance between the two-dimensional original majority class 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 signals. The calculation formula is: ； ； Among them, represents a set of data points of the two-dimensional original majority class signal, represents the data points of the two-dimensional original minority class signal set, is at the th nearest neighbor, represents the Euclidean distance, is the (K + 1)-th smallest nearest neighbor value.

4. The intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling according to claim 1, characterized in that The two-stage feature learning random vector functional link network TSFL-RVFL includes a random projection layer, a deep feature enhancement layer, a regularization mechanism, and a classification output layer; In the random projection stage, based on the input dimension d and the number of hidden layer units h, project the input data into a high-dimensional space through a fixed random mapping; In the adaptive refinement stage, cascade and train the fully connected layer, and also use activation functions and introduce regularization constraints.

5. The intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling according to claim 1, wherein, The dynamic threshold-aware focus balance loss function DTAFBLoss in step 3 is based on the prediction probability and the adaptive threshold to divide the loss function into a hard sample branch loss function and an easy sample branch loss function , and the expression is as follows: ； ； Among them, is the amplification coefficient of the hard sample loss, which controls the amplification degree of the hard sample loss and adjusts the influence degree of the important local features of the time series captured by the ODLC-SMOTE algorithm on classification; is the focal loss adjustment factor, which controls the adjustment intensity of the focal loss; is the predicted probability of the time step category is the predicted probability distribution of the model, is a very small constant to prevent the derivative from being 0; Combined with the class balance weight and the focal loss branch, the final loss function is defined as: ； Among them, 1 / is the true class of sample i is the class balance weight, is the predicted probability of sample i, is the adaptive threshold of sample i, is the frequency at which the deep fault feature appears, and C is the number of samples.

6. The intelligent real-time diagnosis method for reconstructing bearing unbalance impact faults by wavelet phase space oversampling according to claim 1, wherein The real-time diagnosis test process in the above Step 3 is: First, obtain the initial weights and the covariance matrix ; Vertically stack the historical output matrix and the current output matrix to construct an augmented matrix , and the corresponding label is augmented as ; Calculate the gain matrix by combining historical data and current data , and introduce a regularization term : ； By means of a forgetting factor to control the decay rate of historical information and subtract the uncertainty in the estimation of the compression parameter; the covariance matrix is updated as follows: ； Weight matrix Updated to: ; Among them, is the historical output matrix, is the current output matrix, is the transposed augmented matrix.

Images