Feature projection contrast distillation and memory playback incremental learning bearing fault diagnosis method

Abstract

The invention discloses a bearing fault diagnosis method based on feature projection contrast distillation and memory playback incremental learning, and relates to the technical field of fault diagnosis, and the method comprises the specific steps: firstly carrying out data processing and initial model training, converting a bearing vibration signal into an RGB time-frequency image, and carrying out the training of an initial model; training an initial model by using a ResNet18 network and a dynamic extensible classifier, and constructing a memory bank; incremental training is carried out, and a student model is trained by comparing distillation loss through feature projection; and finally performing memory playback, updating the memory bank, and repeating the steps to complete bearing fault diagnosis of a new increment task when a new task arrives. According to the method, by adopting feature projection comparison distillation and dynamic extensible classifier design, old fault knowledge is effectively reserved, and new fault categories are flexibly adapted; and meanwhile, a distillation loss weight strategy is dynamically adjusted by utilizing memory bank sample balance updating and cosine annealing, so that the incremental training stability and the sample utilization efficiency are improved, and the diagnosis precision consistency is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically to a bearing fault diagnosis method based on feature projection comparison distillation and memory playback incremental learning. Background Technology

[0002] In the field of industrial equipment operation and maintenance, bearings are core transmission components, and the timeliness and accuracy of their fault diagnosis directly affect the stable operation of the production system. As industrial scenarios become increasingly complex, bearing fault modes are becoming more diverse and dynamically emerging. Traditional fault diagnosis models are mostly trained based on fixed fault categories. When a new fault type appears, all fault data must be collected again to retrain the model, which not only consumes a lot of time and data acquisition costs, but also makes it difficult to adapt to the needs of dynamic expansion of fault types in actual production. Therefore, bearing fault diagnosis technology with incremental learning capabilities has become a research focus in this field.

[0003] Current mainstream incremental fault diagnosis technologies for bearings have several limitations. On the one hand, traditional incremental learning models tend to significantly forget their existing fault identification capabilities when learning new fault categories, leading to a marked decline in the diagnostic accuracy of older fault categories. On the other hand, traditional classifier structures are difficult to flexibly adapt to new categories, often requiring model restructuring and adjustments, which increases the complexity of technology implementation. Furthermore, most solutions do not dynamically optimize the sample distribution and loss weights during training, which can easily lead to training imbalance and ultimately affect the stability and reliability of the overall diagnostic performance. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a bearing fault diagnosis method based on feature projection contrastive distillation and memory playback incremental learning. This method proposes an innovative approach for bearing fault diagnosis through a feature projection contrastive distillation and memory playback incremental learning strategy. First, the vibration signal is converted into an RGB time-frequency image, and an initial model is trained using a ResNet18 network and a dynamic classifier. In the incremental stage, a student model is trained using feature projection contrastive distillation loss, retaining old knowledge while learning new categories. The memory playback mechanism updates the memory bank, ensuring sample balance. This scheme improves the stability and diagnostic accuracy of incremental learning and adapts to dynamically changing industrial environments.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a bearing fault diagnosis method based on feature projection comparison distillation and memory playback incremental learning, the specific steps of which are as follows:

[0006] S1 Data Processing and Initial Model Training Stage: The collected bearing vibration signal is converted into a 224×224 pixel RGB time-frequency image through wavelet transform; a ResNet18 network is used as a feature extractor, and the output of the feature extractor is connected to a dynamic scalable classifier. The initial model is trained using initial fault category samples; a memory bank with limited capacity is constructed, and a preset number of samples are randomly selected from the initial samples and stored in the memory bank.

[0007] S2 Incremental Training Phase: The model trained in the previous round is used as the teacher model, and the model to be trained in this round is used as the student model. The 512-dimensional high-dimensional features output by the feature extractors of the teacher model and student model are projected to 128-dimensional low-dimensional features through the feature projection head and normalized. The feature similarity matrix between the teacher model and student model is calculated, and the feature projection comparison distillation loss is constructed by using KL divergence in the row and column directions. The total loss function is formed by combining the cross-entropy classification loss. The weights of the distillation loss are dynamically adjusted using the cosine annealing method. The student model is trained by combining new task samples with historical samples in the memory bank, and the weight matrix and bias vector of the dynamically expandable classifier are updated.

[0008] S3 Memory Replay Stage: After the incremental training is completed, a preset number of samples are randomly selected from the new task samples to update the memory bank, keeping the capacity of the memory bank constant; when a new incremental task arrives, steps S2 to S3 are repeated to complete the bearing fault diagnosis corresponding to the new incremental task.

[0009] Furthermore, the dynamically scalable classifier targets the first... The mathematical expression for each task is: ,in, This is the sequence number of the currently executing incremental task. For the first The classifier output results for each task Let be the d-dimensional feature vector output by the feature extractor, where d is the dimension of the output feature vector. For the first The classifier weight matrix corresponding to each task is: , For the first The number of historical fault categories learned before each task is executed. For the first The number of new fault categories added per task For the first The classifier bias vector corresponding to each task; the weight matrix By splicing together historical weight matrices With the newly added category weight matrix To obtain, that is: ,in, This is the classifier weight matrix corresponding to the historical fault categories. The classifier weight matrix corresponding to the newly added fault category; the bias vector By splicing historical bias vectors With the newly added class bias vector To obtain, that is: ,in, This represents the classifier bias vector corresponding to the historical fault categories. This is the classifier bias vector corresponding to the newly added fault category.

[0010] Furthermore, the structure of the feature projection head is as follows: it sequentially includes a linear layer with an input dimension of 512 and an output dimension of 256, a ReLU activation layer with an input dimension of 256 and an output dimension of 256, a linear layer with an input dimension of 256 and an output dimension of 128, and a LayerNorm normalization layer with an input dimension of 128 and an output dimension of 128.

[0011] Furthermore, the calculation process of the feature similarity matrix is ​​as follows: Let the feature matrix output by the feature projection head be H, and the output dimension of the feature projection head be... The temperature coefficient is Then the feature similarity matrix of the teacher model Feature similarity matrix with student model Calculated using the following formula: ,in, The first element in the similarity matrix Line number Column elements, The first element in the characteristic matrix H The element in row d and column d.

[0012] Furthermore, the feature projection contrasts with distillation loss. The calculation process is as follows: calculate the similarity matrix row directions respectively. Divergence loss: and column direction Divergence loss: The total distillation loss is obtained by averaging the row and column losses: ,in, For the row direction of the similarity matrix Divergence loss, Let be the row vector of the feature similarity matrix of the student model. Let the row vectors of the teacher model feature similarity matrix be... For the column direction of the similarity matrix Divergence loss, This is the column vector of the student model feature similarity matrix. This is the column vector of the teacher model feature similarity matrix. The total loss of distillation is compared with the feature projection.

[0013] Furthermore, the formula for calculating the total loss function is as follows: ,in, The total loss during model training. Cross-entropy classification loss, As a weight for distillation losses, The total loss of distillation is compared with the feature projection.

[0014] Furthermore, the expression for adjusting the distillation loss weight using the cosine annealing method is as follows: ,in, For the first Distillation loss weights corresponding to each training epoch. This is the current training epoch number. This is the final value for the distillation loss weight. This is the initial value for the distillation loss weight. The total number of epochs trained for a single task. It is a cosine function.

[0015] Furthermore, during the memory bank update process, a sample balancing strategy is employed to ensure a balanced ratio between the number of new task samples and historical samples in the training set; simultaneously, label space remapping is performed, merging the label space corresponding to the new samples into the original label space to obtain the updated label space. ,in, For the updated tag space, For the label space corresponding to historical fault categories, This is the label space corresponding to the fault category of the new task.

[0016] Furthermore, the vibration signal processing procedure is as follows: the collected vibration signal is segmented, with each segment having a sample length of 1024. The segmented samples are then subjected to either overlapping sampling or direct sampling based on their data characteristics, and finally converted into a time-frequency image through wavelet transform.

[0017] Furthermore, the capacity of the memory bank is a preset fixed value, and a replacement update strategy is adopted during updates, selecting samples from new tasks to replace some historical samples in the memory bank to keep the capacity of the memory bank constant.

[0018] Compared with existing technologies, this feature projection contrast distillation and memory playback incremental learning bearing fault diagnosis method has the following advantages:

[0019] I. This invention, through the design of a feature projection contrast distillation combined with a dynamic scalable classifier, effectively retains existing fault knowledge and flexibly adapts to the recognition requirements of newly added fault categories in the incremental learning scenario of bearing fault diagnosis. Traditional incremental diagnosis methods often lose the ability to identify old faults due to model updates, while this scheme strengthens the memory effect of historical fault features by leveraging the feature similarity constraints between the teacher model and the student model. At the same time, the dynamic classifier, through the concatenation of weights and biases, can quickly expand the categories without reconstructing the model, significantly improving the adaptability and iterative efficiency of the diagnostic model in multi-fault scenarios.

[0020] Second, this invention improves the stability and sample utilization efficiency of incremental training through a strategy of balancing and updating the sample memory and dynamically adjusting the distillation loss weights using cosine annealing. The memory is updated after each incremental task and maintains a constant size, continuously retaining representative historical fault samples and avoiding imbalance in the distribution of training data. The dynamic and gradual change of the distillation loss weights allows the model to focus on learning historical knowledge in the early stages of training and on newly added fault features in the later stages, effectively balancing the learning priority of old and new knowledge and further ensuring the consistency of diagnostic accuracy for different fault categories.

[0021] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the overall fault diagnosis method;

[0024] Figure 2 For feature projection comparison of distillation process flow chart;

[0025] Figure 3 This is a flowchart of the dynamic scalable classifier and memory update process. Detailed Implementation

[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0027] Example 1:

[0028] This embodiment is applied to the factory inspection of automotive engine bearings after production and the operation and maintenance diagnosis of motor bearings in vehicles in use. In this scenario, the bearings need to be adapted to the operating requirements of engines with different displacements. In the initial stage, it is necessary to accurately identify three core fault types: bearing assembly deviation, ball wear, and inner and outer ring deformation. At the same time, it is necessary to deal with the new fault identification requirements brought about by subsequent production process optimization or vehicle model upgrades.

[0029] The overall process of this invention is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0030] In the data processing and initial model training phase, high-precision piezoelectric vibration sensors were installed at key inspection stations on the bearing manufacturing production line and near in-use vehicle engine bearings. The sensor sampling frequency was set according to the engine's rated speed to ensure the capture of vibration characteristics under different operating conditions. For bearings inspected on the production line, vibration signals were collected during 30 seconds of no-load operation. For in-use vehicle bearings, vibration signals were collected under multiple operating conditions, including normal driving, idling, and acceleration, to avoid feature loss due to single operating conditions. The collected vibration signals were preprocessed to remove redundant information such as power interference and mechanical noise. The signals were then segmented according to a preset length, ensuring each segment completely contained the features of at least one bearing rotation cycle. Wavelet transform was then used to convert the segmented signals into RGB time-frequency images, which clearly show the time-domain and frequency-domain feature correlation of the vibration signals, facilitating fault mode identification by the feature extraction network. A ResNet18 network was selected as the feature extractor, as it is efficient and has good generalization capabilities in image feature extraction. Its output is connected to a dynamically scalable classifier, with initial weights randomly initialized and a bias vector set to zero. The dynamically scalable classifier is used for the first... The mathematical expression for each task is: ,in, This is the sequence number of the currently executing incremental task. For the first The classifier output results for each task Let be the d-dimensional feature vector output by the feature extractor, where d is the dimension of the output feature vector. For the first The classifier weight matrix corresponding to each task is: , For the first The number of historical fault categories learned before each task is executed. For the first The number of new fault categories added per task For the first The classifier bias vector corresponding to each task; the weight matrix By splicing together historical weight matrices With the newly added category weight matrix To obtain, that is: ,in, This is the classifier weight matrix corresponding to the historical fault categories. The classifier weight matrix corresponding to the newly added fault category; the bias vector By splicing historical bias vectors With the newly added class bias vector To obtain, that is: ,in, This represents the classifier bias vector corresponding to the historical fault categories. This is the classifier bias vector corresponding to the newly added fault category. Using sample data of three fault types and normal states accumulated over three years from the production line, covering bearing samples from different production batches and of different materials, a training dataset is constructed. A stochastic gradient descent optimizer is used to train the initial model. During training, the fault identification accuracy of the model on the validation set is monitored. Training of the initial model is stopped when the accuracy does not significantly improve for 10 consecutive training cycles. Simultaneously, a fixed-capacity memory is constructed. Selection criteria include the clarity of fault features, completeness of operating condition coverage, and data quality stability. A certain number of representative samples are selected from each fault type and normal state sample and stored in the memory to ensure that the memory comprehensively covers all fault modes and operating state characteristics in the initial stage.

[0031] During the incremental training phase, with the upgrading of automotive engines, a new type of failure, bearing high-temperature sintering, has emerged. This failure is caused by insufficient optimization of the engine cooling system and is prone to occur under high-load operating conditions. At this point, the previously trained initial model is used as the teacher model, which already possesses mature identification capabilities for three types of failures: assembly deviation, ball wear, and inner / outer ring deformation. A newly built model with a structure completely identical to the teacher model is used as the student model, and its initial parameters are also obtained through random initialization. The high-dimensional features output by the feature extractors of both models are processed using a feature projection head. The multi-layered structure of the feature projection head effectively compresses feature dimensions while retaining key information. After converting high-dimensional features to low-dimensional features, normalization is performed to eliminate dimensional differences between different feature dimensions, making the new and old fault features more comparable in the same feature space. The feature similarity matrix between the teacher and student models is calculated. This matrix reflects the degree of feature correlation between samples. A similarity constraint loss is constructed based on the row and column directions of the matrix to ensure that the student model can not only learn the features of individual samples but also grasp the feature distribution relationship between samples. The core process of this feature projection comparison distillation is as follows: Figure 2 As shown. The calculation process of the feature similarity matrix is ​​as follows: Let the feature matrix output by the feature projection head be H, and the output dimension of the feature projection head be... The temperature coefficient is Then the feature similarity matrix of the teacher model Feature similarity matrix with student model Calculated using the following formula: ,in, The first element in the similarity matrix Line number Column elements, The first element in the characteristic matrix H The element in row d and column d. This constraint loss is combined with the cross-entropy classification loss to form the total loss function. The cross-entropy classification loss ensures accuracy in classifying new fault categories, while the bidirectional similarity constraint loss maintains the ability to identify historical fault categories. The formula for calculating the total loss function is: ,in, The total loss during model training. Cross-entropy classification loss, As a weight for distillation losses, The total distillation loss is compared using feature projection. A cosine annealing method is used to dynamically adjust the weights of the distillation loss. The expression for adjusting the weights of the distillation loss using cosine annealing is: ,in, For the first Distillation loss weights corresponding to each training epoch. This is the current training epoch number. This is the final value for the distillation loss weight. This is the initial value for the distillation loss weight. The total number of epochs trained for a single task. The model uses a cosine function. In the early stages of training, a higher distillation loss weight is set to allow the student model to fully learn the feature extraction logic and recognition experience of the teacher model regarding the three original faults, avoiding the loss of existing capabilities due to learning new faults. As training progresses, the distillation loss weight is gradually reduced, shifting the training focus to feature learning of the newly added high-temperature sintering faults, ensuring the recognition accuracy of the new fault categories. During training, 500 newly collected high-temperature sintering fault samples are combined with historical samples from the memory bank to form the training set, which is divided into training and validation sets in an 8:2 ratio. After each training cycle, the model performance is evaluated on the validation set, and the learning rate is adjusted based on the evaluation results. The dynamically scalable classifier adapts to new fault categories by concatenating the weight matrix and bias vector corresponding to the new high-temperature sintering fault categories, without requiring reconstruction of the overall model structure, thus reducing the technical difficulty and time cost of model iteration.

[0032] During the memory playback phase, after incremental training, a comprehensive analysis is performed on the training samples of newly added high-temperature sintering faults. Samples with different sintering degrees, different operating conditions, and high feature recognition are selected. Following a sample balancing strategy, the ratio of existing fault samples to newly added fault samples in the memory is ensured to be balanced, avoiding model bias caused by sample distribution imbalance during subsequent training. Some historical samples with weak feature representativeness and high operating condition repetition in the memory are replaced, maintaining a constant memory capacity. The updated memory simultaneously contains typical feature samples of the original three fault types, normal state, and newly added high-temperature sintering faults, providing reliable historical feature support for subsequent incremental training of newly added faults. Figure 3 As shown, when other new bearing failures occur in the automotive parts processing plant due to improvements in production processes or changes in raw materials, the above incremental training and memory playback steps are repeated to continuously expand the fault identification range of the model and ensure that the diagnostic model can always adapt to actual production and maintenance needs.

[0033] Example 2:

[0034] This embodiment is applied to the fault diagnosis of pitch bearings in deep-sea wind farms. In this scenario, the bearings are in a harsh environment with high humidity, high salt spray and large wind speed fluctuations for a long time. In addition, the single unit capacity is large and the operation and maintenance costs are high. In the initial stage, it is necessary to identify three common fault types: rolling element pitting, cage fracture and raceway scratches. At the same time, it is necessary to deal with the new fault challenges brought about by the intensified corrosion of the marine environment.

[0035] During the data processing and initial model training phase, two corrosion-resistant and interference-resistant vibration sensors were symmetrically installed at the upper and lower end caps of the pitch bearing of each offshore wind turbine. The sensors feature a waterproof and sealed design, enabling them to withstand the high humidity and high salt spray environment at sea. Bearing vibration signals were collected under different wind speeds (3-25 m / s) and wave heights. Three months of continuous operational data were collected for each turbine to ensure the sample covers the complex and variable operating conditions at sea. Considering the low-speed, heavy-load operation characteristics of the pitch bearing, the vibration signals were segmented using an overlapping sampling method. The overlap ratio was determined based on the bearing speed to ensure that subtle fault characteristics during low-speed operation were not missed. The segmented signals were filtered to remove interference signals from wave impacts and wind load fluctuations. Then, wavelet transform was used to convert the vibration signals into an RGB time-frequency image, which effectively highlights the abnormal manifestations of fault characteristics in both the time and frequency domains. An integrated feature extraction and classification model was constructed, using the ResNet18 network as the feature extractor. The depth and computational cost of this network balance the computational limitations of edge computing nodes in offshore wind turbines. A dynamically scalable classifier was connected to the feature extractor's output. The initial weights and biases of the classifier were set using the Xavier initialization method to improve the convergence speed of model training. A training dataset was constructed using sample data from multiple turbines, covering three fault types and normal states. Outliers and invalid data were removed. An adaptive moment estimation optimizer was used to train the initial model. An early stopping strategy was employed during training; training was stopped when the fault identification accuracy on the validation set decreased to avoid overfitting. When constructing the memory, differences in geographical location, operating time, and maintenance records among different turbines were comprehensively considered. A certain number of samples were selected from the samples of each turbine and stored in the memory to ensure the diversity and representativeness of the memory samples. The memory capacity was set to a fixed value, and solid-state drives were used for storage to meet the storage requirements of the humid offshore environment.

[0036] During the incremental training phase, as the offshore wind farms age, some turbine pitch bearings exhibit a new type of fault, raceway corrosion, caused by salt spray corrosion. This fault initially presents with weak characteristics and is easily confused with existing raceway scratch faults. At this point, the already trained initial model is used as the teacher model, which can accurately distinguish the three original fault types. A newly initialized model with the same structure as the teacher model is used as the student model, with parameter initialization consistent with the initial model. The high-dimensional features output by both models are reduced using a feature projection head. The multi-level linear transformation and activation function of the feature projection head effectively extract key features. The reduced low-dimensional features are then normalized to achieve a more uniform feature distribution and improve the accuracy of similarity calculation. The feature similarity matrix between the teacher and student models is calculated. This matrix quantifies the strength of feature associations between samples. A bidirectional similarity constraint loss is constructed based on the row and column directions of the matrix. This loss forces the student model to learn the feature extraction logic of the teacher model, maintaining its ability to identify historical faults. The total loss function is formed by combining bidirectional similarity constraint loss and cross-entropy classification loss. The cross-entropy classification loss is used to optimize the classification boundary for newly added raceway corrosion faults. A cosine annealing method is used to dynamically adjust the distillation loss weights. In the early stages of training, the distillation loss weights are set to higher values, allowing the student model to fully absorb the feature recognition experience of the teacher model regarding the three original faults, ensuring that it does not forget its existing capabilities when learning new faults. As the training cycle progresses, the distillation loss weights are gradually reduced according to the cosine function, shifting the training focus to feature learning of newly added raceway corrosion faults. Training with a large number of new samples allows the model to master the unique characteristics of corrosion faults. During training, 800 newly collected raceway corrosion fault samples are combined with historical samples in the memory bank for model training. The training batch is dynamically adjusted according to the computing power of the edge computing nodes to ensure the stability of the training process. The dynamically scalable classifier achieves rapid adaptation to new fault categories by concatenating the weight matrix and bias vector corresponding to the new raceway corrosion fault category, without modifying the model's network structure and forward propagation logic, significantly shortening the model iteration cycle.

[0037] During the memory playback phase, after incremental training, newly added raceway corrosion fault samples are systematically analyzed. They are categorized and screened according to different corrosion degrees, wind speed conditions, and turbine numbers to select representative samples that comprehensively reflect the characteristics of raceway corrosion faults. Following a sample balancing strategy, the proportion of existing fault samples, normal state samples, and newly added raceway corrosion fault samples in the memory is ensured to avoid sample bias in subsequent incremental training. When updating the memory, historical samples with high feature repetition, single operating conditions, and atypical fault degrees are prioritized for replacement, maintaining a constant memory capacity. The updated memory not only retains the core features of the original faults and normal states but also adds key features of raceway corrosion faults, providing high-quality historical samples to support incremental training for subsequent new faults. When other new types of bearing faults occur in offshore wind farms due to environmental changes, equipment aging, or other reasons, the above incremental training and memory playback steps are repeated to continuously expand the model's fault identification capabilities, meeting the long-term, efficient, and low-cost operation and maintenance diagnostic needs of deep-sea wind farms and reducing turbine downtime losses caused by bearing faults.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A bearing fault diagnosis method based on feature projection comparison distillation and memory playback incremental learning, characterized in that... The specific steps of this method are as follows: S1, Data Processing and Initial Model Training Stage: The collected bearing vibration signal is converted into a 224×224 pixel RGB time-frequency image through wavelet transform; a ResNet18 network is used as a feature extractor, and the output of the feature extractor is connected to a dynamic scalable classifier. The initial model is trained using initial fault category samples; a memory bank with limited capacity is constructed, and a preset number of samples are randomly selected from the initial samples and stored in the memory bank. S2, Incremental Training Phase: The model trained in the previous round is used as the teacher model, and the model to be trained in this round is used as the student model; the 512-dimensional high-dimensional features output by the feature extractors of the teacher model and student model are projected to 128-dimensional low-dimensional features through the feature projection head and normalized; the feature similarity matrix between the teacher model and student model is calculated, and the feature projection comparison distillation loss is constructed by using KL divergence in the row and column directions; the total loss function is formed by combining cross-entropy classification loss; the weights of the distillation loss are dynamically adjusted using the cosine annealing method, and the student model is trained by combining new task samples with historical samples in the memory bank, updating the weight matrix and bias vector of the dynamically expandable classifier; S3, Memory Replay Stage: After the incremental training is completed, a preset number of samples are randomly selected from the new task samples to update the memory bank, keeping the capacity of the memory bank constant; when a new incremental task arrives, steps S2 to S3 are repeated to complete the bearing fault diagnosis corresponding to the new incremental task.

2. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The dynamically scalable classifier is for the first... The mathematical expression for each task is: ,in, This is the sequence number of the currently executing incremental task. For the first The classifier output results for each task Let be the d-dimensional feature vector output by the feature extractor, where d is the dimension of the output feature vector. For the first The classifier weight matrix corresponding to each task is: , For the first The number of historical fault categories learned before each task is executed. For the first The number of new fault categories added per task For the first The classifier bias vector corresponding to each task; the weight matrix By splicing together historical weight matrices With the newly added category weight matrix To obtain, that is: ,in, This is the classifier weight matrix corresponding to the historical fault categories. The classifier weight matrix corresponding to the newly added fault category; the bias vector By splicing historical bias vectors With the newly added class bias vector To obtain, that is: ,in, This represents the classifier bias vector corresponding to the historical fault categories. This is the classifier bias vector corresponding to the newly added fault category.

3. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The structure of the feature projection head is as follows: it includes a linear layer with an input dimension of 512 and an output dimension of 256, a ReLU activation layer with an input dimension of 256 and an output dimension of 256, a linear layer with an input dimension of 256 and an output dimension of 128, and a LayerNorm normalization layer with an input dimension of 128 and an output dimension of 128.

4. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The calculation process of the feature similarity matrix is ​​as follows: Let the feature matrix output by the feature projection head be H, and the output dimension of the feature projection head be... The temperature coefficient is Then the feature similarity matrix of the teacher model Feature similarity matrix with student model Calculated using the following formula: ,in, The first element in the similarity matrix Line number Column elements, The first element in the characteristic matrix H The element in row d and column d.

5. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The characteristic projection compares distillation loss The calculation process is as follows: calculate the similarity matrix row directions respectively. Divergence loss: and column direction Divergence loss: The total distillation loss is obtained by averaging the row and column losses: ,in, For the row direction of the similarity matrix Divergence loss, Let be the row vector of the feature similarity matrix of the student model. Let the row vectors of the teacher model feature similarity matrix be... For the column direction of the similarity matrix Divergence loss, This is the column vector of the student model feature similarity matrix. This is the column vector of the teacher model feature similarity matrix. The total loss of distillation is compared with the feature projection.

6. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The formula for calculating the total loss function is as follows: ,in, The total loss during model training. Cross-entropy classification loss, As a weight for distillation losses, The total loss of distillation is compared with the feature projection.

7. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The expression for adjusting the distillation loss weight using the cosine annealing method is as follows: ,in, For the first Distillation loss weights corresponding to each training epoch. This is the current training epoch number. This is the final value for the distillation loss weight. This is the initial value for the distillation loss weight. The total number of epochs trained for a single task. It is a cosine function.

8. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, During the memory bank update process, a sample balancing strategy is adopted to ensure that the number of new task samples and historical samples in the training set is balanced; at the same time, label space remapping is performed, merging the label space corresponding to the new samples into the original label space to obtain the updated label space. ,in, For the updated tag space, For the label space corresponding to historical fault categories, This is the label space corresponding to the fault category of the new task.

9. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The vibration signal processing procedure is as follows: the collected vibration signal is divided into segments, each segment having a sample length of 1024. The segmented samples are then subjected to either overlapping sampling or direct sampling based on their data characteristics, and finally converted into a time-frequency image through wavelet transform.

10. The feature projection comparison distillation and memory playback incremental learning bearing fault diagnosis method according to claim 1, characterized in that, The memory bank has a preset fixed capacity. When updating, a replacement update strategy is adopted, in which samples from new tasks are selected to replace some historical samples in the memory bank to keep the memory bank capacity constant.

Images