Small sample intelligent diagnosis method for rolling bearing based on improved conditional generative adversarial network

Abstract

The application belongs to the technical field of intelligent fault diagnosis of small samples of rotating machinery, and relates to an intelligent diagnosis method for small samples of rolling bearings based on an improved conditional generative adversarial network, which is suitable for intelligent fault diagnosis of key components of mechanical equipment such as rolling bearings. The method comprises the following steps: taking the latent features extracted from real fault samples by an encoder as the semantic prior of the generator, so as to improve the process controllability of the generator and the quality of the generated data; a label coupling mechanism is used to establish the internal relationship between the features and the labels, and the generator is effectively guided to approximate the distribution of different fault modes through multi-layer embedding; under the dual constraints of the authenticity of the discriminator and the classification labels, the reliability of the generated data in the distribution and category levels is ensured. The method combines the feature expression ability of the autoencoder and the category constraint advantage of the auxiliary classification generative adversarial network, realizes high-quality data augmentation through the generation of highly consistent simulated fault samples, and effectively improves the accuracy of the diagnosis model under the condition of small samples.

Description

Technical Field

[0001] This invention belongs to the field of small-sample intelligent fault diagnosis technology for rotating machinery, specifically involving a small-sample intelligent diagnosis method for rolling bearings based on an improved conditional generative adversarial network. It is mainly applicable to intelligent fault diagnosis of key components of mechanical equipment such as rolling bearings. Background Technology

[0002] The stable and reliable operation of mechanical equipment is a fundamental prerequisite for ensuring the safety and continuity of industrial processes. As equipment service life extends, the cumulative damage and potential failure probability of critical components such as rolling bearings increase significantly. If such potential threats are not identified and addressed in a timely manner, they can easily escalate into major accidents, causing irreversible equipment failures or even personal injury. Therefore, conducting research on intelligent fault diagnosis for rolling bearings is of great significance for improving the proactive maintenance capabilities of complex mechanical systems, extending equipment lifespan, and reducing the risk of unplanned downtime.

[0003] In real-world industrial scenarios, equipment managers typically require immediate shutdown and inspection of machinery exhibiting abnormal symptoms. However, collecting and accumulating fault samples is extremely difficult, easily leading to a small sample problem where fault samples constitute a very low percentage of the total available training samples and the data category distribution is severely imbalanced. This issue has become one of the core bottlenecks restricting the diagnostic accuracy and generalization ability of intelligent fault diagnosis methods in practical engineering.

[0004] By leveraging Generative Adversarial Networks (GANs) to fully learn the true feature distribution of fault data, and then synthesizing simulated data with statistical similarity to real fault samples through a trained generator, the current main solution for small-sample diagnosis is to expand the minority class of samples. This type of method can expand the scale of fault samples at the data level, achieve dataset balance, and greatly alleviate the problem of insufficient diagnostic model performance caused by imbalanced fault samples.

[0005] However, existing GAN-based few-sample diagnostic methods still face several technical shortcomings. First, GANs and most of their variants are prone to training oscillations and pattern collapse during training, resulting in insufficient realism and diversity of generated samples. Second, the generation mapping mechanism is ambiguous during data synthesis, leading to low controllability and poor data quality. Third, category labels are usually simply embedded in the first layer of the generator, making it difficult to achieve precise and effective conditional control. Therefore, there is an urgent need to design a novel generative adversarial network structure to improve the fidelity, diversity, and category controllability of synthesized data under few-sample conditions, thereby improving the performance of intelligent diagnostic models under conditions of scarce fault samples. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention provides a small-sample intelligent diagnostic method for rolling bearings based on an improved conditional generative adversarial network.

[0007] The technical solution of the present invention is as follows: First, an encoder is constructed and trained to obtain latent feature vectors that characterize the distribution patterns of real fault data. These latent feature vectors serve as semantic prior information for subsequent generation processes, constraining the generator's input space and thus enhancing the controllability of the data synthesis process and ensuring the quality of the generated data.

[0008] Secondly, the latent feature vector output by the encoder is fused with the corresponding fault label information at the feature level. Based on the fused features, the generator synthesizes high-quality simulated fault samples that match the specified fault category through multiple deconvolution operations. During this process, the label information is embedded into multiple deconvolution layers of the generator to achieve precise conditional control over the generated sample category.

[0009] Furthermore, a discriminator with multiple constraint mechanisms is designed. This discriminator simultaneously distinguishes between the authenticity and fault category of the input sample. By jointly optimizing the discriminant loss and classification loss, it forces the generator's generated data to maintain a high degree of consistency with the real data in terms of data distribution characteristics and fault category.

[0010] Finally, the trained generator is used to synthesize a specified number of simulated fault samples to form a training set that is class-balanced and of sufficient size. Based on this, an independently deployed intelligent diagnostic network is trained to complete the fault mode recognition task in a small sample scenario.

[0011] Through the aforementioned collaborative adversarial training mechanism of encoder-generator-discriminator, this invention can generate high-fidelity, highly diverse, and category-controllable simulated fault data in scenarios where fault samples are extremely scarce. This alleviates the performance degradation problem of diagnostic models caused by data imbalance and improves the recognition accuracy and robustness of intelligent diagnostic methods for rolling bearings in small-sample engineering scenarios.

[0012] A method for intelligent diagnosis of rolling bearings with few samples based on improved conditional generative adversarial networks, the specific steps of which are as follows: Step 1: Collect bearing vibration data under different fault conditions of mechanical equipment and perform normalization processing to construct an initial unbalanced dataset.

[0013] Step 2: Construct an encoder consisting of multiple concatenated convolutional blocks. This method extracts latent feature representations from real bearing vibration data, providing highly relevant semantic information to the generator during data generation and enabling proactive feature guidance, thereby enhancing the controllability of the data generation process. The expression for the latent features is as follows: in, Indicates input encoder Real bearing vibration data samples; Indicates that it is made by the encoder Extracted latent features; Indicates encoder right The output result after feature extraction; This indicates batch normalization processing; , They represent encoders respectively. The first in Each convolutional layer and its corresponding activation function, superscript e This indicates that it belongs to the encoder category. The encoder of this invention All use the LeakyReLU activation function.

[0014] Step 3: Identify latent features Category tags and noise vector Together as generators The input is fed into a generator consisting of multiple cascaded deconvolution operations. In the process, the generated bearing vibration data corresponding to the specified fault category is gradually obtained. .

[0015] in, , Representing generators respectively The first in Each deconvolutional layer and its corresponding activation function, superscript This indicates that it belongs to the generator. Tag coupling mechanism This refers to the Label Coupling mechanism, used to combine category labels. With latent features In the generator Coupling is performed during the multi-layer deconvolution process to enhance the class discrimination and semantic consistency of the generated samples.

[0016] The mechanism works as follows: (1) Let the output feature of a certain deconvolution layer be the output feature of the deconvolution layer. The deconvolutional layer outputs features. as a latent feature Through generator The intermediate features obtained after mapping the preceding deconvolutional layers are compared with the latent features in step 3. This involves a pass-through relationship. First, a conditional batch normalization operation is introduced before the subsequent deconvolutional layer, causing the output features of the deconvolutional layer to... According to category labels Normalization and adaptive adjustment are performed to maintain the separability of features from different categories during the generation process, thereby effectively enhancing the quality of category condition generation. The specific calculation is as follows: Among them, the output features of the deconvolution layer The intermediate features before normalization and the latent features after normalization are represented. Indicates the output features of the deconvolution layer Conditional normalization function Processed features; conditional batch normalization function Represents the conditional batch normalization function; characteristic mean Characteristic variance These represent the output features of the deconvolution layer, respectively. Mean and variance; scaling parameters Offset parameters It is based on category labels Dynamically adjusted scaling and offset parameters.

[0017] (2) Using category label vectors as query vector To normalize latent features as a key vector Sum value vector For category labels With normalized latent features By aggregating and weighting the data, and explicitly establishing the interaction relationship between the two, we can obtain the latent feature semantic vector of the associated category information. and will As input to subsequent deconvolutional layers.

[0018] in, , , These represent the query vector, key vector, and value vector, respectively. These are the corresponding weight matrices; This represents the attention mechanism; This represents the normalized exponential function, used to calculate the attention weights between the query vector and the key vector.

[0019] (3) Before the generator After each deconvolutional layer, a label coupling mechanism is added to ensure that label information can continuously participate in guiding and playing a role in all stages of data generation, thereby achieving more precise and effective condition control.

[0020] Step 4: Construct a discriminator consisting of multiple cascaded convolutional layers. and real bearing vibration data and generating bearing vibration data Enter them separately Discriminator Front-end convolutional structure and encoder They have the same number of convolutional layers and channel settings, but their functions differ: encoder Used to output latent features Discriminator Used to output the discrimination result. Discriminator At the output, two parallel linear layers are added at the end. The first linear layer outputs the authenticity judgment result. The first linear layer is used to determine whether the input data is real data or generated data; the second linear layer outputs the category prediction label. This is used to determine the fault category to which the input data belongs. In addition, the discriminator... Spectral normalization is used after each convolutional layer. , and encoder Use batch normalization Therefore, the two are not entirely identical in structure; spectral normalization Through constraint discriminator The spectral norm of the weights is used to control the Lipschitz constant, thereby stabilizing the adversarial training process.

[0021] Step 5: Based on the discriminator in Step 4 Output authenticity judgment results and category prediction results In the discriminator Multiple loss constraint mechanisms are applied at the end, and the constraint generator is used. Output and generate bearing vibration data This enables the generation of bearing vibration data. It is highly similar to real fault samples in terms of data authenticity and category consistency.

[0022] Step 5.1: Determine the loss Judging the results by their authenticity For the computational object, constrain the similarity between generated data and real data from the perspective of data distribution.

[0023] in, Represents actual bearing vibration data Follows the true data distribution Expectations at that time; Representing noise variables Follows noise distribution Expectations at that time; This indicates the discriminator's judgment result on the authenticity of the data; Represents generator The output generates bearing vibration data.

[0024] Step 5.2: Classification Loss Based on category prediction results For computational objects, used to ensure the consistency of categories between generated data and real data.

[0025] in, This represents the category prediction result output by the discriminator; This indicates the category label corresponding to the input data.

[0026] Step 6: Model training, based on the discriminant loss obtained in Step 5. and classification loss Construct discriminator loss separately With generator loss Discriminator loss From step 5 and The combination is used to update the discriminator. To make the discriminator Simultaneously improve the ability to determine authenticity and class recognition; generator loss Use the bearing vibration data generated in step 5 The corresponding authenticity criteria and classification criteria are used to jointly update the encoder. and generator .

[0027] Step 6.1: Freeze the encoder and generator The parameters will be the actual bearing vibration data. and generating bearing vibration data Input discriminator ,according to Calculate the discriminator loss And update the discriminator The network parameters are adjusted to enhance its comprehensive ability to identify data authenticity and fault types.

[0028] Step 6.2: Freeze the discriminator The network parameters will be used by the encoder. Extracted latent features Category tags and noise vector Input generator Obtain the generated bearing vibration data Then, based on the bearing vibration data generated in step 5... The generator loss is calculated based on the corresponding authenticity discrimination and classification terms. and jointly update the encoder through backpropagation. and generator The network parameters prompt the encoder The output is more suitable for the generator Potential features used and prompt the generator Synthesized data that is highly similar to real data and consistent with the categories.

[0029] in, To generate bearing vibration data The corresponding authenticity criteria, To generate bearing vibration data The corresponding category.

[0030] Step 6.3: Alternately implement steps 6.1 and 6.2, by first updating the discriminator. Then jointly update the encoder and generator In this way, the generator is guaranteed In latent features and category labels Guided by this, the output is a specified category sample that is highly consistent with the real data in both distribution and category dimensions.

[0031] Step 7: After completing the alternating training in Step 6, the discriminator is obtained. Constrained encoder and generator Among them, the discriminator Used for authenticity judgment and category prediction feedback during the training phase, it does not participate in subsequent sample generation. A pre-trained encoder is used. and generator Real bearing vibration data Input encoder Obtain latent features Within the same fault category, potential characteristics Perform feature interpolation and center traction processing, and then input the processed latent features into the trained generator. In order to obtain high-quality and diverse simulated fault samples.

[0032] Step 7.1: Randomly select two samples from the potential features of the same fault type. and And obtain the hybrid features by linear interpolation. .

[0033] Among them, interpolation coefficients Used to control the blending characteristics in and The degree of centering between them.

[0034] Step 7.2: To avoid interpolated features By removing it from its original category space and applying a central pull, new latent features are obtained. And use it to guide the generator to synthesize more diverse fault samples.

[0035] Among them, the traction coefficient Used to regulate latent features With class center The degree of deviation; This represents the feature center of each category, which is the mean vector of all potential features of that category.

[0036] Step 7.3: Transferring potential features after central traction Its corresponding category label noise vector Input into the trained generator In step 3, a specified number and category of simulated fault samples are obtained.

[0037] Step 8: Merge the simulated fault samples obtained in Step 7 with the original imbalanced dataset, supplement the minority class samples, and obtain a dataset with a balanced number of samples in each class.

[0038] Step 9: Construct an independent intelligent diagnostic network consisting of four levels of one-dimensional convolutional feature extraction units and two levels of linear discriminant units, and train the intelligent diagnostic network based on the class-balanced dataset obtained in Step 8. Each convolutional feature extraction unit includes a convolutional layer, a batch normalization layer, and a ReLU activation layer, used to extract fault features from bearing vibration data; the first linear mapping layer is used to map the fault features and form class discriminant features; the second linear classification layer is used to output the predicted probability of each fault category. After training, the test data is input into the intelligent diagnostic network, and the category with the highest predicted probability is determined as the fault diagnosis result. At the same time, the diagnostic performance of the intelligent diagnostic network is evaluated.

[0039] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention proposes a data generation method based on latent features, which introduces the semantic information of real samples into the data synthesis process, enhances the controllability of the generation process and improves the quality of the generated data, and avoids the problems of using random noise as the main input, fuzzy mapping mechanism and insufficient training stability in the traditional method.

[0040] (2) The present invention designs a tag coupling mechanism to deeply explore the coupling relationship between potential features and category tags, embeds it into the multi-layer structure of the generator, improves the problem of tag information being weakened during network transmission, and realizes more accurate feature guidance and condition control.

[0041] (3) The present invention introduces spectral normalization in the discriminator to stabilize the training process, and at the same time constructs a loss function with multiple loss collaborative constraints at the discriminator end to ensure that the generated data maintains a high degree of consistency with the real data in terms of distribution consistency and class discriminability. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the implementation process of the present invention.

[0043] Figure 2 This is a schematic diagram of the network structure of the present invention.

[0044] Figure 3 The comparison results between the data generated by this invention and the real data are shown; where (a) and (b) are the waveform and spectrum of the real data, respectively, and (c) and (d) are the waveform and spectrum of the generated sample, respectively.

[0045] Figure 4 The figures show the accuracy curves before and after data augmentation using the method of the present invention; where (a) is the training result of the original imbalanced dataset, and (b) is the training result after data augmentation and balancing using the method of the present invention.

[0046] Figure 5The result is a confusion matrix after data augmentation using the method of this invention. Detailed Implementation

[0047] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0048] The process of this invention is as follows Figure 1 As shown, the specific network structure is as follows: Figure 2 As shown. In this embodiment, a rotating machinery vibration test bench was used to collect bearing vibration data. This test bench includes components such as a drive motor, a rotating shaft, and a torque sensor. Specifically, in this embodiment, pitting defects were created on the outer ring, rolling elements, and inner ring of a deep groove ball bearing of model 6205-2RS. The bearing's pitch diameter is 39.04 mm. mm There are a total of 9 with a diameter of 7.94. mm The rolling element has a contact angle of 0°.

[0049] In the embodiment, the labels on the inner ring with fault diameters of 7 mils, 14 mils, and 21 mils are named I7, I14, and I21, respectively; the labels on the rolling elements with pitting diameters of 7 mils, 14 mils, and 21 mils are named B7, B14, and B21, respectively; similarly, the labels on the outer ring with pitting diameters of 7 mils, 14 mils, and 21 mils are named O7, O14, and O21, respectively.

[0050] This embodiment presents a small-sample intelligent diagnostic method for rolling bearings based on an improved conditional generative adversarial network, the overall steps of which include: S1. Two accelerometers are respectively placed on the bearing housings at the fan end and drive end of the electric motor, with a sampling frequency set to 12 kHz. Under the condition of motor speed of 1797 r / min and load of 0 Hp, vibration data of different fault types of the equipment are collected using the fan end sensor. The number of normal samples is set to 200 sets, and the number of samples for the other 9 types of faults is set to a limited 20 sets each, constructing a diagnostic scenario with an imbalance ratio of 10:1 between normal and fault samples.

[0051] S2. Construct an encoder consisting of 5 concatenated convolutional blocks. The basic structure of each convolutional block is "convolutional layer - batch normalization layer - LeakyReLU activation", and the data of each fault type is input into the encoder. In the process, latent feature representations are extracted; S3. Construct a generator consisting of 5 cascaded deconvolutional blocks. The basic structure of each convolutional block is "deconvolutional layer - batch normalization layer - LeakyReLu activation", which inputs latent features, label information, and noise vectors into the generator. In, and in the generator A label coupling mechanism is set after the first four deconvolution blocks to obtain simulated fault samples; S4. Construct a system consisting of 5 convolutional layers and spectral normalization layers. SN A discriminator consisting of a LeakyReLU activated convolutional block and two parallel linear layers. The two parallel linear layers output the authenticity judgment result and the category prediction result, respectively, which serve as the basis for calculating the subsequent discrimination loss and classification loss. S5. Input the real data and the generated samples into the discriminator respectively. In the middle, the discrimination loss is calculated based on the authenticity discrimination result. Calculate classification loss based on category prediction results ; S6, Fixed Encoder and generator The parameters are based on real data and generated samples in the discriminator. The discriminator loss is calculated based on the authenticity judgment results and category prediction results. First, the discriminator parameters are updated. Second, with the discriminator parameters fixed, the latent features, class labels, and noise vectors extracted by the encoder are input into the generator to obtain generated samples. The generator loss is calculated based on the authenticity judgment result and class prediction result of the generated samples in the discriminator. The parameters of the encoder and generator are jointly updated through gradient propagation; the above process is repeated alternately until the model converges.

[0052] S7. After completing the alternating training, fix the discriminator. The encoder after training constraints and generator The fault data for each category is input into the trained encoder. The process involves obtaining latent feature representations of real fault samples, performing feature interpolation and centering within the same fault category to obtain 180 sets of processed latent features. These processed latent features, corresponding category labels, and noise vectors are then input into a pre-trained generator. , to obtain synthetic samples of the specified category, with 180 sets of samples generated for each fault type.

[0053] S8. Use the generated data to expand the original imbalanced dataset so that the number of samples for each of the nine types of faults can reach 200, thereby constructing a dataset with a balanced number of samples for each category.

[0054] S9. Construct an independent intelligent diagnostic network consisting of four levels of one-dimensional convolutional feature extraction units and two levels of linear discriminant units cascaded together. Train the intelligent diagnostic network for 200 rounds based on the balanced dataset, obtain diagnostic results, and evaluate the network performance.

[0055] Taking fault label B7 as an example, the comparison results between the generated data and the actual data in this embodiment are as follows: Figure 3 As shown in the figure, comparative analysis reveals that the generated data for this fault type maintains a high degree of similarity to the real data of the corresponding category in terms of overall waveform morphology, amplitude variation trend, and spectrum distribution pattern. This also demonstrates that the method of the present invention has high generation quality.

[0056] Accuracy variation curves under different datasets are shown below Figure 4 As shown, while the training accuracy curve converges quickly when training directly using the original imbalanced dataset, the validation accuracy curve fluctuates significantly and exhibits poor convergence, indicating a noticeable overfitting phenomenon. After data augmentation and dataset balancing, the training curve shows faster convergence and higher stability, while effectively mitigating the model overfitting problem, demonstrating the feasibility of the proposed method in data generation and augmentation.

[0057] Ultimately, through high-quality data augmentation, the method of this invention improves the average diagnostic accuracy of the model from 94.53% to 99.80% in diagnostic scenarios with highly imbalanced data distribution. The confusion matrix of the diagnostic results is as follows: Figure 5 As shown, even with a limited number of available fault samples, the method of this invention can still fully learn and fit the characteristic patterns of various rolling bearing faults, thereby more effectively mitigating the adverse effects of data imbalance on diagnostic performance.

Claims

1. A method for intelligent diagnosis of rolling bearings with few samples based on an improved conditional generative adversarial network, characterized in that, The specific steps are as follows: Step 1: Collect bearing vibration data under different fault conditions of mechanical equipment and perform normalization processing to construct an initial unbalanced dataset; Step 2: Construct an encoder consisting of multiple concatenated convolutional blocks. It extracts latent feature representations from real bearing vibration data, provides highly relevant semantic information to the generator during the data generation process, and achieves positive feature guidance, thereby enhancing the controllability of the data generation process; Step 3: Identify latent features Category tags and noise vector Together as generators The input is fed into a generator consisting of multiple cascaded deconvolution operations. In the process, the generated bearing vibration data corresponding to the specified fault category is gradually obtained. ; Step 4: Construct a discriminator consisting of multiple cascaded convolutional layers. and real bearing vibration data and generating bearing vibration data Enter them separately Discriminator Front-end convolutional structure and encoder They have the same number of convolutional layers and channel settings, but their functions differ: encoder Used to output latent features Discriminator Used to output the discrimination result; Discriminator At the output, two parallel linear layers are added at the end. The first linear layer outputs the authenticity judgment result. This is used to determine whether the input data is real data or generated data; The second linear layer outputs the predicted category label. This is used to determine the fault category to which the input data belongs; in addition, the discriminator Spectral normalization is used after each convolutional layer. , and encoder Use batch normalization ; Step 5: Based on the discriminator in Step 4 Output authenticity judgment results and category prediction results In the discriminator Multiple loss constraint mechanisms are applied at the end, and the constraint generator is used. Output to generate bearing vibration data ; Step 6: Model training, based on the discriminant loss obtained in Step 5. and classification loss Construct discriminator loss separately generator loss Discriminator loss From step 5 and The combination is used to update the discriminator. Generator loss Use the bearing vibration data generated in step 5 The corresponding authenticity criteria and classification criteria are used to jointly update the encoder. and generator ; Step 7: After completing the alternating training in Step 6, the discriminator is obtained. Constrained encoder and generator Among them, the discriminator Used for authenticity judgment and category prediction feedback during the training phase, it does not participate in subsequent sample generation; the pre-trained encoder is fixed. and generator Real bearing vibration data Input encoder Obtain latent features Within the same fault category, potential characteristics Perform feature interpolation and center traction processing, and then input the processed latent features into the trained generator. In order to obtain simulated fault samples; Step 8: Merge the simulated fault samples obtained in Step 7 with the original imbalanced dataset, supplement the minority class samples, and obtain a dataset with a balanced number of samples in each class. Step 9: Construct an independent intelligent diagnostic network consisting of four levels of one-dimensional convolutional feature extraction units and two levels of linear discriminant units, and train the intelligent diagnostic network based on the class-balanced dataset obtained in Step 8. Each convolutional feature extraction unit includes a convolutional layer, a batch normalization layer, and a ReLU activation layer, used to extract fault features from bearing vibration data. The first linear mapping layer is used to map the fault features and form class discriminant features. The second linear classification layer is used to output the predicted probability of each fault category. After training, the test data is input into the intelligent diagnostic network, and the category with the highest predicted probability is determined as the fault diagnosis result. At the same time, the diagnostic performance of the intelligent diagnostic network is evaluated.

2. The method for intelligent diagnosis of rolling bearings based on improved conditional generative adversarial networks according to claim 1, characterized in that, In step 2, the expression for the latent features is as follows: in, Indicates input encoder Real bearing vibration data samples; Indicates that it is made by the encoder Extracted latent features; Indicates encoder right The output result after feature extraction; This indicates batch normalization processing; , They represent encoders respectively. The first in Each convolutional layer and its corresponding activation function, superscript e This indicates that it belongs to the encoder category. .

3. The method for intelligent diagnosis of rolling bearings based on improved conditional generative adversarial networks according to claim 1, characterized in that, In step 3, bearing vibration data is generated. The expression is as follows: in, , Representing generators respectively The first in Each deconvolutional layer and its corresponding activation function, superscript This indicates that it belongs to the generator. Tag coupling mechanism This refers to the Label Coupling mechanism, used to combine category labels. With latent features In the generator Coupling is performed during the multi-layer deconvolution process to enhance the class discrimination and semantic consistency of the generated samples.

4. The method for intelligent diagnosis of rolling bearings based on improved conditional generative adversarial networks according to claim 3, characterized in that, The mechanism works as follows: (1) Let the output feature of a certain deconvolution layer be the output feature of the deconvolution layer. The output features of this deconvolution layer as a latent feature Through generator The intermediate features obtained after mapping the preceding deconvolutional layers are first addressed by introducing conditional batch normalization before subsequent deconvolutional layers, thus making the output features of the deconvolutional layers more consistent. According to category labels Normalization and adaptive adjustment are performed to maintain the separability of features from different categories during the generation process. The specific calculation is as follows: Among them, the output features of the deconvolution layer The intermediate features before normalization and the latent features after normalization are represented. Indicates the output features of the deconvolution layer Conditional normalization function Processed features; conditional batch normalization function Represents the conditional batch normalization function; characteristic mean Characteristic variance These represent the output features of the deconvolution layer, respectively. Mean and variance; scaling parameters Offset parameters It is based on category labels Dynamically adjusted scaling and offset parameters; (2) Using category label vectors as query vector To normalize latent features as a key vector Sum value vector For category labels With normalized latent features By aggregating and weighting the data, and explicitly establishing the interaction relationship between the two, we can obtain the latent feature semantic vector of the associated category information. and will As input to subsequent deconvolutional layers; in, , , These represent the query vector, key vector, and value vector, respectively. These are the corresponding weight matrices; This represents the attention mechanism; This represents the normalized exponential function, used to calculate the attention weights between the query vector and the key vector; (3) Before the generator After each deconvolutional layer, a label coupling mechanism is added to ensure that label information can continuously participate in each stage of data generation.

5. The method for intelligent diagnosis of rolling bearings based on improved conditional generative adversarial networks according to claim 1, characterized in that, Step 5 is as follows: Step 5.1: Determine the loss Judging the results by their authenticity For the computational object, constrain the similarity between generated data and real data from the perspective of data distribution; in, Represents actual bearing vibration data Follows the true data distribution Expectations at that time; Representing noise variables Follows noise distribution Expectations at that time; This indicates the discriminator's judgment result on the authenticity of the data; Represents generator The output is generated bearing vibration data; Step 5.2: Classification Loss Based on category prediction results For calculation purposes, this is used to ensure the consistency of the generated data with the categories of the real data; in, This represents the category prediction result output by the discriminator; This indicates the category label corresponding to the input data.

6. The method for intelligent diagnosis of rolling bearings based on improved conditional generative adversarial networks according to claim 1, characterized in that, Step 6 is as follows: Step 6.1: Freeze the encoder and generator The parameters will be the actual bearing vibration data. and generating bearing vibration data Input discriminator ,according to Calculate the discriminator loss And update the discriminator Network parameters; Step 6.2: Freeze the discriminator The network parameters will be used by the encoder. Extracted latent features Category tags and noise vector Input generator Obtain the generated bearing vibration data Then, based on the bearing vibration data generated in step 5... The generator loss is calculated based on the corresponding authenticity discrimination and classification terms. and jointly update the encoder through backpropagation. and generator The network parameters prompt the encoder The output is more suitable for the generator Potential features used ; in, To generate bearing vibration data The corresponding authenticity criteria, To generate bearing vibration data Corresponding category items; Step 6.3: Alternately perform steps 6.1 and 6.2, first updating the discriminator. Then jointly update the encoder and generator .

7. The method for intelligent diagnosis of rolling bearings based on improved conditional generative adversarial networks according to claim 1, characterized in that, Step 7 is as follows: Step 7.1: Randomly select two samples from the potential features of the same fault type. and And obtain the hybrid features by linear interpolation. ; Among them, interpolation coefficients Used to control the blending characteristics in and The degree of centering between them; Step 7.2: Feature Analysis After applying central traction, new potential features are obtained. ; Among them, traction coefficient Used to regulate latent features With class center The degree of deviation; This represents the feature center of each category, which is the mean vector of all latent features of that category; Step 7.3: Transferring potential features after central traction Its corresponding category label noise vector Input into the trained generator In step 3, a specified number and category of simulated fault samples are obtained.

Images