Rectified current-based bearing fault data generation method and system

Abstract

The invention discloses a rectification current bearing fault data generation method and system based on space-time condition fusion and non-classifier guidance, and belongs to the field of industrial fault diagnosis. The method comprises the following steps: converting a one-dimensional vibration signal into a two-dimensional time-frequency diagram through continuous wavelet transform, inputting the two-dimensional time-frequency diagram into a UNet network, and embedding a residual Seaformer module to fuse time-space condition information; a rectification flow path consistency loss function is adopted to train a network, a high-fidelity synthesis sample is generated through non-classifier guidance and Euler integration in the reasoning stage, and enhanced data is obtained through cosine similarity screening. According to the system, under the 1: 400 extreme unbalance condition, the accuracy rate is larger than or equal to 99% in a CWRU data set, the SEU data set is reduced to 81.15% due to complex working conditions, the reasoning time is only 6.7% of that of DDPM, and the system can be deployed on edge equipment and is suitable for scenes such as electric power, mines and rail transit.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis at the intersection of Industrial Internet of Things and intelligent manufacturing. Specifically, it relates to a spatiotemporal condition fusion rectified flow high-fidelity equilibrium sample generation method and system for bearing fault diagnosis of key equipment such as aircraft engines, industrial machine tools, and wind turbines. Background Technology

[0002] In modern industrial production, bearings are core components of critical equipment such as aircraft engines, industrial machine tools, and wind turbines. Their failure can cause abnormal vibrations and noise, and even lead to serious safety accidents and huge economic losses. Beyond these scenarios, bearings also play a crucial role in power systems (such as generators and motors), industrial automated production lines (such as conveyor equipment and robotic arms), mining machinery, and rail transportation. For example, in power systems, generator bearing failure can lead to grid outages; in automated production lines, conveyor equipment bearing failure can cause the entire production line to shut down, resulting in significant economic losses. Therefore, the need for bearing fault diagnosis in different industrial scenarios is becoming increasingly urgent.

[0003] However, these scenarios generally suffer from problems such as scarce fault samples, complex operating conditions, and limited data acquisition, making it difficult for traditional diagnostic methods to meet the requirements of high accuracy, real-time performance, and robustness (Liu D. et al., “RotatingMachinery Fault Diagnosis Under Time-Varying Speeds: A Review,” IEEE SensorsJournal, Vol.23, No.24, pp.29969-29990, 2023.). Therefore, accurate bearing fault diagnosis is crucial for extending equipment life and ensuring industrial safety.

[0004] Traditional bearing fault diagnosis methods rely on domain expert experience and manual feature engineering, resulting in high costs in terms of manpower and time. While existing deep learning methods can automatically extract features, they are overly dependent on large-scale balanced datasets. In real-world scenarios, bearing fault conditions are rare, and fault signal data is far less than normal signal data, leading to a severe class imbalance problem. This causes the model to be biased towards predicting normal states, resulting in a significant decrease in diagnostic accuracy (Vashishtha G. et al., “A roadmap to fault diagnosis of industrial machines via machine learning: A brief review,” Measurement, 2025.).

[0005] To address the problem of data scarcity, researchers have proposed data augmentation techniques: Non-generative methods (such as oversampling and SMOTE) rely on existing samples, resulting in limited diversity and generalization ability of generated data; Generative methods (such as GANs and VAEs) suffer from training instability (GANs mode collapse) and fuzzy generated samples (VAEs latent distribution constraints); While mainstream diffusion models can generate high-quality samples, their inference process relies on random sampling, resulting in high computational costs. Furthermore, they require auxiliary classifiers to achieve conditional generation, leading to high training complexity. They also tend to ignore the spatiotemporal and conditional dependencies in signals, making it difficult to meet the needs of real-time industrial diagnostics (Han P. et al., IEEE Transactions on Automation Science and Engineering, 2025).

[0006] Therefore, how to improve model training efficiency and inference speed while ensuring data synthesis quality and diagnostic accuracy, and at the same time fully explore the spatiotemporal correlation of fault signals, has become an urgent problem to be solved in the field of bearing fault diagnosis. Summary of the Invention

[0007] This disclosure provides at least one method for generating bearing fault diagnosis data based on spatiotemporal condition information fusion. In a first aspect, embodiments of this disclosure provide a bearing fault diagnosis method based on a spatiotemporal-conditional information fusion classifier-free guided flow model (SCIF-CFGRF), the method comprising: Data preprocessing stage: Normalize the bearing vibration signal, use a sliding window to segment the samples, and construct training datasets with different balance ratios. Model training phase: The UNet backbone network of SCIF-CFGRF is trained using the constructed dataset. This network includes a Continuous Wavelet Transform Feature Extraction Network block (CFEN-block) and a Residual Seaformer block (RS-block) to fuse the spatiotemporal features of the signal with conditional information. An optimized loss function designed specifically for Rectified Flow (RF) is used to accelerate convergence and improve training stability.

[0008] Data synthesis stage: The trained UNet network is embedded into the inference sampler. Spatiotemporal conditional information fusion without a classifier guidance mechanism and the Quality Enhancer block (QE-block) are used, combined with the Euler synthesis method of rectified flow, to efficiently synthesize high-quality bearing fault data. Diagnostic model evaluation: Synthetic data is added to the original dataset to balance samples from each class, used to train downstream classifiers (such as VGG), and metrics such as accuracy (ACC), F1 score, and G-mean are evaluated on the test set. In an optional implementation, the working mechanism of the CFEN-block includes: The continuous wavelet transform (CWT) is used to convert a one-dimensional time-domain signal into a two-dimensional time-frequency domain feature. Two-dimensional time-frequency feature heatmaps are generated through convolutional layers. The rectified flow velocity field is generated by combining the embedded vector (fusion of time step and category label) and then a high-dimensional feature representation is extracted. In one alternative implementation, the RS-block is used to enhance feature extraction and information fusion: The fusion embedding of reception conditions and time steps. Support for auxiliary information integration is achieved through processing with multiple custom residual blocks (RS-blocks). By leveraging Seaformer to capture global context and fine-grained local features, a high-dimensional rectified flow feature representation that integrates spatiotemporal condition information is output. In one optional implementation, the optimization loss function is designed based on the path velocity characteristics of the rectified flow, and its mathematical expression is as follows:

[0009] in For the initial noise distribution, For the target data distribution, For network learning speed field.

[0010] In one optional implementation, the SCIF-CFGRF operates as follows during the inference phase: Using the trained UNet network, input the real class labels respectively. and mask label Obtain conditional prediction and unconditional prediction . The final generated feature representation is calculated using the following formula:

[0011] in This is the guiding coefficient. In one optional implementation, the QE-block is used to filter defective samples, and its working principle is as follows:

[0012] in denoted as the threshold ratio, and CS as the cosine similarity.

[0013] Secondly, this disclosure also provides a bearing fault diagnosis device based on SCIF-CFGRF, the device comprising: Data preprocessing module: used for signal normalization, segmentation, and dataset construction. Model training module: Includes the UNet backbone network, integrating CFEN-block and RS-block for high-dimensional feature extraction and spatiotemporal conditional information fusion. Inference sampling module: Embeds a pre-trained UNet network and combines SCIF-CFGRF and QE-block to achieve efficient data synthesis. Evaluation module: Used for synthetic data quality evaluation (CS, PCC) and downstream classifier performance testing (ACC, F1-score, G-mean). Thirdly, embodiments of this disclosure also provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method described above. Fourthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method described above. The beneficial effects of this invention are as follows: the bearing fault diagnosis method and device based on SCIF-CFGRF significantly improves synthesis efficiency and quality by introducing rectified flow to replace the traditional diffusion model; CFEN-block and RS-block effectively fuse the spatiotemporal features and conditional information in the signal, enhancing the authenticity and diversity of the generated data; the specially designed loss function accelerates model convergence, while the SCIF-CFGRF mechanism avoids the use of explicit classifiers, reducing computational costs and alleviating data sparsity problems. Experiments show that under extreme imbalance conditions (class balance ratio of 1:5), this invention achieves a cosine similarity of over 0.94 and a diagnostic accuracy of 99.50% on the CWRU and SEU datasets, while the inference time is only 6.7% of that based on the DDPM method, significantly outperforming existing mainstream methods. Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This disclosure provides a training framework diagram for a rectified flow model based on spatiotemporal conditional information fusion, corresponding to the process of S100 (data preprocessing), S120 (data synthesis), S130 (performance evaluation), and S140 (ablation evaluation) in Embodiment 1. In the diagram, S110 is data acquisition and normalization, S120 is SCIF-CFGRF UNet model training, S130 is the quality evaluation of synthesized data and classification performance, and S140 is the component and hyperparameter ablation experiment. Figure 2 This disclosure provides a flowchart of a high-quality sample generation method based on the SCIF-CFGRF model, corresponding to steps S210 (loading the UNet model), S220 (no classifier guidance), S230 (generating samples using Euler integrals), and S240 (QE-block quality screening) in Embodiment 2. Figure 3 This is a diagram illustrating the overall architecture of SCIF-CFGRF provided in this embodiment of the disclosure. Figure 4 A schematic diagram of a fault diagnosis system based on SCIF-CFGRF is provided for embodiments of this disclosure; each module corresponds to the preprocessing module, data generation module, and diagnosis module of claim 6; the display and alarm module is used to trigger equipment maintenance warnings when the diagnostic accuracy is lower than 80%; Detailed Implementation To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] As used herein, the phrases “in one embodiment,” “according to one embodiment,” “in some embodiments,” etc., generally refer to the fact that a particular feature, structure, or characteristic following the phrase can be included in at least one embodiment of this disclosure. Therefore, a particular feature, structure, or characteristic can be included in more than one embodiment of this disclosure, such that these phrases do not necessarily refer to the same embodiment. As used herein, the terms “example,” “exemplary,” etc., are used to “serve as an example, instance, or illustration.” Any implementation, aspect, or design described herein as “example” or “exemplary” is not necessarily to be construed as preferred or superior to other implementations, aspects, or designs. Rather, the use of the terms “example,” “exemplary,” etc., is intended to present concepts in a specific manner.

[0018] To facilitate understanding of this embodiment, a model construction method disclosed in this invention will first be described in detail. This invention relates to a method, apparatus, medium, and electronic equipment for constructing a rectified flow model based on spatiotemporal condition information fusion. It is mainly applied to fault diagnosis and data augmentation tasks for industrial equipment (such as bearings, gearboxes, and motors) under conditions of extreme data imbalance. This method, by constructing an SCIF-CFGRF, achieves high-fidelity modeling and efficient synthesis of fault signals, significantly improving the accuracy and robustness of downstream diagnostic models. Example 1: Model Construction and Training Methods See Figure 1 This embodiment provides a rectified flow model training method based on spatiotemporal condition information fusion, including the following steps: See Figure 1 This embodiment provides a training method for a rectified flow model (SCIF-CFGRFUNet) based on spatiotemporal conditional information fusion. The specific steps are as follows: S100: Data Preprocessing. This step uses the equipment fault datasets from CWRU (Case Western Reserve University) and SEU (Southeast University) as the core data source, replacing single industrial equipment signal acquisition. The specific operation is as follows: 1. Data Acquisition: Extract raw vibration signal data (such as accelerometer signals) from industrial equipment from CWRU and SEU datasets. The sampling frequency is adapted to 12kHz, 30kHz, etc. according to the characteristics of the dataset to ensure the diversity and representativeness of data sources. 2. Standardization processing: Normalize the original vibration signal to eliminate dimensional differences between different equipment and operating conditions, unify the data distribution range, and lay the foundation for subsequent feature extraction; 3. Sliding window and segmentation: The standardized signal is segmented into samples of fixed length using the sliding window method, ensuring that each sample contains complete fault feature information and avoiding feature fragmentation; 4. Training / Test Set Splitting: The samples are split into 80% training set and 20% test set to ensure that the test set and training set data are independently distributed and to objectively reflect the model's generalization ability; 5. Construct an imbalanced training set: Based on the fault type label, construct multiple training subsets with different imbalance ratios, such as 1:400, 1:200, 1:100, 1:50, 1:25, 1:10, 1:5, etc., to simulate the real imbalanced scenario where fault samples are scarce in industrial settings.

[0019] S120: Data Synthesis: This step revolves around the SCIF-CFGRF UNet model, generating high-quality synthetic data through model training and inference sampling to supplement the scarce fault samples in the imbalanced training set. The SCIF-CFGRF UNet model backbone network is constructed. The model is based on UNet and integrates three key modules to achieve deep fusion of spatiotemporal conditional information and flow model: 1. CFEN-block: The preprocessed one-dimensional time-domain vibration signal is converted into a two-dimensional time-frequency map through continuous wavelet transform (CWT), which maps the time-domain features to the time-frequency domain and highlights the local correlation of fault features. High-dimensional time-frequency features are further extracted through 3 convolutional layers (3×3 kernel size, stride 1, padding=1) to filter out noise interference; Generate initial feature representations for rectified flow modeling, and output feature maps with dimensions adapted to the UNet downsampling layer for use by subsequent modules.

[0020] 2. RS-block: Receive time-frequency feature map from CFEN-block, and simultaneously input conditional embedding information (including fault category label and signal acquisition time step information); By fusing spatiotemporal features and conditional information through residual connections (to avoid gradient vanishing) and Seaformer networks (to capture long-range spatiotemporal dependencies), the model's ability to learn the correlation between fault type, signal features, and time dimension is enhanced. The output is a fused high-dimensional feature representation, which is used to construct the velocity field of the rectified flow, ensuring that the direction of the flow model generation is consistent with the distribution of the real data.

[0021] 3. Embedding Mechanism: The fault category label (one-hot encoding) and time step information (normalized timestamp) are merged into a conditional embedding vector through a concatenation operation; This embedding vector is inserted into the downsampling layer, bottleneck layer, and upsampling layer of UNet to enable the propagation of conditional information across all layers of the network, thereby enhancing the model's ability to distinguish signal features of different fault categories and different time stages.

[0022] Custom RF-based Loss construction: To adapt to the characteristics of the rectified flow model, a loss function based on rectified flow path consistency is proposed. By minimizing the difference between the model's predicted velocity field and the actual path velocity field, this ensures that the generated path aligns with the target data distribution, thereby improving training stability and convergence speed. The mathematical expression is:

[0023] in, Indicates from real samples To noise sample The linear interpolation path, The velocity field predicted by the SCIF-CFGRF UNet network. The time steps are uniformly distributed in [0,1].

[0024] Model training and inference sampling: 1. Model Training: Input the imbalanced training set constructed in S100 into the SCIF-CFGRF UNet model, use the AdamW optimizer, set the learning rate to 0.0001, batch size to 64, and training epochs to 500; calculate the loss value through the above custom loss function, backpropagate to update the network parameters, so that the model gradually learns the mapping relationship between noise samples and real samples, while integrating spatiotemporal and conditional information.

[0025] 2. Inference Sampling: After training, the SCIF-CFGRF and Quality Enhancement (QE) inference sampler is enabled. With noisy samples as input, combined with the conditional embedding vector corresponding to the target fault category, synthetic data consistent with the feature distribution of real samples is generated. 3. Data Supplementation: The generated high-quality synthetic data is added to the original imbalanced training set to balance the number of samples in each fault category and alleviate the data imbalance problem.

[0026] S130: Data and Model Performance Evaluation. This step evaluates the model from two dimensions: the quality of the synthesized data and the performance of downstream classification, to ensure model effectiveness. 1. Quality assessment of synthetic data: The correlation coefficient (PCC) and cosine similarity (CS) are used as evaluation indicators to calculate the similarity between synthetic data and real data at the feature level. PCC reflects the degree of linear correlation between the two, and CS reflects the consistency of the feature vector direction. The closer the index value is to 1, the higher the quality of synthetic data. 2. Downstream classification model training: The original training set and synthetic data are used as a new training set. Three classic convolutional neural networks, VGG-11, VGG-13, and VGG-16 (as baseline classifiers), are input to train the downstream fault classification model. 3. Classification performance evaluation: Using the 20% test set divided in S100, the accuracy (ACC), F1 score (F1-score), and geometric mean (G-mean) are used as evaluation metrics. ACC reflects the overall classification accuracy, F1-score balances precision and recall, and G-mean measures the balance of classification performance across different categories, comprehensively evaluating the model's performance in the fault classification task.

[0027] S140: Ablation Evaluation. To verify the effectiveness of each component and hyperparameter of the SCIF-CFGRF UNet model, an ablation experiment was designed: 1. Hyperparameter ablation: For the key hyperparameters of the model, adjust the parameter values ​​one by one, repeat the S120-S130 process, evaluate the impact of hyperparameters on model training efficiency, synthetic data quality, and classification performance, and determine the optimal parameter combination; 2. Component ablation: Remove key components from the model one by one, construct a "model missing a single component", and compare the performance difference between the complete model and the ablated model; 3. Ablation performance evaluation: The evaluation indicators (ACC, F1-score, G-mean, PCC, CS) in S130 are used to quantitatively analyze the contribution of each hyperparameter and component, providing a basis for model structure optimization.

[0028] Example 2: Reasoning and Sample Generation Methods See Figure 2 This embodiment provides a high-quality sample generation method based on the SCIF-CFGRF model, including the following steps: S210: Load the trained UNet model. Load the UNet model trained in Example 1 into the inference device to prepare for sample synthesis. S220: Execute Classifier-Free Guidance (CFG), input complete class labels, and obtain the conditional prediction feature vpred_cond; Input the mask category label (e.g., an all-zero vector) to obtain the unconditional prediction feature vpred_uncond; The fusion features are generated using the following formula:

[0029] Here, scale is the guiding factor, and the recommended value is 1.0. S230: Samples are generated based on the Euler integral method. The Euler method is used to integrate the fused features to generate a two-dimensional time-frequency feature map. The integration formula is:

[0030] Time step defined as

[0031] in, The step size is 0.1, with a recommended value of 0.1, and the total number of integration steps is 10. S240: QE-block, which evaluates the quality of generated samples by comparing them with real samples using cosine similarity (CS), setting a threshold to filter out low-quality samples and retain high-fidelity synthetic data. Example 3: Downstream Classification and Diagnostic Methods See Figure 3 This embodiment provides a fault diagnosis method based on synthetic samples, including the following steps: S310: Construct a balanced training set by fusing the high-quality synthetic samples generated in Example 2 with real samples to construct a class-balanced training dataset. S320: Training the downstream classifier, using a convolutional neural network (such as VGG-11, VGG-13, VGG-16) as the downstream classifier. The input is a two-dimensional time-frequency graph, and the output is the fault category. The training process uses the cross-entropy loss function, the optimizer is Adam, the learning rate is 0.001, and the batch size is 64. S330: Evaluate classification performance by assessing the classifier's accuracy, F1-score, and G-mean on the test set. Experimental results show that even under an extreme imbalance of 1:400, this method still achieves over 80% classification accuracy, significantly outperforming mainstream methods such as DDPM, DDIM, and ACGAN. Example 4: System Device and Hardware Implementation See Figure 4 This embodiment provides a fault diagnosis system device based on SCIF-CFGRF, including: a data acquisition module for acquiring vibration signals from industrial equipment; a preprocessing module for signal normalization, slicing, and dataset construction; a model training module for training the SCIF-CFGRF backbone network; a sample generation module for generating synthetic samples during the inference stage; a classification and diagnosis module for training downstream classifiers and outputting diagnostic results; and a display and alarm module for displaying diagnostic results and triggering maintenance warnings. The above modules can be deployed on edge computing devices, industrial PCs, or cloud servers, supporting real-time diagnostics and offline training. Example 5: Experimental Verification and Performance Comparison To verify the effectiveness of the invention, the inventors conducted extensive experiments on two publicly available bearing fault datasets, CWRU and SEU, covering 14 imbalance ratios (1:400 to 1:5), and compared the results with mainstream methods such as DDPM, DDIM, ACGAN, SAGAN, and LSGAN. In terms of synthesis quality, the samples generated by this method show a cosine similarity (CS) exceeding 0.94 and a Pearson correlation coefficient (PCC) exceeding 0.83 compared to the comparative methods, both superior to the comparison methods. Regarding diagnostic accuracy, this method achieves over 80% accuracy even under extreme imbalance conditions (class balance ratio of 1:400), significantly higher than other methods. In terms of synthesis efficiency, this method requires only 2.86 seconds to generate 64 samples, far less than DDPM's 42.57 seconds, representing an approximately 15-fold improvement in inference speed.

[0032] Abstract of experimental results:

[0033] Example 6: Application of generator bearing fault diagnosis in power systems Data source: Vibration data of bearings from a 330MW steam turbine generator unit in a power plant, with a sampling frequency of 10kHz; Operating conditions: Simulate different loads (50%, 75%, 100%) and different fault types (inner ring, outer ring, rolling element); imbalance ratio, set to 1:200 (fault: normal); Experimental results: The cosine similarity between the synthetic samples and the real samples reached 0.945; the VGG-16 classification accuracy reached 98.35%; and the inference time was only 6.9% of DDPM. This invention has excellent diagnostic performance and real-time performance in power system scenarios, and is suitable for monitoring the status of critical equipment.

[0034] Example 7: Application of bearing fault diagnosis in industrial conveying equipment Data source: Vibration signal of bearings in conveyor equipment on a car assembly line; Operating conditions: Simulate different operating speeds (0.5m / s, 1.0m / s, 1.5m / s) and loads (no load, full load); imbalance ratio, set to 1:100; Experimental results: The PCC of the synthetic samples reached 0.921, and the CS reached 0.836; the downstream classifier accuracy was 97.68%; the model converged in only about 60 training rounds. This invention is applicable to bearing fault diagnosis under multiple working conditions and complex industrial environments, and has good generalization ability.

[0035] In summary, this invention provides a method and system for constructing a rectified flow model based on spatiotemporal conditional information fusion. This scheme introduces a CFEN-block to convert one-dimensional vibration signals into two-dimensional time-frequency feature maps, utilizes an RS-block to achieve efficient fusion of spatiotemporal conditional information, and innovatively employs a rectified flow path consistency loss function to improve training stability. During the inference phase, a classifier-free guidance mechanism and the Euler integral method are used to achieve efficient synthesis of high-quality samples, significantly improving the fault diagnosis accuracy under extreme imbalance conditions. Experimental results show that our proposed method significantly outperforms traditional methods on the CWRU and SEU bearing datasets: under extreme imbalance conditions of 1:400, the cosine similarity between the synthesized samples and the real samples exceeds 0.94, and the downstream diagnostic accuracy reaches 80.27% (CWRU) and 81.15% (SEU), while the inference efficiency is approximately 15 times higher than that of the DDPM method. These advantages make this invention of significant application value in the field of industrial equipment fault diagnosis.

[0036] Furthermore, without departing from the scope of this disclosure, the discrete or individual technologies, systems, subsystems, and methods described and illustrated in the various embodiments may be combined or integrated with other systems, modules, technologies, or methods. Other items shown or discussed as coupled may be directly connected or indirectly coupled or communicated via some interface, device, or intermediate component in an electrical, mechanical, or other manner. Those skilled in the art can identify other examples of changes, substitutions, and modifications without departing from the spirit and scope of this disclosure.

Claims

1. A method for generating bearing fault data, characterized in that, include: One-dimensional vibration signals are converted into two-dimensional time-frequency maps; the two-dimensional time-frequency maps are input into a UNet network, which fuses spatiotemporal condition information and outputs a rectified flow velocity field; the velocity field is conditionally sampled based on a classifier-free guidance mechanism to obtain synthetic samples; the synthetic samples are quality-screened, and samples with similarity to real samples higher than a threshold are retained to obtain augmented data.

2. The method according to claim 1, wherein, The two-dimensional time-frequency graph is obtained by continuous wavelet transform (CWT), with the wavelet basis being the Morlet function and the scale number not less than 64.

3. The method according to claim 1 or 2, wherein, The rectified flow model uses the Euler method for integration in 10 steps with a step size ε=0.

1.

4. The method according to any one of claims 1 to 3, wherein, The classifier-free guidance mechanism obtains conditional predictions and unconditional predictions by inputting the conditional embedding vector and the mask zero vector into the network, respectively, and then fuses them according to the guidance coefficient.

5. The method according to any one of claims 1 to 4, wherein, The quality screening module sets a cosine similarity threshold of ≥0.8, and synthetic samples below the threshold are discarded.

6. A bearing fault diagnosis system, characterized in that, include: The preprocessing module is used to normalize the original vibration signal and perform sliding window segmentation; The data generation module is configured to perform the method according to any one of claims 1 to 5 and output a balanced dataset; the diagnosis module uses a convolutional neural network to train the balanced dataset and outputs the fault category.

7. The system according to claim 6, wherein, The diagnostic module is one of the VGG-11, VGG-13, or VGG-16 networks.

8. The system according to claim 6, wherein, The system achieves a diagnostic accuracy of ≥80% when using the VGG-16 classifier under a class imbalance ratio of 1:

400.

9. The system according to any one of claims 6 to 8, wherein, The system is deployed on edge computing devices for real-time data acquisition and diagnosis.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 5.

Images