Crossed roller bearing data augmentation and fault diagnosis method based on diffusion generation

Abstract

This invention discloses a data augmentation and fault diagnosis method for crossed roller bearings based on diffusion generation. The method involves inputting the vibration signal to be detected into a fault diagnosis classification network trained on an enhanced training dataset to obtain the corresponding operating state or fault category. Specifically, it acquires vibration signals of the crossed roller bearing under different operating states and constructs a few-sample original dataset based on these vibration signals. The vibration signals in the few-sample original dataset are mapped to high-dimensional feature representations. A diffusion generation model performs forward noise addition and backward noise reduction sampling on the high-dimensional feature representations to generate high-dimensional feature samples corresponding to the conditions. The high-dimensional feature samples output by the diffusion generation model are mapped to vibration signal samples, and the generated vibration signal samples are merged with the few-sample original dataset to obtain the enhanced training dataset. This invention achieves effective augmentation of the fault dataset under extremely limited sample conditions, improving the stability and recognition capability of the diagnostic model.

Description

Technical Field

[0001] This invention belongs to the field of bearing fault diagnosis technology, specifically relating to a method for data enhancement and fault diagnosis of cross roller bearings based on diffusion generation. Background Technology

[0002] Crossed roller bearings are widely used in high-precision rotating systems such as industrial robot joints, CNC rotary tables, precision machining equipment, and aerospace equipment due to their advantages of compact structure, high load-bearing capacity, high rotational accuracy, and high rigidity. Crossed roller bearings typically operate under complex loads and continuous running conditions for extended periods, and their operating status directly affects the accuracy, reliability, and safety of the entire system.

[0003] In practical engineering applications, obtaining fault samples for crossed roller bearings presents significant challenges. On one hand, bearing faults are sudden and unrepeatable, making large-scale data collection difficult under laboratory conditions. On the other hand, crossed roller bearings require high assembly precision, and disassembly and inspection are costly; large-scale fault data collection is typically not feasible on-site, resulting in a very limited number of fault samples available for model training. Traditional supervised learning-based fault diagnosis methods usually rely on large amounts of labeled data, which can lead to unstable model training and insufficient generalization ability under limited sample conditions.

[0004] In recent years, some studies have attempted to alleviate the data shortage problem through feature extraction or few-sample learning methods. However, most of these methods focus on discriminative modeling and fail to model and expand the fault vibration signals from the perspective of data distribution. Therefore, they are still insufficient to significantly improve diagnostic performance under very limited sample conditions. Thus, there is an urgent need for a technical solution that can effectively expand the cross roller bearing fault dataset and improve the stability and accuracy of fault diagnosis under limited sample conditions.

[0005] The "AUV Thruster Fault Diagnosis Method Based on Wavelet Entropy and APN" disclosed by Luo Guangyu et al. in Chinese patent document CN121144793A enhances the original signal through optimal wavelet reconstruction scale and combines attention mechanism and prototype network to achieve fault diagnosis under small sample conditions, thus improving the diagnostic performance in small sample scenarios to a certain extent. However, the diagnostic model of this invention application still belongs to the discriminative learning framework, whose core objective is to learn the classification boundary from a limited number of samples, without modeling the generation distribution of fault vibration signals. When the number of fault samples is extremely small, this method is difficult to fundamentally expand the scale of training data, and the model performance is still limited by the number and distribution characteristics of the original samples. At the same time, this method mainly relies on feature space measurement and prototype construction, and its adaptability to new operating conditions or scenarios where the sample distribution changes is limited. Summary of the Invention

[0006] To address the issues of unstable training and limited diagnostic performance in existing cross-roller bearing fault diagnosis methods under conditions of extremely limited fault samples, this invention provides a cross-roller bearing data augmentation and fault diagnosis method based on diffusion generation. This method enhances the stability and recognition capability of the fault diagnosis model in scenarios with few samples by modeling and generating vibration signals of the cross-roller bearing under limited sample conditions, effectively enhancing the original fault dataset. The enhanced dataset is then used to improve the recognition capability and stability of the fault diagnosis model.

[0007] To achieve the objectives of this invention, the present invention provides a method for enhancing and diagnosing cross-roller bearing data based on diffusion generation. This method inputs the vibration signal to be detected into a fault diagnosis classification network trained using the enhanced training dataset to obtain the corresponding operating state or fault category. The enhanced training dataset is obtained through the following steps: Vibration signals of crossed roller bearings under different operating conditions are obtained, and a small sample original dataset is constructed based on the vibration signals; Map vibration signals from a small sample original dataset into a high-dimensional feature representation; A diffusion generation model containing an asymmetric U-Net network structure is used to perform forward noise addition and reverse noise reduction sampling processes on the high-dimensional feature representation to generate high-dimensional feature samples with corresponding conditions. The high-dimensional feature samples output by the diffusion generation model are deconvolved to map them into vibration signal samples. The generated vibration signal samples are then merged with the original dataset with fewer samples to obtain the enhanced training dataset.

[0008] This invention addresses the challenges of obtaining fault samples in crossed roller bearings and the extremely limited number of samples in each category by introducing a diffusion generation model to model and generate vibration signals from crossed roller bearings. By learning the latent distribution characteristics of the vibration signals under limited sample conditions, it generates synthetic samples that accurately reflect the characteristics of actual faults, thus effectively enhancing the original dataset. By combining the generated samples with the original samples to construct an enhanced training dataset, this invention significantly improves the training stability and recognition capability of the fault diagnosis model without relying on a large amount of newly labeled data. This makes it more suitable for intelligent operation and maintenance scenarios in practical engineering where crossed roller bearing fault samples are scarce and acquisition costs are high.

[0009] Furthermore, when constructing the few-sample original dataset, vibration signal data of crossed roller bearings under different operating conditions are collected. Data segments of the same length are extracted from continuous vibration signals as samples to construct the few-sample original dataset. The number of vibration signal samples corresponding to each type of operating condition is small, and each sample contains at least one data point of the rotating machinery completing one full rotation to ensure the integrity and comparability of the samples.

[0010] Furthermore, in the process of mapping the vibration signals in the few-sample original dataset to a high-dimensional feature representation, wavelet kernel convolution processing is performed on the one-dimensional vibration signals in the few-sample original dataset, and the Mexican hat wavelet is used as the wavelet basis function to map the original vibration signals to a high-dimensional feature representation with 96 channels, so as to enhance the ability to express local impact features and multi-scale time-frequency information in the vibration signals.

[0011] Furthermore, the diffusion generation model includes a downsampling path, a central module, and an upsampling path. The input is the high-dimensional feature representation. Under the constraint of the fault category condition, the high-dimensional feature representation is subjected to forward noise addition and reverse noise reduction sampling processes. The downsampling path, the central module, and the upsampling path are all formed through residual modules. The downsampling path and the upsampling path have the same number of layers. At the same resolution level, the number of residual modules in the upsampling path is greater than the number of residual modules in the downsampling path, so as to enhance the ability to reconstruct detailed features during the generation process, thereby generating high-dimensional feature samples consistent with the corresponding operating state.

[0012] Preferably, two residual modules are set for each layer of the upsampling path, while one residual module is set for each layer of the downsampling path, in order to enhance the ability to reconstruct details.

[0013] Furthermore, a deconvolution layer is set at the output of the diffusion generation model to map the generated high-dimensional feature samples into one-dimensional vibration signals. The generated one-dimensional vibration signal samples are then merged with the original small number of vibration signal samples to construct an enhanced training dataset, thereby effectively expanding the number of fault samples.

[0014] Furthermore, a one-dimensional deconvolution layer is set at the output of the diffusion generation model to map the generated high-dimensional feature samples into one-dimensional vibration signal samples, so as to ensure that the generated samples are consistent with the dimensions of the original input signals.

[0015] Furthermore, the enhanced training dataset is used to train the fault diagnosis classification network. By minimizing the difference between the output of the classification model and the true state of the sample, the model parameters are optimized to complete the training of the fault diagnosis classification network.

[0016] Furthermore, in the forward noise addition and reverse noise reduction sampling process, the forward noise addition process will reduce the original high-dimensional feature samples... Convert to noisy features Their relationship satisfies:

[0017] in, For random noise that follows a standard normal distribution, For diffusion time step Changing noise scheduling parameters; In the reverse denoising process, the noise or equivalent residual is predicted using a neural network to obtain the denoising result and generate it by sampling.

[0018] Furthermore, during the training phase of the diffusion generation model, the network parameters are optimized using noise prediction loss, preferably mean squared error loss.

[0019] Furthermore, the generated one-dimensional vibration signal sample set is merged with the small sample original dataset to construct an enhanced training dataset, thereby expanding the training sample size without adding a large amount of manual collection and annotation costs.

[0020] Furthermore, an enhanced training dataset is used to train the fault diagnosis classification network. This makes its output class prediction probability , This represents the sample feature data input to the fault diagnosis classification network. The model parameters are adjusted by minimizing the classification loss. Optimization is performed using cross-entropy loss. After training, the vibration signal to be tested is input into the fault diagnosis classification network, which outputs the corresponding operating status / fault category identification results.

[0021] The present invention also provides a data enhancement and fault diagnosis device for cross roller bearings based on diffusion generation.

[0022] The present invention also provides a computer device.

[0023] The present invention also provides a computer-readable storage medium.

[0024] Compared with the prior art, the beneficial effects of the present invention are at least as follows: (1) This invention addresses the problems of difficulty in obtaining fault samples for crossed roller bearings and unstable training of fault diagnosis models under conditions of extremely small numbers of fault samples for each type. It proposes a few-sample crossed roller bearing fault diagnosis method based on a diffusion generation model, thus solving the limitation of existing technologies in few-sample scenarios from a data perspective. Compared with existing methods that rely on discriminative learning models for direct fault classification, this invention introduces a diffusion generation model to model and generate the vibration signals of crossed roller bearings. This enables the learning of the potential distribution characteristics of vibration signals under few-sample conditions, generating synthetic samples consistent with real fault characteristics, effectively enhancing the original dataset, and significantly improving the training stability and recognition accuracy of the fault diagnosis model.

[0025] (2) This invention combines wavelet kernel convolution with an asymmetric U-Net network structure to effectively model the local impact components and multi-scale time-frequency features in the vibration signals of crossed roller bearings, reducing reliance on manual feature design and expert experience. By generating samples and constructing an enhanced training dataset together with the original samples, the fault diagnosis model can obtain more sufficient training data without increasing the additional labeling cost, thus improving the method's adaptability to sample scarcity and changing operating conditions in practical engineering. At the same time, this invention does not require changing the acquisition method of the original vibration signal, is easy to integrate with existing online monitoring systems, and is suitable for condition monitoring and intelligent operation and maintenance of various rotating machinery such as bearings, gearboxes, and motors, with good engineering application value. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the diffusion-based cross roller bearing data enhancement and fault diagnosis method provided in the embodiments of the present invention.

[0027] Figure 2 This is a flowchart illustrating the steps in an embodiment of the present invention.

[0028] Figure 3 This is a network structure diagram of the diffusion model in an embodiment of the present invention.

[0029] Figure 4 This is a flowchart of the fault diagnosis classification in an embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram of the confusion matrix of fault diagnosis classification results in an embodiment of the present invention. Detailed Implementation

[0031] To make the technical solutions and objectives of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific implementation steps described herein are only used to better illustrate the application of the present invention, but the technical features involved in the embodiments of the present invention are not limited thereto.

[0032] Please see Figure 2This invention provides a data augmentation and fault diagnosis method for crossed roller bearings based on diffusion generation, comprising five stages: The first stage is the data acquisition and sample construction stage, which collects vibration signals of the crossed roller bearing under different operating conditions to construct a small-sample original dataset for model training; the second stage is the feature mapping stage, which processes the original one-dimensional vibration signal through wavelet kernel convolution to map the vibration signal into a high-dimensional feature representation to enhance the time-frequency features of the local impact component in the vibration signal; the third stage is the diffusion generation stage, which constructs a diffusion generation model containing an asymmetric U-Net network structure, and performs forward noise addition and backward noise reduction sampling on the high-dimensional feature representation under fault category constraints to generate feature samples consistent with the corresponding operating conditions; the fourth stage is the data augmentation stage, which maps the generated high-dimensional feature samples into one-dimensional vibration signals through deconvolution, and merges the generated samples with the original small-sample dataset to construct an enhanced training dataset; the fifth stage is the fault diagnosis stage, which uses the enhanced training dataset to train a fault diagnosis classification network to achieve fault identification of the crossed roller bearing.

[0033] Specifically, the method includes the following steps: Step 1: Data Acquisition and Construction of a Small Sample Dataset. Vibration signals of the crossed roller bearing under different operating conditions were collected. , Represent continuous physical time variables and construct a small sample original dataset for model training.

[0034] This step includes the following sub-steps: Step 1.1: Obtain the one-dimensional vibration acceleration signal during the operation of the crossed roller bearing using a vibration sensor. Divide the continuous one-dimensional vibration acceleration signal into equal segments of fixed length to form sample segments. The expression is: , In the formula, For sample length, Indicates the sample segment number; Indicates the first The initial sampling time of each sample segment on the vibration signal time axis; This indicates the amplitude of the vibration signal at the corresponding sampling time.

[0035] In some embodiments of the present invention, to ensure that each sample segment contains information about at least one complete rotation of the rotating machinery, the sample length satisfies the following requirements: ,in Sampling frequency, For frequency conversion.

[0036] Step 1.2: Normalize the sample segment using zero-mean, unit-variance standardization:

[0037] in, This represents a data sample segment standardized with zero mean and unit variance, as opposed to a sample segment directly truncated from the original signal. ; and These are the sample mean and standard deviation, respectively. To prevent division by zero of small constants.

[0038] Step 1.3: Under the condition of few samples, only select one of each type of running state. A few sample segments are used to construct a small sample original dataset. ,in, Labels for operating status or fault category. This represents the total number of sample segments in the original dataset with fewer samples. ,in This indicates the total number of operating states or fault categories. In some embodiments of the invention, .

[0039] In some embodiments of the present invention, in order to facilitate the construction of the subsequent diffusion generation model structure, the collected vibration acceleration signal is divided into multiple sample segments of length 1024. A small amount of data from each class is taken to form the training samples of the subsequent diffusion generation model, thus forming a small sample original dataset. 300 sample segments from each class are taken to form the test data of the subsequent fault diagnosis classification network.

[0040] Step 2: Wavelet kernel convolution feature mapping. Wavelet kernel convolution is applied to the vibration signals in the few-sample original dataset to map the vibration signals into a high-dimensional feature representation.

[0041] The Mexican hat wavelet is used as the wavelet basis function, and its expression is: , The wavelet function, known as the Mexican hat wavelet function, enhances the time-frequency feature representation of local impact components in vibration signals. Wavelet kernel convolution maps a one-dimensional vibration signal to a high-dimensional feature representation. , For the number of channels, a high-dimensional feature representation The first in The characteristics of each channel are:

[0042] In the formula, The first wavelet kernel convolution mapping is used to obtain the second wavelet kernel. Each feature channel in time index The characteristic response value at that location; This indicates the original one-dimensional vibration signal at the time index. The signal amplitude at that location.

[0043] Characteristic response value In the Specific instances on each sample segment for:

[0044] In the formula, Indicates the first Each sample segment has a time index relative to the current time. Forward offset The signal amplitude at each sampling point; It is a local discrete index used to traverse the effective support interval of the convolution kernel; For convolution kernel, convolution kernel Mexican hat wavelet function By constructing scale and translation parameters, the enhanced expression of multi-scale local impact components and time-frequency information can be achieved.

[0045] In some embodiments of the present invention, the convolution kernel The expression is:

[0046] In the formula, is the scaling parameter, used to control the scaling of the wavelet function; The translation parameter controls the position of the wavelet function on the time axis. Different scale parameters can be set. This allows for the acquisition of multi-scale wavelet convolution kernels, thereby enhancing the ability to express the local impact components and time-frequency features of vibration signals.

[0047] In this embodiment of the invention, wavelet kernel convolution is used to process the one-dimensional vibration signal. The Mexican hat wavelet is used as the wavelet basis function. Through multi-scale convolution operation, the original one-dimensional vibration signal is mapped into a high-dimensional feature representation, so as to enhance the expressive ability of local impact components and multi-scale time-frequency features in the vibration signal.

[0048] Step 3: Diffusion Generative Model Construction and Feature Generation. A diffusion generative model with an asymmetric U-Net network structure, including downsampling paths, a central module, and upsampling paths, is constructed. The model input is the high-dimensional feature representation obtained in Step 2. Under fault category constraints, forward noise addition and backward denoising sampling processes are performed on the high-dimensional feature representation to generate high-dimensional feature samples corresponding to the conditions.

[0049] In this embodiment of the invention, the downsampling path, the central module, and the upsampling path together constitute an "encode-bottleneck-decode" neural network structure for implementing the feature modeling and reconstruction process in the diffusion generation model. This structure corresponds to the U-Net network architecture, named after the U-shaped topology of the network. The downsampling path extracts features and compresses resolution layer by layer from the input vibration signal, expanding the receptive field level by level to obtain multi-scale high-level fault features. The central module, located at the lowest resolution layer of the network structure, performs global modeling and information reconstruction on the deep features extracted by the downsampling path, i.e., the multi-scale high-level fault features, to enhance the overall consistency and semantic expressive power of the features, outputting high-level semantic features. The upsampling path restores feature resolution level by level, mapping the high-level semantic features output by the central module back to the high-resolution feature space and fusing them with features from the corresponding layer of the downsampling path, thereby restoring the signal detail information. At each resolution layer... In this process, the features extracted from the downsampling path are transferred to the corresponding level of the upsampling path through the feature copying operation, and are spliced ​​or fused with the features in the upsampling process to form a cross-level information transmission mechanism, thereby realizing the joint modeling of local details and global semantic information.

[0050] In some embodiments of the present invention, the same resolution level In the following case, the number of residual modules in the upsampling path is greater than the number of residual modules in the downsampling path. For example, the number of residual modules set in the downsampling path is... The number of residual modules set in the upsampling path is And satisfy:

[0051] In some embodiments of the present invention, at at least one resolution level In the process, only one residual module is set in each layer of the downsampling path, while two residual modules are set in each layer of the upsampling path to enhance the reconstruction capability of high-frequency detail features during the generation process, thereby forming an asymmetric U-Net structure.

[0052] In some embodiments of the present invention, please refer to the network structure of the diffusion generation model. Figure 3 The downsampling and upsampling paths each have three layers, and the central module has one layer. Each layer of the downsampling path has one residual module, while each layer of the central module and the upsampling path has two residual modules, thus forming an asymmetric U-Net structure to enhance the reconstruction capability of high-frequency detail features during the generation process. It contains two residual modules.

[0053] The number of layers in the downsampling and upsampling paths is determined comprehensively based on the length of the input vibration signal, the target's lowest feature resolution, the requirements for multi-scale feature extraction, and the network's computational complexity. The number of layers in the downsampling path controls the progressive compression of feature resolution and the expansion of the receptive field, while the number of layers in the upsampling path controls the progressive recovery of feature resolution. Preferably, the upsampling path's number of resolution layers corresponds to the number of layers in the downsampling path, facilitating feature replication and fusion connections between corresponding layers. The central module is located between the downsampling and upsampling paths, at the lowest resolution layer, and is used for global modeling and information reconstruction of deep features. Its number of layers is not required to be the same as the downsampling or upsampling paths. In some embodiments of this invention, both the downsampling and upsampling paths are set to three layers, and the central module is set to one layer, thus forming a symmetrical resolution layer correspondence and a bottleneck modeling structure in the middle. The asymmetric U-Net in these embodiments is mainly reflected in the fact that, at the same resolution layer, the upsampling path has more residual modules than the downsampling path, rather than in the difference in the number of resolution layers between the upsampling and upsampling paths.

[0054] In some embodiments of the present invention, the residual module includes a one-dimensional group normalization layer, a Swish activation function, two one-dimensional convolutional layers, a temporal embedding mapping layer, and a Dropout layer. The one-dimensional group normalization layer is used to stabilize the feature distribution, the Swish activation function is used to enhance the nonlinear expressive power, the two one-dimensional convolutional layers are used to extract and fuse one-dimensional temporal features, the temporal embedding mapping layer is used to introduce diffusion time step conditional information, and the Dropout layer is used to suppress overfitting. The above components work together to form a residual module with good stability and expressive power.

[0055] In some embodiments of the present invention, taking the input layer size (8, 96, 1024) as an example (where 8 is the number of samples, 96 is the number of channels, and 1024 is the sample length), the structural parameters of the residual module are shown in Table 1: Table 1 Residual Module Structural Parameters

[0056] In Table 1, p represents the probability that a neuron is randomly set to zero during training. This is used to randomly mask some feature responses during the feature learning phase, thereby reducing the risk of model overfitting and improving the network's generalization ability.

[0057] Furthermore, during the training of the diffusion generation model, forward noise addition and reverse noise reduction are performed on the high-dimensional feature representation. Only the running state or fault category is used as conditional information. Conditional feature vectors are generated through label embedding and multilayer perceptron mapping. These vectors are then introduced into the residual module of the diffusion generation model and fused with multi-scale high-level fault features to achieve feature generation under conditional constraints, thereby generating high-dimensional feature samples corresponding to the conditions.

[0058] In one embodiment of the present invention, the forward noise addition process transforms the original high-dimensional feature samples... Convert to noisy features Their relationship satisfies:

[0059] in, For random noise that follows a standard normal distribution, For diffusion time step Varying noise scheduling parameters.

[0060] In the reverse denoising process, the parameters are used as Neural networks under fault category conditions Predicting noise under constraints yields denoised features. It is represented as:

[0061] Represents a parameter The conditional neural network mapping function, as represented, is used to perform a given diffusion time step. and fault category conditions Under the constraints, the noisy features at the current moment Perform noise modeling and denoising inference.

[0062] In one embodiment of the present invention, during the training phase of the diffusion generation model, noise prediction loss is used to optimize the network parameters, preferably mean squared error loss: ,in This represents the predicted output of the diffusion generation model for noise. This represents the expected value of the joint distribution of the random variables corresponding to their subscripts; This represents the random noise variable injected during the forward diffusion process, which follows a standard normal distribution.

[0063] In one embodiment of the present invention, the diffusion generation model, during the model training phase, uses the Adam optimizer, has a batch size of 8 samples, 1000 noise-adding steps, 800 training epochs, and a learning rate of 0.0005.

[0064] Step 4: Deconvolution Reconstruction and Data Augmentation. The high-dimensional feature samples output by the diffusion generation model are deconvolved to generate high-dimensional feature samples. Mapped to one-dimensional vibration signal samples A one-dimensional vibration signal sample set is obtained. To ensure consistency between the generated samples and the original vibration signals in terms of dimension and data format, This represents a one-dimensional transpose convolution operation. For the transposed convolution kernel parameters, a one-dimensional vibration signal sample The dimension of the sampled data is consistent with the dimension of the original vibration signal sampled in step one; the generated one-dimensional vibration signal sample set is then used. Compared with the small sample original dataset in step 1 The datasets are merged to create an enhanced training dataset. .

[0065] In this embodiment of the invention, a one-dimensional deconvolution layer is set at the output of the diffusion generation model to map the generated high-dimensional feature samples into one-dimensional vibration signal samples, so as to ensure the consistency of the generated samples with the original vibration signals in terms of dimension and data format. Furthermore, the generated one-dimensional vibration signal samples are merged with a small number of original vibration signal samples from step one to obtain an enhanced training dataset, thereby effectively expanding the number of training samples.

[0066] Step 5: Fault Diagnosis Classification Network Training and Recognition. Utilizing the enhanced training dataset... Training a fault diagnosis classification network The fault diagnosis classification network outputs the category prediction probability. , These are the network parameters for the fault diagnosis and classification network. The fault diagnosis and classification network is trained by minimizing the classification loss function. After training, the network is input with all real vibration signals and outputs the corresponding operating status or fault category.

[0067] In one embodiment of the present invention, the fault diagnosis classification network adopts a one-dimensional residual convolutional neural network (ResNet1D) structure, including a feature extractor and a classifier. The feature extractor includes a shallow feature extraction module, a multi-level residual feature extraction module, and a global feature aggregation module. The classifier includes a classification decision module. The fault diagnosis classification network first performs preliminary feature extraction on the original vibration signal through a one-dimensional convolutional layer (i.e., the shallow feature extraction module). The convolution kernel size is 7 and the stride is 2, which is used to capture local temporal patterns within a large time receptive field. Batch normalization and ReLU activation function are used to stabilize the training process and enhance nonlinear expression ability. Subsequently, max pooling operation is used to further compress the time dimension and suppress high-frequency noise. Subsequently, the network sequentially undergoes four residual feature extraction stages (i.e., multi-level residual feature extraction modules). Each stage is composed of multiple stacked one-dimensional residual blocks. As the network deepens, the number of feature channels increases progressively while the temporal resolution gradually decreases, enabling the fault diagnosis classification network to gradually transition from low-level local vibration modes to high-level fault discrimination features. The introduction of residual connections allows the network to maintain good gradient propagation capabilities while deepening, which is beneficial for learning subtle feature changes caused by faults. In some stages, stride convolution and downsampling branches are used to align the feature dimensions to ensure the rationality of residual addition. After deep residual feature extraction, the global feature aggregation module of the fault diagnosis classification network adopts adaptive one-dimensional global average pooling to globally statistically aggregate features in the temporal dimension, thereby obtaining a high-level semantic representation independent of the input length and enhancing the robustness of the model under different operating conditions. Finally, the classification decision module maps the aggregated features to the fault category space through a fully connected layer, outputting the discrimination results of various faults, and realizing the automatic identification and classification of the operating status of mechanical equipment. The fault diagnosis classification network used in this embodiment is existing, and this embodiment does not improve upon it. The enhanced training dataset in this embodiment of the invention can be combined with any fault diagnosis classification network to improve classification accuracy, and is not limited to the network structure used here. In other embodiments, any network such as Wide-Width-Depth Convolutional Neural Network (WDCNN), Convolutional Neural Network (CNN), or Multi-Scale Convolutional Neural Network (MSCNN) can also be used as the fault diagnosis classification network.

[0068] In some embodiments of the present invention, such as Figure 4 As shown, the enhanced training dataset is used to train the fault diagnosis classification network. The model parameters are optimized by minimizing the difference between the output of the fault diagnosis classification network and the actual operating state of the samples. The loss function is the cross-entropy loss. , For the true class label of the sample, This represents the predicted probability of the fault diagnosis classification network for the corresponding category. After the fault diagnosis classification network is trained, the vibration signal to be detected is input into the trained fault diagnosis classification network, which outputs the corresponding operating status or fault category, thereby realizing fault identification of rotating machinery.

[0069] In one embodiment of the present invention, during the model training phase, the fault diagnosis classification network uses the Adam optimizer, with a batch size of 64 samples, 200 training epochs, and a learning rate of 0.001.

[0070] In one embodiment of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the foregoing embodiments.

[0071] In one embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the foregoing embodiments.

[0072] The present invention will be further described below with reference to the accompanying drawings and experimental examples.

[0073] To evaluate the performance of the method proposed in this embodiment, cross roller bearing data under different service conditions were collected for experimental verification. The bearing model used in the cross roller bearing dataset was Kymco-HG20, and its basic parameters and experimental conditions are shown in Table 2. No other running bearings were present on the same platform during signal acquisition. The collected signal data included normal bearing data (H0) and faulty bearing data. The faults were artificially implanted and included roller (roller) faults, inner ring raceway faults, and outer ring raceway faults. For roller faults, rectangular recesses were machined on the rolling surface of the cross roller bearing to simulate roller peeling faults. The machining position was 0.6 mm from the left side of the rectangular recess (taking the average left side distance of the faulty part). For inner and outer ring raceway faults, rectangular recesses were machined at the center of one side of the inner and outer ring raceways to simulate raceway peeling faults. The machining position was at the center of one side of the inner and outer ring raceways (machined only on one side of the raceway). The sampling duration ensured sufficient samples. Specific implantation failure information is shown in Table 3.

[0074] Table 2. Basic parameters and experimental conditions of the Kymco-HG20 crossed roller bearing.

[0075] Table 3 Data Types Collected

[0076] To construct the training dataset for the diffusion generation model, for each bearing condition, each sample was divided into 1024 data points. Ten samples each from normal operating conditions, roller failure conditions, and inner and outer raceway failure conditions of the crossed roller bearing were used as training data for the feature extractor. Wavelet kernel convolution was used to perform feature mapping on the original one-dimensional vibration signal to obtain a high-dimensional feature representation. Based on this high-dimensional feature representation, a diffusion generation model containing an asymmetric U-Net structure was trained. After the diffusion generation model was trained, a reverse denoising sampling process was performed under the constraints of each operating state to generate vibration signal samples consistent with the characteristics of the corresponding operating state. The vibration signal samples generated by the diffusion generation model were merged with a small number of original real vibration signal samples to construct an enhanced training dataset. The enhanced training dataset contained 10 real samples and 190 generated samples for each state type, totaling 200 samples per state type. This enhanced training dataset was then used to train the fault diagnosis classification network. Subsequently, 300 test vibration signals of each category collected from the rotating machinery under different operating conditions were input into the trained fault diagnosis classification network to obtain the corresponding classification results. The classification accuracy statistics are shown in Table 4. As can be seen from Table 4, this embodiment achieves a very high classification accuracy.

[0077] Table 4 Fault Diagnosis Model Classification Results

[0078] In this embodiment, the Adam optimizer is used in the training phase of the diffusion generation model, with a batch size of 8, a diffusion step size of 1000, a training epoch size of 800, and a learning rate of 0.0005; the Adam optimizer is also used in the training phase of the fault diagnosis classification network, with a batch size of 64, a training epoch size of 200, and a learning rate of 0.001.

[0079] To verify the effectiveness of the method in this embodiment under limited sample conditions, a confusion matrix visualization analysis was performed on the fault diagnosis classification results, and the results are as follows: Figure 5 As shown. By Figure 5 It can be seen that after introducing the data augmentation method based on the diffusion generation model, most test samples can be correctly classified, and the proportion of the main diagonal elements of the confusion matrix is ​​relatively high, indicating that the fault diagnosis classification network has a good ability to identify various operating states.

[0080] Further analysis reveals a significant reduction in misclassification of samples across different operating states, particularly in categories with very few samples. These categories demonstrate good stability, exhibiting no obvious class bias or collapse. This indicates that the vibration signal samples generated by the diffusion generation model maintain good consistency with real samples in terms of statistical characteristics, effectively compensating for the inadequacy of training data under limited sample conditions.

[0081] Furthermore, the confusion between different categories in the confusion matrix shows that the fault diagnosis classification network can still maintain a high degree of discrimination for operating states with similar vibration characteristics. This indicates that the constructed enhanced training dataset helps the fault diagnosis classification network learn more discriminative feature representations, thereby improving the overall performance of cross roller bearing fault diagnosis under conditions with few samples.

[0082] Based on the above analysis, the confusion matrix results verify that the method described in this embodiment can effectively improve the recognition accuracy and robustness of the fault diagnosis classification network under conditions of very few samples, further proving the feasibility and effectiveness of the proposed data augmentation method in the fault diagnosis scenario of rotating machinery.

[0083] Experimental results show that the data augmentation method based on the diffusion generation model can effectively expand the scale of training data under conditions of very few samples, improve the training stability and recognition accuracy of the fault diagnosis model, and achieve effective differentiation of different operating states of crossed roller bearings, thus achieving the expected purpose of this invention.

[0084] This invention, through the introduction of a diffusion generation model, enhances the vibration signals of crossed roller bearings with a small sample size, reducing reliance on large amounts of labeled fault data and improving the applicability of the rotating machinery fault diagnosis model in practical engineering applications. After model training, the trained fault diagnosis model can be used to quickly identify the operating status of rotating machinery, aiding in condition monitoring and maintenance decisions for crossed roller bearing equipment in industrial settings.

[0085] It should be noted that although the present invention has been described in detail with reference to specific embodiments, those skilled in the art should understand that any modifications, equivalent substitutions or improvements made to the method of the present invention without departing from the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation, characterized in that, The vibration signal to be detected is input into a fault diagnosis classification network trained using the enhanced training dataset to obtain the corresponding operating status or fault category; the enhanced training dataset is obtained through the following steps: Vibration signals of crossed roller bearings under different operating conditions are obtained, and a small sample original dataset is constructed based on the vibration signals; Map vibration signals from a small sample original dataset into a high-dimensional feature representation; A diffusion generation model containing an asymmetric U-Net network structure is used to perform forward noise addition and reverse noise reduction sampling processes on the high-dimensional feature representation to generate high-dimensional feature samples with corresponding conditions. The high-dimensional feature samples output by the diffusion generation model are deconvolved to map them into vibration signal samples. The generated vibration signal samples are then merged with the original dataset with fewer samples to obtain the enhanced training dataset.

2. The method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation according to claim 1, characterized in that, When constructing a small sample original dataset, the original vibration signal is truncated to a preset length, and the truncated length of the vibration signal sample corresponding to each type of running state is the same, so as to ensure that each sample contains at least one data point that has completed one full rotation.

3. The method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation according to claim 1, characterized in that, In the process of mapping vibration signals in a few-sample original dataset to a high-dimensional feature representation, wavelet kernel convolution processing is performed on the vibration signals in the few-sample original dataset for mapping.

4. The method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation according to claim 1, characterized in that, The diffusion generation model includes a downsampling path, a central module, and an upsampling path. All three are formed by residual modules. The downsampling path and the upsampling path have the same number of layers, and at the same resolution level, the number of residual modules in the upsampling path is greater than the number of residual modules in the downsampling path.

5. The method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation according to claim 4, characterized in that, The diffusion generation model generates vibration signals using operating status or fault category as conditional information. The conditional information is represented by label embedding and mapped by a multilayer perceptron to generate conditional feature vectors. The conditional feature vectors are introduced into the residual module of the diffusion generation model and fused with the multi-scale high-level fault features extracted by the downsampling path to realize the diffusion generation process under conditional constraints and generate high-dimensional feature samples corresponding to the conditions.

6. The method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation according to claim 1, characterized in that, A deconvolutional layer is set at the output of the diffusion generation model to map the high-dimensional feature samples output by the diffusion generation model into vibration signals with the same dimension as the initially acquired vibration signals.

7. The method for data enhancement and fault diagnosis of crossed roller bearings based on diffusion generation according to any one of claims 1-6, characterized in that, In the forward noise addition and reverse noise reduction sampling process, the forward noise addition process will reduce the original high-dimensional feature samples. Convert to noisy features Their relationship satisfies: in, For random noise that follows a standard normal distribution, For diffusion time step Changing noise scheduling parameters; In the reverse denoising process, the parameters are used as Neural networks under fault category conditions Noise is predicted under constraints to obtain denoised features.

8. A data enhancement and fault diagnosis device for cross roller bearings based on diffusion generation, characterized in that, For implementing the method according to any one of claims 1-7, the apparatus comprises the following modules: The few-sample acquisition module is used to acquire vibration signals of crossed roller bearings under different operating conditions, and construct a few-sample original dataset based on the vibration signals; The mapping module is used to map vibration signals in a small sample original dataset into a high-dimensional feature representation; The diffusion generation module is used to perform forward noise addition and reverse noise reduction sampling processes on the high-dimensional feature representation through a diffusion generation model containing an asymmetric U-Net network structure to generate high-dimensional feature samples under corresponding conditions. The data augmentation module is used to perform deconvolution processing on the high-dimensional feature samples output by the diffusion generation model, map the high-dimensional feature samples to vibration signal samples, and merge the generated vibration signal samples with the original dataset with fewer samples to obtain the augmented training dataset. The diagnostic module is used to input the vibration signal to be detected into the fault diagnosis classification network trained with the enhanced training dataset to obtain the corresponding operating status or fault category.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

Images