Rolling bearing fault diagnosis method based on improved OfficientNetV2 network

Abstract

The invention provides a rolling bearing fault diagnosis method based on an improved OfficientNetV2 network, and belongs to the technical field of state monitoring and fault prediction of mechanical equipment, and the method comprises the steps: obtaining a full-life vibration signal of a bearing, generating a time-frequency image through continuous wavelet transform, and constructing an image classification data set based on an end-to-end definition method; constructing a rolling bearing fault diagnosis model based on an improved OfficientNetV2 network, embedding an ECA attention mechanism into a Fused-MBConv module in an early stage of feature extraction, and introducing a CBAM attention mechanism into a deep MBConv module in a core stage of feature extraction; and carrying out collaborative correction on a result obtained by the fault diagnosis model and a K-means clustering result, and judging health and fault states according to'or 'logic fusion. Compared with the prior art, the method can effectively improve the precision, generalization ability and robustness of bearing health stage division.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring and fault prediction technology, specifically a rolling bearing fault diagnosis method based on an improved EfficientNetV2 network. Background Technology

[0002] As a core transmission component of rotating machinery, rolling bearings directly determine the overall safety, reliability, and maintenance costs of the equipment, playing an irreplaceable role in industrial production, rail transportation, and energy equipment. With the upgrading of industrial intelligence, higher demands are placed on the health monitoring and fault diagnosis of rolling bearings. This requires not only accurate fault identification but also reduced diagnostic costs and improved adaptability to varying operating conditions. In particular, it is crucial to effectively detect early, minor faults (such as micro-cracks and slight wear) to provide a reliable basis for predictive maintenance of equipment.

[0003] Currently, rolling bearing fault diagnosis technology is mainly divided into three categories, but it still has significant shortcomings in practical industrial applications, making it difficult to simultaneously meet the core requirements of "low cost, high precision, and strong robustness":

[0004] The first category is the traditional diagnostic method based on human experience thresholds. This type of method relies on the experience of domain experts or historical fault statistics, setting fixed discrimination thresholds for monitoring parameters such as bearing vibration and temperature. Status determination is achieved by comparing real-time monitored values ​​with these thresholds. Its advantages lie in its ease of deployment and fast response time, making it suitable for fault alarms under simple operating conditions. However, its limitations are significant: threshold setting depends on expert subjective judgment, resulting in extremely poor adaptability to variable operating conditions such as load changes and speed fluctuations; it is insensitive to signals from weak faults such as early spalling and micro-cracks, often indicating that the fault has already entered a late stage of development when the threshold is triggered, thus failing to achieve early warning functionality and making it difficult to meet the needs of predictive maintenance for equipment.

[0005] The second category is diagnostic methods based on manual features combined with machine learning. This method is currently the mainstream technology in industrial scenarios, achieving diagnosis through a two-step approach: "signal processing for feature extraction—machine learning for modeling and classification." In the signal processing stage, techniques such as Fourier Transform (FFT), Short-Time Fourier Transform (STFT), and Wavelet Packet Decomposition (WPD) are commonly used to extract manually designed features from vibration signals, including time-domain (peak value, kurtosis), frequency-domain (center frequency), and time-frequency-domain (wavelet entropy). In the modeling stage, models such as Support Vector Machine (SVM), Random Forest (RF), and Extreme Learning Machine (ELM) are used for classification. Compared to traditional threshold methods, its diagnostic accuracy is improved, but it relies heavily on expert selection and optimization of features, resulting in high manual costs. Furthermore, in high-noise environments such as those with electromagnetic interference from motors, or in scenarios with varying loads and speeds, the robustness of manually designed features is insufficient, leading to a high rate of missed early fault detections and making it difficult to adapt to the diverse needs of complex industrial scenarios.

[0006] The third category is intelligent diagnostic methods based on deep learning. This method achieves end-to-end mapping from "raw data input to fault classification output" through deep learning models, eliminating the reliance on manual features and becoming a research hotspot in recent years. The mainstream technical paths include two categories: one is "signal-image" conversion methods based on convolutional neural networks (CNNs), such as converting vibration signals into time-frequency maps via continuous wavelet transform (CWT) and then inputting them into network models like ResNet and VGG for automatic feature extraction; the other is direct modeling of time-series data based on recurrent neural networks (RNNs) or Transformers, such as using long short-term memory networks (LSTM) and temporal convolutional networks (TCNs) to process one-dimensional vibration sequences. Although this type of method has significantly improved adaptability compared to the previous two types, it still faces three major pain points in industrial applications: First, the labeled data requires experts to confirm the fault type and stage on a sample-by-sample basis, and the labeling cost of the entire life cycle data of a single piece of equipment is extremely high, making it difficult to promote on a large scale; Second, existing deep learning models (such as the original EfficientNet and ResNet) are insufficient in capturing features of weak faults such as early microcracks, and feature extraction lacks specificity, resulting in a low early fault identification rate; Third, the diagnostic results of a single model are easily affected by fluctuations in operating conditions, have limited generalization ability, and are difficult to accurately locate the fault initiation time, affecting the timeliness and accuracy of maintenance decisions.

[0007] Furthermore, some existing diagnostic methods rely on fixed operating parameters of specific datasets, resulting in insufficient adaptability to multi-load and multi-speed scenarios. Simultaneously, the lack of a collaborative optimization mechanism between model predictions and unsupervised learning results hinders further improvement in the reliability of fault state determination. Therefore, there is an urgent need for a rolling bearing fault diagnosis method that can reduce annotation costs, enhance early fault feature extraction, and improve generalization ability under varying operating conditions to overcome the shortcomings of existing technologies. Summary of the Invention

[0008] To address the technical pain points of existing rolling bearing fault diagnosis methods, such as high annotation costs, insufficient ability to identify early and subtle faults, limited generalization ability under varying operating conditions, and difficulty in accurately locating the fault initiation time, this invention provides a rolling bearing fault diagnosis method based on an improved EfficientNetV2 network. By optimizing the dataset construction method, improving the network feature extraction mechanism, and employing a multi-model fusion judgment strategy, this method achieves low-cost, high-precision, and robust fault diagnosis and fault initiation time location, providing reliable technical support for predictive maintenance of equipment.

[0009] The core technical solution of this invention includes four steps: dataset construction, improvement of the EfficientNetV2 network model, model training, and multi-result fusion and judgment, as detailed below:

[0010] Vibration signals throughout the entire life cycle of rolling bearings were obtained using the publicly available PHM2012 rolling bearing dataset (collected by FEMTO-ST Research Institute on the PRONOSTIA platform). This dataset is divided into three operating conditions based on different loads and accelerations (Operating Condition 1: Load 4000N, Acceleration 1800rpm; Operating Condition 2: Load 4200N, Acceleration 1650rpm; Operating Condition 3: Load 5000N, Acceleration 1500rpm).

[0011] The above vibration signal is subjected to continuous wavelet transform (CWT) to generate time-frequency images corresponding to each sampling time, and an image-based feature dataset (denoted as fn, where fn is the time-frequency image of the nth sampling time and n is the total number of sampling times) is constructed. This dataset contains 300 data points under three working conditions.

[0012] Based on the first and last boundary method, the sample labels were divided. The data of the first 50 moments in the initial stage of bearing operation were labeled as healthy state, and the data of the last 20 moments in the final stage of operation were labeled as fault state, thus forming an image classification dataset. Then, the training set and test set were divided in a 7:3 ratio for model training and performance verification.

[0013] An improved EfficientNetV2 network model is constructed, comprising a Stem layer, a shallow feature extraction module, a deep feature extraction module, and a Head layer. The functions and structure of each module are as follows:

[0014] 1. Stem layer: Receives a 224×224 time-frequency image input, and completes initial feature extraction and downsampling through 3×3 convolution operation, laying the foundation for subsequent feature extraction.

[0015] 2. Shallow Feature Extraction Module: This module comprises three stages, consisting of stacked 2-layer, 4-layer, and 4-layer improved Fused-MBConv modules, respectively. The specific processing flow of the improved Fused-MBConv module is as follows: It receives the feature map X output from the upstream module. F-in Perform 1×1 convolution and batch normalization (BN) operations to output the upgraded feature map X. F1 ; For X F1 The SiLU activation function is applied to enhance the nonlinear expressive power of the features, and the activated feature map X is output. F2 ; For X F2 Perform 3×3 convolution and batch normalization (BN) operations to output feature map X. F3 ; For X F3 Embed an efficient channel attention module (ECA) to output a weighted feature map X. F4 ; For X F4 Perform 1×1 convolution and batch normalization (BN) operations to output the dimensionality-reduced feature map X.F5 ; For X F5 Perform Dropout regularization and finally output the processed X. F-out .

[0016] 3. Deep Feature Extraction Module: This module comprises three stages, consisting of stacked 6-layer, 9-layer, and 15-layer improved MBConv modules, respectively. The specific processing flow of the improved MBConv module is as follows: It receives the feature map X output from the upstream module. F-out Perform 1×1 convolution, batch normalization (BN), and Swish activation function operations to output the upscaled feature map X. M1 ; For X M1 Perform 3×3 convolution, batch normalization (BN), and Swish activation function operations to output feature map X. M2 ; For X M2 Embedded convolutional block attention module (CBAM) outputs weighted feature map X. M3 ; For X M3 Perform 1×1 convolution and batch normalization (BN) operations to output the dimensionality-reduced feature map X. M4 ; For X M4 Perform Dropout regularization and finally output the processed X. M-out .

[0017] 4. Feature Fusion and Head Layer: The multi-scale features output by the shallow feature extraction module and the deep feature extraction module are fused and then input into the Head layer; after 1×1 convolution to compress the feature dimension and global pooling to extract global features, the health / fault binary classification prediction result is output through a fully connected layer.

[0018] The improved EfficientNetV2 network was trained using the image classification dataset constructed above. During the training process, a cross-validation strategy was used to evaluate the model performance, and the training loss and classification accuracy were monitored in real time. The optimal model parameters were saved to obtain a fault diagnosis model with stable diagnostic capabilities.

[0019] The trained fault diagnosis model is used to predict the time-frequency image corresponding to the bearing vibration signal, obtaining the first healthy state sequence. Simultaneously, K-means unsupervised binary classification is performed on the features of the same bearing vibration signal after CWT transformation, obtaining the second healthy state sequence. The two state sequences are summarized and statistically analyzed using "OR" logic. For any sampling time, if either the fault diagnosis model or the K-means clustering algorithm determines it as a fault, then that time is ultimately classified as a fault state; only when both sequences are classified as healthy is that time classified as a healthy state. In the final output state classification result, the position where three consecutive sampling times are classified as faulty is determined as the starting time of the bearing fault, and all times before this position are classified as healthy stages.

[0020] Compared with the prior art, the present invention has the following significant advantages:

[0021] 1. Significantly reduce annotation costs: The dataset is constructed using the beginning and end definition method, eliminating the need for experts to annotate the fault type and stage for each sample. The dataset can be constructed simply by defining "initial health" and "final failure". Combined with multi-condition data support from public datasets, the manpower and time costs of data annotation are significantly reduced.

[0022] 2. Enhance early fault identification capability: By embedding the ECA attention mechanism in the shallow feature extraction module and introducing the CBAM attention mechanism in the deep feature extraction module, dual feature focusing in the channel and spatial dimensions is achieved, which effectively enhances the feature response of weak faults such as early microcracks and improves the sensitivity of fault identification.

[0023] 3. Improve generalization ability and diagnostic reliability: Improve the multi-scale feature fusion design of EfficientNetV2 network to adapt to feature differences under different working conditions. Combine supervised model prediction and unsupervised K-means clustering with "OR" logic fusion strategy to reduce the risk of misjudgment by a single model and improve the diagnostic robustness under varying working conditions.

[0024] 4. Precisely pinpoint the fault initiation time: By using the "fault at three consecutive sampling times" judgment rule, the initiation time of the fault can be accurately pinpointed, providing a clear time node reference for equipment maintenance decisions and further improving the practicality of predictive maintenance. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the improved fault diagnosis method for the EfficientNetV2 network in an embodiment of the present invention.

[0026] Figure 2 This is an overall architecture diagram of the improved fault diagnosis method of EfficientNetV2 network in an embodiment of the present invention;

[0027] Figure 3 This is a network structure diagram of the improved Fused-ECA Conv module in an embodiment of the present invention;

[0028] Figure 4 This is a network structure diagram of the improved CBAM Conv module in an embodiment of the present invention;

[0029] Figure 5 These are the loss curve and accuracy curve results after model training in this embodiment of the invention; Detailed Implementation

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0031] This invention provides a rolling bearing fault diagnosis method based on an improved EfficientNetV2 network, belonging to the field of mechanical equipment condition monitoring and fault prediction technology. The method includes: acquiring the bearing's full-life vibration signal, generating a time-frequency image through continuous wavelet transform, and constructing an image classification dataset based on the head-tail delimitation method; constructing a rolling bearing fault diagnosis model based on the improved EfficientNetV2 network, embedding an ECA attention mechanism in the Fused-MBConv module in the early stage of feature extraction, and introducing a CBAM attention mechanism in the deep MBConv module in the core stage of feature extraction; co-correcting the results obtained from the partitioning model with the K-means clustering results, and determining the health and fault states based on "OR" logic fusion.

[0032] This embodiment presents a rolling bearing fault diagnosis method based on an improved EfficientNetV2 network, comprising the following steps:

[0033] Step 1. Data preprocessing and dataset construction.

[0034] This embodiment selects a publicly released rolling bearing dataset as the training and validation data source. The data comes from the publicly available PHM2012 rolling bearing dataset, which was collected by FEMTO-ST Research Institute on the PRONOSTIA platform. The experimental data is divided into three operating conditions according to different loads and accelerations. The training set is the complete life-cycle data of two bearings under each operating condition (from healthy operation to failure), and the test set is the truncated data of the remaining bearings under each operating condition (only providing partial degradation data before failure).

[0035] In the data preprocessing process, the original vibration data is first converted into a time-frequency graph using continuous wavelet transform and Morlet wavelet, thus constructing an image-based feature dataset F = (f1, f2, f3, ..., f n ), f nThis is the time-frequency diagram of the bearing at each sampling moment, where n is the total number of sampling moments.

[0036] The image feature dataset F is standardized by uniformly adjusting all images to 224×224 pixels to eliminate potential size differences. Then, the PIL-formatted RGB images are converted to tensor form with dimensions arranged as (C×H×W). Finally, zero-mean unit variance normalization is performed using the mean and standard deviation of the ImageNet dataset, as shown in the formula: Where x is the original pixel value, x norm These are the normalized pixel values. Based on the first-to-last delimitation method, the data from the initial stage of bearing operation (the first 50 time points) are labeled as healthy (label 1 for healthy), and the data from the final stage of bearing operation (the last 20 time points) are labeled as faulty (label 0 for faulty). The image classification dataset is re-divided, with a single sample format of image tensor (3×224×224) + class label (1 / 0). The batch size is set to 32, and the iterator is constructed after randomly shuffling the data order to avoid order bias during training.

[0037] Step 2. Construct a deep learning model for rolling bearing fault diagnosis based on the improved EfficientNetV2 network.

[0038] like Figure 2As shown, the network model constructed in this embodiment is based on an improved EfficientNetV2 network, including a Stem layer (stage0), shallow feature extraction modules (stages1-3), deep feature extraction modules (stages4-6), and a Head layer (stage7). In the Stem layer, the 3×224×224 feature C1 is transformed into a 24×112×112 feature C2 through 3×3 convolutions; then it enters stage1, which consists of two stacked layers of improved FusedMBConv modules, outputting C3; then it enters stage2, which consists of four stacked layers of improved FusedMBConv modules, outputting a 48×56×56 feature C4; then it enters stage3, which consists of four stacked layers of improved FusedMBConv modules, outputting a 64×28×28 feature C5; then it enters stage4, which consists of... The system consists of 6 layers of improved MBConv modules stacked together, outputting a 128×14×14 feature C6; then it enters stage 5, which consists of 9 layers of improved MBConv modules stacked together, outputting a 160×14×14 feature C7; then it enters stage 6, which consists of 9 layers of improved MBConv modules stacked together, outputting a 256×7×7 feature C8; then it enters stage 7, where the feature is converted to 1280×7×7 and further compressed to 1280×1×1 through average pooling. Through a fully connected layer, it is linearly mapped to a 2D output, which divides the state into faulty or healthy.

[0039] like Figure 3 As shown, this paper introduces an improved Fused-MBConv module, embedding an ECA channel attention mechanism after the BN layer of Fused-MBConv to enhance the channel response of low-level fault features. The feature map C×H×W is compressed to C×1×1 using adaptive average pooling; then, local cross-channel interaction is employed, using one-dimensional convolution to capture local dependencies between channels, outputting the channel weights corresponding to C×1×1; and the channel attention weights α are generated using the Sigmoid function. c Then, the attention weights are multiplied channel by channel by channel of the original feature map, using the following formula:

[0040]

[0041] in For broadcast operations, 1 1×H×W It is a full 1 tensor.

[0042] like Figure 4As shown, this paper introduces an improved MBConv module, employing a CBAM attention mechanism after a 3×3 convolution to enhance the dual channel and spatial attention of mid-to-high-level fault features. Specifically, the channel attention branch first compresses C×H×W to C×1×1 using adaptive average pooling, then performs 1×1 convolution for dimensionality reduction, ReLU activation, and 1×1 convolution for dimensionality increase, generating C×1×1 channel weights. Finally, the channel attention weights M are generated using the Sigmoid function. c (x), then perform channel weighting:

[0043] x channel =x×M c (x)

[0044] The spatial attention branch first performs channel-dimensional pooling and calculates x. channel The channel mean and maximum values ​​are concatenated into a 2×H×W array, then spatial convolution activation is performed, and a 1×H×W spatial weight M is generated using the sigmoid function. S (x), then perform spatial weighting:

[0045] x final =x channel ×M s (x)

[0046] Step 3. The image classification dataset is used to train the model using the improved network.

[0047] First, the network is encapsulated as a binary classification network adapted for bearing fault diagnosis, with the output dimension set to 2. This ensures that when the input batch is a B×3×224×224 image tensor, it can output a B×2 tensor for classifying health and fault. The training process is then streamlined, with 30 training epochs. A StepLR learning rate scheduler is introduced (the initial learning rate of 0.001 is decayed by a factor of 10 every 10 epochs, successively decreasing to 0.0001 and 0.00001). In each training epoch, an image tensor of batch size 32 (32×3×224×224) is input into the network for forward propagation, outputting a 32×2 predicted score tensor. The batch loss value is then calculated using the cross-entropy loss function, and the gradient of the loss with respect to all network parameters is calculated using backpropagation. The parameters are updated using the Adam optimizer, and the loss value and classification accuracy for each batch and each epoch are statistically analyzed in real time.

[0048] Step 4. Evaluation and Visualization of Training Results

[0049] To visually evaluate the training convergence characteristics of the improved EfficientNetV2 network, this step visualizes the loss and accuracy data recorded during 30 training epochs, creating a combined visualization chart that includes the training loss curve and the accuracy curve, such as... Figure 5As shown in the figure, the loss curve is plotted with the number of training epochs (1-30) on the horizontal axis and the average loss value on the vertical axis, while the accuracy curve is plotted with the number of training epochs (1-30) on the horizontal axis and the classification accuracy (%) on the vertical axis. This clearly demonstrates the advantages of the hierarchical attention network in that it converges quickly in the early stage of training and does not have obvious overfitting in the later stage, providing an intuitive basis for the quantitative evaluation of the model training effect.

[0050] Step 5. Fusion of Fault Diagnosis Model and K-means Clustering Health Status

[0051] A multimodal fusion judgment process is constructed by introducing the K-means clustering algorithm. By combining supervised model prediction and unsupervised clustering results, the robustness of bearing health status judgment is improved. First, the trained bearing fault diagnosis model is invoked to predict the image samples corresponding to the bearing vibration signals one by one in batches, outputting a health / fault label sequence for each sample (denoted as the first healthy state sequence, with a sequence length consistent with the number of test samples N, and each element representing health / fault). Then, continuous wavelet transform (CWT) processing is performed on the bearing vibration signals of the same batch to extract time-frequency domain features and construct a feature matrix (with dimensions N×D, where D is the feature dimension after CWT transformation). Based on this feature matrix, K-means unsupervised binary clustering is performed (with the number of cluster centers set to 2), and the clustering results are mapped to a health / fault label sequence (denoted as the second healthy state sequence, also with a length of N, and each element representing health / fault). Finally, the first and second healthy state sequences are summarized and statistically analyzed, and a fusion judgment is completed using an "OR" logic rule: that is, for any bearing sample, if the first sequence determines it to be faulty or the second sequence determines it to be faulty, then the sample is ultimately determined to be in a faulty state; only when both sequences are determined to be healthy is the sample ultimately determined to be in a healthy state, and the final fault diagnosis result is output.

[0052] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the embodiments listed above. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for diagnosing rolling bearing faults based on an improved EfficientNetV2 network, characterized in that, Includes the following steps: Step 1. Obtain the vibration signal of the rolling bearing throughout its entire life cycle, perform continuous wavelet transform (CWT) on the vibration signal to generate the corresponding time-frequency image, and mark the continuous moment data in the early stage of bearing operation as the healthy state and the continuous moment data in the late stage of bearing operation as the fault state based on the head-tail definition method to form an image classification dataset. Step 2. Construct a rolling bearing fault diagnosis model based on the improved EfficientNetV2 network. The model includes a Stem layer, a shallow feature extraction module, a deep feature extraction module, and a Head layer. The Stem layer receives the time-frequency image and performs initial feature extraction and downsampling through a 3×3 convolution. The shallow feature extraction module includes three stages, each consisting of stacked 2-layer, 4-layer, and 4-layer improved Fused-MBConv modules, with the improved Fused-MBConv module embedding an efficient channel attention mechanism (ECA). The deep feature extraction module includes three stages, each consisting of 6-layer, 9-layer, and 15-layer improved MBConv modules stacked together. The improved MBConv module introduces a convolutional block attention mechanism (CBAM). The features output by the shallow feature extraction module and the deep feature extraction module are fused at multiple scales (including feature dimension alignment and weighted summation). The fused features are input into the Head layer and then processed by 1×1 convolution, global pooling and fully connected layers to obtain the prediction classification result. Step 3. Train the improved EfficientNetV2 network using the image classification dataset, evaluate the model performance using cross-validation, save the optimal model parameters, and obtain the trained fault diagnosis model. Step 4. Use the trained fault diagnosis model to predict the time-frequency image corresponding to the bearing vibration signal to obtain the first health state sequence; perform K-means unsupervised binary classification on the flattened features of the time-frequency image after CWT transformation of the same bearing vibration signal to obtain the second health state sequence; fuse the first health state sequence and the second health state sequence according to "OR" logic to determine the health and fault states, and output the result.

2. The rolling bearing fault diagnosis method based on the improved EfficientNetV2 network according to claim 1, characterized in that, The processing flow of the improved Fused-MBConv module includes: Receive the feature map X of the upstream output F-in Perform 1×1 convolution and batch normalization (BN) operations to output the upgraded feature map X. F1 ; For the X F1 The SiLU activation function is applied to enhance the nonlinear representation of features, and the activated feature map X is output. F2 ; For the X F2 Perform 3×3 convolution and batch normalization (BN) operations to output feature map X. F3 ; For the X F3 An efficient channel attention module (ECA) is embedded to enhance key feature responses through channel weighting, and the weighted feature map X is output. F4 ; For the X F4 Perform 1×1 convolution and batch normalization (BN) operations to output the dimensionality-reduced feature map X. F5 ; For X F5 Perform Dropout regularization to output the final processed feature map X. F-out .

3. The rolling bearing fault diagnosis method based on the improved EfficientNetV2 network according to claim 1, characterized in that, The processing flow of the improved MBConv module includes: Receive the feature map X of the upstream output F-out Perform 1×1 convolution, batch normalization (BN), and Swish activation function operations to output the upscaled feature map X. M1 ; For the X M1 Perform 3×3 convolution, batch normalization (BN), and Swish activation function operations to output feature map X. M2 ; For the X M2 An embedded convolutional block attention module (CBAM) focuses on fault feature regions through dual channel and spatial weighting, outputting a weighted feature map X. M3 ; For the X M3 Perform 1×1 convolution and batch normalization (BN) operations to output the dimensionality-reduced feature map X. M4 ; The X M4 Perform Dropout regularization to output the final processed feature map X. M-out .

4. The rolling bearing fault diagnosis method based on the improved EfficientNetV2 network according to claim 1, characterized in that, In step 1: The rolling bearing's full life cycle vibration signal includes data collected under multiple operating conditions, which are divided according to differences in load and acceleration. In the beginning and end definition method, the number of continuous moments in the initial stage of bearing operation is 50, and the number of continuous moments in the final stage of operation is 20. The image classification dataset is divided into training set and test set in a 7:3 ratio.

5. The rolling bearing fault diagnosis method based on the improved EfficientNetV2 network according to claim 1, characterized in that, In step 4: For any sampling time, if the fault diagnosis model or the K-means clustering algorithm determines it as a fault, then that time is finally classified as a fault state. In the final output division result, the position where the first three consecutive sampling times are determined as faults is identified as the starting time of the bearing fault, and all times before the starting time are classified as healthy stages.

Images