Intelligent diagnostic data generation method and diagnostic method for fracturing truck

Abstract

The invention relates to a method for generating intelligent diagnostic data and a diagnostic method for a fracturing truck, and belongs to the technical field of diagnosis and detection of fracturing trucks. In order to solve the problem that the existing intelligent diagnosis method for fracturing trucks based on neural networks has poor diagnostic effect due to the limited number of samples and imbalanced samples. The present invention converts the collected vibration signal data of the fracturing truck into image data as a real data sample; the real data sample is input into the DiffGAN model to obtain generated data, and the DiffGAN model includes a generator and a discriminator; the generator adopts an improved multi-scale residual U-net structure, and the discriminator is used to determine whether the data generated by the generator is a real sample in the adversarial training stage. The generated data samples are used to train the fault detection model constructed based on the neural network, and the fault detection model constructed based on the neural network is used for intelligent diagnosis of fracturing trucks.

Description

Technical Field

[0001] The present invention belongs to the technical field of diagnosis and detection of fracturing vehicles, and relates to a method for generating diagnostic data for a fracturing vehicle and a diagnostic method. Background Art

[0002] As the development of unconventional oil and gas resources such as shale gas and tight oil accelerates in deep (burial depth > 3,500 meters) and ultra-deep (> 6,000 meters) areas around the world, the operating reliability of fracturing trucks, as the core power equipment for hydraulic fracturing operations, is directly related to the reservoir transformation effect and development economic benefits.

[0003] Currently, the industry primarily uses two types of diagnostic methods. The first is a vibration feature analysis method based on mechanism modeling. This method extracts time-domain statistical indicators (peak-to-peak value, kurtosis coefficient, etc.) by establishing a kinematic model of the plunger pump (such as the dynamic equation of the crank slider mechanism). However, this type of method has three limitations: First, the modeling process relies on precise mechanical parameters, and complex on-site working conditions cause model parameter drift; second, single-point analysis methods have difficulty resolving the spatiotemporal correlation characteristics of multi-sensor signals. For example, vibration waves caused by eccentric wear of the plunger can be transmitted along the pump body structure to the pump head sensor 1.2 meters away; third, feature engineering relies on expert experience and has poor adaptability to new fault modes (such as high-frequency electromagnetic interference from electric-driven fracturing vehicles). The second type is an end-to-end diagnostic solution based on deep learning. Although progress has been made in the field of rotating machinery, direct application to fracturing vehicles still faces severe challenges. Existing neural network recognition methods generally rely on the representativeness of the data space of the collected samples. If the data space of the collected samples can represent the actual data space distribution, then relatively good detection effects will be achieved with good robustness. However, once the sampled data cannot fully represent the actual data space distribution, the neural network often cannot achieve good detection effects, especially when the working conditions of the fracturing truck change, the detection effect of the neural network will be seriously degraded.

[0004] Therefore, neural network training often requires collecting large amounts of sample data to fully characterize the distribution of the data space. However, for fracturing trucks, collecting large amounts of sample data is unrealistic, especially due to the imbalance in the detection signal data of fracturing trucks. This seriously affects the effectiveness of neural networks in fracturing truck diagnosis and testing, and severely limits their application in this field. Summary of the Invention

[0005] The present invention aims to solve the problem of poor diagnostic effect caused by limited sample number and sample imbalance in the existing neural network-based intelligent diagnosis method for fracturing trucks.

[0006] A method for generating intelligent diagnostic data for a fracturing truck comprises the following steps:

[0007] The collected vibration signal data of the fracturing truck is converted into image data as a real data sample; the real data sample is input into the DiffGAN model to obtain generated data; the DiffGAN model includes a generator and a discriminator;

[0008] The generator adopts an improved multi-scale residual U-net structure, which includes 5 multi-scale residual blocks and 4 upsampling convolution modules, which are sequentially recorded as the first to fifth multi-scale residual blocks and the first to fourth upsampling convolution modules in the order of data processing;

[0009] The multi-scale residual block includes multiple parallel convolution branches, which use convolution layers with convolution kernels of different sizes to perform convolution processing; the upsampling convolution modules each include an upsampling layer and a convolution unit;

[0010] A maximum pooling layer is provided after each of the first to fourth multi-scale residual blocks. The outputs of the first to fourth multi-scale residual blocks pass through the maximum pooling layer and serve as the input of the next multi-scale residual block; the output of the fifth multi-scale residual block serves as the input of the first upsampling convolution module; at the same time, the output of the first multi-scale residual block passes through the convolution layer and is respectively connected to the upsampling layers of the fourth upsampling convolution module to the first upsampling convolution module and is then sent to the corresponding convolution unit; the output of the second multi-scale residual block passes through the convolution layer and is respectively connected to the upsampling layers of the third upsampling convolution module to the first upsampling convolution module and is then sent to the corresponding convolution unit; the output of the third multi-scale residual block passes through the convolution layer and is respectively connected to the upsampling layers of the second upsampling convolution module to the first upsampling convolution module and is then sent to the corresponding convolution unit; the output of the fourth multi-scale residual block passes through the convolution layer and is connected to the upsampling layer of the first upsampling convolution module and is then sent to the corresponding convolution unit;

[0011] The discriminator is used to determine whether the data generated by the generator is a real sample during the adversarial training phase.

[0012] Furthermore, in the process of converting the collected vibration signal data of the fracturing truck into image data, wavelet transform is used to convert the data into 2D RGB data.

[0013] Furthermore, the multi-scale residual block includes 5 parallel convolution branches, wherein the first branch includes a convolution layer with a convolution kernel size of 1×1, and a BN layer is provided after the convolution layer; the remaining 4 parallel convolution branches each include two convolution layers, and a BN layer is provided after the convolution layer. The BN layer corresponding to the first convolution layer is followed by an activation function ReLU, and the convolution kernel sizes of the convolution layers of the four parallel branches are 1×1, 3×3, 5×5 and 7×7 respectively; the outputs of the four parallel branches after ADD processing are connected with the output of the BN layer of the first branch and then sent to the activation function ReLU for activation as the final output of the multi-scale residual block.

[0014] Furthermore, the convolution unit in the upsampling convolution module includes a convolution layer, a BN layer and an activation function ReLU.

[0015] Furthermore, the discriminator adopts an improved multi-scale residual network, including a Stem module, an Inception-Residual module, a Reduction module, a Dropout layer and a fully connected layer. Finally, it performs identification based on the features output by the fully connected layer, and makes a probability prediction that the identification data is a real sample. Among them, the Stem module adopts a multi-branch network structure for preliminary feature extraction and fusion. After preliminary processing by the Stem module, the data enters the Inception-Residual module, which adopts a four-branch network structure for multi-scale feature extraction. After the data enters the Reduction module through the Inception-Residual module, the Reduction module adopts a three-branch network structure for feature dimensionality reduction processing.

[0016] Furthermore, the processing of the Stem module is as follows:

[0017] The input first passes through two convolutional layers and then is divided into two branches. Each branch uses one convolutional layer. The output channels of the two branches are superimposed and then pass through the FC layer before being further divided into two branches for processing. One branch has two convolutional layers and the other has four convolutional layers. The output channels of the two branches are superimposed and then pass through the FC layer before being further divided into two branches for processing. Each branch uses one convolutional layer. The output channels of the two branches are superimposed and then pass through the FC layer before being sent to the BN layer. All convolutional layers use a convolutional kernel of 3×3.

[0018] Furthermore, the processing process of the Inception-Residual module is as follows:

[0019] The input is sent to four branches for processing. The first branch is the original input and is not processed. One of the other three branches uses a 1×1 convolution layer, one branch uses a 1×1 convolution layer and a 3×3 convolution layer, and one branch uses a 1×1 convolution layer, a 3×3 convolution layer, and a 3×3 convolution layer. The output channels of the three branches are superimposed and sent to a 1×1 convolution layer, then processed by the Dropout layer and connected to the input residual of the first path, and then activated by the activation function ReLU and output.

[0020] Furthermore, the processing of the Reduction module is as follows:

[0021] The input is sent to three branches for processing. One branch performs 3×3 maximum pooling, one branch performs 1×1 convolution, and one branch performs 1×1 convolution, 3×3 convolution, and 3×3 convolution. The output channels of the three branches are superimposed and sent to the FC layer and BN layer, and then activated by the activation function ReLU before output.

[0022] Furthermore, the number of Stem modules, Inception-Residual modules, and Reduction modules in the discriminator is set to 1, 2, and 1 respectively.

[0023] A fracturing truck intelligent diagnosis method converts the collected vibration signal data of the fracturing truck into image data, which is then fed into a fault detection model built based on a neural network to identify the type of fault in the fracturing truck, thereby realizing intelligent diagnosis of the fracturing truck.

[0024] The fault detection model constructed based on the neural network is pre-trained, and the sample data in the training set used in the training process includes generated samples obtained according to the method for generating intelligent diagnostic data for a fracturing truck.

[0025] Beneficial effects:

[0026] The present invention uses DiffGAN for data augmentation, effectively addressing limited sample size and sample imbalance, further improving classification accuracy across categories and significantly reducing inter-category misclassification rates. Therefore, DiffGAN can effectively alleviate data imbalance, enhance the model's ability to learn fault categories, and improve the accuracy and stability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the DiffGAN model structure.

[0028] Figure 2 Schematic diagram of the multi-scale residual U-net structure.

[0029] Figure 3 Schematic diagram of multi-scale residual block.

[0030] Figure 4 Schematic diagram of the discriminator structure. DETAILED DESCRIPTION Specific implementation method one:

[0032] This embodiment is a method for generating intelligent diagnostic data and a diagnostic method for a fracturing truck, which includes the steps of generating data (step S 3) and the actual diagnostic steps (steps S 4). The specific process includes:

[0033] S 1. Collect vibration signal data from the fracturing truck and use wavelet transform to convert the data into 2D RGB data as real data samples;

[0034] For fault diagnosis of fracturing trucks, multiple types of data can be collected, such as vibration signals and temperature signals. However, considering the actual operating conditions and signal data collection difficulty of fracturing trucks, as well as the noise impact of collected signals, the present invention ultimately uses fracturing truck vibration signal data for detection. Furthermore, considering the representativeness of the vibration signals collected from the fracturing truck's structure and components, as well as the vibration amplitude and impact of the vibration signals of different components, the present invention uses bearing vibration signals as a representative example for subsequent signal processing and detection.

[0035] S 2. Divide the real data samples into training set and test set;

[0036] S 3. Input the training set sample data into the DiffGAN model for training to obtain the generated data to supplement the data set.

[0037] The core idea of ​​DiffGAN is to use the noise addition mechanism of the diffusion model to provide conditional input for the generator to improve the data generation quality and model stability. DiffGAN mainly includes three stages: noise addition process, data generation and adversarial training. Figure 1 As shown in the model training and data generation section, it includes:

[0038] First, we use the idea of ​​Diffusion Model to gradually add noise to the real data samples. Specifically, the real data samples are recorded as the original data , Gaussian noise is introduced through a predefined noise scheduling strategy to evolve it into samples with different noise levels ,Right now: ;in, is the noise attenuation coefficient, which controls the change of noise with time step t, is independent and identically distributed Gaussian noise. This noise addition process simulates the forward diffusion process of the diffusion model to a certain extent, making the input data samples gradually evolve towards a standard Gaussian distribution.

[0039] In the traditional GAN ​​structure, the generator usually samples from a completely random noise distribution to synthesize data, while DiffGAN introduces noisy samples. As conditional input to the generator: ,in, represents the generator network, is a trainable parameter, This conditional generation strategy allows the generator to learn based on the existing data structure without having to directly infer the data distribution from completely random noise. This strategy significantly reduces the difficulty of training and helps improve the quality and stability of generated data.

[0040] Since it is still very difficult for the generator to receive noisy data and ultimately generate high-fidelity data, the present invention sets up a special improved multi-scale residual U-net structure, which combines the U-Net's skip connection (SkipConnection), multi-scale feature extraction (Multi-ScaleFeatureExtraction) and residual learning (ResidualLearning) mechanism to enhance the network's expressive ability and improve the quality of generated samples.

[0041] In some implementations, the improved multi-scale residual U-net has a specific structure such as Figure 2 As shown, it includes 5 multi-scale residual blocks (Multi-Scale Residual Blocks) and 4 upsampling convolution modules, which are recorded as the first multi-scale residual block to the fifth multi-scale residual block and the first upsampling convolution module to the fourth upsampling convolution module in the order of data processing;

[0042] like Figure 3 As shown, the multi-scale residual block includes multiple parallel convolution branches, and the parallel convolution branches use convolution layers with convolution kernels of different sizes for convolution processing; the upsampling convolution modules each include an upsampling layer and a convolution unit, and the convolution unit includes a convolution layer, a BN layer and an activation function ReLU;

[0043] A maximum pooling layer is provided after each of the first to fourth multi-scale residual blocks. The outputs of the first to fourth multi-scale residual blocks pass through the maximum pooling layer and serve as the input of the next multi-scale residual block; the output of the fifth multi-scale residual block serves as the input of the first upsampling convolution module;

[0044] At the same time, the output of the first multi-scale residual block passes through the convolution layer and is connected to the upsampling layers of the fourth upsampling convolution module to the first upsampling convolution module and then sent to the corresponding convolution unit;

[0045] The output of the second multi-scale residual block passes through the convolution layer and is connected to the upsampling layers of the third upsampling convolution module to the first upsampling convolution module and then sent to the corresponding convolution unit;

[0046] The output of the third multi-scale residual block passes through the convolution layer and is connected to the upsampling layers of the second upsampling convolution module to the first upsampling convolution module and then sent to the corresponding convolution unit;

[0047] The output of the fourth multi-scale residual block passes through the convolution layer, is connected to the upsampling layer of the first sampling convolution module, and then sent to the corresponding convolution unit;

[0048] Furthermore, in some embodiments, the multi-scale residual block includes 5 parallel convolution branches, wherein the first branch includes a convolution layer with a convolution kernel size of 1×1, and a BN layer is provided after the convolution layer; the remaining 4 parallel convolution branches each include two convolution layers, and a BN layer is provided after the convolution layer. The BN layer corresponding to the first convolution layer is followed by an activation function ReLU, and the convolution kernel sizes of the convolution layers of the four parallel branches are 1×1, 3×3, 5×5 and 7×7 respectively; the outputs of the four parallel branches after ADD processing are connected with the output of the BN layer of the first branch and then sent to the activation function ReLU for activation as the final output of the multi-scale residual block.

[0049] Because deep networks are prone to the vanishing gradient problem, and information at different scales is crucial for generating high-quality data, U-Net employs a multi-level skip connection mechanism to fuse shallow features with deep features. Specifically, shallow features are selectively transferred via gated multi-scale residual blocks to preserve key information and suppress redundant information. The U-Net generator architecture includes a multi-scale downsampling module that extracts features using convolutional kernels with different receptive fields to enhance the model's generalization. Small receptive field kernels of 1×1 and 3×3 capture local details and microscopic features, while large receptive field kernels of 5×5 and 7×7 extract global structural information, enhancing the consistency of data generation. Multi-scale residual blocks are constructed by parallelly connecting convolutional kernels of different sizes, effectively improving the model's sensitivity to features at different scales.

[0050] Compared to the traditional U-Net architecture, the combination of multi-scale convolution enables the network to simultaneously capture local and global information, improving the quality of detail in the generated data. The introductory control mechanism reduces redundant information interference and enhances the expressiveness of key features. Combined with residual connections, it effectively alleviates the vanishing and exploding gradient problems of deep networks, improving training stability.

[0051] During the adversarial training phase, the discriminator It is necessary to determine whether the input data is a real sample. The optimization goal is: ;in, represents the real data distribution, Represents the generated data distribution. The generator minimizes the adversarial loss of the discriminator to make the distribution of generated samples as close as possible to the real data distribution.

[0052] In the DiffGAN framework, the discriminator adopts Figure 4 The improved multi-scale residual network shown in Figure 3 enhances the ability to discriminate generated data and improves the training stability and generalization ability of the model. Figure 4 (a) is the improved multi-scale residual network structure diagram of the discriminator, which includes the Stem module, Inception-Residual (IR) module, Reduction module, as well as the Dropout layer and the fully connected layer (Fully Connected, FC). Finally, based on the features output by the fully connected layer, the input data is judged to be a probability prediction of a real sample.

[0053] The Stem module, Inception-Residual (IR) module, and Reduction module are used to achieve efficient feature extraction, nonlinear enhancement, and optimization of computational complexity.

[0054] In some embodiments, the Stem module, Inception-Residual (IR) module, and Reduction module of the discriminator are set to 1, 2, and 1 respectively;

[0055] The discriminator first performs preliminary feature extraction and fusion through the Stem module. This module uses multiple branches to increase the network depth and improve the feature expression ability; its specific structure is as follows Figure 4As shown in (b), the input first passes through two convolutional layers and then is divided into two branches. Each branch uses one convolutional layer. The output channels of the two branches are superimposed and then pass through the FC layer before being further divided into two branches for processing. One branch has two convolutional layers and the other has four convolutional layers. The output channels of the two branches are superimposed and then pass through the FC layer before being further divided into two branches for processing. Each branch uses one convolutional layer. The output channels of the two branches are superimposed and then pass through the FC layer before being sent to the BN layer. All convolutional layers use a convolutional kernel of 3×3.

[0056] After preliminary processing by the Stem module, the data enters the Inception-Residual (IR) module, which adopts a four-branch structure to enhance the multi-scale feature extraction capability and stabilize the gradient flow. The shallow feature retention branch directly passes the input data to retain the low-level feature information and ensure the integrity of the local structure; the two deep feature extraction branches are stacked through multiple layers of convolution to model the complex relationship between data and improve the feature expression capability; the residual connection branch adopts the identity mapping mechanism to alleviate the gradient disappearance problem and improve the stability of model training; finally, the output of each branch is scaled after splicing to suppress gradient explosion and enhance the stability of training. The specific structure of the Inception-Residual module is as follows Figure 4 As shown in (c), the input is sent to four branches for processing. The first branch is the original input without processing. One of the remaining three branches uses a 1×1 convolution layer, one branch uses a 1×1 convolution layer and a 3×3 convolution layer, and one branch uses a 1×1 convolution layer, a 3×3 convolution layer, and a 3×3 convolution layer. The output channels of the three branches are superimposed and sent to a 1×1 convolution layer, then processed by the Dropout layer and connected to the input residual of the first path, and then activated by the activation function ReLU and output.

[0057] After the data enters the Reduction module through the Inception-Residual (IR) module, the Reduction module further reduces the dimension to reduce the computational burden and extract more abstract information. The specific structure of the Reduction module is as follows Figure 4 As shown in (d), the input is sent to three branches for processing, one branch performs 3×3 maximum pooling, one branch performs 1×1 convolution, and one branch performs 1×1 convolution, 3×3 convolution, and 3×3 convolution; the output channels of the three branches are superimposed and sent to the FC layer and BN layer, and then output after activation by the activation function ReLU.

[0058] Subsequently, the feature vector passes through the Dropout layer to reduce the risk of overfitting and is input into two fully connected layers. The final output is judged by the classifier to predict the probability of whether the input data is a real sample.

[0059] Compared to traditional GANs, which generate data directly from random noise, DiffGAN uses noisy samples to provide data structure information, enabling the generator to more effectively learn the data distribution and improve the authenticity of the generated data. The proposed DiffGAN model, by introducing a diffusion model-based noise addition mechanism, provides more reasonable input priors for the GAN generation process, improving data generation quality and model training stability.

[0060] S 4. The generated samples and the original training samples are input into the Resnet model for training to obtain a trained fault detection model, which is used for intelligent diagnosis of fracturing trucks.

[0061] In the experimental verification, different fault data from fracturing truck bearings were selected for testing. The experiment used sliding window sampling of the original vibration signals of five different health states, with each fault signal type sampled 100 times using a window length of 1024. Subsequently, the one-dimensional vibration signals were converted into two-dimensional images of 64×64 pixels using a wavelet transform, generating a total of 100 images. During the data augmentation training phase, 40 images of each fault type were randomly selected to train the data augmentation model, and the remaining samples were used for subsequent quality assessment. To visually demonstrate the effects of data augmentation, augmentation training was performed using the same training rounds. Ultimately, 200 augmented samples were generated for each fault type, for a total of 4×200 samples.

[0062] The quality of generated data plays a key role in the original training set for fault diagnosis, especially its similarity to the original data. When the quality of generated data is low, the fault diagnosis model may extract features that differ significantly from the original data, affecting or even reducing diagnostic accuracy. The feature quantification metric J quantifies the quality of generated data by calculating the intra-class covariance and inter-class covariance:

[0063]

[0064]

[0065]

[0066] in represents the within-class covariance matrix, represents the between-class covariance matrix, represents the number of categories, Indicates the The number of samples of the class, represents the feature vector of the i-th sample in the c-th class, represents the mean eigenvector of the category, represents the overall mean eigenvector of all samples.

[0067] Intra-class covariance measures the degree of aggregation of samples within the same category, while inter-class covariance reflects the discrimination between different categories. When the J value is large, the inter-class covariance is relatively high, indicating that the distribution of samples from different categories is clear, while samples within the same category are more compact. This means that the generated data not only maintains consistency within the category but also ensures good discrimination between categories. Conversely, a low J value indicates weak inter-class separation or a more dispersed distribution of samples within the same category. This may lead to insufficient discrimination of samples after data augmentation, or even the introduction of more noise or outliers.

[0068] The quality of data augmentation was evaluated by calculating a quantitative evaluation metric, J, for the generated data. This metric quantifies the separability of the generated data across different categories and the consistency within the same category. The J metric for the data generated using DiffGAN was 4.75. This indicates that the data generated by DiffGAN maintains the compactness of data within a category (smaller intra-class covariance) while enhancing the distinction between data from different categories (larger inter-class covariance). This indicates that the data generated by DiffGAN not only better matches the original data distribution but also creates clearer decision boundaries between categories, helping to improve the classification performance of subsequent models.

[0069] Quantitative evaluation of features can measure the intrinsic properties of data, the separability and consistency of categories. In order to evaluate the effect of generated data in actual fault diagnosis, the present invention uses the ResNet model to accurately evaluate the quality of generated data in real fault diagnosis. In order to simulate the training process after data imbalance and data enhancement, the data is distributed as follows: samples in normal healthy state do not require data enhancement, so 60 real samples are randomly selected from the healthy state category for training and 40 for testing. For the four fault states, 60 real samples are randomly selected, of which 10 are used for training and 50 are used for testing; in addition, in order to alleviate the problem of insufficient fault category samples, all 200 generated samples of these four types of faults are used for training to improve the model's ability to recognize fault modes.

[0070] The training parameters for the ResNet model were set to: batch size 128, learning rate 0.0001, and training epochs 100. While accuracy can reflect the overall predictive power of a model, it may not fully reflect the model's prediction performance for minority classes when working with unbalanced datasets. Therefore, in addition to accuracy, the present invention also uses the F1 value as an evaluation metric to more comprehensively assess the actual performance of the model. In the ResNet model training experiment, the F1 value was 95.32 and the ACC value was 95.17.

[0071] In order to verify the role of the data enhancement method of the present invention in fault diagnosis tasks, especially the improvement of the fault category recognition ability under data imbalance, the present invention uses the confusion matrix and cluster diagram to analyze the classification effects of different data enhancement methods on the ResNet model in the experiment. In the case of no data enhancement, the diagonal elements of the confusion matrix are unevenly distributed, the classification accuracy of some fault categories is low, and they are easily confused with other categories, indicating that when the amount of data is insufficient, the model's recognition ability of fault samples is weak and it is difficult to effectively capture their characteristics. After using DiffGAN for data enhancement, the diagonal elements of the confusion matrix are most prominent, the classification accuracy of each category is further improved, and the misclassification rate between categories is significantly reduced, indicating that DiffGAN can effectively alleviate the problem of data imbalance, enhance the model's learning ability for fault categories, and improve the accuracy and stability of fault diagnosis.

[0072] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for generating intelligent diagnostic data for a fracturing truck, characterized in that: The following steps are involved: The collected vibration signal data of the fracturing truck is converted into image data as a real data sample; the real data sample is input into the DiffGAN model to obtain generated data; the DiffGAN model includes a generator and a discriminator; The generator adopts an improved multi-scale residual U-net structure, which includes 5 multi-scale residual blocks and 4 upsampling convolution modules, which are sequentially recorded as the first to fifth multi-scale residual blocks and the first to fourth upsampling convolution modules in the order of data processing; The multi-scale residual block includes multiple parallel convolution branches, which use convolution layers with convolution kernels of different sizes to perform convolution processing; the upsampling convolution modules each include an upsampling layer and a convolution unit; A maximum pooling layer is provided after each of the first to fourth multi-scale residual blocks. The outputs of the first to fourth multi-scale residual blocks pass through the maximum pooling layer and serve as the input of the next multi-scale residual block; the output of the fifth multi-scale residual block serves as the input of the first upsampling convolution module; at the same time, the output of the first multi-scale residual block passes through the convolution layer and is respectively connected to the upsampling layers of the fourth upsampling convolution module to the first upsampling convolution module and is then sent to the corresponding convolution unit; the output of the second multi-scale residual block passes through the convolution layer and is respectively connected to the upsampling layers of the third upsampling convolution module to the first upsampling convolution module and is then sent to the corresponding convolution unit; the output of the third multi-scale residual block passes through the convolution layer and is respectively connected to the upsampling layers of the second upsampling convolution module to the first upsampling convolution module and is then sent to the corresponding convolution unit; the output of the fourth multi-scale residual block passes through the convolution layer and is connected to the upsampling layer of the first upsampling convolution module and is then sent to the corresponding convolution unit; The discriminator is used to determine whether the data generated by the generator is a real sample during the adversarial training phase.

2. The method for generating intelligent diagnostic data of a fracturing truck according to claim 1, characterized in that: In the process of converting the collected vibration signal data of the fracturing truck into image data, wavelet transform is used to convert the data into 2D RGB data.

3. The method for generating intelligent diagnostic data of a fracturing truck according to claim 1, characterized in that: The multi-scale residual block includes 5 parallel convolution branches, of which the first branch includes a convolution layer with a convolution kernel size of 1×1, and a BN layer is provided after the convolution layer; the remaining 4 parallel convolution branches each include two convolution layers, and a BN layer is provided after the convolution layer. The BN layer corresponding to the first convolution layer is followed by an activation function ReLU, and the convolution kernel sizes of the convolution layers of the four parallel branches are 1×1, 3×3, 5×5 and 7×7 respectively; the outputs of the four parallel branches after ADD processing are connected with the output of the BN layer of the first branch and then sent to the activation function ReLU for activation as the final output of the multi-scale residual block.

4. The method for generating intelligent diagnostic data of a fracturing truck according to claim 1, characterized in that: The convolution unit in the upsampling convolution module consists of a convolution layer, a BN layer, and an activation function ReLU.

5. The method for generating intelligent diagnostic data for a fracturing truck according to any one of claims 1 to 4, characterized in that: The discriminator adopts an improved multi-scale residual network, including a Stem module, an Inception-Residual module, a Reduction module, a Dropout layer and a fully connected layer. Finally, it makes a judgment based on the features output by the fully connected layer, and predicts the probability that the data is a real sample. Among them, the Stem module adopts a multi-branch network structure for preliminary feature extraction and fusion. After preliminary processing by the Stem module, the data enters the Inception-Residual module, which adopts a four-branch network structure for multi-scale feature extraction. After the data enters the Reduction module through the Inception-Residual module, the Reduction module adopts a three-branch network structure for feature dimensionality reduction.

6. The method for generating intelligent diagnostic data of a fracturing truck according to claim 5, characterized in that: The processing process of the Stem module is as follows: The input first passes through two convolutional layers and then is divided into two branches. Each branch uses one convolutional layer. The output channels of the two branches are superimposed and then pass through the FC layer before being further divided into two branches for processing. One branch has two convolutional layers and the other has four convolutional layers. The output channels of the two branches are superimposed and then pass through the FC layer before being further divided into two branches for processing. Each branch uses one convolutional layer. The output channels of the two branches are superimposed and then pass through the FC layer before being sent to the BN layer. All convolutional layers use a convolutional kernel of 3×3.

7. The method for generating intelligent diagnostic data of a fracturing truck according to claim 5, characterized in that: The processing process of the Inception-Residual module is as follows: The input is sent to four branches for processing. The first branch is the original input and is not processed. One of the other three branches uses a 1×1 convolution layer, one branch uses a 1×1 convolution layer and a 3×3 convolution layer, and one branch uses a 1×1 convolution layer, a 3×3 convolution layer, and a 3×3 convolution layer. The output channels of the three branches are superimposed and sent to a 1×1 convolution layer, then processed by the Dropout layer and connected to the input residual of the first path, and then activated by the activation function ReLU and output.

8. The method for generating intelligent diagnostic data of a fracturing truck according to claim 5, characterized in that: The processing of the Reduction module is as follows: The input is sent to three branches for processing. One branch performs 3×3 maximum pooling, one branch performs 1×1 convolution, and one branch performs 1×1 convolution, 3×3 convolution, and 3×3 convolution. The output channels of the three branches are superimposed and sent to the FC layer and BN layer, and then activated by the activation function ReLU before output.

9. The method for generating intelligent diagnostic data of a fracturing truck according to claim 5, characterized in that: The number of Stem modules, Inception-Residual modules, and Reduction modules in the discriminator is set to 1, 2, and 1 respectively.

10. An intelligent diagnosis method for a fracturing truck, characterized in that: The collected vibration signal data of the fracturing truck is converted into image data, which is then fed into a fault detection model built based on a neural network to identify the type of fault in the fracturing truck and realize intelligent diagnosis of the fracturing truck. The fault detection model constructed based on the neural network is pre-trained, and the sample data in the training set used in the training process includes the generated samples obtained according to the method for generating intelligent diagnostic data for a fracturing truck according to any one of claims 1 to 9.

Images