Fracturing truck intelligent diagnosis data generation method and diagnosis method

Abstract

The invention discloses a fracturing truck intelligent diagnosis data generation method and a diagnosis method, and belongs to the technical field of fracturing truck diagnosis and detection. The objective of the invention is to solve the problem of poor diagnosis effect caused by limited sample quantity and unbalanced samples in the existing intelligent diagnosis mode of the fracturing truck based on the neural network. The method comprises the following steps: converting collected vibration signal data of the fracturing truck into image data as a real data sample; inputting the real data sample into a DiffGAN model to obtain generated data, wherein the DiffGAN model comprises a generator and a discriminator; the generator adopts an improved multi-scale residual U-net structure, and the discriminator is used for discriminating whether data generated by the generator is a real sample or not in an adversarial training stage. The generated data sample is used for training a fault detection model constructed based on a neural network, and the fault detection model constructed based on the neural network is used for intelligent diagnosis of the fracturing truck.

Description

Technical Field

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

[0002] With the accelerated advancement of the development of unconventional oil and gas resources such as global shale gas and tight oil towards deep (burial depth > 3500 meters) and ultra-deep ( > 6000 meters) fields, as the core power equipment for hydraulic fracturing operations, the operational reliability of fracturing trucks is directly related to the reservoir stimulation effect and development economic benefits.

[0003] Currently, the industrial community mainly adopts two types of diagnostic methods. The first type is the vibration feature analysis method based on mechanism modeling, which extracts time-domain statistical indicators (peak-to-peak value, kurtosis coefficient, etc.) by establishing a kinematic model of the piston pump (such as the dynamic equation of the crank-slider mechanism). However, this type of method has three limitations: First, the modeling process depends on accurate mechanical parameters, and the complex on-site working conditions lead to model parameter drift; Second, the single-measurement-point analysis method is difficult to analyze the spatio-temporal correlation characteristics of multi-sensor signals. For example, the vibration wave caused by piston eccentric wear will be transmitted along the pump body structure to the pump head body sensor 1.2 meters away; Third, feature engineering depends on expert experience and has poor adaptability to new fault modes (such as high-frequency electromagnetic interference of electric drive fracturing trucks). The second type is the end-to-end diagnostic scheme based on deep learning. Although progress has been made in the field of rotating machinery, its direct application to fracturing trucks still faces severe challenges. The existing neural network recognition methods generally rely on the data space representativeness of the collected samples. If the data space of the collected samples can represent the real data space distribution, then relatively good detection effects can be achieved, and it has good robustness; but once the sampled data cannot fully represent the real data space distribution, then 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 degrade severely.

[0004] Therefore, usually for the training process of the neural network, a large number of sample data are collected to try to fully represent the distribution of the data space. However, for fracturing trucks, it is unrealistic to collect a large number of sample data. Especially, there is an imbalance problem in the fracturing truck detection signal data, which seriously affects the detection effect of the neural network in the field of fracturing truck diagnosis and detection and is not rational, and also severely restricts the application of the neural network in the field of fracturing truck diagnosis and detection. Summary of the Invention

[0005] The present invention aims to solve the problem that the existing intelligent diagnosis method for fracturing trucks based on neural networks has poor diagnosis effects due to limited sample quantity and sample imbalance.

[0006] A method for generating intelligent diagnosis data for a fracturing truck includes the following steps: The collected vibration signal data of the fracturing truck is converted into image data as real data samples; the real data samples are 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. The improved multi-scale residual U-net includes 5 multi-scale residual blocks and 4 upsampling convolutional modules, which are sequentially denoted as the first multi-scale residual block to the fifth multi-scale residual block, and the first upsampling convolutional module to the fourth upsampling convolutional module according to the data processing order; Each multi-scale residual block includes multiple parallel convolutional branches, and the parallel convolutional branches are processed by convolutional layers with different sizes of convolutional kernels; each upsampling convolutional module includes an upsampling layer and a convolutional unit; A max-pooling layer is set after the first multi-scale residual block to the fourth multi-scale residual block. The output of the first multi-scale residual block to the fourth multi-scale residual block is used as the input of the next multi-scale residual block after passing through the max-pooling layer; the output of the fifth multi-scale residual block is used as the input of the first upsampling convolutional module; at the same time, the output of the first multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the fourth upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units; the output of the second multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the third upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units; the output of the third multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the second upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units; the output of the fourth multi-scale residual block after passing through the convolutional layer is connected to the upsampling layer of the first sampling convolutional module and then sent into the corresponding convolutional unit; The discriminator is used to determine whether the data generated by the generator is a real sample during the adversarial training stage.

[0007] Further, 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.

[0008] Further, each multi-scale residual block includes 5 parallel convolutional branches. The first branch includes a convolutional layer with a convolutional kernel size of 1×1, and a BN layer is set after the convolutional layer; the remaining 4 parallel convolutional branches each include two convolutional layers, and a BN layer is set after each convolutional layer. An activation function ReLU is set after the BN layer corresponding to the first convolutional layer. The convolutional kernel sizes of the convolutional layers of the four parallel branches are 1×1, 3×3, 5×5, and 7×7 respectively; the output after the ADD processing of the outputs of the four parallel branches is connected to the output of the BN layer of the first branch and then sent into the activation function ReLU for activation to be used as the final output of the multi-scale residual block.

[0009] Furthermore, the convolutional unit in the upsampling convolutional module includes a convolutional layer, a BN layer, and an activation function ReLU.

[0010] 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, based on the features output by the fully connected layer, discrimination is performed to predict the probability that the data is a real sample; among them, the Stem module uses a multi-branch network structure for preliminary feature extraction and fusion; after the preliminary processing by the Stem module, the data enters the Inception-Residual module, which uses 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 uses a three-branch network structure for feature dimensionality reduction processing.

[0011] Furthermore, 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 a 1-layer convolutional layer. After the output channels of the two branches are stacked and passed through an FC layer, they are further divided into two branches for processing. One branch is set with two convolutional layers, and the other is set with four convolutional layers; after the output channels of the two branches are stacked and passed through an FC layer, they are further divided into two branches for processing. Each branch uses a 1-layer convolutional layer. After the output channels of the two branches are stacked and passed through an FC layer, they are sent to a BN layer; all convolutional layers use convolutional layers with a convolution kernel of 3×3.

[0012] Furthermore, 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 without processing. One of the remaining three branches uses a 1×1 convolutional layer, one branch uses a 1×1 convolutional layer and a 3×3 convolutional layer, and one branch uses a 1×1 convolutional layer, a 3×3 convolutional layer, and a 3×3 convolutional layer. After the output channels of the three branches are stacked, they are sent to a 1×1 convolutional layer, then processed through a Dropout layer, and then connected to the input residual of the first path, and then activated by the activation function ReLU and output.

[0013] Furthermore, the processing process of the Reduction module is as follows: The input is sent to three branches for processing. One branch performs 3×3 max pooling, one branch performs 1×1 convolution, and one branch performs 1×1 convolution, 3×3 convolution, and 3×3 convolution; after the output channels of the three branches are stacked, they are sent to an FC layer and a BN layer, and then activated by the activation function ReLU and output.

[0014] Further, one Stem module, two Inception-Residual modules, and one Reduction module are respectively set in the discriminator.

[0015] A smart diagnosis method for a fracturing truck, which collects the vibration signal data of the fracturing truck and converts it into image data, and then sends it to a fault detection model constructed based on a neural network for identifying the fault type of the fracturing truck, so as to realize the smart 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 above-mentioned smart diagnosis data generation method for the fracturing truck.

[0016] Beneficial effects: After the data enhancement is carried out by using DiffGAN in the present invention, the problems of limited sample quantity and sample imbalance can be effectively solved, and further, the classification accuracy of each category is further improved, and the misclassification rate between categories is significantly reduced. Therefore, DiffGAN of the present invention can effectively alleviate the data imbalance problem, enhance the learning ability of the model for fault categories, and improve the accuracy and stability of fault diagnosis. Description of the Drawings

[0017] Figure 1 It is a schematic diagram of the DiffGAN model structure.

[0018] Figure 2 It is a schematic diagram of the multi-scale residual U-net structure.

[0019] Figure 3 It is a schematic diagram of the multi-scale residual block.

[0020] Figure 4 It is a schematic diagram of the discriminator structure. Detailed Embodiments Detailed Embodiment 1: This embodiment is a smart diagnosis data generation method and diagnosis method for a fracturing truck, which includes the steps of data generation (Step S 3) and the actual diagnosis step (Step S 4). The specific process includes: S 1. Collect the vibration signal data of the fracturing truck, and use wavelet transform to convert the data into 2D RGB data as real data samples. For the fault diagnosis of a fracturing truck, multiple types of data can be collected, such as vibration signals, temperature signals, etc. Considering the actual working conditions of the fracturing truck, the difficulty of data acquisition of signals, and the noise impact of the acquired signals, the vibration signal data of the fracturing truck is finally used for detection in the present invention. In addition, considering the representativeness of the vibration signals of the structure and components of the fracturing truck, as well as the vibration amplitude and influence of the vibration signals of different components, the vibration signal of the bearing is used as a representative in the present invention for subsequent signal processing and detection.

[0022] S 2. Divide the real data samples into a training set and a test set; S 3. Input the training set sample data into the DiffGAN model for training to obtain a generated data supplementary dataset.

[0023] The core idea of the 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. The DiffGAN mainly includes three stages: the noise addition process, data generation, and adversarial training, as Figure 1 shown in the model training and data generation part, including: First, draw on the idea of the diffusion model (DiffusionModel) and perform step-by-step noise addition processing on the real data samples. Specifically, record the real data samples as the original data , and introduce Gaussian noise through a predefined noise scheduling strategy to evolve it into samples at different noise levels , that is: ; where is the noise attenuation coefficient, which controls the change of noise with the time step t, is independent and identically distributed Gaussian noise. This noise addition process to a certain extent simulates the forward diffusion process of the diffusion model, making the input data samples gradually evolve towards the standard Gaussian distribution.

[0024] In the traditional GAN structure, the generator usually samples from a completely random noise distribution to synthesize data, while the DiffGAN introduces the noise-added samples as the conditional input of the generator: , where represents the generator network, are trainable parameters, is the generated sample. Through this conditional generation strategy, the generator can learn based on the existing data structure without directly inferring the data distribution from completely random noise. This strategy significantly reduces the training difficulty and helps to improve the quality and stability of the generated data.

[0025] Since it is still quite difficult for the generator to receive noisy data and finally generate high-fidelity data, the present invention provides a specially improved multi-scale residual U-net structure, which combines the skip connection, multi-scale feature extraction, and residual learning mechanisms of U-Net to enhance the network's expressive ability and improve the quality of the generated samples.

[0026] In some embodiments, the specific structure of the improved multi-scale residual U-net is as Figure 2 shown. It includes 5 multi-scale residual blocks and 4 upsampling convolutional modules, which are sequentially denoted as the first multi-scale residual block to the fifth multi-scale residual block, and the first upsampling convolutional module to the fourth upsampling convolutional module according to the data processing order. As Figure 3 shown, the multi-scale residual block includes multiple parallel convolutional branches, and the parallel convolutional branches are processed by convolutional layers with different-sized convolutional kernels; each upsampling convolutional module includes an upsampling layer and a convolutional unit, and the convolutional unit includes a convolutional layer, a BN layer, and an activation function ReLU. A max-pooling layer is provided after each of the first multi-scale residual block to the fourth multi-scale residual block, and the output of the first multi-scale residual block to the fourth multi-scale residual block after passing through the max-pooling layer serves 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 convolutional module. At the same time, the output of the first multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the fourth upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units. The output of the second multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the third upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units. The output of the third multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the second upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units. The output of the fourth multi-scale residual block after passing through the convolutional layer is connected to the upsampling layer of the first sampling convolutional module and then sent into the corresponding convolutional unit. Further, in some embodiments, the multi-scale residual block includes five parallel convolutional branches. The first branch includes a convolutional layer with a kernel size of 1×1, and a BN layer is arranged after the convolutional layer. The remaining four parallel convolutional branches each include two convolutional layers, and a BN layer is arranged after each convolutional layer. An activation function ReLU is arranged after the BN layer corresponding to the first convolutional layer. The kernel sizes of the convolutional layers of the four parallel branches are 1×1, 3×3, 5×5, and 7×7 respectively. The output after the ADD process of the outputs of the four parallel branches is connected to the output of the BN layer of the first branch and then fed into the activation function ReLU for activation, which is used as the final output of the multi-scale residual block.

[0027] Since the deep network is prone to the problem of gradient disappearance, and information of different scales is crucial for generating high-quality data, U-Net adopts a multi-level skip connection mechanism to fuse shallow features and deep features. Specifically, the shallow features are selectively transmitted through the gated multi-scale residual block to retain key information and suppress redundant information. The U-Net structure of the generator contains a multi-scale downsampling module, which extracts features through convolutional kernels with different receptive fields to enhance the generalization ability of the model. The convolutional kernels with small receptive fields of 1×1 and 3×3 capture local details and microscopic features; the convolutional kernels with large receptive fields of 5×5 and 7×7 extract global structural information and enhance the consistency of data generation. Multi-scale residual blocks (Multi-Scale Residual Blocks) are constructed through parallel connection using convolutional kernels of different sizes, effectively improving the sensitivity of the model to features of different scales.

[0028] Compared with the traditional U-Net structure, combining multi-scale convolution enables the network to capture local and global information simultaneously, improving the detail quality of the generated data. Incorporating the gating mechanism reduces the interference of redundant information and enhances the expression ability of key features. Combining residual connections effectively alleviates the problems of gradient disappearance or gradient explosion in the deep network and improves the training stability.

[0029] In the adversarial training stage, the discriminator needs to determine whether the input data is a real sample, and the optimization objective is: ; where represents the real data distribution, and represents the generated data distribution. The generator minimizes the adversarial loss of the discriminator to make the distribution of the generated samples as close as possible to the real data distribution.

[0030] In the DiffGAN framework, the discriminator adopts the improved multi-scale residual network as shown in Figure 4 to enhance the discrimination ability of the generated data and improve the training stability and generalization ability of the model. Figure 4Among them, (a) is the improved multi-scale residual network structure diagram of the discriminator. The discriminator includes a Stem module, an Inception-Residual (I-R) module, a Reduction module, as well as a Dropout layer and a fully connected layer (FC). Finally, based on the features output by the fully connected layer, discrimination is performed, and the probability prediction of the input data being a real sample is made.

[0031] The Stem module, Inception-Residual (I-R) module, and Reduction module are used to achieve efficient feature extraction, non-linear enhancement, and optimization of computational complexity.

[0032] In some embodiments, the discriminator has 1 Stem module, 2 Inception-Residual (I-R) modules, and 1 Reduction module respectively; The discriminator first performs preliminary feature extraction and fusion through the Stem module. This module uses multiple branches to increase the network depth and enhance the feature expression ability; its specific structure is as Figure 4 shown in (b) among them. The input first passes through two convolutional layers, and then is divided into two branches. Each branch uses 1 convolutional layer. After the output channels of the two branches are stacked and passed through the FC layer, it is continued to be divided into two branches for processing. One branch has two convolutional layers, and the other has four convolutional layers; after the output channels of the two branches are stacked and passed through the FC layer, it is continued to be divided into two branches for processing. Each branch uses 1 convolutional layer. After the output channels of the two branches are stacked and passed through the FC layer, it is sent to the BN layer; all convolutional layers use convolutional layers with a convolution kernel of 3×3.

[0033] After the preliminary processing by the Stem module, the data enters the Inception-Residual (I-R) module. This module uses a four-branch structure to enhance the multi-scale feature extraction ability and stabilize the gradient flow. Among them, 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 stack multiple convolutional layers to model the complex relationships between data and enhance the feature expression ability; the residual connection branch uses the Identity Mapping mechanism to alleviate the gradient vanishing problem and improve the stability of model training; finally, the outputs of each branch are scaled after being concatenated to suppress gradient explosion and enhance the stability of training. The specific structure of the Inception-Residual module is as Figure 4As shown in (c), the input is fed into four branches for processing. The first branch is the original input without processing. One of the remaining three branches uses a 1×1 convolutional layer, one branch uses a 1×1 convolutional layer and a 3×3 convolutional layer, and one branch uses a 1×1 convolutional layer, a 3×3 convolutional layer, and another 3×3 convolutional layer. The output channels of the three branches are stacked and then fed into a 1×1 convolutional layer. After passing through the Dropout layer, it is connected to the input residual of the first path, and then activated by the ReLU activation function and output.

[0034] After the data enters the Reduction module through the Inception-Residual (I-R) 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 Figure 4 shown in (d). The input is fed into three branches for processing. One branch performs 3×3 max pooling, one branch performs 1×1 convolution, and one branch performs 1×1 convolution, 3×3 convolution, and another 3×3 convolution. The output channels of the three branches are stacked and then fed into the FC layer and the BN layer, and then activated by the ReLU activation function and output.

[0035] 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.

[0036] Compared with the traditional GAN that directly generates data from random noise, DiffGAN provides data structure information through the noisy samples, enabling the generator to learn the data distribution more effectively and improving the authenticity of the generated data. The DiffGAN model proposed in the present invention provides a more reasonable input prior for the generation process of GAN by introducing the noise addition mechanism of the diffusion model, improving the data generation quality and the stability of model training.

[0037] S 4. Input the generated samples and the original training samples into the Resnet model for training to obtain a trained fault detection model, which is used for the intelligent diagnosis of the fracturing truck.

[0038] In the experimental verification, different fault data of the fracturing truck bearings were selected for experiments. The experiments performed sliding window sampling on the original vibration signals of five different health states, and each type of fault signal was sampled 100 times with a window length of 1024. Subsequently, the one-dimensional vibration signals were converted into two-dimensional images of 64×64 pixels through wavelet transform, and a total of 100 images were generated. In the data augmentation training stage, 40 images were randomly selected from each type of fault sample for training the data augmentation model, and the remaining samples were used for subsequent quality assessment. To visually demonstrate the effect of data augmentation and perform augmentation training under the same number of training rounds. Finally, 200 augmented samples were generated for each type of fault, totaling 4×200 samples.

[0039] The quality of the generated data plays a crucial role in the original training set for fault diagnosis, especially the similarity with the original data. When the quality of the generated data is low, the fault diagnosis model may extract features that are significantly different from the original data, thereby affecting or even reducing the diagnostic accuracy. The feature quantitative index J calculates the quantified quality of the generated data through the within-class covariance and between-class covariance: where represents the within-class covariance matrix, represents the between-class covariance matrix, represents the number of classes, represents the number of samples in the c-th class, represents the feature vector of the i-th sample in the c-th class, represents the mean feature vector of this class, represents the overall mean feature vector of all samples.

[0040] The within-class covariance measures the aggregation degree of samples within the same class, while the between-class covariance reflects the discrimination degree between different classes. When the J value is large, the between-class covariance is relatively high, indicating that the sample distributions of different classes are clear, and at the same time, the samples within the same class are more compact. This means that the generated data can not only maintain the consistency within the class but also ensure good discrimination between classes. On the contrary, a lower J value indicates weaker between-class separation or more dispersed sample distributions within the same class, which may lead to insufficient discrimination of the samples after data augmentation and even introduce more noise or outliers.

[0041] The quality of data augmentation is evaluated by calculating the feature quantitative evaluation index J of the generated data, which quantifies the separability between different categories and the consistency within the same category of the generated data. The quantitative evaluation index J of the data generated by DiffGAN is 4.75. The data generated by DiffGAN can maintain the compactness of the data within the category (smaller within-class covariance), while enhancing the discrimination between different categories of data (larger between-class covariance). This shows that the data generated by DiffGAN can not only better match the original data distribution, but also form a clearer decision boundary between different categories, which helps to improve the classification performance of the subsequent model.

[0042] Feature quantitative evaluation can measure the intrinsic properties of data, the separability and consistency of categories. To evaluate the effect of the generated data in actual fault diagnosis, the ResNet model is used in the present invention to accurately evaluate the quality of the generated data in real fault diagnosis. To simulate the training process after data imbalance and data augmentation, the data is allocated as follows: Samples in the normal healthy state do not require data augmentation, 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 respectively, of which 10 are used for training and 50 for testing. In addition, to alleviate the problem of insufficient samples in the fault categories, 200 generated samples of these four types of faults are all used for training to improve the model's ability to identify fault patterns.

[0043] The training parameters of the ResNet model are set as follows: the batch size is 128, the learning rate is 0.0001, and the number of training epochs is 100. Although the accuracy can reflect the overall prediction ability of the model, when dealing with an imbalanced dataset, the accuracy may not fully reflect the prediction effect of the model on the minority class. Therefore, in addition to the accuracy, the F1 value is also used as an evaluation index in the present invention to more comprehensively evaluate the actual performance of the model. In the experiment of training the ResNet model, the F1 value is 95.32 and the ACC value is 95.17.

[0044] To verify the role of the data augmentation method of the present invention in the fault diagnosis task, especially the improvement of the fault category recognition ability under data imbalance, the present invention analyzed the classification effects of different data augmentation methods in the experiment on the ResNet model through a confusion matrix and a clustering diagram. Without data augmentation, the diagonal elements of the confusion matrix are unevenly distributed, the classification accuracy of some fault categories is low, and it is easy to be confused with other categories, indicating that when the data volume is insufficient, the model has a weak ability to recognize fault samples and is difficult to effectively capture their features. After using DiffGAN for data augmentation, the diagonal elements of the confusion matrix are the 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 data imbalance problem, enhance the model's learning ability for fault categories, and improve the accuracy and stability of fault diagnosis.

[0045] The above examples of the present invention are only for explaining in detail the calculation model and calculation process of the present invention, rather than limiting the implementation manner of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made on the basis of the above description. It is impossible to list all the implementation manners here. Any obvious changes or modifications derived from the technical solution of the present invention still fall within the protection scope of the present invention.

Claims

1. A method for generating intelligent diagnosis data of a fracturing truck, characterized in that, It includes the following steps: Collect the vibration signal data of the fracturing truck and convert it into image data as real data samples; input the real data samples 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. The improved multi-scale residual U-net includes 5 multi-scale residual blocks and 4 upsampling convolutional modules, which are sequentially denoted as the first multi-scale residual block to the fifth multi-scale residual block, and the first upsampling convolutional module to the fourth upsampling convolutional module according to the data processing order. The multi-scale residual block includes multiple parallel convolutional branches, and the parallel convolutional branches are processed by convolutional layers with different sizes of convolutional kernels; each upsampling convolutional module includes an upsampling layer and a convolutional unit. A max-pooling layer is set after the first multi-scale residual block to the fourth multi-scale residual block. The output of the first multi-scale residual block to the fourth multi-scale residual block after passing through the max-pooling layer is used as the input of the next multi-scale residual block; the output of the fifth multi-scale residual block is used as the input of the first upsampling convolutional module; at the same time, the output of the first multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the fourth upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units; the output of the second multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the third upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units; the output of the third multi-scale residual block after passing through the convolutional layer is respectively connected to the upsampling layers of the second upsampling convolutional module to the first sampling convolutional module and then sent into the corresponding convolutional units; the output of the fourth multi-scale residual block after passing through the convolutional layer is connected to the upsampling layer of the first sampling convolutional module and then sent into the corresponding convolutional 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 diagnosis data of a fracturing truck according to claim 1, wherein During the process of collecting the vibration signal data of the fracturing truck and converting it into image data, wavelet transform is used to convert the data into 2D RGB data.

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

4. A method for generating intelligent diagnosis data of a fracturing vehicle according to claim 1, characterized in that, The convolutional unit in the upsampling convolutional module includes a convolutional layer, a BN layer, and an activation function ReLU.

5. A method for generating intelligent diagnosis data of a fracturing vehicle 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 judged data is a real sample. Among them, the Stem module uses 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 uses 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 uses a three-branch network structure for feature dimensionality reduction processing.

6. A method for generating intelligent diagnosis 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. After the output channels of the two branches are stacked and passed through the FC layer, it is continued to be divided into two branches for processing. One branch is set with two convolutional layers, and the other is set with four convolutional layers. After the output channels of the two branches are stacked and passed through the FC layer, it is continued to be divided into two branches for processing. Each branch uses one convolutional layer. After the output channels of the two branches are stacked and passed through the FC layer, it is sent to the BN layer. All convolutional layers use convolutional layers with a convolution kernel of 3×3.

7. A method for generating intelligent diagnosis 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 without processing. One of the other three branches uses a 1×1 convolutional layer, one branch uses a 1×1 convolutional layer and a 3×3 convolutional layer, and one branch uses a 1×1 convolutional layer, a 3×3 convolutional layer, and a 3×3 convolutional layer. After the output channels of the three branches are stacked, they are sent to a 1×1 convolutional layer, then processed by the Dropout layer, and then connected to the input residual of the first path. After activation by the ReLU activation function, the output is obtained.

8. A method for generating intelligent diagnosis data of a fracturing vehicle according to claim 5, characterized in that, The processing process of the Reduction module is as follows: The input is sent to three branches for processing. One branch performs 3×3 max pooling, one branch performs 1×1 convolution, and one branch performs 1×1 convolution, 3×3 convolution, and 3×3 convolution. After the output channels of the three branches are stacked, they are sent to the FC layer and the BN layer, and then output after activation by the ReLU activation function.

9. A method for generating intelligent diagnosis data of a fracturing truck according to claim 5, characterized in that, There is 1 Stem module, 2 Inception-Residual modules, and 1 Reduction module in the discriminator respectively.

10. An intelligent diagnosis method for a fracturing vehicle, characterized in that, The vibration signal data of the fracturing truck collected is converted into image data, and then sent to a fault detection model constructed based on a neural network for identifying the fault types of the fracturing truck, realizing the intelligent diagnosis of the fracturing truck. The fault detection model constructed based on the neural network is pre-trained. The sample data in the training set used in the training process includes the generated samples obtained according to the method for generating fracturing truck intelligent diagnosis data described in any one of claims 1 to 9.

Images