A tracked vehicle engine small sample fault diagnosis method based on a generative one-sided adversarial network

Abstract

The application discloses a tracked vehicle engine small sample fault diagnosis method based on a generated unilateral adversarial network. First, data samples of a tracked vehicle engine under different fault states are collected; then, a generated unilateral adversarial network (GUAN) is trained using a small number of fault samples; after all fault types of the tracked vehicle engine obtain a good performance generator network, a large number of false data are generated and mixed with real data samples; finally, a gated recurrent unit (GRU) fault diagnosis network is trained using the mixed data, so that small sample fault diagnosis of the tracked vehicle engine is realized. The application can generate high-quality false samples, expand small sample data to a sufficient data set required by a neural network, save the cost required for collecting tracked vehicle engine fault data, and enable the GRU fault diagnosis network to be applied to tracked vehicle engines lacking sufficient fault samples.

Description

Technical Field

[0001] This invention relates to the field of engine fault diagnosis for tracked vehicles, and specifically to a method for small-sample fault diagnosis by using a generated one-sided adversarial network to expand engine fault samples. Background Technology

[0002] Because tracked vehicles typically operate in harsh environments, their engine systems are subjected to complex conditions such as strong vibrations and extreme temperatures over long periods, resulting in a high probability of failure. Therefore, fault diagnosis measures are needed to facilitate maintenance personnel. Traditional fault diagnosis methods mainly rely on threshold alarm sensor monitoring and the experience and sensory judgment of maintenance personnel, but these methods suffer from low accuracy and poor adaptability. Therefore, researchers have proposed using a gated recurrent neural network (GRU) for real-time fault diagnosis of tracked vehicle engines. However, training a GRU neural network requires a large number of samples, hence some scholars have proposed using a generative adversarial network (GAN) for sample generation.

[0003] Reference [1] proposes a gradient-penalized generative adversarial network to generate a smaller number of samples of different types, and then uses BIGRU to achieve fault diagnosis; Reference [2] proposes a self-attention mechanism to assist classification generative adversarial network to ensure the accuracy of feature extraction; Reference [3] proposes a railway unit fault diagnosis model based on an improved residual network and a generative adversarial network, which improves the residual network and integrates the generative adversarial network to establish a fault diagnosis model to improve feature extraction efficiency.

[0004] However, none of the aforementioned generative adversarial networks can avoid the problems of training instability and overfitting of the discriminator under small sample conditions caused by their network characteristics.

[0005] [1] Wang Peng, Shi Weiping, Cai Shuangle, et al. Fault diagnosis method for circuit breakers based on WGAN-GP-BiGRU with few samples [J]. Electrical Engineering, 2025, (09): 51-56. [2]Zhang Q, Lian [3] Wang Juan, Cai Yonghui. A railway locomotive fault diagnosis model based on improved residual network and generative adversarial network [J]. Microcomputer Applications, 2025, 41(8):236-239. Summary of the Invention

[0006] To address the problem of insufficient fault samples for neural network training in actual production, this invention provides a novel method for small-sample fault diagnosis of tracked vehicle engines, based on a generative unilateral adversarial network (GUAN) to expand the limited number of engine fault samples. This invention generates data based on the GUAN: first, the GUAN is pre-trained using the original data to obtain basic generator parameters; then, only the generator network is retained, and a new round of training and parameter updates are performed using a combination of multiple loss functions.

[0007] The present invention adopts the following technical solution: the method for small-sample fault diagnosis of tracked vehicle engine based on generative one-sided adversarial network includes the following steps: Step 1: Collect historical vibration data of the tracked vehicle engine under six operating conditions. This includes six operating conditions identifiable by output vibration signals: normal operating condition, cylinder knocking fault, ignition failure fault, valve knocking fault, piston knocking fault, and bearing wear fault; use... The original sample set representing the normal working state. The original sample set representing cylinder block knocking failure. The original sample set representing ignition failures. The original sample set representing valve knocking faults. The original sample set representing piston knocking failure. The original sample set representing bearing wear failure; For the original sample set ~ Normalization is performed to map the data to the range [-1, 1], resulting in sample sets. ; n=1,2,...,6; Next, from A specified number of samples are extracted from the dataset to form a training dataset. The remaining sample data constitute the validation dataset. ; Step 2: When a certain type of fault sample set is in the training dataset T When the amount of data is insufficient, the Generative One-Sided Adversarial Network (GUAN) is used for learning. The sample distribution is used to obtain the generator network; the generator one-sided adversarial network refers to performing one-sided training in generative adversarial learning, fixing the discriminator network, and updating the generator network only through adversarial loss; Using fault samples with insufficient data Obtain the corresponding generator network Then, let each generator network Generate a sufficient number of fake samples corresponding to the faults and compare them with the real sample set. The dataset is mixed to obtain a hybrid dataset, which is then used to replace the real sample set for that fault category; the resulting dataset forms a new training dataset. ; Step 3: Construct a tracked vehicle engine fault diagnosis model using GRU network layers, Dropout network layers, and fully connected layers. Set the parameters for the corresponding networks and use the dataset obtained in Step 2. The validation dataset V obtained in step one is used as the training set and fed into the model for training. The model with the smallest validation loss value is saved during the iteration process, and the final model is used for fault diagnosis of tracked vehicle engine.

[0008] Specifically, in step two, a generative one-sided adversarial network is used for learning. The method to obtain the generator network from the sample distribution is as follows: Step 2.1: Construct a generative one-sided adversarial network, training it using a training set T and a validation set V to obtain the generator network, denoted as . ; Step 2.2: Use the multi-kernel maximum mean difference (MK-MMD) algorithm to quantify the data distribution of real and fake samples, and calculate the average data distribution distance of all real samples: Calculate the data distribution distance between the fake and real sample sets: Calculate the average data distribution distance for all fake samples: in, This represents the MK-MMD algorithm; and These represent individual real and fake samples, respectively. Indicates and Similar fake datasets; ; ; and These represent the total number of real and fake input samples, respectively. Step 2.3, using , , and The quality of fake samples is determined by a loss function. This involves integration to adjust the direction of the generator's parameter updates; The formula is as follows: in, This represents the maximum similarity loss value used during training, determined by... The similarity is obtained by calculating the similarity between all fake and real samples in each training batch; Step 2.4: Set the optimization algorithm to Adaptive Moment Estimation Adam, set the corresponding parameters and training epochs, and save the generator network when the loss function is minimized; Step 2.5: Repeat steps 2.1 to 2.4 to obtain the corresponding generator network for each type of fault sample with insufficient data.

[0009] Specifically, step 2.5 uses the trained generator network to generate a specified number of fake samples for the corresponding fault. The density values ​​of fake samples and real samples are used to determine whether the generated samples meet the standards. A density threshold is set. When the density of a certain type of fake sample does not reach the threshold, the one-sided adversarial network is retrained to regenerate the corresponding type of fake sample.

[0010] The advantages of this invention are: This invention utilizes a generative one-sided adversarial network (GAN) to generate fault data, addressing the problem of insufficient fault data. It effectively avoids the poor quality of generated samples caused by training instability and overfitting in traditional GANs, thereby improving subsequent fault diagnosis accuracy. This invention uses a density metric to evaluate the generator's performance, ensuring that the fault samples generated by the model are of good quality, allowing the fault diagnosis model to demonstrate good fault diagnosis accuracy on the test set.

[0011] Because this invention can generate high-quality fake samples, it expands small sample data into a sufficient dataset required by the neural network, saving the cost required to collect tracked vehicle engine fault data. Therefore, the GRU fault diagnosis network can be applied to tracked vehicle engines that lack sufficient fault samples. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a pre-trained generator.

[0013] Figure 2This is a schematic diagram illustrating the principle of training and generating a one-sided adversarial network.

[0014] Figure 3 These are the parameters for generating a one-sided adversarial network generator.

[0015] Figure 4 These are the parameters for generating the discriminator of the one-sided adversarial network.

[0016] Figure 5 This is the overall flowchart of the present invention.

[0017] Figure 6 This is a structural diagram of the GRU fault diagnosis network.

[0018] Figure 7 This is a graph showing the loss values ​​of a GRU model trained using small sample data.

[0019] Figure 8 This is a graph showing the loss values ​​of a GRU model trained using mixed data. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments, so that the present invention can be more easily understood by those skilled in the art.

[0021] like Figure 5 As shown, the implementation steps of the present invention are as follows.

[0022] Step 1: Building the training and test datasets.

[0023] Historical output vibration data of tracked vehicle engines under different operating conditions were collected and preprocessed.

[0024] 1.1 Data Acquisition.

[0025] To simulate real-world fault conditions, the tracked vehicle engine data in this embodiment was obtained from historical fault data collected under different operating conditions. For each of the six engine operating conditions, one hour of data was collected. The sampling frequency was 0.01 seconds per sampling.

[0026] The dataset contains experimental data on the engine under different operating conditions. It is divided into normal operating conditions, cylinder knocking faults, ignition failure faults, valve knocking faults, piston knocking faults, and bearing wear faults.

[0027] A sliding window was used to create a sequence of 300 time points, with 1200 samples for each working state. The working state of each sample was selected as the sample label, resulting in the following data samples for each working state: in, This represents a sample of raw data with a length of 300 time points. The original sample set representing the normal working state. The original sample set representing cylinder block knocking failure. The original sample set representing ignition failures. The original sample set representing valve knocking faults. The original sample set representing piston knocking failure. The original sample set representing bearing wear failure.

[0028] 1.2 Data preprocessing.

[0029] For the original data sample Perform normalization processing, Elements normalized to the range [-1, 1]: in, This represents the new data sample after normalization. This represents the maximum value in the data. This represents the minimum value in the data. This represents the average value of the data.

[0030] The sample obtained after normalization is as follows: in, It is a data sample with a length of 300 time points. This represents the sample set after preprocessing under normal working conditions. This represents the sample set after preprocessing for cylinder block knocking faults. This represents the sample set after preprocessing for ignition failure. This represents the sample set after preprocessing for valve knocking faults. This represents the sample set after preprocessing for piston knocking faults. This represents the sample set after preprocessing for bearing wear faults.

[0031] 1.3. Divide the training set and the validation set.

[0032] In this embodiment, 12 samples are selected for each type of fault. Since the number of samples in the normal working state is much larger than the number of fault samples, 112 samples in the normal working state and samples of each type of fault are selected to form the training set. : in, This represents a sample set consisting of 112 normally functioning samples. This represents a sample set consisting of 12 samples for each type of fault. This indicates that the sample does not exist.

[0033] Then, all remaining samples from the six data categories are used to form a validation set. This is used to verify the model's effectiveness.

[0034] Step 2: Data augmentation based on generating a one-sided adversarial network.

[0035] When a certain type of dataset is in the original training set T When the number is small, this invention uses generative one-sided adversarial networks for learning. The sample distribution is determined to obtain a generator network that can effectively generate usable data samples; the generated fake samples and real samples are mixed to obtain a hybrid dataset. Replace the original dataset with less data to form a new training set. This is the training set that will be used for model training.

[0036] The generator and discriminator of the one-sided adversarial network are both composed of 5 layers of one-dimensional convolutional neural networks. The parameters of the generator and discriminator are as follows: Figure 3 and Figure 4 As shown.

[0037] The generator network maps its output features to the interval [-1, 1] using the hyperbolic tangent (Tanh) activation function, thus aligning with the range of data normalization. The objective function of the generator unilateral adversarial network is... for: in, These are real sample data. These are the generated sample data. and Let these represent the probability distributions of the real sample data and the generated sample data, respectively. Represents the mathematical expectation. express obey distributed, express obey distributed, It is the output of the discriminator D, the gradient penalty term. , It is the penalty coefficient. It involves random interpolation sampling between generated samples and real samples. , This represents a random value within the interval [0, 1). It is the gradient of the discriminator D.

[0038] The specific steps for generating fault samples using the above-described one-sided adversarial network include: Step 2.1: Construct a generative one-sided adversarial network, such as... Figure 1 As shown, the generator network is trained using a training set T and a validation set V to obtain the generator network, denoted as . ; Step 2.2: Quantize the data distribution of real and fake samples using the MK-MMD algorithm, such as... Figure 2 As shown, the formula is expressed as follows: in, This represents the average data distribution distance across all real samples; This represents the data distribution distance between the fake and real sample sets; This represents the average data distribution distance of all fake samples; This represents the MK-MMD algorithm; and These represent individual real and fake samples, respectively. Indicates and Similar fake datasets; ; ; and These represent the total number of real and fake input samples, respectively.

[0039] Step 2.3, using , , and The quality of fake samples is determined by a loss function. This allows for integration, thereby adjusting the direction of the generator's parameter updates.

[0040] The formula is shown below: in, This represents the maximum similarity loss value used during training, determined by... The similarity is obtained by calculating the similarity between all fake and real samples in each training batch.

[0041] Step 2.4: Set the optimization algorithm to Adam, with parameters weight_decay = 0.001, betas = (0.0, 0.999), discriminator learning rate of 0.0001, batch_size = 6, and Gaussian white noise. The format is (batch_size, 80); then the training iterations are set to 6000, and the generator with the smallest loss value is saved starting from the 5000th iteration. The trained generator will master the data distribution, generate data, and the generated data will have the same data distribution as the data in the small sample.

[0042] Step 2.5: Repeat steps 2.1 to 2.4 to obtain a corresponding generator network for each fault.

[0043] Use the trained model to generate 100 fake samples for each type of fault. The density value of the generated sample is compared with that of the real sample to determine whether the generated sample meets the standard. The parameter is set as: nearest neighbor distance. =5, actual sample size =12. The formula for calculating the density index is: in, Indicates the index function, Indicates a fake sample. Indicates real data The surrounding radius is spherical region For a set of real samples, This represents the distance from each real data point to its corresponding real data point. The nearest neighbor distance.

[0044] Substitute the actual values ​​from the implementation examples into the density index calculation formula: The density threshold is set to 0.9. When the density of a certain type of fake sample generated does not reach the threshold, the one-sided adversarial network is retrained to regenerate the corresponding type of fake sample.

[0045] Finally, the density values ​​of the corresponding types of fault samples generated by each generator network in the embodiment are shown in Table 1.

[0046] Table 1 After generating a sufficient number of fake samples with clear labels, they are mixed with real samples to obtain the dataset required for training the fault diagnosis model.

[0047] In the example, 100 fake samples are generated for each type of fault. Its formula is as follows: in, This represents a sample set consisting of 112 samples in normal working condition. This represents a mixed sample set consisting of 12 real fault samples and 100 fake samples.

[0048] Step 3: Engine fault diagnosis for tracked vehicles based on the GRU algorithm.

[0049] By studying the engine data of tracked vehicles, it can be seen that the relationship between the data and time is precise. Therefore, GRU is used as the fault diagnosis model for tracked vehicle engines.

[0050] The architecture of the fault diagnosis model for tracked vehicle engines is as follows: Figure 6 As shown, the model uses GRU to extract features from engine parameter data and implements fault classification through Dropout and fully connected layers. The input feature dimension of GRU is set to 4, the feature dimension of the hidden layer to 64, Dropout=0.25, the dimensions of the two fully connected layers are 32 and 6 respectively, the model iterations are 600, the training batch is 64 samples per batch, the validation batch is 32 samples per batch, the optimization algorithm is Adam, the learning rate is set to 0.0001, and other parameters are at their default values.

[0051] Then, the dataset obtained in step two... As training set, validation set The two datasets obtained in step one are then used to train the fault diagnosis model. When the validation loss function is minimized during the iteration process, the model data is saved, resulting in the final usable fault diagnosis model.

[0052] The effectiveness of the method of the present invention is verified by obtaining the diagnostic accuracy through experiments.

[0053] Firstly, for small training sets and verification set Directly using a GRU neural network for fault diagnosis, the loss curve is as follows: Figure 7 As shown, it is clear that the fault diagnosis under small sample conditions exhibits overfitting, making the model difficult to apply in actual industrial production.

[0054] Then, the sample after adding the generated fault samples. and verification set The model loss curve is shown below when the GRU neural network is trained again. Figure 8 As shown, the neural network can be trained normally, demonstrating the effectiveness of the generated fault data.

[0055] The finally trained fault diagnosis model was applied to the signal acquisition of an actual vehicle engine, and its diagnostic accuracy was tested over ten days, as shown in Table 2. It can be seen that the model can effectively identify various fault states.

[0056] Table 2

Claims

1. A method for small-sample fault diagnosis of tracked vehicle engines based on generative one-sided adversarial networks, characterized in that, Includes the following steps: Step 1: Collect historical vibration data of the tracked vehicle engine under six operating conditions. These six operating conditions include: normal operation, cylinder knocking, ignition failure, valve knocking, piston knocking, and bearing wear. The original sample set representing the normal working state. The original sample set representing cylinder block knocking failure. The original sample set representing ignition failures. The original sample set representing valve knocking faults. The original sample set representing piston knocking failure. The original sample set representing bearing wear failure; For the original sample set ~ Normalization is performed to map the data to the range [-1, 1], resulting in sample sets. ; n=1,2,...,6; Next, from A specified number of samples are extracted from the dataset to form a training dataset. The remaining sample data constitute the validation dataset. ; Step 2: When a certain type of fault sample set is in the training dataset T When the amount of data is insufficient, generative one-sided adversarial networks are used for learning. The sample distribution is used to obtain the generator network; the generator one-sided adversarial network refers to performing one-sided training in generative adversarial learning, fixing the discriminator network, and updating the generator network only through adversarial loss; Using fault samples with insufficient data Obtain the corresponding generator network Then, let each generator network Generate a sufficient number of fake samples corresponding to the faults and compare them with the real sample set. The dataset is mixed to obtain a hybrid dataset, which is then used to replace the real sample set for that fault category; the resulting dataset forms a new training dataset. ; Step 3: Construct a fault diagnosis model for the tracked vehicle engine using GRU network layers, randomly deactivated network layers, and fully connected layers. Set the parameters for the corresponding networks and use the dataset obtained in Step 2. The validation dataset V obtained in step one is used as the training set and fed into the model for training. The model with the smallest validation loss value is saved during the iteration process, and the final model is used for fault diagnosis of tracked vehicle engine.

2. The method for small-sample fault diagnosis of tracked vehicle engines based on generative one-sided adversarial networks as described in claim 1, characterized in that, Step two uses a generative one-sided adversarial network for learning. The method to obtain the generator network from the sample distribution is as follows: Step 2.1: Construct a generative one-sided adversarial network, training it using a training set T and a validation set V to obtain the generator network, denoted as . ; Step 2.2: Use the multi-kernel maximum mean difference (MK-MMD) algorithm to quantify the data distribution of real and fake samples, and calculate the average data distribution distance of all real samples: Calculate the data distribution distance between the fake and real sample sets: Calculate the average data distribution distance for all fake samples: in, This represents the MK-MMD algorithm; and These represent individual real and fake samples, respectively. Indicates and Similar fake datasets; ; ; and These represent the total number of real and fake input samples, respectively. Step 2.3, using , , and The quality of fake samples is determined by a loss function. This involves integration to adjust the direction of the generator's parameter updates; The formula is as follows: in, This represents the maximum similarity loss value used during training, determined by... The similarity is obtained by calculating the similarity between all fake and real samples in each training batch; Step 2.4: Set the optimization algorithm to Adaptive Moment Estimation Adam, set the corresponding parameters and training epochs, and save the generator network when the loss function is minimized; Step 2.5: Repeat steps 2.1 to 2.4 to obtain the corresponding generator network for each type of fault sample with insufficient data.

3. The method for small-sample fault diagnosis of tracked vehicle engines based on generative one-sided adversarial networks as described in claim 2, characterized in that, Step 2.5 Use the trained generator network to generate a specified number of fake samples for the corresponding fault. The density values ​​of fake samples and real samples are used to determine whether the generated samples meet the standards. A density threshold is set. When the density of a certain type of fake sample does not reach the threshold, the one-sided adversarial network is retrained to regenerate the corresponding type of fake sample.

Images