Pipeline Valve Internal Leakage Fault Diagnosis Method Based on DCGAN-DCNN

Abstract

The present invention relates to the technical field of pipeline valves, and particularly to a method for diagnosing internal leakage faults of pipeline valves based on DCGAN-DCNN, including collecting acoustic emission data of pipeline valves under normal working conditions and different leakage conditions; exporting the acoustic emission data as voltage data and performing preprocessing; constructing a DCGAN network model, optimizing the objective function, training the model using batch normalization, and updating the parameters using the stochastic gradient descent method to expand the fault signal data; evaluating the accuracy of the DCGAN network model through a similarity comparison algorithm; training a DCNN model using the expanded data; and selecting the proportion of generated data based on the model recall rate, precision rate, and F1 score. The present invention aims to solve the problem of poor fault diagnosis effect of existing valve fault diagnosis models in the case of insufficient sample data.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline valves, and particularly to a method for diagnosing internal leakage faults of pipeline valves based on DCGAN-DCNN. Background Art

[0002] A valve is a control device for regulating the internal medium of a pipeline or opening and closing a pipeline channel, and is an important part of an urban gas pipeline network. Due to reasons such as operation errors and untimely valve replacement, internal leakage of the valve may occur. If not dealt with in time, serious accident consequences will be caused.

[0003] The acoustic emission technology is widely used in the research on valve internal leakage diagnosis. Since the acoustic emission signal is an unpredictable non-stationary sudden instantaneous signal, the collected signal contains fault signals and a large amount of noise signals, making data processing and feature extraction relatively difficult.

[0004] Currently established fault diagnosis models are mainly obtained through training with a large amount of data sets. However, in actual working conditions, usually for safety reasons, only a small amount of fault data can be collected. Small sample category data is likely to cause problems such as instability during training, vanishing gradients, and easy collapse of the model. Summary of the Invention

[0005] In view of the deficiencies of the existing algorithms, the present invention aims to solve the problem of poor fault diagnosis effect of the existing valve fault diagnosis model under insufficient sample data. A DCGAN structure under a small sample data set is designed, a pipeline valve internal leakage fault diagnosis model based on DCGAN-DCNN is constructed, and it is verified whether the data generated by DCGAN can improve the diagnosis effect of the DCNN valve internal leakage fault diagnosis model.

[0006] The technical solution adopted by the present invention is: a method for diagnosing internal leakage faults of pipeline valves based on DCGAN-DCNN, including the following steps:

[0007] Step 1: Collect acoustic emission data of pipeline valves under normal working conditions and different leakage conditions;

[0008] Further, the pipeline valves under different leakage conditions are set by adjusting the opening of the regulating gate valve.

[0009] Step 2: Export the acoustic emission data as voltage data and perform preprocessing;

[0010] Further, the preprocessing includes: removing the singular values of the data, adding category labels to the data, and normalizing the data.

[0011] Step 3: Construct a DCGAN network model, optimize the objective function, train the model using batch normalization, update the parameters of the DCGAN network using the stochastic gradient descent method, and expand the fault signal data.

[0012] Furthermore, the generator of the DCGAN network model consists of six transposed convolutional layers and one fully connected layer. The generator consists of six transposed convolutional layers and one fully connected layer. First, Gaussian noise is input into the fully connected layer, and the input noise is mapped to the dimension for transposed convolution in the DCGAN network. The first layer maps the input of a large dimension to the output of a small dimension, and the subsequent layers all map the data dimension of the input to twice the previous one as the output. The dimension of the data increases by a factor of two layer by layer until the output dimension is 512. The second to sixth layers all perform upsampling of the acoustic emission signal through transposed convolution operations, gradually expanding the dimension of the output data to realize the generation of the acoustic emission signal. The discriminator consists of six convolutional layers and one fully connected layer. First, the preprocessed acoustic emission signal and the signal generated by the generator are used as the input of the model. The convolutional layer is used to perform downsampling on the signal. The dimensions of the first six layers are halved layer by layer. The last layer maps the input of the previous layer to a value of 1*1 dimension through the fully connected layer as the output of the discriminator network. Finally, through the fully connected layer, the features extracted by the convolutional kernel are mapped to a value of 1*1 as the output of the discriminator.

[0013] Furthermore, the formula for optimizing the objective function is:

[0014]

[0015] In the formula, D(X) is the output of the real data X on the discriminator D; G(z) is the output of the random vector z in the random noise in the generator G, and D[G(z)] is the probability that the discriminator judges whether the generated signal G(z) is real, P data is the random sampling of real data, and P G is the prior distribution of the vector z.

[0016] Furthermore, when training the model using batch normalization, the parameters γ and β are introduced.

[0017]

[0018] β = E[X]

[0019] Among them, β is the average value of the input values of each neuron; γ is the standard deviation of the input values of each neuron.

[0020] Step 4: Evaluate the accuracy of the DCGAN network model through a similarity comparison algorithm;

[0021] Furthermore, it specifically includes:

[0022] Compare the similarity between the image data after training the DCGAN network model and the original image, and calculate the correlation distance between the images through the Pearson correlation coefficient algorithm.

[0023] Step 5: Train the DCNN model using the augmented data;

[0024] Further, it specifically includes:

[0025] The DCNN model consists of four convolutions: an input layer, a convolutional layer, a pooling layer, and a fully connected layer. Each convolution includes 2 convolutional layers and a max pooling layer. The stride of the convolutional kernel of the first 3 convolutions is set to 2, and the stride of the convolutional kernel of the 4th unit is set to 1; The padding method is set to Same, padding 0 for the data edges to prevent the loss of data edge features; The activation function is set to ReLU. An average pooling layer is added at the end of the 4th convolution to further compress the features and Dropout is set. Finally, the probability distribution of the categories is output through the fully connected layer.

[0026] Step 6: Select the proportion of the generated data based on the model recall rate, precision rate, and F1 score.

[0027] Advantages of the present invention:

[0028] 1. Aiming at the problem of low accuracy of the pipeline valve CNN model caused by small sample datasets, the valve fault diagnosis model based on DCGAN and DCNN expands the sample set and solves problems such as poor model diagnosis effect in the case of small samples.

[0029] 2. In DCGAN, adding a transfer BN layer can mitigate the ICS problem, enabling a larger learning rate to be used when training the neural network, making the network converge faster, and thus reducing the need for regularization penalty terms when training the neural network. Brief Description of the Drawings

[0030] Figure 1 is the flow chart of the pipeline valve internal leakage fault diagnosis method based on DCGAN-DCNN of the present invention;

[0031] Figure 2 is the structure diagram of the generator and discriminator of the present invention;

[0032] Figure 3 is the BN operation diagram of the present invention;

[0033] Figure 4 is the structure diagram of the DCNN model of the present invention;

[0034] Figure 5 is the original fault diagram of the gate valve;

[0035] Figure 6 is the DCGAN generated data diagram;

[0036] Figures 7(a) and (b) are the loss function curves of the training set and the validation set respectively;

[0037] Figures 8(a) and 8(b) are the accuracy curves of the training set and the validation set respectively. Detailed implementation manners

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only schematically shows the basic structure of the present invention, so it only shows the components related to the present invention.

[0039] As Figure 1 shown, the pipeline valve internal leakage fault diagnosis method based on DCGAN-DCNN includes the following steps:

[0040] Step 1: Collect pipeline valve internal leakage fault signal data under different leakage conditions; in this embodiment, a 20# seamless carbon steel pipe with a diameter of 25 mm and a wall thickness of 5 mm is selected to simulate a gas pipeline to build an experimental platform, and pipeline valve internal leakage fault signal data is collected;

[0041] Adjust the gate valve opening to fully closed, 1 turn, 3 turns, and 5 turns; among them, the fully closed state is the normal state of the gate valve; successively conduct gate valve fault simulation experiments, each situation is collected for 90 seconds, and the acoustic emission data under each leakage condition of the gate valve is saved;

[0042] Step 2: Export the collected acoustic emission signals into voltage value data through acoustic emission software, and use the Numpy and Pandas modules in Python to preprocess the data, mainly removing the singular values of the data, adding category labels to the data, and normalizing the data.

[0043] Step 3: Build a DCGAN network model: As Figure 2 , a network model for generating valve fault signals is built according to the characteristics of valve acoustic emission signals; this model consists of a generator and a discriminator, and both the generator and the discriminator are composed of convolutional neural networks; among them, the generator consists of six transposed convolutional layers and a fully connected layer; first, Gaussian noise is used as the input of the model's fully connected layer, and the input noise is mapped to a dimension suitable for neural network transposed convolution. The first layer maps the large-dimensional input to a small-dimensional output, and the subsequent layers all map the input data dimension to twice the previous one as the output, and the dimension of the data increases by one time layer by layer until the output dimension is 512, which is the same as the original data dimension; the subsequent layers all realize the upsampling of the acoustic emission signal through transposed convolution operations, and gradually expand the dimension of the output data to realize the generation of the acoustic emission signal.

[0044] The discriminator consists of six convolutional layers and one fully connected layer. First, the preprocessed acoustic emission signal and the signal generated by the generator are used as the inputs of the model. The convolutional layers are used to downsample the signals. The dimensions of the first six layers are halved layer by layer. The last layer maps the input of the previous layer to a value with a dimension of 1*1 through the fully connected layer as the output of the discriminator network. Finally, through the fully connected layer, the features extracted by the convolutional kernels are mapped to a value of 1*1 as the output of the discriminator.

[0045] Step 31: The generation network G is responsible for generating data from random noise, and the discriminant network D is responsible for distinguishing the generated data from the real data as much as possible, that is, maximizing the objective function V(D, G). The generation network G is responsible for continuously generating data that can fool the discriminator, that is, minimizing the objective function V(D, G).

[0046] Step 32: The objective function that needs to be continuously optimized for adversarial training is as follows:

[0047]

[0048] In the formula, D(X) is the output of the real data X on the discriminator D; G(z) is the output of the random vector z in the random noise in the generator G, and D[G(z)] is the probability that the discriminator judges whether the generated signal G(z) is real, P data is randomly sampled real data, P G is the prior distribution of the vector z.

[0049] Step 33: The transfer function uses the Sigmoid function, and the batch normalization (BN) technology is used to optimize the model training. Its core operations are as Figure 3 shown. The size of the batch is n, that is, the inputs X1, X2, …, X n , X n are multiplied by the weights W n to obtain the corresponding output values S n . In the ordinary neural network structure, after obtaining the output values, they will be passed to the corresponding activation functions. After being processed by the activation functions, they will be passed to the next layer. But here, the BN operation is embedded. First, the mean μ and variance σ corresponding to S1, S2, …, S n will be calculated, and then the current value S n is subtracted by the mean μ and then divided by the square root of the variance σ squared plus ∈ to obtain the new S n , where ∈ is just a tiny positive number to avoid the case where the variance σ is 0. At this time, the distribution of the calculated S n will generally be restricted to a normal distribution with a mean of 0 and a variance of 1, which will reduce the fitting ability of the network. To avoid this situation, two parameters γ and β are introduced, and these two parameters are learned automatically during model training. The formulas for both are as follows:

[0050]

[0051] β = E[X]

[0052] β is the average value of the input values of each neuron when there is a batch of training data; γ refers to the standard deviation of the input values of each neuron when there is a batch of training data.

[0053]

[0054] S n = γ·S n +β ⑶

[0055] Step 34: During the model training process, use the stochastic gradient descent method to update the parameters. The judgment result is used as a feedback for the generator network and the discriminator network, and the two form an adversarial mechanism to perform cross-training for multiple rounds until reaching or approaching the Nash equilibrium. The training ends until the DCGGAN can generate a large number of pseudo-samples highly similar to the original data. Then, use the python tool to convert the data into images.

[0056] According to the generator and discriminator network structures of the DCGAN, use the preprocessed acoustic emission signals of the valve as the training data set, and use the data sets under the four states of the valve to train the generative adversarial network respectively. To match the Sigmoid transfer function, the learning rate is set to a relatively small value of 0.01. Slice the acoustic emission data with a length of 5000 into data with a length of 512, as Figure 5 shown; based on the principle of small sample data augmentation, only take 200 pieces of data with a length of 512 for each working condition for DCGAN training. Each time, import 10 training samples, and each sample uses the gradient independently. During the training process, the number of sample iteration steps randomly sampled in each round is set to 30 steps. After training 3000 times, the data generated by the DCGAN is as Figure 6 shown.

[0057] Step Four: Compare the similarity between the images obtained after each training and the original images. Among them, the image distance is called the correlation distance. When the correlation distance between the generated images and the original images of each category reaches the lowest, the finally obtained is a trained DCGAN fault data generation model; the present invention uses the correlation distance (Pearson correlation coefficient) to measure the similarity between the original signal curve and the generated signal curve in each state.

[0058] The relevant formula is as follows:

[0059] Among them, In the formula, N is the number of sample points in the data set, σ x 、σ yThey are the average value and standard value difference of variables x and y respectively. r represents the correlation coefficient. The closer the absolute value of r is to 1, the stronger the similarity. On the contrary, the closer r is to 0, the weaker the similarity. Usually, 0.9 - 1 indicates extremely strong similarity.

[0060] The calculation results are shown in the table as follows:

[0061] It can be seen from Table 1 that the signals generated under each category are extremely similar to the original signal, verifying that the overall generation effect of the model of the present invention has reached the requirements.

[0062] Table 1 Similarity between the generated signal and the original signal

[0063]

[0064] Such as Figure 6 It can be seen from the visualization output by DCGAN that the categories from top to bottom are no fault, mild fault, moderate fault, and severe fault respectively. The generated signal context is clear on the abscissa, and the variation range of the voltage value on the ordinate is basically consistent with that of Figure 5 the original signal.

[0065] Step 5: Build a DCNN model: It consists of 4 segments of convolution, each segment of convolution includes 2 convolutional layers and a max pooling layer. Set the moving step of the convolutional kernel of the first 3 units to 2, and the convolutional kernel step of the 4th unit to 1 to ensure better feature extraction; The padding method is set to Same, padding 0 for the data edge to prevent the loss of data edge features; To accelerate the convergence speed, the activation function is set to ReLU. Add an average pooling layer at the end of the 4th unit to continue compressing features and set Dropout to 0.3 to randomly discard 30% of the neurons to prevent the model from overfitting. Finally, output the probability distribution of the category through the fully connected layer.

[0066] The operation process of each neuron in DCNN is as follows:

[0067] f(x) = act(Σθ (n-i) x ij + b), (5)

[0068] Among them, act() represents the activation function, θ is the weighted value of the neuron, and b is the added bias.

[0069] The activation function of the DCNN model is Relu. In the prior art, the Sigmoid activation function is often selected. The function of the Sigmoid activation function is to compress the value of the function. If the input is a particularly large positive number, it will output 1; if the input is a very large negative number, it will output 0. This is the characteristic of the Sigmoid function and also its drawback. If the input value is very large or very small, after being output by the Sigmoid function, the gradient of the neuron approaches 0, resulting in the neuron being in a state of gradient disappearance. If most neurons are in a state of gradient disappearance, it will be very difficult for the neural network to converge.

[0070] The ReLU function cuts off the part below the x-axis on the basis of the function graph, making the ReLU function only take numbers greater than or equal to 0 as input and output values greater than or equal to 0, having a certain unilateral inhibitory property; the ReLU function has good non-linear characteristics, making the calculation of the gradient relatively simple and the training process relatively stable.

[0071] 150 groups of data collected under laboratory conditions are extracted for each working condition, and 50 groups of data generated by the improved DCGAN are extracted for each working condition after preprocessing. A total of 800 groups of real data and generated data are respectively input into the 1DCNN while keeping the proportion of the generated data in the total data at 25%, 50%, 75%, and 100%. The model structure and parameters are kept the same as before, and after training for 50 generations, the loss function curves in the training set and validation set are shown in Figures 7(a) and (b), and the accuracy curves are shown in Figures 8(a) and (b).

[0072] Step Six: Select the proportion of the generated data based on the model recall rate, precision rate, and F1 score.

[0073] Analysis of the loss function graph:

[0074] It can be seen from the loss function curves of the training set and the validation set that in each case, the loss function of the model drops rapidly in the first 10 generations. When the proportion of generated data is 25%, the loss function of the validation set fluctuates between 0.2 and 0.4 between the 10th generation and the 20th generation, and the entire model tends to stabilize after the 20th generation; when the proportion of generated data is 50%, there are small fluctuations in the curve all the time. The model reaches the minimum value in the 8th generation and then fluctuates between 0 and 0.2 all the time. Although it tends to stabilize, the overall situation of the loss function still shows a downward trend; when the proportion of generated data is 75%, the model reaches the minimum value in the 8th generation and then fluctuates between 0 and 0.2 all the time. The convergence speed is relatively fast, but there are violent fluctuations in the loss function of the validation set in the 37th generation. The loss function of the validation set rises to 0.49, and the loss function of the training set also reaches 0.23 at this point. After the 40th generation, the model becomes stable again; when the proportion of generated data is 100%, that is, all training data are generated data, the model gradually tends to stabilize after the 30th generation. The convergence speed is slower than the previous situation, and the fluctuations of the model between the 20th and 30th generations are more obvious. The loss function of the validation set rises above 0.3 many times; when the proportion of generated data is 0, that is, all experimental data are used, the loss function and the accuracy rate tend to stabilize after the 10th generation. The average accuracy rate of the validation set reaches 93.72%. Among them, the accuracy rate reaches the highest 99.58% in the 45th generation, and the average accuracy rate of the validation set between the 10th generation and the 50th generation reaches 97.14%.

[0075] Analysis of the accuracy rate function image:

[0076] It can be seen from the accuracy rate curve that all models have achieved relatively high recognition accuracy rates in the training set and the validation set, and the model fitting degree is good. However, when the proportion of generated data is 75%, the accuracy rate of the validation set of the model drops to 0.81 in the 37th generation, and the accuracy rate curve of the training set also drops to 0.89. After the 40th generation, it becomes stable again, indicating that the model cannot maintain a relatively high accuracy rate for data recognition in the 37th generation. When the proportion of generated data is 100%, that is, all training data are generated data, the accuracy rate curve keeps rising before the 14th generation and fluctuates violently between the 14th generation and the 30th generation. The lowest accuracy rate of the validation set has dropped to 85.67%, and the highest has reached 97.83%. After the 30th generation, the model tends to stabilize, and the accuracy rate of the validation set is stable between 94% and 98%, indicating that although the generated data has a slower convergence speed for the training of the model, as the number of training generations increases, the model will eventually tend to stabilize, and finally the accuracy rate of the validation set reaches the highest 98.17% in the 44th generation.

[0077] Performance evaluation of the model:

[0078] To further verify the reliability of the model, 400 groups of untrained test data sets were input into each model for classification and recognition. Metrics such as recall rate, precision rate, and F1 score calculated based on the obtained confusion matrix were used to further evaluate the performance of the model on the test set.

[0079] (1) Model recall rate

[0080] As an important metric for evaluating deep learning models, the recall rate refers to the proportion of correctly predicted samples for a certain type of fault among all samples of that type of fault. Its purpose is to evaluate the recognition effect of the model for each type of fault. The summary results of the model's recall rate are shown in Table 2.

[0081] Table 2: Comparison of recall rates of generated data with different proportions

[0082]

[0083] According to the data in the table, it can be seen that the model has a good recognition effect on the normal and severe fault conditions of the valve, and a poor recognition effect on the other two cases. When the model is trained entirely with data under laboratory conditions, the average recall rate of the model for the 4 types of faults reaches 97.25%. As the proportion of generated data increases, the average recall rate of the model decreases. Finally, when the proportion of generated data is 100%, the average recall rate drops to 88.5%.

[0084] (2) Model precision rate

[0085] The precision rate refers to the proportion of correctly predicted samples for a certain type of fault among all samples predicted as that type of fault. The summary results of the model's precision rate are shown in Table 3:

[0086] Table 3: Comparison of precision rates of generated data with different proportions

[0087]

[0088] According to Table 3, it can be seen that when the proportion of generated data does not exceed 50%, the average precision rate of the model can reach over 95%. When the proportion of generated data is 75%, the recognition precision rate of the model for moderate faults drops to 87.5%. The model misidentifies other working conditions as moderate faults in a relatively large number, resulting in a larger base number and a decrease in the precision rate. When the proportion of generated data is 100%, the model only has a relatively high recognition precision rate for the normal working condition, and the other 3 types of faults do not reach over 90%.

[0089] (3) Model F1 score

[0090] The F1 score combines the values of recall and precision and can reflect the output effect of the model. It is mainly obtained by calculating the harmonic mean of recall and precision, and its value range is between 0 and 1. The closer it is to 1, the better the model's effect. The summary results of the F1 scores of the model are shown in Table 4.

[0091] Table 4 Comparison of F1 scores of generated data with different ratios

[0092]

[0093] As can be seen from Table 4, when the proportion of generated data does not exceed 50%, the output effect of the model is better; among them, the output effects of normal working conditions and severe faults are the best. When the proportion of generated data is 75%, the result is the same as that of other previous indicators. Only the output effect of the model in medium faults is relatively low, but it is also close to 0.9; when the proportion of generated data is 100%, the average F1 score of the model drops to 0.89, but the output effect of the model for normal working conditions still remains the best.

[0094] Considering all aspects of the model's indicators, for the problems of class imbalance and small sample data sets in fault diagnosis, the small sample data set is expanded by improving GAN. When the proportion of the expanded generated data in the total data does not exceed 75%, the recognition effect of training 1DCNN is not much different from that of training with all original data. Once the proportion of generated data exceeds 75% or more, the output effect of the model will decrease. Therefore, the method of the present invention can effectively solve the problems of class imbalance and small sample data set for valve internal leakage fault diagnosis on the premise that the proportion of the generated data set does not exceed 75%. Finally, a DCGAN-DCNN model with an accuracy higher than 90% can be obtained, which has a higher precision rate compared with the diagnosis of a single DCNN model.

[0095] Inspired by the ideal embodiments of the present invention described above, through the above description, relevant staff can make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for diagnosing internal leakage faults of pipeline valves based on DCGAN-DCNN, characterized in that It includes the following steps: Step 1: Collect the acoustic emission data of pipeline valves under normal conditions and different leakage conditions. Step 2: Export the acoustic emission data as voltage data and perform preprocessing. Step 3: Construct a DCGAN network model, optimize the objective function, train the model using batch normalization, update the parameters of the DCGAN network using the stochastic gradient descent method, and augment the fault signal data. The generator of the DCGAN network model consists of six transposed convolution layers and a fully connected layer. The discriminator also consists of six transposed convolution layers and a fully connected layer. First, Gaussian noise is input into the fully connected layer, and the input noise is mapped to the dimension for transposed convolution in the DCGAN network. The first layer maps the large-dimensional input to a small-dimensional output, and each subsequent layer maps the input data dimension to twice the previous one as the output, with the data dimension increasing by a factor of two layer by layer until the output dimension is 512. The second to sixth layers all perform upsampling of the acoustic emission signal through transposed convolution operations, gradually expanding the dimension of the output data to generate the acoustic emission signal. The discriminator consists of six convolution layers and a fully connected layer. First, the preprocessed acoustic emission signal and the signal generated by the generator are used as the input of the model. The convolution layer is used to downsample the signal. The dimensions of the first six layers are halved layer by layer. The last layer maps the input of the previous layer to a value with a dimension of 1*1 through the fully connected layer as the output of the discriminator network. Finally, through the fully connected layer, the features extracted by the convolution kernel are mapped to a value of 1*1 as the output of the discriminator. The formula for optimizing the objective function is: ； Wherein, is the real data X output on the discriminator ; is the random vector in the random noise output in the generator ; is the probability that the discriminator determines whether the generated signal is real or not, is the randomly sampled real data, is the vector prior distribution; Step 4: Evaluate the accuracy of the DCGAN network model through a similarity comparison algorithm. Step 5: Train the DCNN model using the augmented data. Step 5 specifically includes: The DCNN model consists of four convolutions including an input layer, convolution layers, pooling layers, and a fully connected layer. Each convolution includes 2 convolution layers and a max pooling layer. The moving step of the convolution kernels of the first 3 convolutions is set to 2, and the convolution kernel step of the 4th unit is set to 1. The Padding method is set to Same. The activation function is set to ReLU. An average pooling layer is added at the end of the 4th unit to further compress the features and Dropout is set. Finally, the probability distribution of the categories is output through the fully connected layer. Step 6: Select the proportion of generated data based on the model recall rate, precision rate, and F1 score.

2. The method for diagnosing internal leakage faults of pipeline valves based on DCGAN-DCNN according to claim 1, wherein The pipeline valves under different leakage conditions are set by adjusting the opening of the regulating gate valve.

3. The method for diagnosing internal leakage faults of pipeline valves based on DCGAN-DCNN according to claim 1, characterized in that The preprocessing includes: removing the singular values of the data, adding category labels to the data, and normalizing the data.

4. The pipeline valve internal leakage fault diagnosis method based on DCGAN-DCNN according to claim 1, wherein, Step 4 specifically includes: comparing the similarity between the image data after training the DCGAN network model and the original images, and calculating the correlation distance between the images through the Pearson correlation coefficient algorithm.

Images