Centrifugal pump rotor fault diagnosis method employing cwgan-GP and two-stream CNN models

Abstract

Disclosed is a centrifugal pump rotor fault diagnosis method employing CWGAN-GP and two-stream CNN models. The method comprises: first, collecting vibration displacement signals, carrying out reconstruction and then obtaining new vibration displacement signals, and forming a CWGAN-GP model training set; inputting the training set and label information corresponding to various faults into a CWGAN-GP model and performing training, then inputting random noise and the label information corresponding to various faults into the trained CWGAN-GP model, generating data corresponding to fault types, adding the generated data to vibration displacement signal sets for rotor imbalance faults and rotor misalignment faults, and obtaining an equalized data set; and training a two-stream CNN model by using the equalized data set, and then inputting vibration displacement signals and two-dimensional time frequency image data into the trained two-stream CNN model, so as to classify different types of faults of a centrifugal pump rotor. The present invention improves the efficiency and accuracy of fault diagnosis for a centrifugal pump rotor.

Description

A centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model Technical Field

[0001] The present invention belongs to the research field of rotating machinery fault signal analysis technology, and specifically relates to a centrifugal pump rotor fault diagnosis method based on CWGAN-GP (conditional gradient penalty Wasserstein generative adversarial network) and dual-stream CNN (convolutional neural network) models. Background Art

[0002] Imbalance and misalignment are common rotor failures in centrifugal pumps. When these occur, the safety of the entire unit is seriously threatened. Therefore, research into centrifugal pump rotor fault diagnosis technology is crucial for ensuring industrial production safety, reducing energy consumption and maintenance costs, and improving production efficiency. Currently, with the rapid development of technologies such as big data, cloud computing, and artificial intelligence (AI), centrifugal pump rotor fault diagnosis has gradually shifted from traditional reliance on expert experience to intelligent diagnosis based on deep learning.

[0003] At present, the centrifugal pump rotor fault diagnosis method based on deep learning still has the following problems to be solved: (1) Fault signal preprocessing. Since the vibration displacement signal is often accompanied by a large amount of noise components, it is difficult to directly extract obvious rotor fault features from it. Therefore, how to eliminate the noise components in the original signal and extract useful fault feature information has become one of the research difficulties in centrifugal pump rotor fault diagnosis. (2) Fault identification. Fault identification based on deep learning models is an important part of centrifugal pump rotor fault diagnosis. Deep learning models usually have high complexity and parameter quantity, and effective model optimization and parameter adjustment are required to improve the generalization ability and diagnostic accuracy of the model. (3) Fault sample acquisition is difficult. Since the centrifugal pump is in a healthy state for a long time, the collected vibration displacement data are mostly healthy state samples, and there are few fault samples, which causes the problem of imbalance in the number of samples of different categories in the data set. Therefore, how to obtain enough fault samples has become one of the bottlenecks in centrifugal pump rotor fault diagnosis.

[0004] Summary of the Invention

[0005] In view of the shortcomings in the prior art, the present invention provides a centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model, thereby improving the efficiency and accuracy of centrifugal pump rotor fault diagnosis.

[0006] The present invention achieves the above technical objectives through the following technical means.

[0007] A centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model:

[0008] The vibration displacement signals of the centrifugal pump rotor under normal state, rotor imbalance fault, and rotor misalignment fault are collected and reconstructed to obtain new vibration displacement signals, which constitute the training set of the CWGAN-GP model.

[0009] Inputting the training set and label information corresponding to each type of fault into a CWGAN-GP model for training to obtain a trained CWGAN-GP model; then inputting the label information corresponding to each type of fault and random noise into the trained CWGAN-GP model to generate data corresponding to the fault type, and adding the generated data to the vibration displacement signal set of the rotor imbalance fault and the rotor misalignment fault to obtain a balanced data set;

[0010] The dual-stream CNN model is trained using the balanced data set, and then the vibration displacement signal and the two-dimensional time-frequency image data are input into the trained dual-stream CNN model to achieve the classification of different types of centrifugal pump rotor faults.

[0011] Furthermore, the CWGAN-GP model is established based on the CGAN and WGAN-GP models: the Wasserstein distance is used to replace the cross-entropy loss function in the traditional GAN; GP is used to penalize the discriminator to force it to meet the Lipschitz continuity condition; and additional data label information is added.

[0012] Furthermore, the discriminator in the CWAGN-GP model consists of 12 convolutional layers, 2 maximum pooling layers, 6 batch normalization layers, and 1 fully connected layer. The convolution kernel size of each convolutional layer is 5×1, and the number of convolution kernels of each convolutional layer is 8, 8, 16, 16, 32, 32, 64, 64, 128, 128, 256, 256, respectively. The step size of the convolution calculation of convolution layer 1, convolution layer 3, convolution layer 5, convolution layer 7, convolution layer 9, and convolution layer 11 is 2, and the step size of convolution layer 2 and convolution layer 3 is 2. The step size of the convolution calculation of layer 4, convolution layer 6, convolution layer 8, convolution layer 10, and convolution layer 12 is 1, and all convolution layers use the LeakyReLU activation function; the maximum pooling layer is arranged after BN layer 2 and BN layer 4, and the pooling kernel size of maximum pooling layer 1 and maximum pooling layer 2 is 5×1 and the step size is 2; the fully connected layer deletes the activation function and does not add bias calculation; the BN layer is arranged after convolution layer 2, convolution layer 4, convolution layer 6, convolution layer 8, convolution layer 10, and convolution layer 12.

[0013] Furthermore, the generator in the CWAGN-GP model consists of 1 fully connected layer, 13 convolutional layers, 3 upsampling layers and 6 batch normalization layers; the fully connected layer consists of 256 neurons; convolutional layer 1 and convolutional layer 2 are composed of 256 convolution kernels of size 5×1, convolutional layer 3 and convolutional layer 4 are composed of 128 convolution kernels of size 5×1, convolutional layer 5 and convolutional layer 6 are composed of 64 convolution kernels of size 5×1, and convolutional layer 7 and convolutional layer 8 are composed of 32 convolution kernels of size 5×1. Convolutional layer 9 and convolutional layer 10 are composed of 16 convolution kernels of size 5×1, convolutional layer 11 and convolutional layer 12 are composed of 8 convolution kernels of size 5×1, and convolutional layer 13 is composed of 1 convolution kernel of size 5×1. The convolution step size of all convolutional layers is 1. The Than activation function is used after convolutional layer 13, and the ReLu activation function is used after the remaining convolutional layers. The size of the upsampling layer is set to 2. The BN layer is set after convolutional layer 2, convolutional layer 4, convolutional layer 6, convolutional layer 8, convolutional layer 10 and convolutional layer 12.

[0014] Furthermore, the dual-stream CNN model consists of an input layer, an improved 1D LeNet-5 model, an improved 2D LeNet-5 model, a connection layer, a fully connected layer, and an output layer.

[0015] Furthermore, the improved 1D LeNet-5 model includes:

[0016] The size of the input layer is changed to 1024×1;

[0017] The two-dimensional convolution kernels of 3×3 and 5×5 sizes are changed to one-dimensional convolution kernels of 3×1 and 5×1 sizes. 64 convolution kernels are used in the first convolution layer and 128 convolution kernels are used in the second convolution layer.

[0018] Furthermore, the improved 2D LeNet-5 model includes:

[0019] Select a 64×64 RGB image, use 64 convolution kernels of 3×3 size in the first convolution layer, and use 128 convolution kernels of 5×5 size in the second convolution layer;

[0020] Delete the second fully connected layer, add a Dropout operation after the fully connected layer, and add a BN operation after each activation layer;

[0021] The ReLu activation function is used instead of the Sigmoid activation function in the traditional LeNet-5.

[0022] Furthermore, the specific process of reconstructing the vibration displacement signal is as follows:

[0023] 1) The IPSO algorithm is used to optimize the parameters K and a when VMD decomposes the vibration displacement signal to obtain the optimal parameter combination;

[0024] 2) Use the optimized VMD to decompose the original vibration displacement signal and obtain K IMF components;

[0025] 3) Calculate the KLD value of each IMF component and sort the KLD values ​​in ascending order;

[0026] 4) The IMF component with the smallest KLD value is selected as the effective IMF component to reconstruct the new vibration displacement signal.

[0027] Furthermore, the IPSO algorithm is an algorithm that improves the Tent chaotic map in the PSO algorithm and adjusts the inertia weight and learning factor.

[0028] Further:

[0029] The Tent chaotic map is improved using the following iterative formula:

[0030] Among them, X i It is represented as the i-th particle, α and β are random numbers between [0, 0.1];

[0031] Adjust the inertia weight and learning factor:

[0032] Among them, w is the inertia weight, c1 is the individual learning factor, c2 is the group learning factor, δ1, δ2, δ3 are constants, ε is the specified number of iterations, k is the current number of iterations, n max is the maximum number of iterations, c 1up 、c 1low is the upper and lower bounds of the particle individual learning factor, c 2up 、c 2low are the upper and lower bounds of the particle group learning factor, and ρ is a random number between [0,1].

[0033] The beneficial effects of the present invention are:

[0034] (1) The present invention extracts centrifugal pump rotor fault features from the original vibration displacement signal based on IPSO-VMD-KLD, realizes the effective reconstruction of the centrifugal pump rotor fault signal, and improves the accuracy of rotor fault feature extraction.

[0035] (2) The present invention inputs the label information and random noise corresponding to each type of fault into the trained CWGAN-GP model to generate data corresponding to the fault type, and adds the generated data to the vibration displacement signal set of rotor imbalance fault and rotor misalignment fault to obtain a balanced data set; it solves the problem that the number of normal state samples and fault state samples in the data set is unbalanced due to the limited and difficult acquisition of centrifugal pump rotor fault data samples.

[0036] (3) The present invention uses a balanced data set to train a dual-stream CNN model, and then inputs the vibration displacement signal and two-dimensional time-frequency image data into the trained dual-stream CNN model to classify different types of centrifugal pump rotor faults. This solves the problem that traditional fault diagnosis methods are time-consuming and labor-intensive, and that existing deep learning models can only extract features from a single dimension. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] FIG1 is a schematic structural diagram of the CWAGN-GP model of the present invention;

[0038] FIG2 is a schematic diagram of the structure of the discriminator according to the present invention;

[0039] FIG3 is a schematic diagram of the structure of the generator according to the present invention;

[0040] FIG4 is a flow chart of generating centrifugal pump rotor fault data based on CWGAN-GP according to the present invention;

[0041] FIG5 is a schematic diagram of the structure of the dual-stream CNN fault diagnosis model according to the present invention;

[0042] FIG6 is a flow chart of the centrifugal pump rotor fault diagnosis and verification solution based on CWGAN-GP and dual-flow CNN according to the present invention;

[0043] FIG7 is a diagram showing the fault diagnosis results of the dual-stream CNN model after the dataset expansion according to the present invention;

[0044] FIG8 is a diagram showing the fault diagnosis results of the dual-stream CNN model of the present invention on the expanded and balanced data set;

[0045] In the figure: In the figure: 1-1D input layer, 2, 5-1D convolution layer, 3, 6-1D BN layer, 4, 7-1D maximum pooling layer, 8, 17-Flatten layer, 9, 18, 20-fully connected layer, 10-2D input layer, 11, 14-2D convolution layer, 12, 15-2D BN layer, 13, 16-2D maximum pooling layer, 19-connected layer, 21-output layer. DETAILED DESCRIPTION

[0046] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0047] The present invention provides a centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model, which specifically includes the following steps:

[0048] Step 1: Extract centrifugal pump rotor fault characteristics from the original vibration displacement signal based on IPSO-VMD-KLD

[0049] (1) In order to improve the convergence speed and convergence accuracy of the PSO algorithm, on the one hand, the Tent chaotic map in the PSO algorithm is improved, and on the other hand, the inertia weight and learning factor in the PSO algorithm are adjusted. The algorithm after the above improvements and adjustments is called the IPSO algorithm. Specifically:

[0050] The Tent chaotic map is improved using the following iterative formula:

[0051] Among them, X i It is represented as the i-th particle, α and β are random numbers between [0, 0.1]. i +1>1, the return value is equal to 1, when X i When +1<0, the return value is equal to 0. This improves the quality of the particles in the PSO algorithm population after initialization.

[0052] Adjust the inertia weight and learning factor as shown in (2)-(4):

[0053] Among them, w is the inertia weight, c1 is the individual learning factor, c2 is the group learning factor, δ1, δ2, δ3 are constants, ε is the specified number of iterations, k is the current number of iterations, n max is the maximum number of iterations, c 1up 、c 1low is the upper and lower bounds of the particle individual learning factor, c 2up 、c 2low are the upper and lower bounds of the particle group learning factor, and ρ is a random number between [0,1].

[0054] The IPSO algorithm process specifically includes:

[0055] 1) Set the size of the particle swarm, the maximum number of iterations, the specified number of iterations, and the learning factor in the IPSO algorithm, and initialize the speed and position of the particles in the particle swarm based on the improved Tent chaos map;

[0056] 2) Calculate the particle inertia weight, individual learning factor and group learning factor according to equations (2) to (4);

[0057] 3) Calculate the fitness value of each particle in the particle swarm according to the fitness function to be solved;

[0058] 4) Update the historical optimal position and group optimal value of each particle;

[0059] 5) Determine whether the IPSO algorithm has reached the termination condition. If so, stop the iterative calculation and return the optimal value. If not, return to 2) and continue iterating until the termination condition is met.

[0060] (2) The IPSO algorithm is used to automatically optimize the parameters of VMD (variational mode decomposition) to achieve adaptive decomposition of vibration displacement signals. The specific process includes:

[0061] 1) Initialize the parameter values ​​of the IPSO algorithm and the value range of VMD parameters K and a, generate the initial particle population, and use the permutation entropy as the fitness function of the IPSO algorithm;

[0062] 2) Perform VMD decomposition on the original vibration displacement signal to obtain K IMF components;

[0063] 3) Calculate the permutation entropy of K IMF components, sort the calculated permutation entropies, obtain the minimum permutation entropy of the particle population, and obtain the position of the particle with the minimum permutation entropy;

[0064] 4) Update the position and velocity of particles in the particle swarm, as well as the historical optimal position of the particles and the optimal position of the swarm;

[0065] 5) Repeat steps 2) to 4) until the iteration termination condition is reached;

[0066] 6) After the iteration is completed, the optimal particle position is output, which is the optimal parameter combination (K, a).

[0067] (3) Based on KLD (relative entropy), an effective IMF component screening criterion is established, and the rotor fault characteristics in the vibration displacement signal are extracted through IPSO-VMD-KLD.

[0068] The specific criteria for screening effective IMF components are: select the IMF component with the smallest KLD value, which is the one closest to the original vibration displacement signal and contains rich fault characteristics, as the effective IMF component for reconstructing the new vibration displacement signal. The calculation formula of KLD is shown in formula (5):

[0069] Where p(x) is the target distribution, q(x) is the approximate distribution, M is the length of the distribution, and X k is a time domain sequence, i=1, 2, 3, ..., n, and n is the sequence size.

[0070] IPSO-VMD-KLD is used to extract rotor fault features from vibration displacement signals, including:

[0071] 1) The IPSO algorithm is used to optimize the parameters K and a when VMD decomposes the vibration displacement signal to obtain the optimal parameter combination;

[0072] 2) Use the optimized VMD to decompose the original vibration displacement signal and obtain K IMF components;

[0073] 3) Calculate the KLD value of each IMF component and sort the KLD values ​​in ascending order;

[0074] 4) The IMF component with the smallest KLD value is selected as the effective IMF component to reconstruct the new vibration displacement signal, and finally the purpose of extracting the rotor fault characteristics is achieved.

[0075] Step 2: Establish the CWGAN-GP model to improve the imbalance between fault feature data and non-fault data

[0076] (1) Combining CGAN (Conditional Generative Adversarial Network) and WGAN-GP models to establish the Conditional Gradient Penalty Wasserstein Generative Adversarial Network (CWGAN-GP)

[0077] First, the Wasserstein distance is used to replace the cross entropy loss function in the traditional GAN ​​(generative adversarial network) to optimize the training of the generator and discriminator;

[0078] Then, GP is used to penalize the discriminator, forcing it to meet the Lipschitz continuity condition, so that the discriminator can better distinguish between real data (original vibration displacement signal) and generated data (data produced by the CWGAN-GP model), reducing the problem of mode collapse;

[0079] Finally, additional data label information is added to enable CWGAN-GP to better control and guide the data generated by the generator. The structural diagram of the CWAGN-GP model is shown in Figure 1, and the objective function is:

[0080] Among them, D is the discriminator, G is the generator, V(D,G) is the target function, x is the real data, is the linear interpolation of the real data x and the generated data G(z), p r is the real data distribution, p g To generate data distribution, z represents random noise, λ is the gradient penalty coefficient, y is the data label (for example, normal is 0, imbalanced data is 1), E[] represents the mathematical expectation, is the gradient operator.

[0081] The input layer of the discriminator in the CWAGN-GP model is 1024-dimensional data consisting of 1024-dimensional real data or generated data fused with the label information of each fault type. The output is the discrimination result of the real data and generated data. The structure of the discriminator is shown in Figure 2. Specifically, the discriminator consists of 12 convolutional layers, 2 maximum pooling layers, 6 batch normalization layers, and 1 fully connected layer. To extract more local information from the input data, the convolutional kernel size is 5×1 in all convolutional layers. The number of convolutional kernels in each layer is 8, 8, 16, 16, 32, 32, 64, 64, 128, 128, 256, and 256, respectively. The stride size of the convolution calculations in Convolutional Layers 1, 3, 5, 7, 9, and 11 is 2, and the stride size of the convolution calculations in Convolutional Layers 2, 4, 6, 8, 10, and 12 is 1. All convolutional layers use the LeakyReLU activation function. After BN Layers 2 and 4, a max pooling layer is used to reduce the feature dimensionality to reduce the difficulty of discriminator training. The pooling kernel size of Max Pooling Layers 1 and 2 is 5×1 and the stride size is 2. The final fully connected layer removes the activation function and does not include bias calculation. In order to speed up the training of the generator model, BN layers are added after convolutional layers 2, 4, 6, 8, 10, and 12.

[0082] The input layer of the generator is 140-dimensional data composed of 140-dimensional uniformly distributed random noise and the fusion of each fault type label. The final output is 1024-dimensional generated data. The structure of the generator is shown in Figure 3. Specifically, the fully connected layer consists of 256 neurons, which expands the random noise data to 256 dimensions. Convolution layer 1 consists of 256 convolution kernels of size 5×1, convolution layer 2 consists of 256 convolution kernels of size 5×1, convolution layer 3 consists of 128 convolution kernels of size 5×1, convolution layer 4 consists of 128 convolution kernels of size 5×1, convolution layer 5 consists of 64 convolution kernels of size 5×1, convolution layer 6 consists of 64 convolution kernels of size 5×1, and convolution layer 7 consists of 32 convolution kernels of size 5×1. Convolutional layer 8 consists of 32 5×1 convolutional kernels, convolutional layer 9 consists of 16 5×1 convolutional kernels, convolutional layer 10 consists of 16 5×1 convolutional kernels, convolutional layer 11 consists of 8 5×1 convolutional kernels, convolutional layer 12 consists of 8 5×1 convolutional kernels, and convolutional layer 13 consists of 1 5×1 convolutional kernel. The convolution stride of all convolutional layers is 1. The Than activation function is used after convolutional layer 13, and the ReLu activation function is used after all other convolutional layers. The size of upsampling layers 1, 2, and 3 is set to 2. Similar to the discriminator, batch normalization layers are added after convolutional layers 2, 4, 6, 8, 10, and 12.

[0083] (2) The training set and the label information corresponding to each type of fault are input into the CWGAN-GP model for training. The input label information and random noise are input into the generator to generate data corresponding to the fault type, thus completing the data generation task.

[0084] The original vibration displacement signals of different fault types are reconstructed to obtain new vibration displacement signals, which are then used to construct the CWGAN-GP model training set.

[0085] The training set and the labels corresponding to various fault types were input into the CWGAN-GP model for training. During training, the RMSprop optimizer was used for parameter optimization. The initial learning rates of the generator and the discriminator were set to 0.0005 and 0.0001, respectively. The coefficient λ of the model gradient penalty term was set to 10. Separate training of the generator and discriminator was adopted. First, the generator was set to an untrainable state, and the discriminator was trained to update its parameters. The discriminator was then fixed and the generator was trained. Each iteration trained the discriminator eight times and the generator twice. Through continuous iterative training, the discriminator and generator eventually reached a Nash equilibrium, resulting in a trained CWGAN-GP model.

[0086] The label information and random noise are input into the generator to generate data corresponding to the fault type, completing the data generation task. The data generation process of the centrifugal pump rotor based on the CWGAN-GP model is shown in Figure 4.

[0087] Step 3: Build a two-stream CNN model

[0088] (1) Improved 2D LeNet-5 model and 1D LeNet-5 model

[0089] 1) Improve the 2D LeNet-5 model, including:

[0090] First, change the input layer size and increase the number of convolution kernels. Specifically, choose a 64×64 RGB image. Use 64 convolution kernels of 3×3 size in the first convolution layer, and 128 convolution kernels of 5×5 size in the second convolution layer.

[0091] Then, the number of fully connected layers is reduced. Specifically, the second fully connected layer is deleted, and a Dropout operation is added after the fully connected layer, and a Batch Normalization (BN) operation is added after each activation layer.

[0092] Finally, the ReLu activation function is used to replace the Sigmoid activation function in the traditional LeNet-5.

[0093] 2) Improve the 1D LeNet-5 model, including:

[0094] First, change the size of the input layer. Specifically, change the size of the input layer to 1024×1.

[0095] Then, the size of the convolution kernel is changed. Specifically, the two-dimensional convolution kernels of 3×3 and 5×5 sizes are changed to one-dimensional convolution kernels of 3×1 and 5×1 sizes. The number of convolution kernels in each convolution layer is the same as that of the improved 2D LeNet-5 model.

[0096] (2) Using a dual-stream architecture to establish a dual-stream CNN fault diagnosis model

[0097] Based on the improved 1D LeNet-5 model and the improved 2D LeNet-5 model, a two-stream CNN fault diagnosis model was established using a two-stream architecture. The schematic diagram is shown in Figure 5, and the detailed structure is shown in Table 1. The two-stream CNN model primarily consists of an input layer, an improved 1D LeNet-5 model, an improved 2D LeNet-5 model, a connection layer, a fully connected layer, and an output layer. The input layer's data consists of one-dimensional vibration displacement data (raw vibration displacement signal) and two-dimensional time-frequency image data, respectively. The connection layer, consisting of a fully connected layer, is used to fuse the depth-invariant features extracted by the improved 1D LeNet-5 model and the improved 2D LeNet-5 model. The connection layer integrates features from different dimensions in a two-stream fusion, taking into account the different dimensions of the data, extracting more feature information from the data and improving the model's fault diagnosis accuracy. The output layer uses a Softmax classifier to classify different fault types, ultimately achieving the fault diagnosis task.

[0098] Table 1. Structural parameters of the dual-stream CNN model

[0099] (3) Set hyperparameters such as optimizer, learning rate, training batch, and Dropout value

[0100] Specifically, the optimizer is Adam, the learning rate is 0.0001, the training batch size is 8, and the Dropout value is 0.4.

[0101] Step 4: Combine the CWGAN-GP model with the dual-flow CNN model to realize the fault diagnosis of centrifugal pump rotor under data imbalance

[0102] The original unbalanced data set (original vibration displacement signal) is expanded through the trained CWGAN-GP model to achieve a balance between the quantity of various fault data and normal state data.

[0103] Different unbalanced data sets and expanded balanced data sets are used to train the dual-stream CNN model, and the trained model is used to diagnose centrifugal pump rotor faults.

[0104] An example of the verification of the centrifugal pump rotor fault diagnosis effect based on CWGAN-GP and dual-flow CNN is shown in Figure 6:

[0105] (1) The CWGAN-GP model is trained using the fault training set in the unbalanced data set, and the model reaches the Nash equilibrium state through continuous iterative training;

[0106] (2) Generate fault data of specified types through the trained CWGAN-GP model and add it to the original unbalanced dataset to expand the original unbalanced dataset so that the quantity of various types of fault data and normal state data is balanced;

[0107] (3) Build a two-stream CNN model and select the best hyperparameter combination to initialize the two-stream CNN model;

[0108] (4) Using different unbalanced datasets and expanded balanced datasets to train the two-stream CNN model;

[0109] (5) The trained two-stream CNN model is verified through the validation set (Table 3), and the fault diagnosis accuracy of different unbalanced data sets and the expanded balanced data sets are obtained, and a comparative analysis is performed.

[0110] The centrifugal pump rotor fault simulation test is as follows:

[0111] During the centrifugal pump rotor failure simulation experiment, the centrifugal pump was operated at its rated speed (2900r / min), and the operating flow of the centrifugal pump was adjusted to 7.42m 3 / h、9.01m 3 / h、10.6m 3 / h、12.72m 3 / h. At each operating flow rate, the signal acquisition system collected vibration displacement signals for normal, rotor imbalance, and rotor misalignment conditions. To comply with the Nyquist principle, which requires the sampling frequency to be greater than twice the analysis frequency, the sampling frequency in this experiment was set to 25,600 Hz and the sampling time to 1 second. The simulation protocol for rotor imbalance and rotor misalignment faults is as follows. The amount of data collected for each fault type is shown in Table 2.

[0112] (1) Rotor imbalance fault simulation scheme

[0113] The rotor imbalance failure simulation experiment simulates the rotor imbalance failure by adding gaskets at the connecting bolts of the elastic coupling. The weights of the gaskets are 2.5g, 6.5g, and 9g respectively. After adding the gaskets, the elastic coupling bolts are tightened to simulate the rotor imbalance failure generated during the operation of the centrifugal pump.

[0114] (2) Rotor misalignment fault simulation scheme

[0115] The rotor misalignment fault simulation experiment simulates the rotor parallel misalignment fault by moving the position of the motor end and using a micrometer to measure the rotor parallel misalignment amount, causing the rotor parallel misalignment fault during the operation of the centrifugal pump. The parallel misalignment amounts are 30 wires, 40 wires, and 50 wires respectively.

[0116] Table 2 Data collection quantity for centrifugal pump rotor fault simulation experiment

[0117] Construct an unbalanced dataset:

[0118] According to the unbalanced data ratios of 30:1, 12:1, 6:1, 3:1, and 2:1, a total of five data sets A, B, C, D, and E (original vibration displacement signals are collected and reconstructed) are set to simulate different degrees of data imbalance and construct data sets with different unbalanced data ratios, as shown in Table 3.

[0119] Table 3 Imbalanced dataset

[0120] Verification based on CWGAN-GP, dual-stream CNN and fault diagnosis results:

[0121] During the experiments, the two-stream CNN model was trained and validated 10 times independently on both the unbalanced dataset and the dataset augmented by the CWGAN-GP model. The results of the 10 experiments using the two-stream CNN model on the unbalanced dataset are shown in Figure 7. The results of the 10 experiments using the dataset augmented by the CWGAN-GP model, denoted as A', B', C', D', and E', are shown in Figure 8.

[0122] Figures 7 and 8 show that after augmenting the imbalanced dataset with generated data, the validation accuracy of the two-stream CNN model has significantly improved, and the fluctuation range has been significantly reduced. Compared to the imbalanced dataset, the maximum validation accuracy of the model after the dataset augmentation increased by 16%, 17.33%, 8.66%, 4.33%, and 1.33%, respectively.

[0123] The embodiments described are preferred implementations of the present invention, but the present invention is not limited to the above implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention are within the scope of protection of the present invention.

Claims

1. A centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model, characterized by: The vibration displacement signals of the centrifugal pump rotor under normal state, rotor imbalance fault and rotor misalignment fault are collected, and new vibration displacement signals are obtained after reconstruction to form the training set of the CWGAN-GP model. Inputting the training set and label information corresponding to various faults into the CWGAN-GP model for training to obtain a trained CWGAN-GP model; then inputting the label information and random noise corresponding to various faults into the trained CWGAN-GP model to generate data corresponding to the fault type, and adding the generated data to the vibration displacement signal set of the rotor unbalance fault and the rotor misalignment fault to obtain a balanced data set; The dual-stream CNN model is trained using the balanced data set, and then the vibration displacement signal and the two-dimensional time-frequency image data are input into the trained dual-stream CNN model to achieve the classification of different types of centrifugal pump rotor faults.

2. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 1 is characterized in that: The CWGAN-GP model is established based on the CGAN and WGAN-GP models: the Wasserstein distance is used to replace the cross entropy loss function in the traditional GAN; the GP is used to penalize the discriminator to force the discriminator to meet the Lipschitz continuity condition; and additional data label information is added.

3. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 2 is characterized in that: The discriminator in the CWAGN-GP model consists of 12 convolutional layers, 2 maximum pooling layers, 6 BN layers and 1 fully connected layer; the convolution kernel size of the convolutional layer is 5×1, and the number of convolution kernels of each convolutional layer is 8, 8, 16, 16, 32, 32, 64, 64, 128, 128, 256, 256 respectively; the step size of the convolution calculation of convolutional layer 1, convolutional layer 3, convolutional layer 5, convolutional layer 7, convolutional layer 9 and convolutional layer 11 is 2, and the step size of convolutional layer 2, convolutional layer 4, The step size of convolution calculation of convolution layer 6, convolution layer 8, convolution layer 10, and convolution layer 12 is 1, and all convolution layers use LeakyReLU activation function; the maximum pooling layer is arranged after BN layer 2 and BN layer 4, and the pooling kernel size of maximum pooling layer 1 and maximum pooling layer 2 is 5×1 and the step size is 2; the fully connected layer deletes the activation function and does not add bias calculation; the BN layer is arranged after convolution layer 2, convolution layer 4, convolution layer 6, convolution layer 8, convolution layer 10, and convolution layer 12.

4. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 2 is characterized in that: The generator in the CWAGN-GP model consists of 1 fully connected layer, 13 convolutional layers, 3 upsampling layers and 6 BN layers; the fully connected layer consists of 256 neurons; convolutional layer 1 and convolutional layer 2 are composed of 256 convolutional kernels of size 5×1, convolutional layer 3 and convolutional layer 4 are composed of 128 convolutional kernels of size 5×1, convolutional layer 5 and convolutional layer 6 are composed of 64 convolutional kernels of size 5×1, convolutional layer 7 and convolutional layer 8 are composed of 32 convolutional kernels of size 5×1, and convolutional layer 9 is composed of 32 convolutional kernels of size 5×1.

9. Convolutional layer 10 is composed of 16 convolutional kernels of size 5×1, convolutional layer 11 and convolutional layer 12 are composed of 8 convolutional kernels of size 5×1, and convolutional layer 13 is composed of 1 convolutional kernel of size 5×1. The convolution step size of all convolutional layers is 1. The Than activation function is used after convolutional layer 13, and the ReLu activation function is used after the other convolutional layers; the size of the upsampling layer is set to 2; the BN layer is set after convolutional layer 2, convolutional layer 4, convolutional layer 6, convolutional layer 8, convolutional layer 10 and convolutional layer 12.

5. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 1 is characterized in that: The dual-stream CNN model consists of an input layer, an improved 1D LeNet-5 model, an improved 2D LeNet-5 model, a connection layer, a fully connected layer and an output layer.

6. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 5 is characterized in that: The improved 1D LeNet-5 model includes: The size of the input layer is changed to 1024×1; The two-dimensional convolution kernels of 3×3 and 5×5 are changed to one-dimensional convolution kernels of 3×1 and 5×1. 64 convolution kernels are used in the first convolution layer and 128 convolution kernels are used in the second convolution layer.

7. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 5 is characterized in that: The improved 2D LeNet-5 model includes: Select a 64×64 RGB image, use 64 convolution kernels of 3×3 size in the first convolution layer, and use 128 convolution kernels of 5×5 size in the second convolution layer; Delete the second fully connected layer, add a Dropout operation after the fully connected layer, and add a BN operation after each activation layer; The ReLu activation function is used to replace the Sigmoid activation function in the traditional LeNet-5.

8. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 1 is characterized in that: The specific process of reconstructing the vibration displacement signal is as follows: 1) The IPSO algorithm is used to optimize the parameters K and a when VMD decomposes the vibration displacement signal to obtain the best parameter combination; 2) Use the optimized VMD to decompose the original vibration displacement signal and obtain K IMF components; 3) Calculate the KLD value of each IMF component and sort the KLD values ​​in ascending order; 4) Filter out the IMF component with the smallest KLD value as the effective IMF component for reconstructing the new vibration displacement signal.

9. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 8, characterized in that: The IPSO algorithm is an algorithm that improves the Tent chaotic map in the PSO algorithm and adjusts the inertia weight and the learning factor.

10. The centrifugal pump rotor fault diagnosis method based on CWGAN-GP and dual-flow CNN model according to claim 8, characterized in that: The Tent chaotic map is improved by the following iterative formula: Among them, X i It is represented as the i-th particle, α and β are random numbers between [0, 0.1]; Adjust the inertia weight and learning factor: Among them, w is the inertia weight, c1 is the individual learning factor, c2 is the group learning factor, δ1, δ2, δ3 are constants, ε is the specified number of iterations, k is the current number of iterations, n max is the maximum number of iterations, c 1up 、c 1low is the upper and lower bounds of the particle individual learning factor, c 2up 、c 2low are the upper and lower bounds of the particle group learning factor, and ρ is a random number between [0,1].

Images