Seawater circulating pump impeller fault diagnosis method based on multivariate fusion network

Abstract

The invention discloses a seawater circulating pump impeller fault diagnosis method based on a multivariate fusion network, and belongs to the technical field of pump equipment fault diagnosis. The seawater circulating pump impeller fault diagnosis method comprises the steps that impeller vibration signals of a water pump are collected through a vibration acceleration sensor; performing data set division on the collected impeller vibration signals to obtain a training set, a verification set and a test set; constructing a fault diagnosis model comprising a feature extraction module, a fault classification module and an output module; training and verifying the model to obtain a trained fault diagnosis model; the model can accurately identify four states of normal, unbalanced, inlet blade breakage and outlet blade breakage of the impeller. According to the seawater circulating pump impeller fault diagnosis method based on the multivariate fusion network, the accuracy is still very high under strong noise interference, the problems that the seawater circulating pump impeller fault recognition precision is low and the anti-interference capability is weak are solved, and reliable technical support is provided for equipment operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of pump equipment fault diagnosis technology, and in particular to a method for diagnosing seawater circulating pump impeller faults based on a multi-element fusion network. Background Technology

[0002] Seawater circulating pumps are core power equipment in marine engineering and ship systems, undertaking critical tasks such as cooling and transportation. Their operating status directly affects the reliability and safety of the entire system. As the core working component of the seawater circulating pump, the impeller is constantly subjected to complex conditions such as seawater corrosion, media erosion, and load fluctuations, making it prone to malfunctions such as imbalance and blade breakage. If these malfunctions are not identified in time, they may lead to increased pump vibration, increased energy consumption, and even serious accidents such as equipment shutdown and seawater leakage, resulting in huge economic losses and safety risks.

[0003] Existing seawater circulating pump impeller fault diagnosis technologies are mainly divided into traditional signal processing methods and methods based on simple machine learning. Traditional methods, such as spectrum analysis and wavelet transform, rely on manual feature extraction and have poor adaptability to multi-source interference signals under complex operating conditions, making it difficult to accurately capture fault features and time-series correlation information. Simple machine learning methods, such as support vector machines and backpropagation neural networks, have limited ability to learn features from high-dimensional time-series vibration signals, and their diagnostic accuracy drops significantly under strong noise interference, failing to meet the high-precision requirements of actual operation and maintenance.

[0004] Therefore, in view of the problems of weak anti-interference ability, low diagnostic accuracy and insufficient feature learning in the existing technology, there is an urgent need for a seawater circulating pump impeller fault diagnosis method based on multi-dimensional fusion network that can adaptively extract multi-scale fault features and strengthen the weight of key information to improve the reliability and robustness of fault identification. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method for seawater circulating pump impellers based on multi-scale fusion networks. This method overcomes the shortcomings of existing seawater circulating pump impeller fault diagnosis methods, such as poor adaptability to complex working conditions, insufficient fault feature extraction, and low diagnostic accuracy under strong noise. The invention provides a fault diagnosis method and system that integrates multi-scale feature extraction and time-series modeling, enabling accurate and efficient identification of various types of impeller faults.

[0006] To achieve the above objectives, this invention provides a method for diagnosing impeller faults in a seawater circulating pump based on a multi-element fusion network, comprising the following steps: S1. The impeller vibration signal of the water pump is collected by a vibration acceleration sensor. The impeller vibration signal includes: normal signal, unbalance signal, inlet blade breakage signal, and outlet blade breakage signal. S2. Divide the collected impeller vibration signals into a dataset to obtain a training set, a validation set, and a test set; S3. Construct an MCNN-BiGRU-Attention fault diagnosis model, including a feature extraction module, a fault classification module, and an output module; S4. Based on the partitioned dataset, the MCNN-BiGRU-Attention fault diagnosis model is trained and validated to obtain the trained fault diagnosis model. S5. Input the impeller vibration signals to be diagnosed in the test set into the trained fault diagnosis model, output the probability distribution of impeller fault types, and realize the accurate identification of impeller faults in seawater circulating pumps. In S3, the feature extraction module includes a multi-scale convolutional neural network (MCNN). The fault classification module includes a bidirectional gated recurrent unit (BiGRU), an attention mechanism, and a fully connected layer.

[0007] Preferably, in S2, the training set is divided using a sliding window method, with a sample size of 1024 and a step size of 200. The validation set and the test set are partitioned using a non-overlapping sequential partitioning method, with a sample size of 1024. The ratio of the training set, validation set, and test set is 6:2:2.

[0008] Preferably, the specific working content of the feature extraction module includes: S311. Receive the one-dimensional raw time-series signal of impeller vibration, and extract features through three parallel convolution branches to obtain output features; S312. Perform feature fusion on the output features to obtain a multi-scale fused feature matrix; The convolutional branch includes a convolutional layer and a pooling layer.

[0009] Preferably, the operational expression of the convolutional layer is: ; in, Indicates the first Convolutional layers at position The output feature value; Represents all elements of the convolution kernel The summation operation is performed within the size range of the convolution kernel; It is the first Convolutional layers at position The values ​​of the input features; It is the first The convolution kernel of the convolutional layer is located at... The weight matrix; It is the first l Bias terms of convolutional layers; The nonlinear activation function of the convolutional layer; The mathematical expression for the max pooling of the pooling layer is: ; in, For the position after pooling The output feature value; For pooled window regions; This pools all elements within the window. Before pooling The value of the input feature at location; The operation represents the maximum value of all elements within the window.

[0010] Preferably, the specific working content of the fault classification module includes: S321. The multi-scale fusion feature matrix is ​​converted into a temporal feature sequence, and the temporal feature sequence is input into a bidirectional gated recurrent unit (BiGRU). S322. Extract historical temporal dependency features through the forward hidden layer GRU and extract future temporal association features through the backward hidden layer GRU, and integrate them to obtain a temporal feature vector containing the complete hidden state. S323. The attention mechanism performs a linear transformation on the temporal feature vector output by the bidirectional gated recurrent unit to obtain the query matrix. Q Key matrix K Sum matrix V ; S324, Calculation Q and K The correlation is calculated using the attention softmax function to obtain the preliminary attention weight vector of the temporal feature vector; S325. The initial attention weight vector is input into the fully connected layer for integration, and the output of the fully connected layer is normalized using the probability distribution Softmax function vector formula to obtain the standardized attention weights, and the probability distribution of the impeller state is output.

[0011] Preferably, the expression for the attention softmax function is: ; in, T It is the transpose symbol.

[0012] The formula for the probability distribution Softmax function vector is: ; in, For the first j The probability distribution vectors of each category, This is the original score vector output by the fully connected layer. e It is a natural constant, and the number of original fractional vectors is equal to the number of impeller state categories. C , To C The index results for each category are summed.

[0013] Preferably, the specific content of S4 includes: S401. Conduct model training based on the training set to generate an initial version of the fault diagnosis model; S402. Validate the initial version of the fault diagnosis model based on the validation set, under a preset number of iterations: If the verification results meet the verification requirements, the training is complete, the model parameters are saved, and the trained fault diagnosis model is obtained. If the verification result does not meet the verification requirements, the model parameters of the initial version of the fault diagnosis model are adjusted, and the verification is repeated for the preset number of iterations until the verification result meets the verification requirements, and the trained fault diagnosis model is obtained. The model parameters include: batch size, learning rate, and input signal size; The verification requirements include: accuracy. No less than 95% accuracy No less than 90%, recall rate No less than 95%, and an F1 score of no less than 0.92.

[0014] Preferably, the formula for calculating the accuracy is: ; The formula for calculating the accuracy is: ; The formula for calculating the recall rate is: ; The formula for calculating the F1 score is: ; in, TP , TN , FP and FN These represent the number of true positives, true negatives, false positives, and false negatives, respectively; a true positive indicates that the impeller is actually faulty and the model also predicts it to be faulty; a true negative indicates that the impeller is actually normal and the model also predicts it to be normal. A false positive indicates that the impeller is actually functioning normally, but the model predicts it to be faulty; a false negative indicates that the impeller is actually faulty, but the model predicts it to be functioning normally.

[0015] Therefore, the above-mentioned seawater circulating pump impeller fault diagnosis method based on multi-element fusion network, adopted in this invention, has the following advantages compared with the prior art: (1) The model can automatically extract features from the original signal and no longer rely on expert experience; multi-channel data fusion can effectively improve the accuracy of the model. In normal environment, using three-channel fused vibration data, the model achieves 100% accuracy in identifying impeller faults.

[0016] (2) The bidirectional structure of BiGRU can effectively extract the historical and future temporal dependencies of vibration signals, overcome the shortcomings of traditional convolutional networks in making insufficient use of temporal information, and improve the correlation between fault characteristics and working conditions. (3) The attention mechanism can adaptively strengthen the weight of key fault features and suppress noise interference in the seawater environment. Under single-channel low noise conditions, the diagnostic accuracy reaches 98.92%, and the robustness is better. It meets the actual operation and maintenance needs for high-precision diagnosis, and can effectively reduce false alarms and missed alarms, and reduce equipment operation and maintenance costs.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a structural diagram of the MCNN-BiGRU-Attention model for a seawater circulating pump impeller fault diagnosis method based on a multi-element fusion network according to the present invention. Figure 2 This is a fault diagnosis flowchart of a seawater circulating pump impeller fault diagnosis method based on a multi-element fusion network according to the present invention. Figure 3 This is a comparison chart of the accuracy of an embodiment of the seawater circulating pump impeller fault diagnosis method based on a multi-element fusion network according to the present invention; Figure 4 This is a comparison chart of loss rates of an embodiment of the seawater circulating pump impeller fault diagnosis method based on a multi-element fusion network according to the present invention; Figure 5 This is a performance comparison chart of four models of an embodiment of the seawater circulating pump impeller fault diagnosis method based on multi-element fusion network of the present invention; Figure 6 This is a graph showing the test results of noise impact on four models in an embodiment of a seawater circulating pump impeller fault diagnosis method based on a multi-element fusion network according to the present invention. Figure 7This is a visualization result of the original data obtained by using the t-SNE dimensionality reduction method in an embodiment of the seawater circulating pump impeller fault diagnosis method based on multi-element fusion network of the present invention. Figure 8 This is a visualization of the output results of each layer of a seawater circulating pump impeller fault diagnosis method based on a multi-element fusion network according to an embodiment of the present invention, using the t-SNE dimensionality reduction method. Figure 8 (a) is a visualization of the output of the MCNN layer. Figure 8 (b) is a visualization of the output of the BiGRU layer. Figure 8 (c) is a visualization of the output of the Attention layer. Figure 8 (d) is a visualization of the output of the fully connected layer. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] Example 1 This embodiment establishes a seawater circulating pump fault simulation test bench. For four typical impeller states—normal impeller operation, imbalance, inlet blade breakage, and outlet blade breakage—the MCNN-BiGRU-Attention model is used for fault diagnosis. By collecting vibration signals in three directions, constructing a standardized dataset, and optimizing model training parameters, the diagnostic accuracy and noise resistance of the method are verified, providing a reproducible implementation plan for practical engineering applications.

[0022] S1. The impeller vibration signal of the water pump is collected by a vibration acceleration sensor. The impeller vibration signal includes: normal signal, unbalanced signal, inlet blade breakage signal, and outlet blade breakage signal. The sampling rate is set to 10kHz and the sampling time is set to 20s. The sensor is located in the X, Y, and Z directions near the impeller of the pump body. XY measures the radial signal near the impeller, and the angle between the two measurement points is 90°. Z measures the axial signal near the impeller.

[0023] S2. Divide the collected impeller vibration signals into a dataset, obtaining a training set, a validation set, and a test set: For the model training to be complete, the sample data must contain data from at least one complete pump cycle. Therefore, when determining the small sample size, it can be calculated using the following formula: ; in NRepresents small sample size. i The number of pump revolutions. n Pump speed (rpm) f The sampling frequency is (Hz).

[0024] In planning small sample datasets, two methods were compared: a sliding window method and a sequential partitioning method. Sliding window method, with a fixed sample size and sliding step, expands the dataset as much as possible within a limited total data set by utilizing overlapping data from adjacent small samples, making it suitable for training set partitioning. Sequential partitioning, where there is no overlap between small sample data, results in a smaller number of small samples. This method ensures sample independence, reduces the risk of data leakage, and improves computational and evaluation efficiency, making it suitable for validation and test sets. The overall dataset partition uses a 6:2:2 ratio for training, validation, and test sets, calculated according to the small sample size planning formula. For ease of binary calculation, the training set is partitioned using a sliding window with a small sample size of 1024 and a step size of 200. The validation and test sets also have a sample size of 1024 and are partitioned without overlap to prevent data leakage that could reduce model generalization. The number of small samples in each dataset is shown in Table 1.

[0025] Table 1. Number of small samples in the dataset ;

[0026] As can be seen from the table, the number of small samples obtained by using different partitioning methods for the training set, validation set, and test set differs significantly. There are 595 small samples in the training set for each impeller state, while there are 39 small samples in each of the validation and test sets. The former meets the requirement of a large number of samples for model training, while the latter ensures data independence and evaluation efficiency during validation and testing.

[0027] S3. Construct an MCNN-BiGRU-Attention fault diagnosis model, including a feature extraction module, a fault classification module, and an output module: The feature extraction module uses a multi-scale convolutional neural network (MCNN). Unlike traditional CNNs, this module uses convolutional kernels of different scales to obtain receptive fields of different sizes, thereby focusing on different local feature structures in the input signal and extracting richer multi-level features. MCNNs can usually achieve higher accuracy and stronger robustness when dealing with complex scenes.

[0028] Considering that vibration signals are essentially time-series data, while the features extracted by convolutional neural networks are mainly concentrated at the spatial level, in order to fully explore the temporal dependencies in the signal, such as... Figure 2 As shown, a fault classification module consisting of a bidirectional gated recurrent unit (BiGRU) and an attention mechanism is connected after the feature extraction module.

[0029] The specific tasks of the feature extraction module include: S311. Receive the one-dimensional raw time-series signal of impeller vibration, and extract features through three parallel convolution branches to obtain output features; S312. Perform feature fusion on the output features to obtain a multi-scale fused feature matrix; The convolutional branch includes convolutional layers and pooling layers: The first convolution branch performs a 7×7 convolution, two 2×2 pooling operations followed by a 3×3 convolution, and a 3×3 convolution, with the number of output channels being 16, 16, 16, and 8 respectively. The second convolution branch performs one 5×5 convolution, two 2×2 pooling operations followed by a 5×5 convolution, and one 5×5 convolution in sequence, with the number of output channels being 20, 40, 20, and 10 respectively. The third convolution branch performs one 3×3 convolution, two 2×2 pooling operations followed by a 3×3 convolution, and one 3×3 convolution, with the number of output channels being 24, 48, 24, and 12 respectively.

[0030] This method efficiently and automatically learns and extracts highly discriminative fault features from raw input signals without relying on complex manual feature engineering. Convolutional and pooling layers are stacked alternately as core components, forming the main body of feature extraction. In the convolutional layers, features are extracted by performing convolution operations on the input signal using filters with equal weights. The mathematical expression for the convolution operation is: ; in, Indicates the first Convolutional layers at position The output feature value; Represents all elements of the convolution kernel The summation operation is performed within the size range of the convolution kernel; It is the first Convolutional layers at position The values ​​of the input features; It is the first The convolution kernel of the convolutional layer is located at... The weight matrix; It is the first l Bias terms of convolutional layers; The nonlinear activation function of the convolutional layer; To reduce data dimensionality, simplify network computational complexity, and enhance the robustness of learned features to minor deformations, pooling is added after convolution. Max pooling is the most commonly used method, highlighting the most salient features in each local region. Its mathematical expression is: ; in, For the position after pooling The output feature value; For pooled window regions; This pools all elements within the window. Before pooling The value of the input feature at location; The operation represents the maximum value of all elements within the window.

[0031] The specific tasks of the fault classification module include: S321. The multi-scale fusion feature matrix is ​​converted into a temporal feature sequence, and the temporal feature sequence is input into a bidirectional gated recurrent unit (BiGRU). S322. Extract historical temporal dependency features through the forward hidden layer GRU and extract future temporal association features through the backward hidden layer GRU, and integrate them to obtain a temporal feature vector containing the complete hidden state. S323. The attention mechanism performs a linear transformation on the temporal feature vector output by the bidirectional gated recurrent unit to obtain the query matrix. Q Key matrix K Sum matrix V ; S324, Calculation Q and K The correlation is calculated using the attention softmax function to obtain the preliminary attention weight vector of the temporal feature vector; S325. The initial attention weight vector is input into the fully connected layer for integration, and the output of the fully connected layer is normalized using the probability distribution Softmax function vector formula to obtain the standardized attention weights, and the probability distribution of the impeller state is output.

[0032] GRU is an important variant of recurrent neural networks (RNNs). By introducing a "gating" mechanism, it effectively controls the flow of information, thereby alleviating the gradient vanishing or exploding problems that occur in traditional RNNs during training, enabling it to better learn and capture long-term dependencies in time-series data.

[0033] The attention mechanism can automatically learn and assign different weights to each time step, ignoring redundant or noisy information irrelevant to fault diagnosis, highlighting the key time step features for classification decisions, and improving the model's diagnostic accuracy and robustness. The expression for the attention Softmax function is: ; in, T It is the transpose symbol.

[0034] The formula for the probability distribution Softmax function vector is: ; in, For the first j The probability distribution vectors of each category, This is the original score vector output by the fully connected layer. e It is a natural constant, and the number of original fractional vectors is equal to the number of impeller state categories. C , To C The index results for each category are summed.

[0035] Regarding S4, based on the partitioned dataset, the MCNN-BiGRU-Attention fault diagnosis model is trained and validated to obtain a trained fault diagnosis model. Details are attached. Figure 1 As shown: S401. Conduct model training based on the training set to generate an initial version of the fault diagnosis model; S402. Validate the initial version of the fault diagnosis model based on the validation set, under a preset number of iterations: If the verification results meet the verification requirements, the training is complete, the model parameters are saved, and the trained fault diagnosis model is obtained. If the verification result does not meet the verification requirements, the model parameters of the initial version of the fault diagnosis model are adjusted, and the verification is repeated for the preset number of iterations until the verification result meets the verification requirements, and the trained fault diagnosis model is obtained. The model parameters include: batch size, learning rate, and input signal size; The verification requirements include: accuracy. No less than 95% accuracy No less than 90%, recall rate No less than 95%, and an F1 score of no less than 0.92.

[0036] Preferably, the formula for calculating the accuracy is: ; The formula for calculating the accuracy is: ; The formula for calculating the recall rate is: ; The formula for calculating the F1 score is: ; in, TP , TN , FP and FN These represent the number of true positives, true negatives, false positives, and false negatives, respectively. A true positive indicates that the impeller is actually faulty, and the model also predicts it to be faulty. A true negative indicates that the impeller is actually normal, and the model also predicts it to be normal. A false positive indicates that the impeller is actually normal, but the model predicts it to be faulty. A false negative indicates that the impeller is actually faulty, but the model predicts it to be normal.

[0037] The MCNN-BiGRU-Attention model architecture was implemented based on the PyTorch framework. Experiments were conducted on a computer configured with a 2.5GHz i5-13490f processor, 32GB of RAM, an NVIDIA GeForce RTX 3050 graphics processor, and Windows 11 operating system. Preprocessed data was trained 10 times on the MCNN-BiGRU-Attention model, and the average loss and accuracy of the training and test sets were taken for each trial. For large and complex equipment like seawater circulation pumps, the vibration transmission paths are diverse, and data acquisition using only a single sensor is insufficient to accurately reflect the equipment's health status. To fully utilize the collected data, single-channel and multi-channel fused data were used for comparative training and testing within the model. The accuracy comparison results of the model on the test set under different channels are shown below. Figure 3 As shown, the comparison results of the loss values ​​are as follows: Figure 4 As shown.

[0038] In the figure, vib2, vib3, and vib4 correspond to the vibration signals in the X, Y, and Z directions, respectively. As shown in the figure, the multi-channel fused dataset has higher accuracy and faster convergence speed in model testing compared to the single-channel dataset.

[0039] Table 2. Fault Diagnosis Results of Seawater Circulation Pump Impeller ;

[0040] To more closely compare the differences in various indicators of the model of this invention after training with single-channel and multi-channel data, the fault diagnosis results of the seawater circulating pump impeller are listed in the table above, as shown in Table 2. The single-channel data collected in different directions of the pump body have varying abilities to identify different impeller faults and cannot comprehensively reflect the actual state of the impeller. However, the fault identification accuracy of single-channel data for inlet and outlet blade breakage and imbalance is lower than that of three-channel fused data, indicating that the multi-channel fused data can provide more comprehensive impeller state parameters for model training.

[0041] Example 2 To verify the effectiveness of the proposed MCNN-BiGRU-Attention model in seawater circulating pump impeller fault identification, this embodiment compares the identification accuracy of each type of impeller fault by establishing a Support Vector Machine (SVM) model and commonly used fault diagnosis models CNN and CNN-BiGRU. Single-channel and multi-channel training results show that multi-channel data provides more comprehensive information. Therefore, all four models use the same multi-channel fused dataset, extracting 10 time-domain features and 8 frequency-domain features as input to the SVM. The SVM kernel function is the RBF function, the penalty factor is 1, and the kernel radius is the default value of scikit-learn. The average accuracy of the model after 10 training iterations is taken, and the fault identification results are compared. Figure 5 As shown.

[0042] Depend on Figure 5 It can be seen that the traditional machine learning method SVM has a higher recognition accuracy than the CNN model with shallower layers, but it is much worse than the CNN-BiGRU model. The MCNN-BiGRU-Attention model proposed in this invention combines the advantages of the three models and has better feature extraction and classification capabilities. Compared with other models, the average recognition accuracy can reach up to 99.5%, which is 3.9% higher than CNN-BiGRU. Therefore, the model in this embodiment can more effectively utilize the vibration information of the seawater circulation pump and has better fault identification performance.

[0043] Considering the impact of noise, random noise was added to the original signal. To better simulate the real environment, Gaussian white noise with signal-to-noise ratios of 1, 3, 5, 7, and 9 was added to the original data. The test results of the noise impact on the four models are as follows. Figure 6 As shown.

[0044] The MCNN-BiGRU-Attention-based diagnostic model for seawater circulating pump impellers, as presented in this invention, exhibits significantly higher fault diagnosis accuracy than the other three models under ambient noise conditions. Although the accuracy of all four models decreases with decreasing signal-to-noise ratio (SNR), the model proposed in this embodiment still achieves an accuracy of over 90% even in an environment with an SNR of 1. This demonstrates that the MCNN-BiGRU-Attention-based model outperforms expert knowledge-based methods and other deep learning models in noisy environments, and also indicates its superior robustness.

[0045] To more intuitively describe the model learning process, the t-SNE dimensionality reduction method is used to visualize the output results of each layer in a two-dimensional space. In the figure, labels 0, 1, 2, and 3 correspond to the normal impeller, the unbalanced fault impeller, the inlet damage fault impeller, and the outlet damage fault impeller, respectively.

[0046] like Figure 7 As shown, the fault features in the original signal overlap and cannot be directly distinguished by these features. Figure 8 As shown, after training with the MCNN layer model, some features are clearly classified, but some fault features are still difficult to distinguish. After passing through the BiGRU layer and the Attention layer, various features are clearly distinguished, but the states are relatively scattered. After passing through the classification layer, the fault features corresponding to various labels are well classified and clustered together. This further proves that it can more effectively learn the fault features of seawater circulating pump impellers and achieve fault classification.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing a failure of an impeller of a seawater circulating pump based on a multi-element fusion network, characterized in that, The method comprises the following steps: S1, collecting impeller vibration signals of a water pump by a vibration acceleration sensor, wherein the impeller vibration signals comprise normal signals, unbalance signals, inlet blade breakage signals and outlet blade breakage signals; S2, dividing the collected impeller vibration signals into data sets to obtain a training set, a verification set and a test set; S3, constructing an MCNN-BiGRU-Attention fault diagnosis model, including a feature extraction module, a fault classification module and an output module; S4, training and verifying the MCNN-BiGRU-Attention fault diagnosis model based on the divided data sets to obtain a trained fault diagnosis model; S5, inputting impeller vibration signals to be diagnosed in the test set into the trained fault diagnosis model to output an impeller fault type probability distribution, thereby realizing accurate identification of the impeller fault of the seawater circulating pump. In S3, the feature extraction module comprises a multi-scale convolutional neural network (MCNN). The fault classification module comprises a bidirectional gated recurrent unit (BiGRU), an attention mechanism (Attention) and a fully connected layer. 2.The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 1, characterized in that, In S2, the training set is divided by a sliding window method, the sample capacity is 1024, and the step length is 200. The verification set and the test set are divided by a non-overlapping sequential division method, and the sample capacity is also 1024. The division ratio of the training set, the verification set and the test set is 6:2:

2. 3.The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 1, characterized in that, The specific working content of the feature extraction module comprises: S311, receiving a one-dimensional original time sequence signal of impeller vibration, and extracting features through three parallel convolution branches respectively to obtain output features; S312, fusing the output features to obtain a multi-scale fusion feature matrix; The convolution branch comprises a convolution layer and a pooling layer.

4. The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 3, characterized in that, The operation expression of the convolution layer is: ； wherein, represents the output feature value of the layer convolutional layer at position ; represents the sum operation on all elements of the convolution kernel, which is performed within the size range of the convolution kernel; is the input feature value of the layer convolutional layer at position ; is the weight matrix of the convolution kernel of the layer convolutional layer at position ; is the bias term of the l layer convolutional layer; is the nonlinear activation function of the convolutional layer; The maximum pooling mathematical expression of the pooling layer is: ； wherein, is the value of the output feature at position after pooling; is the region of the pooling window; is all elements within the pooling window; is the value of the input feature at position before pooling; the operation represents taking the maximum value of all elements within the window.

5. The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 3, characterized in that, The specific working content of the fault classification module comprises: S321, converting the multi-scale fusion feature matrix into a time sequence feature sequence, and inputting the time sequence feature sequence into a bidirectional gated recurrent unit (BiGRU); S322, extracting historical time sequence dependent features through a forward hidden layer GRU, and extracting future time sequence correlation features through a backward hidden layer GRU, and integrating to obtain a time sequence feature vector containing complete hidden states; S323、The attention mechanism Attention performs linear transformation on the time sequence feature vector output by the bidirectional gate recurrent unit to obtain a query matrix Q , a key matrix K , and a value matrix V ; S324、calculating Q and K correlation, and the preliminary attention weight vector of the time sequence feature vector is obtained by calculating through the attention Softmax function. S325, inputting the preliminary attention weight vector into a fully connected layer for integration, and using a probability distribution Softmax function vector formula to normalize the output of the fully connected layer to obtain standardized attention weights, and outputting a probability distribution of the impeller state.

6. The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 5, characterized in that, The expression of the attention Softmax function is: ； wherein T is a transposition symbol, The probability distribution Softmax function vector formula is: ; wherein, is the probability distribution vector of the j th class, is the original score vector of the fully connected layer output, e is a natural constant, and the number of the original score vector is the number of impeller state classes C , is the sum of the exponential results of the C th class.

7. The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 1, characterized in that, The specific content of S4 comprises: S401, training a model based on the training set to generate a preliminary fault diagnosis model; S402, verifying the preliminary fault diagnosis model based on the verification set, and under a preset number of iterations: If the verification result meets the verification requirement, the training is completed, the model parameters are saved, and a trained fault diagnosis model is obtained. If the verification result does not meet the verification requirement, the model parameters of the preliminary version of the fault diagnosis model are returned, and the verification is performed again for the preset number of iterations until the verification result meets the verification requirement, and a trained fault diagnosis model is obtained. The model parameters include: batch size, learning rate, and input signal size. The verification requirement includes: accuracy Not less than 95%, precision Not less than 90%, recall Not less than 95%, F1 score not less than 0.

92.

8. The sea water circulating pump impeller fault diagnosis method based on the multi-element fusion network according to claim 7, characterized in that, The calculation formula of the accuracy rate is: ； The calculation formula of the precision rate is: ； The calculation formula of the recall rate is: ； The calculation formula of the F1 score is: ； wherein, TP , TN , FP and FN represent the number of true positive, true negative, false positive and false negative results, respectively. The true positive indicates that the actual is a fault impeller, and the model prediction is also a fault impeller. The true negative indicates that the actual is a normal impeller, and the model prediction is also a normal impeller. The false positive indicates that the actual is a normal impeller, but the model prediction is a fault impeller. The false negative indicates that the actual is a fault impeller, but the model prediction is a normal impeller.

Images