Autonomous ship fault diagnosis method and system based on multi-source data fusion and cross-motor transfer learning, and storage medium
By using a graph Fourier convolutional network and a deep transfer Kolmogorov-Arnold network model with combined dynamic weights, the challenges of multi-source data fusion and cross-motor diagnosis in autonomous ships were solved, achieving efficient fault feature alignment and accurate fault diagnosis.
Patent Information
- Application Number
- CN202511512399.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-22
- Publication Date
- 2026-01-27
AI Technical Summary
In autonomous vessels, the heterogeneity of multi-source sensors makes it difficult to synchronize and fuse data acquisition, and existing transfer learning methods have insufficient generalization performance in cross-motor diagnosis, making it difficult to achieve reliable fault diagnosis.
A graph Fourier convolutional network is used to extract and fuse features from multi-source data. Combined with a deep transfer Kolmogorov-Arnold network model with dynamic weights, cross-motor fault feature alignment and diagnosis are achieved through source domain pre-training and target domain fine-tuning.
It improves the accuracy and robustness of cross-motor fault diagnosis, alleviates the problem of decreased diagnostic performance caused by scarce labeled data and large differences in operating conditions, and enhances the model's generalization ability in real ship operating environments.
Smart Images
Figure CN121412818A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to ship motor fault diagnosis, and in particular to an autonomous ship fault diagnosis method, system, and storage medium based on multi-source data fusion and cross-motor transfer learning. Background Technology
[0002] In the fault diagnosis system of autonomous vessels, data acquisition and intelligent diagnosis are two core components for ensuring equipment health management. Currently, in terms of data acquisition, issues remain, such as inconsistent sampling frequencies and accuracies due to the heterogeneity of multi-source sensors, and difficulties in data synchronization and fusion, increasing the complexity of system integration. Regarding fault diagnosis, although methods such as deep learning have shown potential in feature extraction and state prediction, the short actual operating history of autonomous vessels, the scarcity of fault samples, and the high cost of annotation limit the application of supervised learning methods. Existing transfer learning methods exhibit significantly reduced generalization performance when dealing with scenarios involving significant differences in operating conditions and large distribution shifts between different motors, making it difficult to achieve reliable fault diagnosis across motor equipment.
[0003] Therefore, there is an urgent need to provide an integrated data acquisition and cross-motor transfer learning fault diagnosis method for autonomous ships, which can reliably access and fuse multi-source heterogeneous data, and build a diagnostic model with strong generalization ability under the condition of scarce labeled samples, thereby providing technical support for the long-term stable operation of autonomous ships. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide an autonomous ship fault diagnosis method, system, and storage medium based on multi-source data fusion and cross-motor transfer learning, which can integrate multi-source data and improve the accuracy of cross-motor fault diagnosis when the sample size is insufficient.
[0005] Technical solution: The autonomous ship fault diagnosis method based on multi-source data fusion and cross-motor transfer learning described in this invention includes the following steps:
[0006] S1. Collect known fault operation data of different motors in the ship under different operating conditions through different channels;
[0007] S2. Input all known fault operation data of the original monitored motor and the motor to be monitored into the graph Fourier convolutional network to obtain sample data of comprehensive fault features in the source domain and the target domain respectively, and generate labels corresponding to the faults for the sample data in the source domain and the target domain respectively.
[0008] S3. Use the sample data of the source domain obtained in step S2 as the training set to train the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights. After training, retain the weights of the network model and use the sample data of the target domain to train it again to obtain the target domain network model.
[0009] S4. Input the operating data of the motor to be monitored and diagnosed into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and input the sample data into the target domain network model to obtain the fault diagnosis result of the motor.
[0010] In step S1 of this invention, multi-source data is collected through different channels, covering the operating states of different motors under different working conditions. This effectively overcomes the limitations of single sensor data and improves the comprehensiveness and representativeness of fault features. Step S2 uses a graph Fourier convolutional network to extract features from the source domain and target domain data, i.e., the data of the original monitored motor and the motor to be monitored. Multi-source heterogeneous data (such as vibration, temperature, noise, etc.) are mapped to a unified graph structure for Fourier transform and graph convolution operations, realizing feature alignment and fusion across motor data, providing high-quality and generalizable input samples for subsequent transfer learning. Step S3 uses a deep transfer Kolmogorov-Arnold network model based on combined dynamic weights. It first trains the model using source domain sample data, retains the weights after training, and then retrains the model using target domain sample data. Because the previous weights are retained, the adjustment of weights during retraining is relatively small, thus requiring less sample data. Through the three-stage training strategy of "source domain pre-training + weight reuse + target domain fine-tuning", the problem of scarce target domain labeled data is significantly alleviated. This mechanism, while preserving source domain knowledge, enables rapid model adaptation using a small number of target domain samples, improving the diagnostic accuracy and stability of the model in cross-motor and cross-operating condition scenarios. In summary, this method can achieve multi-source data fusion, and by reusing model weights, it can improve the accuracy of cross-motor diagnosis even with insufficient sample size.
[0011] Preferably, the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights includes a stacked convolutional neural network for feature extraction and a Kolmogorov-Arnold network layer for nonlinear reconstruction. The model is trained using a combined dynamic weight loss function that minimizes the difference between the source and target domains.
[0012] Stacked convolutional neural networks (CNNs) are responsible for multi-level feature extraction, enabling them to automatically learn discriminative fault features from raw signals, thus avoiding the limitations of traditional methods that rely on manual feature engineering. This application uses a Kolmogorov-Arnold network instead of the traditional Multilayer Perceptron (MLP) because the MLP uses linear stacking and activation, while the Kolmogorov-Arnold network, through nonlinear reconstruction, no longer uses fixed linear weights to combine features, allowing it to learn more complex cross-domain nonlinear relationships. Using the Kolmogorov-Arnold network improves nonlinear expressiveness and interpretability. Furthermore, since the Kolmogorov-Arnold network uses fewer parameters than the MLP with the same number of layers, it reduces the vanishing or exploding gradient problem during backpropagation. The introduction of a combined dynamic weight loss function allows the model to adaptively adjust the weights of various losses during training, balancing classification accuracy with inter-domain distribution alignment, thus improving the model's robustness in cross-domain scenarios.
[0013] The autonomous ship fault diagnosis system based on multi-source data fusion and cross-motor transfer learning described in this invention includes:
[0014] Data acquisition module: used to collect known fault operation data of different motors in the ship under different operating conditions through different channels;
[0015] Data fusion module: used to input all known fault operation data of the original monitored motor and the motor to be monitored into the graph Fourier convolutional network to obtain sample data of comprehensive fault features in the source domain and target domain respectively;
[0016] Training module: Used to train a deep transfer Kolmogorov-Arnold network model based on combined dynamic weights using the sample data of the source domain obtained by the data fusion module as the training set. After training, the weights of the network model are retained and the sample data of the target domain are used to train it again to obtain the target domain network model.
[0017] Diagnostic module: It is used to input the operating data of the motor to be monitored and diagnosed into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and input the sample data into the target domain network model to obtain the fault diagnosis result of the motor.
[0018] The computer-readable storage medium for storing one or more programs according to the present invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform any of the methods described above.
[0019] Beneficial effects: This invention achieves feature unification and fusion by fusing and reducing the dimensionality of collected multi-source heterogeneous data through a graph Fourier convolutional network. Then, through the pre-training and weighted repetition of a deep transfer Kolmogorov-Arnold network model based on combined dynamic weights, cross-motor knowledge transfer is achieved, ultimately realizing accurate fault diagnosis of the target motor. This method significantly alleviates the problem of decreased diagnostic performance caused by scarce labeled data, large differences in operating conditions, and distribution offset, and improves the robustness and generalization ability of the model in real ship operating environments. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the method of the present invention;
[0021] Figure 2 This is a diagram of the graph Fourier convolutional network structure used in this invention;
[0022] Figure 3 This is a structural diagram of the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights proposed in this invention;
[0023] Figure 4 This is the classification confusion matrix diagram of the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights in the present invention under condition 1;
[0024] Figure 5 This is the classification confusion matrix diagram of the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights in this invention under working condition 2;
[0025] Figure 6 This is the classification confusion matrix diagram of the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights in this invention under working condition 3. Detailed Implementation
[0026] As shown in the figure, the autonomous ship fault diagnosis method based on multi-source data fusion and cross-motor transfer learning described in this invention includes the following steps:
[0027] S1. Collect known fault operation data of different motors in the ship under different operating conditions through different channels.
[0028] Data can be acquired using an RS485-Modbus data acquisition device, which consists of multiple modules, including: an ESP32 control board, an acceleration sensor, a temperature measurement sensor, a noise sensor, a voltage regulator module, a TTL to 485 converter module, and a power supply.
[0029] Different motors include the original monitoring motor and the motor to be monitored, and the different motors are located in different positions on the ship; different operating conditions include one or two of the following: different wind speeds, different flow rates, different speeds, and different loads.
[0030] The known fault operation data includes triaxial acceleration, temperature and sound data. After data acquisition, the data is uploaded to the IoT platform via the MQTT protocol and stored in the local database through the Python interface.
[0031] S2. Input all known fault operation data of the original monitored motor and the motor to be monitored into the graph Fourier convolutional network to obtain sample data of comprehensive fault features in the source domain and target domain respectively.
[0032] The graph Fourier convolutional network consists of two graph convolutional layers and one ReLU activation layer, used to learn the time-frequency correlation of multi-source signals. Its data fusion specifically includes the following steps:
[0033] S21. Construct graph nodes based on the historical known fault operation data of the original monitored motor and the motor to be monitored.
[0034] S22. Perform Fourier transform on the historical known fault operation data of the original monitored motor and the motor to be monitored, and then put it into the graph convolutional network model to extract comprehensive fault features;
[0035] S23. Perform window sampling and normalization on the extracted comprehensive fault features to keep the sample length uniform.
[0036] The final sample data generated by the Graph Fourier Convolutional Network is one-dimensional data, which realizes the fusion and dimensionality reduction of multi-source data.
[0037] After extracting the sample data, a one-hot label corresponding to the fault is generated for each sample data. The generation of one-hot labels is an operation independent of the graph Fourier convolutional network and is a conventional existing technology. The specific process will not be described in detail in this paper.
[0038] S3. Use the sample data of the source domain obtained in step S2 as the training set to train the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights. After training, retain the weights of the network model and use the sample data of the target domain to train it again to obtain the target domain network model.
[0039] The deep transfer learning Kolmogorov-Arnold network model based on combined dynamic weights consists of stacked convolutional neural networks for feature extraction and Kolmogorov-Arnold network layers for nonlinear reconstruction. The model is trained using a combined dynamic weight loss function that minimizes the difference between the source and target domains.
[0040] The stacked convolutional neural network consists of 5 Cov1D blocks and 1 global average pooling layer. Each Cov1D block includes a convolutional layer, a batch normalization layer, and a max pooling layer.
[0041] The expression for the combined dynamic weight loss function is:
[0042] ,
[0043] in, To combine dynamic weight loss functions, To comprehensively classify losses, Represents a class of balance coefficients, and These are the dynamic distribution adaptive loss and the local maximum mean difference loss, respectively.
[0044] The expression is
[0045] ,
[0046] in, and These are the source domain classification loss and the target domain classification loss, respectively. It is a type II balance coefficient.
[0047] The inclusion of both source and target domain losses is because, during cross-motor diagnostics, tagged source domain data is migrated to untagged target domain data (i.e., the target domain data to be diagnosed, which does not carry the labels corresponding to the faults). This forms the basis of the entire combined dynamic weight loss function. And subsequently... and Based on this, the alignment optimization of the sample data is performed. The more similar the data distribution between the source domain samples and the target domain samples, the better the diagnostic accuracy of the entire algorithm will be. Therefore, the three loss functions are ultimately formed. .
[0048] The expression is
[0049] ,
[0050] in, Represents dynamic weighting factors. Indicators representing distribution differences and These represent sample data from the source and target domains, respectively. and Labels representing sample data from the source and target domains, respectively. For characteristic representation function, For fault category number, Total number of fault categories; Represents a probability distribution. Indicates the first Feature vectors of class-specific faults.
[0051] Traditional dynamic distribution adaptation simply uses the mean or variance, or basic inter-class divergence, neglecting the relationship between classes and features. DDM, however, combines MMD (Maximum Mean Difference) and CORAL (Correlation Alignment) to improve target domain classification accuracy, reduce class confusion, and enhance generalization and transfer stability. The first part of the formula is marginal distribution adaptation, and the second part is conditional distribution adaptation.
[0052] According to the definition of A-distance:
[0053] ,
[0054] In the formula, This represents the A-distance calculated from the edge distribution. This represents the conditional distribution A-distance corresponding to class c.
[0055] The A-distance expression is:
[0056] ,
[0057] in To distinguish between two domains using a linear classifier h and The error.
[0058] The expression is:
[0059] ,
[0060] In the formula, This is the empirical maximum distribution difference index between H distributions in the regenerated Hilbert space. A selection function for separating features associated with type c faults.
[0061] In the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights, the KAN layer (Kolmogorov-Arnold network) outputs a high-dimensional feature vector after nonlinear reconstruction. Then, DG-Softmax is used for fault classification. By providing a decision boundary, DG-Softmax can extract more discriminative features, thereby improving the accuracy of the final classification.
[0062] S4. Input the operating data of the motor to be monitored and diagnosed into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and input the sample data into the target domain network model to obtain the fault diagnosis result of the motor.
[0063] The autonomous ship fault diagnosis system based on multi-source data fusion and cross-motor transfer learning described in this invention includes:
[0064] Data acquisition module: used to collect known fault operation data of different motors in the ship under different operating conditions through different channels;
[0065] Data fusion module: used to input all known fault operation data of the original monitored motor and the motor to be monitored into the graph Fourier convolutional network to obtain sample data of comprehensive fault features in the source domain and target domain respectively;
[0066] Training module: Used to train a deep transfer Kolmogorov-Arnold network model based on combined dynamic weights using the sample data of the source domain obtained by the data fusion module as the training set. After training, the weights of the network model are retained and the sample data of the target domain are used to train it again to obtain the target domain network model.
[0067] Diagnostic module: It is used to input the operating data of the motor to be monitored and diagnosed into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and input the sample data into the target domain network model to obtain the fault diagnosis result of the motor.
[0068] The computer-readable storage medium for storing one or more programs according to the present invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform the method described above.
[0069] To better illustrate the method described in this invention, a specific example is provided below:
[0070] S1 uses an RS485-Modbus data acquisition device to collect triaxial acceleration, temperature, and sound data of the left and right motors under different operating conditions, and uploads them to the Internet of Things platform using the MQTT protocol, and stores them in the local database through the Python interface;
[0071] The fault diagnosis method of the present invention was verified by collecting data sets of motors under different operating conditions. The information of motors under different operating conditions is shown in Table 1.
[0072] Table 1 Definition of Variable Operating Conditions
[0073]
[0074] Motor health status, or comprehensive fault characteristics, includes NOR (normal), BE (bearing failure), ST (stator failure), and RO (rotor failure). Although a large amount of data can be collected and then filtered out for the above four motor health statuses from the data of each operating condition, the sample data of motor failures (motor in BE, ST, or RO condition) are relatively scarce. This would greatly increase the difficulty and cost of data collection. Therefore, it is also possible to manually simulate the motor to be in different health statuses and then collect the corresponding data.
[0075] In this embodiment, the load of the acquisition motor is set to 0HP and the speed to 750rpm.
[0076] The BE condition can be simulated by introducing an impulse pulse and high-frequency modulated noise, as expressed in the following expression:
[0077] ,
[0078] In the formula, Represents vibration signals in the time domain. Indicates the rotor rotation frequency. This represents the pulse signal simulating an impact. High-frequency damped oscillations representing the impact response of a bearing. denoted by Gaussian noise; A is the amplitude of the rotating fundamental signal, representing the intensity of the sinusoidal fundamental wave generated by the rotor rotation; k is the pulse number, representing the kth impact; T is the pulse period, usually related to the rotational speed or the periodicity of the faulty component; N is the total number of pulses; t is the time.
[0079] ST can be simulated by introducing electromagnetic imbalance and low-frequency modulation noise, as expressed by:
[0080] ,
[0081] In the formula, Indicates the power supply frequency. Indicates the modulation frequency. This indicates low-amplitude random noise; The amplitude of the fundamental signal represents the strength of the fundamental frequency component of the power supply. The amplitude of the modulated signal represents the intensity of the additional component caused by factors such as electromagnetic imbalance.
[0082] RO can be simulated by introducing simulated mechanical vibration and subharmonics, as expressed in the following expression:
[0083] ,
[0084] In the formula, B represents mechanical noise; B is the amplitude of the fundamental frequency component, representing the amplitude of the dominant frequency of mechanical vibration; C is the amplitude of the subharmonic component.
[0085] The motor dataset information collected in this embodiment is shown in Table 2:
[0086] Table 2 Dataset Information for Motors Under Changing Operating Conditions
[0087]
[0088] S2, the collected dataset is divided into source domain and target domain according to different motors, and then fed into graph Fourier convolutional networks respectively to obtain one-dimensional sample data of comprehensive fault features, and generate labels corresponding to the faults.
[0089] By training and fusing graphical Fourier convolutional networks and using the data window method and normalization method, a total of 4,000 samples were obtained for each of the four motor health conditions under each working condition, with each sample having 3,072 sampling points.
[0090] S3. Input all source domain sample data for each motor condition into DDTKN (Deep Transfer Kolmogorov-Arnold Network Model Based on Combined Dynamic Weights). The training process includes corresponding fault labels to obtain the source domain operating condition model. During the transfer training, the weights are shared to the target domain sample data for training to obtain the target domain operating condition model.
[0091] The structure of DDTKN is as follows: Figure 2 As shown, it includes stacked convolutional neural networks, combined dynamic weight loss, and Kolmogorov-Arnold layers.
[0092] The stacked convolutional neural network consists of 5 "Cov1D" blocks and 1 GAP layer. Each "Cov1D" block consists of a convolutional layer, a batch normalization (BN) layer, and a max pooling layer.
[0093] The specific training steps for DDTKN are as follows:
[0094] First, a dual-channel deep network structure is constructed, comprising a source domain feature extraction path and a target domain feature extraction path with consistent structure and shared parameters. Here, the dual-channel deep network structure refers to the use of stacked convolutional neural networks for both the source and target domains. Because it is divided into source and target domains, the overall architecture resembles a dual-channel architecture. Both paths include multiple convolutional modules and max-pooling layers, used to extract high-level feature representations from the input source and target domain samples, respectively.
[0095] Embedding a Kolmogorov-Arnold Network layer into the network enhances the model's nonlinear representation capabilities, enabling more accurate identification and representation of complex failure modes.
[0096] During network training, a combined dynamic weight loss function is used to adaptively align the marginal and conditional distributions between the source and target domains, thereby reducing the difference in feature distributions between the two domains and improving domain adaptability.
[0097] During the model optimization phase, the dynamic distribution alignment loss and classification loss are jointly optimized and combined to drive the iterative update of the entire network parameters, ultimately achieving high-precision fault diagnosis under cross-domain operating conditions.
[0098] S4. Input the diagnostic operation data of the motor to be monitored in the target domain into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and then input the sample data into DDTKN to obtain the diagnostic results.
[0099] Based on DDTKN analysis of motor fault information and diagnosis rate, after mutual transfer between the left and right motors, the diagnosis results under three operating conditions are shown in Table 3. The classification confusion matrix of motor faults under each operating condition is as follows: Figure 3 , Figure 4 , Figure 5 .
[0100] Table 3. Fault diagnosis rate of mutual migration between motors on both sides under different operating conditions
[0101]
[0102] As shown in Table 3, a high fault diagnosis rate can be achieved regardless of which side the motor is moved from to the other side, demonstrating the effectiveness of the method of the present invention.
[0103] To verify the superiority of the method of the present invention, the left and right motors were migrated to each other under working condition 1. The experimental results were compared with the current typical variant convolutional network model (VGGNet) and distributed adaptive model (CDA, IJDA, MSSA). The results are shown in Table 4.
[0104] Table 4. Fault diagnosis rates of mutual migration between motors on both sides under different methods under operating condition 1.
[0105]
[0106] As shown in Table 4, the method of the present invention (DDTKN) has the highest fault diagnosis rate compared with the existing methods, regardless of which side of the motor is migrated to the other side. This further verifies that the accuracy of the method of the present invention in cross-motor fault diagnosis is far superior to that of the existing methods.
Claims
1. An autonomous ship fault diagnosis method based on multi-source data fusion and cross-motor transfer learning, characterized in that, Includes the following steps: S1. Collect known fault operation data of different motors in the ship under different operating conditions through different channels; S2. Input all known fault operation data of the original monitored motor and the motor to be monitored into the graph Fourier convolutional network to obtain sample data of comprehensive fault features in the source domain and the target domain respectively, and generate labels corresponding to the faults for the sample data in the source domain and the target domain respectively. S3. Use the sample data of the source domain obtained in step S2 as the training set to train the deep transfer Kolmogorov-Arnold network model based on combined dynamic weights. After training, retain the weights of the network model and use the sample data of the target domain to train it again to obtain the target domain network model. S4. Input the operating data of the motor to be monitored and diagnosed into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and input the sample data into the target domain network model to obtain the fault diagnosis result of the motor.
2. The method according to claim 1, characterized in that: The deep transfer Kolmogorov-Arnold network model based on combined dynamic weights includes a stacked convolutional neural network for feature extraction and a Kolmogorov-Arnold network layer for nonlinear reconstruction. The model employs a combined dynamic weight loss function that minimizes the difference between the source and target domains.
3. The method according to claim 2, characterized in that: The expression for the combined dynamic weight loss function is as follows: ， in, To combine dynamic weight loss functions, To comprehensively classify losses, Represents a class of balance coefficients, and These are the dynamic distribution adaptive loss and the local maximum mean difference loss, respectively. The expression is ， in, and These are the source domain classification loss and the target domain classification loss, respectively. It is a type II balance coefficient.
4. The method according to claim 3, characterized in that: The The expression is ， in, Represents dynamic weighting factors. Indicators representing distribution differences and These represent sample data from the source and target domains, respectively. and Labels representing sample data from the source and target domains, respectively. For characteristic representation function, For fault category number, Total number of fault categories; Represents a probability distribution. Indicates the first Feature vectors of the fault class; The expression is: ， In the formula, This is the empirical maximum distribution difference index between H distributions in the regenerated Hilbert space. A selection function for separating features associated with type c faults.
5. The method according to claim 2, characterized in that: The stacked convolutional neural network includes 5 Cov1D blocks and 1 global average pooling layer. Each Cov1D block includes a convolutional layer, a batch normalization layer, and a max pooling layer.
6. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S21. Construct graph nodes based on the known fault operation data of the original monitored motor and the motor to be monitored. S22. Perform Fourier transform on the known fault operation data of the original monitored motor and the motor to be monitored, and then put it into the graph convolutional network model to extract comprehensive fault features. S23. Perform window sampling and normalization on the extracted comprehensive fault features to keep the sample length uniform.
7. The method according to claim 6, characterized in that: The graph Fourier convolutional network consists of two graph convolutional layers and one ReLU activation layer.
8. The method according to claim 1, characterized in that: The operating conditions include wind speed, flow rate, rotational speed, and load.
9. An autonomous ship fault diagnosis system based on multi-source data fusion and cross-motor transfer learning, characterized in that, include: Data acquisition module: used to collect known fault operation data of different motors in the ship under different operating conditions through different channels; Data fusion module: used to input all known fault operation data of the original monitored motor and the motor to be monitored into the graph Fourier convolutional network to obtain sample data of comprehensive fault features in the source domain and target domain respectively; Training module: Used to train a deep transfer Kolmogorov-Arnold network model based on combined dynamic weights using the sample data of the source domain obtained by the data fusion module as the training set. After training, the weights of the network model are retained and the sample data of the target domain are used to train it again to obtain the target domain network model. Diagnostic module: It is used to input the operating data of the motor to be monitored and diagnosed into the graph Fourier convolutional network to obtain sample data of its comprehensive fault characteristics, and input the sample data into the target domain network model to obtain the fault diagnosis result of the motor.
10. A computer-readable storage medium for storing one or more programs, characterized in that: The program includes one or more instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1 to 8.