A method and device for detecting failure of a propulsion shafting under a zero-failure sample

Abstract

The present application belongs to the technical field of mechanical health state monitoring, and discloses a propeller shaft system fault detection method and device under zero fault sample, which decomposes historical normal samples of multiple key parts of the propeller shaft system into spectrum-localized time domain components, constructs a graph structure based on the correlation between the time domain components to form graph data representing the internal dependency of the signals, and then calculates the reconstruction error by using a graph memory autoencoder, and compares the reconstruction error with a predetermined decision threshold to realize fault detection, so that the present application can realize accurate detection and positioning of the propeller shaft system fault under zero fault sample, effectively overcoming the dependence of the existing propeller shaft system fault detection method on fault data, and having significant engineering application value in ensuring the safe, reliable and stable operation of the ship propulsion system.

Description

Technical Field

[0001] This invention belongs to the technical field of mechanical health condition monitoring, and more specifically, relates to a method and equipment for detecting propulsion shaft system faults under zero-fault conditions. Background Technology

[0002] As a core component of a ship's power transmission system, the propulsion shafting directly affects the ship's sailing performance, safety, and reliability. Its main function is to efficiently transmit the power generated by the engine to the propeller, propelling the ship forward or backward. However, due to long-term exposure to the complex marine environment and continuous high-load operation, the propulsion shafting is highly susceptible to potential failures such as wear, corrosion, and fatigue cracks. Once a failure occurs, it can not only lead to the ship losing power but may also trigger serious safety accidents such as grounding, collisions, or sinking, causing enormous economic losses and environmental damage.

[0003] Traditional fault detection methods mainly rely on periodic shutdowns for maintenance and manual inspections. These experience-based or threshold-based methods have significant limitations: First, they cannot achieve continuous, real-time monitoring of equipment operating status, leading to faults often being discovered only after significant damage has occurred. Second, they lack sufficient sensitivity to early, minor or progressive faults (such as micro-cracks, slight misalignment, or initial bearing wear), making it difficult to provide timely warnings at the nascent stage of a fault. Furthermore, these methods typically require significant manpower and planned shutdowns, resulting in high maintenance costs, low detection efficiency, strong subjectivity, and susceptibility to influence from personnel experience and environmental factors, making it difficult to meet the urgent needs of modern marine propulsion systems for high reliability, high availability, and intelligent operation and maintenance.

[0004] In recent years, with the rapid development of artificial intelligence technology, data-driven methods have made significant progress in the field of fault detection. Typical methods include deep learning models such as convolutional neural networks, recurrent neural networks, and graph neural networks. For example, Guo et al. proposed a deep nonlinear ordered recurrent convolutional network for fault detection in induction motor drive systems; this method effectively suppresses the interference of speed fluctuations on feature extraction by introducing a nonlinear order cyclic spectrum analysis layer to convert the time-domain measurement signal into an angular domain representation. Yang et al. designed a feature-enhanced generative adversarial network that integrates deep embedded category information to alleviate the problem of fault detection under small sample conditions. In addition, Liu et al. combined variational mode decomposition and the Transformer architecture to construct an end-to-end mechanical fault detection network, achieving high-precision fault identification. Although the above methods show excellent performance in specific scenarios, their practical industrial applications still face two key challenges: (1) These methods typically rely on a large amount of labeled fault data to achieve superior performance. However, in real-world operating environments, especially for well-maintained propulsion shaft systems with high reliability requirements, real fault data is often extremely scarce or even completely unavailable.

[0005] (2) These methods mainly focus on fault detection of single components or local locations, lacking the ability to efficiently detect and accurately locate faults in multiple key parts, and are difficult to meet the needs of propulsion shaft systems for comprehensive state perception. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and device for detecting propulsion shaft system faults under zero-fault sample conditions, which aims to solve the problem that the fault detection of existing propulsion shaft systems is limited by fault samples.

[0007] To achieve the above objectives, according to one aspect of the present invention, a method for detecting propulsion shaft system faults under zero-fault sample conditions is provided, comprising the following steps: (1) Decompose the historical normal samples of multiple key parts of the propulsion shaft system into time-domain components with localized spectrum, and construct a graph structure based on the correlation between the time-domain components to form graph data that characterizes the intrinsic dependencies of the signal; (2) Input the graph data corresponding to each key part of the shaft system to be tested into the graph memory autoencoder. The graph memory autoencoder calculates the reconstruction error based on the graph data. If the obtained reconstruction error is greater than the decision threshold, the part corresponding to the graph data is determined to be in a fault state; otherwise, it is in a normal state.

[0008] Furthermore, the graph memory autoencoder includes a graph attention encoder, a memory enhancement module, and a graph attention decoder. The graph attention encoder is used to extract global structural features from graph data; the memory enhancement module calculates the global structural features and each memory vector in the memory matrix based on a self-attention mechanism. The similarity between the graph data is used to obtain the addressing weights. The addressing weight vector is obtained based on the addressing weight vector and the memory matrix to calculate the memory features. The graph attention decoder reconstructs the graph data based on the memory features and then calculates the reconstruction error between the reconstructed data and the graph data.

[0009] Furthermore, the total loss function of the graph memory autoencoder assigns different reconstruction losses to different positions of the propulsion shaft system; all historical normal samples are input into the graph memory autoencoder, which calculates the reconstruction error of each sample, and the maximum value among all calculated reconstruction errors is used as the decision threshold.

[0010] Furthermore, the reconstruction loss function at different locations is weighted and summed with the training weights to obtain the total loss function.

[0011] Furthermore, the reconstruction loss function at different locations is:

[0012] in, Indicates the first Reconstruction loss at each location, It is the number of fault detection locations. Indicates the first The number of training samples at each position. and They represent the first time. Input and output graph data for each location.

[0013] Furthermore, the total loss function The formula is:

[0014] in Indicates the current iteration number. Indicates the first In the nth iteration Reconstruction loss at each location, Indicates the first In the nth iteration Training speed at each position; It is the first Position in the next iteration The adaptive weights of the reconstruction loss are based on Perform dynamic optimization.

[0015] Furthermore, the training weights are calculated based on the training speed; the training speed is calculated based on the reconstruction loss at the corresponding position.

[0016] Furthermore, the graph attention encoder obtains global structural features from the input graph data through two layers of graph attention. The memory enhancement module uses addressing weight vectors. and memory matrix To obtain memory features The graph attention decoder extracts features from memory through two GTN layers. Reconstruct the input graph data:

[0017] in, This represents the reconstructed graph data; Represents the adjacency matrix. This represents the ReLU activation function.

[0018] The present invention also provides a propulsion shaft system fault detection system under zero-fault sample conditions. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the propulsion shaft system fault detection method under zero-fault sample conditions as described above.

[0019] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the propulsion shaft fault detection method under zero-fault sample conditions as described above.

[0020] In summary, compared with the prior art, the propulsion shaft system fault detection method and equipment provided by the present invention under zero-fault sample conditions have the following beneficial effects: 1. This invention decomposes historical normal samples of multiple key components of the propulsion shaft system into time-domain components with localized spectrum, and constructs a graph structure based on the correlation between the time-domain components to form graph data representing the inherent dependencies of the signal. Then, the graph memory autoencoder calculates the reconstruction error, and the defect detection is achieved by comparing the reconstruction error with a predetermined decision threshold. Thus, this invention can achieve accurate detection and location of propulsion shaft system faults even in fault-free samples, effectively overcoming the dependence of existing propulsion shaft system fault detection on fault data, and has significant engineering application value in ensuring the safe, reliable and stable operation of ship propulsion systems.

[0021] 2. This invention achieves accurate detection of propulsion shaft system faults under zero-fault sample conditions. Since the characteristics of fault samples cannot be effectively represented by normal memory, the reconstruction results will be significantly distorted, resulting in a large reconstruction error. This enables high-sensitivity detection of abnormal states.

[0022] 3. The graph memory autoencoder provided by this invention significantly enhances the ability to distinguish between normal and fault states by reconstructing the graph structure of normal states with high fidelity only, thereby achieving accurate fault detection under fault-free samples.

[0023] 4. The total loss function of the graph memory autoencoder is obtained by weighted summation of the reconstruction loss function at different locations and the training weights. By introducing weights dynamically calculated based on the training rate into the loss function, the automatic balancing of each loss term during training is achieved, effectively improving the performance of fault detection and localization.

[0024] 5. Input all historical normal samples into the graph memory autoencoder. The graph memory autoencoder calculates the reconstruction error of each sample. The maximum value among all calculated reconstruction errors is used as the decision threshold. No manual parameter tuning is required, and it effectively ensures a low false alarm rate and improves sensitivity to minor faults. Attached Figure Description

[0025] Figure 1 This is a flowchart of a propulsion shaft system fault detection method under zero-fault sample provided by an embodiment of the present invention; Figure 2 This is a diagram illustrating the graph data construction process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a graph memory autoencoder; Figure 4 This is a block diagram of a fault detection method according to a specific embodiment of the present invention; Figure 5 The failure detection rate results of the method in this embodiment of the invention are shown in the figure, where (a) corresponds to the left bearing support, (b) corresponds to the middle bearing support, and (c) corresponds to the right bearing support. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0027] This invention provides a method for detecting propulsion shaft system faults under zero-fault samples. The detection method can achieve accurate detection and location of propulsion shaft system faults under fault-free samples, effectively overcoming the dependence of existing data-driven methods on fault data, and has significant engineering application value in ensuring the safe, reliable and stable operation of ship propulsion systems.

[0028] Please see Figure 1 The fault detection method mainly includes the following steps: Step 1 involves decomposing historical normal samples from multiple key parts of the propulsion shaft system into time-domain components with localized spectra, and constructing a graph structure based on the correlation between the time-domain components to form graph data that characterizes the intrinsic dependencies of the signal.

[0029] Please see Figure 2 Specifically, the steps for constructing graph data from historical normal samples are as follows: First, the input time-domain signal is transformed using a Fast Fourier Transform (FFT). (i.e., the input historical normal samples) are transformed into the frequency domain to obtain frequency domain complex coefficients. :

[0030] Then, the frequency domain interval Divide into equal intervals Sub-band ,in , It is the sampling frequency; For frequency variables, it represents the oscillation frequency in the frequency domain; The time variable represents the moment of the time-domain signal.

[0031] Next, for each sub-band ,Will After setting zeros outside this frequency band, an inverse fast Fourier transform is performed to convert it back to the time domain, yielding the first... Fourier eigenmode functions and the original signal As an additional mode Thus obtain +1 modal signal, where:

[0032] Among them, if , ;otherwise Therefore, the Fourier decomposition method is used to decompose the input signal. Solve into a set of Fourier eigenmode functions (FIMFs):

[0033] in, Indicates the first One FIMF, Represents the residual components. It is the number of FIMFs.

[0034] After that, with each As a graph node, compute any two nodes and Cosine distance between them:

[0035] Cosine distance is used to measure the distance between nodes. and The correlation between them The first element in the cosine distance matrix lines and Elements in the column.

[0036] Next, to enhance the sparsity of the graph structure, a threshold is applied to the cosine distance matrix. The adjacency matrix is ​​then obtained. Represented as:

[0037] in, It is the th adjacency matrix row and number The column value. If the cosine distance between two nodes is less than the threshold. If they are considered to have important structural connectivity, then they are considered to have nodes. The neighbors were identified as:

[0038] Finally, adopt Calculate the graph Laplacian matrix Defined as:

[0039] in, It is the degree matrix of the graph structure, and its diagonal elements

[0040] Step 2: Input the graph data corresponding to each key part of the shaft system to be tested into the graph memory autoencoder. The graph memory autoencoder calculates the reconstruction error based on the graph data. If the obtained reconstruction error is greater than the decision threshold, the part corresponding to the graph data is determined to be in a fault state; otherwise, it is in a normal state.

[0041] Please see Figure 3 The graph memory autoencoder includes a graph attention encoder, a memory enhancement module, and a graph attention decoder. The graph attention encoder extracts global structural features from graph data; the memory enhancement module calculates the global structural features and each memory vector in the memory matrix based on a self-attention mechanism. The similarity between the graph data is used to obtain the addressing weights. Based on the addressing weights, an addressing weight vector is obtained. Then, the memory features are calculated based on the addressing weight vector and the memory matrix. The graph attention decoder reconstructs the graph data based on the memory features and then calculates the reconstruction error between the reconstructed data and the graph data.

[0042] The graph memory autoencoder assigns different reconstruction losses to different positions of the propulsion shaft. The total loss function is obtained by weighted summation of the reconstruction loss functions at different positions and the training weights. The training weights are calculated based on the training rate. The training rate is calculated based on the reconstruction loss at the corresponding position.

[0043] Specifically, the graph attention encoder obtains global structural features from the input graph data through two graph attention layers (GTN). :

[0044] in, It is the input graph data. Represents the adjacency matrix. This represents the ReLU activation function.

[0045] The memory enhancement module is used to widen the distribution gap between normal and faulty samples by using addressing weight vectors. and memory matrix To obtain memory features :

[0046] In the formula, They are non-negative vectors, and their sum is 1. yes Location The memory vector; It represents the number of entries to remember.

[0047] Among them, addressing weight It is obtained through a content-based self-attention mechanism, focusing on... Moderation and characteristics Highly similar memory vectors are computed from the graph attention module. With each memory vector The similarity between them is then normalized to obtain the addressing weight. Specifically: similarity is measured using cosine similarity.

[0048] Then, by performing a softmax operation on the similarity, the attention weights are calculated. :

[0049] Introducing hard threshold operation to enhance Sparsity in:

[0050] in, It is a sparse weight vector The One element, This indicates the shrinkage threshold.

[0051] Then through The memory features are calculated.

[0052] The graph attention decoder is designed symmetrically with respect to the graph attention module, and it extracts features from memory through two GTN layers. Reconstruct the input graph data:

[0053] in, This represents the reconstructed graph data.

[0054] The total loss function of the graph memory autoencoder is obtained as follows: The reconstruction loss function at different locations is:

[0055] in, Indicates the first Reconstruction loss at each location, It is the number of fault detection locations. Indicates the first The number of training samples at each position. and They represent the first time. Input and output graph data for each location.

[0056] Calculate the training speed at different locations:

[0057] The total loss function is obtained by weighted summation based on the loss function at each position and the training weights. The calculation formula is as follows:

[0058] in Indicates the current iteration number. Indicates the first In the nth iteration Reconstruction loss at each location, Indicates the first In the nth iteration Training speed at each position. When When =0, =1. It is the first Position in the next iteration The adaptive weights of the reconstruction loss are based on Perform dynamic optimization. express The total loss of the model in the next iteration. Smaller. This will result in smaller This can effectively slow down the learning speed at that position and guide the model to learn generalizable features.

[0059] The training weights for different positions are calculated using the following formula:

[0060] The detailed parameters of the graph memory autoencoder are shown in Table 1: Table 1. Detailed parameters of the graph memory autoencoder.

[0061] All historical normal samples are re-inputted into the trained graph memory autoencoder, the reconstruction error of each sample is calculated, and the maximum reconstruction error is used as the decision threshold for fault detection.

[0062] During online fault detection, the samples to be detected from each part are input into a trained graph memory autoencoder, and their reconstruction error is calculated. If the reconstruction error of a sample exceeds the decision threshold, the position is determined to be in a fault state; otherwise, it is in a normal state.

[0063] The present invention will be further described in detail below with reference to specific embodiments.

[0064] Please see Figure 4 and Figure 5 This invention provides a method for detecting and locating propulsion shaft system faults under zero-fault conditions, which includes the following steps: Step 1: Data Collection The experimental data used in this embodiment comes from a self-built propulsion shaft system fault simulation test bench. The test bench consists of a drive motor, three bearing housings, a coupling, and a magnetic powder brake. The drive motor is a YS7124 three-phase asynchronous motor with a rated power of 0.37 kW and an operating speed of 200 r / min. Vibration sensors are installed on all three bearing housings for condition monitoring. Misalignment faults in the bearing housings are simulated by loosening adjacent nuts, thus simulating a single fault. The vibration sensors sample at a frequency of 2000 Hz, and each experiment lasts for 2 minutes.

[0065] Table 2 details the propulsion shaft dataset. In the single-failure mode, only one of the three bearing housings is set to a fault state. In each fault simulation, 100 fault samples and 100 normal samples are collected for each bearing housing, with each sample containing 1024 data points. To reflect the challenge posed by imbalanced normal data in the real world, the training set contains 100, 50, and 10 normal samples from the left, middle, and right bearing housings, respectively, while the test set contains 100 fault samples and 100 normal samples.

[0066] Table 2 Introduction to Propulsion Shaft System Dataset

[0067] Step 2: To demonstrate the superiority of this embodiment, this implementation case evaluates its performance through comparative experiments, defining the following model for reference: VAE: Variational Autoencoder, derived from Auto-Encoding Variational Bayes.

[0068] SAE: Sparse Autoencoder, from A sparse autoencoder-based deep neural network approach for induction motor faults classification.

[0069] MemAE: Memory-augmented deep autoencoder, from Memorizing normality to detect anomaly.

[0070] FGDAE: A new machineryanomaly detection method towards complex operating conditions, from the paper FGDAE: A new machineryanomaly detection method towards complex operating conditions.

[0071] ARGWAE: An unsupervised method for mechanical system fault detection, from the paper "A novel unsupervisedgraph wavelet autoencoder for mechanical system fault detection".

[0072] SGAAEW: The SGAAE method after removing the memory enhancement module from the graph memory encoder.

[0073] SGAAE: Fault detection method according to an embodiment of the present invention.

[0074] Step 3: Define evaluation indicators To comprehensively evaluate the fault detection performance of the proposed method, four evaluation metrics are defined:

[0075] in, This represents the number of correctly detected samples out of the faulty samples. This represents the number of correctly detected samples in a normal sample. This represents the number of false detections in a normal sample. This represents the number of false detections in the faulty sample.

[0076] Step 4: Training the graph memory autoencoder and determining the decision threshold For each sample in the training set, graph data is constructed, and all graph data is then used to train the graph memory autoencoder. All historical training samples are then re-inputted into the trained graph memory autoencoder, and the reconstruction error of each sample is calculated. The maximum reconstruction error is used as the decision threshold for fault detection. Step 5: Online Testing The sample to be tested is input into the trained graph memory autoencoder, and its reconstruction error is calculated. If the reconstruction error of a sample exceeds the threshold, the position is determined to be in a fault state; otherwise, it is in a normal state. All evaluation indicators are calculated according to the formula in step three. Ten completely independent experiments are performed. All experiments are implemented in PyTorch 2.1.2 and executed on an Intel i5-13600K CPU.

[0077] Step Six: Analyze the Test Results Table 3 lists the experimental results of the method of the present invention. The overall accuracy of the method of the present invention reaches 99.50%, the precision reaches 99.03%, the recall reaches 100.00%, and the F-score reaches 99.51%. These results confirm the effectiveness of the method of the present invention in detecting propulsion shaft system faults. Furthermore, Figure 5 The reconstruction error distributions for normal and faulty samples during the test are shown, with a decision threshold of 5.54. For each bearing housing, the median reconstruction error for normal samples remains below the determined threshold, while the median reconstruction error for faulty samples is significantly higher. For example, in the mid-end bearing housing, the prediction scores for normal samples are concentrated around 4.20, while the prediction scores for faulty samples are concentrated around 15.34. This significant difference demonstrates that the method can accurately identify and locate faults in the propulsion shaft system.

[0078] The performance of SGAAE was further compared with the methods described above, and the results are summarized in Table 3. Among all the compared methods, SGAAEW ranked second in performance. Specifically, its overall accuracy, precision, recall, and F-score were 98.21%, 97.82%, 99.33%, and 98.20%, respectively, which were 1.29%, 1.21%, 0.67%, and 1.31% lower than SGAAE. This result shows that removing the graph memory enhancement module significantly weakens the model performance, verifying the effectiveness of removing the graph memory module in enhancing the ability to distinguish between normal and fault states. In summary, the method of this invention exhibits the best overall performance in the fault detection task.

[0079] In terms of computational efficiency, the method of this invention does not have a significant advantage in detection time compared to other methods. However, in industrial applications, a slight increase in computation time that can significantly improve the accuracy of fault detection is generally acceptable.

[0080] Table 3 Fault Detection Experiment Results

[0081] The present invention also provides a propulsion shaft system fault detection system under zero-fault sample conditions. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the propulsion shaft system fault detection method under zero-fault sample conditions as described above.

[0082] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the propulsion shaft fault detection method under zero-fault sample conditions as described above.

[0083] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting propulsion shaft system faults under zero-fault sample conditions, characterized in that, The steps are as follows: (1) Decompose the historical normal samples of multiple key parts of the propulsion shaft system into time-domain components with localized spectrum, and construct a graph structure based on the correlation between the time-domain components to form graph data that characterizes the intrinsic dependencies of the signal; (2) Input the graph data corresponding to each key part of the shaft system to be tested into the graph memory autoencoder, and the graph memory autoencoder calculates the reconstruction error based on the graph data; If the obtained reconstruction error is greater than the decision threshold, the part corresponding to the graph data is determined to be in a fault state; otherwise, it is in a normal state.

2. The propulsion shaft system fault detection method under zero-fault sample as described in claim 1, characterized in that: The graph memory autoencoder includes a graph attention encoder, a memory enhancement module, and a graph attention decoder. The graph attention encoder is used to extract global structural features from graph data. The memory enhancement module calculates the global structural features and each memory vector in the memory matrix based on a self-attention mechanism. The similarity between them is used to obtain the addressing weight, the addressing weight vector is obtained based on the addressing weight, and the memory feature is calculated based on the addressing weight vector and the memory matrix. The graph attention decoder reconstructs graph data based on memory features and then calculates the reconstruction error between the reconstructed data and the graph data.

3. The propulsion shaft system fault detection method under zero-fault sample as described in claim 1, characterized in that: The total loss function of the graph memory autoencoder allows for different reconstruction losses to be configured for different positions in the propulsion shaft system; All historical normal samples are input into the graph memory autoencoder, which calculates the reconstruction error of each sample. The maximum value among all calculated reconstruction errors is used as the decision threshold.

4. The propulsion shaft system fault detection method under zero-fault sample as described in claim 3, characterized in that: The total loss function is obtained by weighted summation of the reconstruction loss function at different locations and the training weights.

5. The propulsion shaft system fault detection method under zero-fault sample as described in claim 4, characterized in that: The reconstruction loss function at different locations is: in, Indicates the first Reconstruction loss at each location, It is the number of fault detection locations. Indicates the first The number of training samples at each position. and They represent the first time. Input and output graph data for each location.

6. The propulsion shaft system fault detection method under zero-fault sample as described in claim 5, characterized in that: Total loss function The formula is: in Indicates the current iteration number. Indicates the first In the nth iteration Reconstruction loss at each location, Indicates the first In the nth iteration Training speed at each position; It is the first Position in the next iteration The adaptive weights of the reconstruction loss are based on Perform dynamic optimization.

7. The propulsion shaft system fault detection method under zero-fault sample as described in claim 4, characterized in that: The training weights are calculated based on the training speed. The training speed is calculated based on the reconstruction loss at the corresponding location.

8. The propulsion shaft system fault detection method under zero-fault sample as described in claim 2, characterized in that: The graph attention encoder obtains global structural features from the input graph data through two layers of graph attention. The memory enhancement module uses addressing weight vectors. and memory matrix To obtain memory features The graph attention decoder extracts features from memory through two GTN layers. Reconstruct the input graph data: in, This represents the reconstructed graph data; Represents the adjacency matrix. This represents the ReLU activation function.

9. A propulsion shaft system fault detection system under zero-fault sample conditions, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the propulsion shaft system fault detection method under zero-fault sample as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the propulsion shaft fault detection method under zero-fault sample as described in any one of claims 1-8.

Images