AUV (Autonomous Underwater Vehicle) fault diagnosis method based on sample playback, medium and equipment

Abstract

The invention provides an AUV (Autonomous Underwater Vehicle) fault diagnosis method, medium and equipment based on sample playback, and relates to the technical field of AUV fault diagnosis, and the method comprises the steps: obtaining the operation state data of an AUV collected by multiple sensors, and carrying out the preprocessing, and obtaining an initial sample; according to the robustness of the initial samples to disturbance, selecting playback samples from the initial samples based on a rainbow memory for storage; extracting a global feature and a local feature of the initial sample, and performing feature fusion to obtain a first fusion feature; when an increment sample of the running state data of the AUV exists, extracting global features and local features of the storage sample of the rainbow memory and the increment sample, and performing feature fusion to obtain a second fusion feature; and inputting the first fusion feature and the second fusion feature into a KAN classifier to obtain an AUV fault diagnosis result. The problem of resource constraint in the incremental learning process can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of AUV fault diagnosis, and in particular to an AUV fault diagnosis method, medium, and equipment based on sample playback. Background Art

[0002] AUVs (Autonomous Underwater Vehicles) are widely used in fields such as ocean exploration, environmental monitoring, and military reconnaissance, but their working environment has the following characteristics: Complex environment: The underwater environment is changeable, with changes in pressure, temperature, salinity, etc.; Limited communication: The underwater communication bandwidth is limited, making real-time monitoring and remote control difficult; Difficult maintenance: Once an AUV fails, the maintenance cost is high and the risk is high. Therefore, AUVs need to have autonomous fault diagnosis capabilities to quickly detect, locate, and repair faults to ensure the smooth completion of the mission. AUV fault diagnosis is a key technology to ensure the reliable operation of AUVs in complex underwater environments. Due to the complex working environment, limited communication, and difficult maintenance of AUVs, fault diagnosis technology is particularly important. AUVs have less fault data, making it difficult to train high-performance diagnostic models. The noise and uncertainty of the underwater environment will also affect the accuracy of sensor data, and AUVs need to achieve real-time fault diagnosis with limited computing resources.

[0003] Incremental learning is particularly susceptible to catastrophic forgetting in terms of model generalization due to its inability to access data from previous tasks. This is because given a fault data stream, the class distribution of each fault will constantly change, which greatly increases the difficulty of incremental learning.

[0004] Methods based on sample replay can store samples from old tasks, effectively avoiding feature drift and improving the accuracy of incremental learning. However, these methods require extracting a portion of old tasks for storage. Because the number of stored samples is far smaller than the number of newly incoming samples, the samples in memory are smaller than the samples from the incoming task, leading to overfitting or being ignored during training. As a straightforward solution, gradually increasing the memory size as samples are passed in would make the problem setting unable to meet the limited fixed memory requirements of incremental learning. Therefore, it is necessary to maintain sufficient information about old classes within a small number of samples. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of resource constraints in the incremental learning process in the fault diagnosis method, and to propose an AUV fault diagnosis method based on sample playback, comprising the following steps:

[0006] S1. Obtain the AUV's operating status data collected by multiple sensors and preprocess it to obtain an initial sample;

[0007] S2. Select playback samples from the initial samples for storage based on the rainbow memory according to the robustness of the initial samples to disturbances;

[0008] S3, extracting global features and local features of the initial sample, and performing feature fusion to obtain a first fused feature;

[0009] S4. When there are incremental samples of the AUV's operating status data, extract the global features and local features of the samples and incremental samples in the rainbow memory, and perform feature fusion to obtain a second fused feature;

[0010] S5. Input the first fusion feature and the second fusion feature into the KAN classifier to obtain the AUV fault diagnosis result.

[0011] Furthermore, the preprocessing method includes: data cleaning and normalization.

[0012] Furthermore, the robustness of the initial sample to the disturbance is determined by calculating the uncertainty of the initial sample relative to the disturbance. The uncertainty of the initial sample relative to the disturbance is expressed as:

[0013]

[0014] Among them, S c Indicates the number of times the c-class sample is the first class of the prediction, T indicates the total number of categories, I c represents a binary class index vector, represents the prior value of the t-th disturbed sample, represents the posterior value of the tth disturbed sample, y represents the label of the sample, represents the estimated value of c, represents the tth perturbation sample, u(x) represents the uncertainty of sample x, c represents the sample category, A represents the number of perturbation samples, Represents the perturbed sample The data distribution defined, represents the perturbed sample, Indicates that under the condition that event x has occurred, event The probability of occurrence, f r (x|θ r ) is a random function, x represents the sample, θ r Represents the hyperparameter of the random factor of the rth perturbation, R represents the maximum value of r, ω r represents a random variable drawn from a categorical bivariate distribution.

[0015] Furthermore, the initial samples are sorted according to their robustness to disturbances, and samples are selected from the initial samples at equal intervals and stored in the rainbow memory.

[0016] Furthermore, global features are extracted through global feature CNN (convolutional neural network), which is expressed as:

[0017] F g =(AvgPool(ReLU(Pool(BN(Conv2d(C i )))))) ×5 ,

[0018] Among them, F g represents the global feature, C i Represents samples, incremental samples or initial samples in the rainbow memory, Conv2d() represents a two-dimensional convolution operation, BN() represents a normalization layer for batch samples, Pool() represents a pooling layer, ReLU represents an activation function, and AvgPool() represents an average pooling operation.

[0019] Furthermore, local features are extracted through local feature CNN, which is expressed as:

[0020] F r =(AvgPool(ReLU(Pool(BN(Conv2d(C i )))))) ×4 ,

[0021] Among them, F r Represents local features, C i Represents samples, incremental samples, or initial samples in the rainbow memory, Conv2d() represents a two-dimensional convolution operation, ReLU represents an activation function, and AvgPool() represents an average pooling operation.

[0022] Furthermore, the global features and local features are fused and expressed as:

[0023] F o =Flatten(F r )⊕Flatten(F g ),

[0024] Among them, F o Represents fusion features, Flatten() represents flattening operation, ⊕ represents splicing operation, F g represents the global feature, F r Represents local features.

[0025] Furthermore, L1 regularization loss is introduced to cut the spline weight parameters of the KAN classifier.

[0026] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned AUV fault diagnosis method based on sample playback is implemented.

[0027] The present invention also proposes an electronic device, comprising a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned AUV fault diagnosis method based on sample playback.

[0028] The beneficial effects brought about by the technical solution provided by the present invention are:

[0029] The present invention proposes an AUV fault diagnosis method based on rainbow memory and optimized KAN classifier. First, an initial sample is obtained based on the AUV's operating status data collected by multiple sensors. Based on the robustness of the initial sample to disturbances, samples are selected from the initial sample for storage based on the rainbow memory. When there are incremental samples of the AUV's operating status data, the fusion features of the initial sample, the rainbow memory stored samples, and the fusion features of the incremental samples are input into the optimized KAN classifier to obtain the AUV fault diagnosis results. Rainbow memory technology can effectively filter and select playback samples from old data. These samples can represent the characteristics and distribution of the data to the greatest extent, solving the problem of resource constraints in the incremental learning process. The KAN classifier with multi-sensor feature fusion can realize feature-level integration of different types of data collected by multiple sensors, fully explore the fault feature information in the data of each sensor, and achieve accurate fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of an AUV fault diagnosis method based on sample playback according to an embodiment of the present invention;

[0031] Figure 2 1 is a diagram of the global features, local features, and feature fusion network structure of an embodiment of the present invention;

[0032] Figure 3 It is a block diagram of an electronic device in an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0034] The flow chart of the AUV fault diagnosis method based on sample playback according to an embodiment of the present invention is as follows: Figure 1 , specifically including the following steps:

[0035] S1. Obtain the AUV's operating status data collected by multiple sensors and preprocess it to obtain an initial sample.

[0036] The complexity of the seabed environment and the diversity of missions place high demands on the real-time and reliability of AUV operational status. We acquire AUV operational status data collected by multiple sensors, including a Single Inertial Navigation System (SINS), Doppler Velocimeter (DVL), and depth gauge. We perform data cleaning to remove outliers and address missing values. We also perform data standardization and normalization, transforming the data and unifying the data format to obtain an initial sample.

[0037] S2. Based on the robustness of the initial samples to disturbances, replay samples are selected from the initial samples for storage based on the Rainbow Memory.

[0038] Specifically:

[0039] Initial samples are converted into feature representations. The feature representations of all initial samples form a feature space. In this feature space, points belonging to the same class are clustered together to form a "class." The samples selected for storage should not only be representative of the corresponding class but also distinguishable from other classes. Generally, samples near the class boundary are the most distinguishing, while samples near the center of the distribution are the most representative. To select samples that meet these two characteristics, different samples in the feature space are sampled.

[0040] To ensure diversity in sample distribution, it's necessary to estimate the relative position of each sample in feature space when collecting samples. However, directly calculating the relative positions of features is computationally expensive because it requires calculating the distances between samples. This paper uses a rainbow memory model to estimate the uncertainty of samples and approximate their relative positions in feature space. This assumes that more representative samples in the rainbow memory model are located closer to the center of the cluster distribution.

[0041] Specifically, a series of data augmentation transformation methods, such as color jittering, shearing, and random occlusion, are used to obtain perturbed samples. The uncertainty of the samples is calculated by measuring the variance of the model output of the perturbed samples. The prior value is When , the uncertainty of the distribution p(y=c|x) is approximately estimated by the Monte Carlo (MC) method. It is defined as a uniform mixture of various disturbances. y and A represent the sample, perturbed sample, sample label, and the number of perturbed samples, respectively. Represents the perturbed sample The data distribution defined by Use the random function f r (g) indicates.

[0042] By calculating the uncertainty of the initial sample relative to the disturbance, the robustness of the initial sample to the disturbance is judged. The uncertainty of the initial sample relative to the disturbance is expressed as:

[0043]

[0044] Among them, S c Indicates the number of times the c-class sample is the first class of the prediction, T indicates the total number of categories, I c represents a binary class index vector, represents the prior value of the t-th disturbed sample, represents the posterior value of the tth disturbed sample, y represents the label of the sample, represents the estimated value of c, represents the tth perturbation sample, u(x) represents the uncertainty of sample x, c represents the sample category, A represents the number of perturbation samples, Represents the perturbed sample The data distribution defined, represents the perturbed sample, Indicates that under the condition that event x has occurred, event The probability of occurrence, f r (x|θ r ) is a random function, x represents the sample, θ r is a hyperparameter representing the random factor of the rth perturbation, R represents the maximum value of r, ω r represents a random variate drawn from a categorical bivariate distribution, where is a discrete probability distribution over a set of categories T.

[0045] All initial samples are sorted according to uncertainty, sampled according to the interval |Dc| / kc, and stored in the rainbow memory. Where |Dc| is the number of initial samples and kc is the number of memory slots for each class.

[0046] S3. Extract global features and local features of the initial sample, and perform feature fusion to obtain a first fused feature.

[0047] Global features are extracted using a CNN network. The global feature CNN is a two-dimensional convolutional neural network composed of one-dimensional convolution kernels. In other words, one channel of a traditional two-dimensional convolution kernel is set to 1. The global feature CNN contains five stacked two-dimensional convolution layers and an average pooling layer. Each two-dimensional convolution layer includes a two-dimensional convolution, a batch normalization (BN) layer, a pooling layer, and a ReLU activation function. The convolution kernels are set to 1, 7, 2, 3, and 3, with a stride of 2 and channels of 3, 8, 16, 32, and 64. It can be expressed as:

[0048] F g =(AvgPool(ReLU(Pool(BN(Conv2d(C i )))))) ×5 ,

[0049] Among them, F g represents the global feature, C i Represents samples, incremental samples or initial samples in the rainbow memory, Conv2d() represents a two-dimensional convolution operation, BN() represents a normalization layer for batch samples, Pool() represents a pooling layer, ReLU represents an activation function, and AvgPool() represents an average pooling operation.

[0050] Local features are extracted using a CNN network, which consists of four stacked two-dimensional convolutional layers and an average pooling layer, and is expressed as follows:

[0051] F r =(AvgPool(ReLU(Pool(BN(Conv2d(C i )))))) ×4 ,

[0052] Among them, F r Represents local features, C i Represents samples, incremental samples or initial samples in rainbow memory, Conv2d n×3 () represents a two-dimensional convolution operation, and ReLU represents the activation function.

[0053] The global feature F g and local features F r Perform Flatten (flattening operation) separately to obtain features F1 and F2, and concatenate features F1 and F2 to obtain fusion features, which are expressed as:

[0054] F o =Flatten(F r )⊕Flatten(F g ),

[0055] Among them, F o Represents fusion features, Flatten() represents flattening operation, ⊕ represents splicing operation, F g represents the global feature, F r Represents local features.

[0056] Global feature extraction, local feature extraction, and feature fusion network block diagram reference Figure 2 . Figure 2In the original data, the parameters (1, 13, 190) indicate the number of image channels, the dimensions of the original data, the number of sensor data points, and the number of sensor data points. In the convolutional layer, the parameters (a, b, c) indicate the number of channels, the size of the convolution kernel (b×c), and the size of the kernel (c×d). The parameters (d, e) in the average pooling layer indicate the size of the convolution kernel (d×e). The parameter (f) in the output layer indicates the number of features.

[0057] S4. When there are incremental samples of the AUV's operating status data, extract the global features and local features of the samples and incremental samples in the rainbow memory, and perform feature fusion to obtain a second fused feature.

[0058] Rainbow Memory Technology (RM) can effectively filter and select playback samples from old data based on specific algorithms and strategies. These samples can best represent the characteristics and distribution of the data.

[0059] S5. Input the first fusion feature and the second fusion feature into the KAN classifier to obtain the AUV fault diagnosis result.

[0060] KAN (Kolmogorov-Arnold Networks) excels at efficiently processing complex data. EKAN (Optimized KAN) builds on KAN, constructed through a combination of linear layers and b-spline transformations. This network is highly efficient in processing complex data. KAN is based on the Kolmogorov-Arnold representation theorem. The Kolmogorov-Arnold representation theorem is as follows:

[0061]

[0062] Function f(x1,K,x n ) can be expressed as a combination of a finite number of continuous functions, unary operations and binary addition operations, x n represents the nth input variable, φ q It is a unary function, q represents the sum index, which is used to process the input p-th vector x p , n represents the number of vectors, external function It is also a unary function used to calculate the internal function φ q Therefore, in the design of a network classifier, multiple univariate functions can be used to perform the classification task.

[0063] An effective KAN is constructed by combining linear layers and b-spline transformations. The construction of the linear layer is described as follows:

[0064] φ(x)=w(b(x)+spline(x)),

[0065] b(x)=SiLu(x)=x / (1+e -x ),

[0066]

[0067] Among them, φ(x) represents the linear layer, B i (x) represents the B-spline transformation, the amplitude of the activation function w() is controllable, and the activation function is a linear combination of the basis function b(x) and the B-spline basis function spline(x), c i is a trainable parameter. By training the shape of the activation function, complex functions can be effectively fitted. In addition, to improve computational efficiency, EKAN introduces an L1 regularization loss for cutting the spline weight parameters. The AUV fault features after multi-sensor feature fusion are used as the input of EKAN with an intermediate layer size of 32 to obtain excellent fault diagnosis capabilities. The time complexity of the KAN layer is described as follows:

[0068] O(L(nk+nm)),

[0069] Where L is the number of linear layers, n is the number of input fusion features, k is the number of b-splines, and m is the number of output features. EKAN utilizes the multi-sensor information of AUVs to achieve accurate fault diagnosis of autonomous underwater vehicles.

[0070] Rainbow Memory selects and stores appropriate samples that contain key information from the old task. It then extracts global and local features from these stored samples and the new task samples. After feature fusion, these features are simultaneously fed into the improved KAN for fault diagnosis. During training for the new task, the feature information from the old task is fully utilized, avoiding the loss of old task features when processing new data. This significantly improves the model's fault diagnosis capabilities when faced with new tasks.

[0071] In an exemplary embodiment, a computer-readable storage medium is included. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the above-mentioned AUV fault diagnosis method based on sample playback is implemented.

[0072] See also Figure 3 In an exemplary embodiment, an electronic device is also included, including at least one processor, at least one memory, and at least one communication bus.

[0073] The memory stores a computer program including computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through a communication bus to execute the above-mentioned AUV fault diagnosis method based on sample playback.

[0074] In order to verify the effectiveness of the method of the present invention, the following experiments were performed.

[0075] AUV fault data is derived from navigation experiments conducted on the Haizhe autonomous underwater vehicle (AUV). The Haizhe AUV is a small, unmanned underwater vehicle (AUV) consisting primarily of a motor, depth sensor, and control room. A depth sensor and an IMU (inertial measurement unit) are installed on the AUV to collect monitoring information for autonomous underwater navigation. Sensor data collected by the Haizhe AUV includes time, control signals, diving depth, pressure, power supply voltage, roll angle, pitch angle, yaw angle, X-axis acceleration, Y-axis acceleration, Z-axis acceleration, roll angular acceleration, pitch angular acceleration, and yaw angular acceleration. In each experiment, a fault was introduced and lasted for 1020 seconds. The Haizhe AUV has five fault states: normal, depth sensor failure, increased load, minor propeller damage, and major propeller damage. For detailed information on the Haizhe dataset, see Table 1.

[0076] Table 1

[0077] Fault status Label Dataset train test Normal data 0 182 146 36 Increased load 1 268 214 54 Depth sensor failure 2 266 213 53 Minor propeller damage 3 260 208 52 Severe propeller damage 4 249 199 50

[0078] During the data preprocessing phase, each data sample is first extracted from the raw data, and the dataset is processed using a sliding window with a window length of 1024. This step aims to capture local features in the data and provide more representative data segments for subsequent feature extraction. In the "Hai Zhe" dataset, 13 key sensor data types, in addition to time data, were selected to cover various state information of AUVs (autonomous underwater vehicles) in different environments. To ensure data consistency and comparability, each sensor data point was standardized, and the length of each data sample was selected to be 190 data points.

[0079] If a sensor data point contains fewer than 190 data points, the last set of data points is replicated until the total number of data points reaches 190. This method effectively mitigates the impact of missing data on model training while maintaining data continuity. Conversely, if the sensor data point exceeds 190, the excess data points are deleted to ensure that all data samples have the same length. This step helps reduce data redundancy and improves the efficiency and accuracy of subsequent model training.

[0080] After unifying the data length, each data sample was normalized to adjust its value range to [-1, 1]. Normalization can eliminate dimensional differences between sensor data, allowing the model to converge better during training and improve its generalization ability.

[0081] Finally, all preprocessed datasets are randomly split into training and test sets. The training set is used for model training and parameter optimization, while the test set is used to evaluate model performance and generalization. This random partitioning ensures that the model does not overfit to a specific data distribution during training, thereby improving the model's reliability in real-world applications.

[0082] The method of the present invention was implemented using Python 3.9 and PyTorch 2.2.1 on NVIDIA RTX2080Ti and Intel(R) i9-10900X on Windows 10. The experimental hyperparameter settings are shown in Table 2.

[0083] Table 2

[0084] Hyperparameters Learning rate Optimizer Loss Function Batch size Number of training rounds Numerical 0.001 Adam Cross Entropy 128 200

[0085] Based on the above AUV public dataset, the proposed method is compared with other incremental learning methods, including: Finetuning (Learning a unified classifier incrementally via rebalancing), eeil (End-to-end incremental learning), Ewc (A novel incremental learning algorithm based on incremental vector support machina and incremental neural network learn++), Freezing (Catastrophic forgetting inconnectionist networks), Lucir (Learning a unified classifier incrementally via rebalancing), Lwf (Learning without forgetting), Lwm (Learning without memorizing), path_integral (A path-integral-based reinforcement learning algorithm for path following of an autoassembly mobile robot), R-walk (Riemannian walk for incremental learning: Understanding forgetting and intransigence), and Icarl (icarl: Incremental classifier and representation learning). Table 3 lists the incremental results of each method on the above dataset.

[0086] Table 3

[0087]

[0088]

[0089] As shown in Table 3, the proposed method achieves optimal values ​​in both accuracy and time metrics. In incremental process 1, 100% accuracy was achieved, demonstrating the proposed method's superior performance in fault classification and its ability to accurately classify and identify various fault types. After incremental process 2, the proposed method's accuracy dropped to 98.1%, a mere 1.9% decrease. The highest accuracy of other incremental learning methods after incremental process 2 was only 97.3%, all lower than the proposed method. This result demonstrates that the proposed network can adapt well to the incorporation of new data and maintain high classification accuracy. After incremental process 3, the proposed method's accuracy dropped only 1.5%, still maintaining an accuracy of 96.6%. This clearly demonstrates that the proposed method not only possesses strong classification capabilities but also exhibits excellent adaptability and learning capabilities. It effectively avoids forgetting old knowledge as new data categories are added, thereby maintaining the network's efficiency and accuracy in fault classification tasks.

[0090] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault diagnosis method for AUV based on sample playback, characterized in that: The following steps are involved: S1. Obtain the AUV's operating status data collected by multiple sensors and preprocess it to obtain an initial sample; S2. Select playback samples from the initial samples for storage based on the rainbow memory according to the robustness of the initial samples to disturbances; S3, extracting global features and local features of the initial sample, and performing feature fusion to obtain a first fused feature; S4. When there are incremental samples of the AUV's operating status data, extract the global features and local features of the rainbow memory stored samples and the incremental samples, and perform feature fusion to obtain a second fused feature; S5. Input the first fusion feature and the second fusion feature into the KAN classifier to obtain the AUV fault diagnosis result.

2. The AUV fault diagnosis method based on sample playback according to claim 1, characterized in that: Preprocessing methods include: data cleaning and normalization.

3. The AUV fault diagnosis method based on sample playback according to claim 1, characterized in that: By calculating the uncertainty of the initial sample relative to the disturbance, the robustness of the initial sample to the disturbance is judged. The uncertainty of the initial sample relative to the disturbance is expressed as: Among them, S c Indicates the number of times the c-class sample is the first class of the prediction, T indicates the total number of categories, I c represents a binary class index vector, represents the prior value of the t-th disturbed sample, represents the posterior value of the tth disturbed sample, y represents the label of the sample, represents the estimated value of c, represents the tth perturbation sample, u(x) represents the uncertainty of sample x, c represents the sample category, A represents the number of perturbation samples, Represents the perturbed sample The data distribution defined, represents the perturbed sample, Indicates that under the condition that event x has occurred, event The probability of occurrence, f r (x|θ r ) is a random function, x represents the sample, θ r Represents the hyperparameter of the random factor of the rth perturbation, R represents the maximum value of r, ω r represents a random variable drawn from a categorical bivariate distribution.

4. The AUV fault diagnosis method based on sample playback according to claim 1, characterized in that: The initial samples are sorted according to their robustness to disturbances, and samples are selected from the initial samples at equal intervals and stored in the rainbow memory.

5. The AUV fault diagnosis method based on sample playback according to claim 1, characterized in that: The global features are extracted by global feature CNN and expressed as: F g =(AvgPool(ReLU(Pool(BN(Conv2d(C i )))))) ×5 , Among them, F g represents the global feature, C i Represents samples, incremental samples or initial samples in the rainbow memory, Conv2d() represents a two-dimensional convolution operation, BN() represents a normalization layer for batch samples, Pool() represents a pooling layer, ReLU represents an activation function, and AvgPool() represents an average pooling operation.

6. The AUV fault diagnosis method based on sample playback according to claim 1, characterized in that: The local features are extracted by local feature CNN and expressed as: F r =(AvgPool(ReLU(Pool(BN(Conv2d(C i )))))) ×4 , Among them, F r Represents local features, C i Represents samples, incremental samples, or initial samples in the rainbow memory, Conv2d() represents a two-dimensional convolution operation, ReLU represents an activation function, and AvgPool() represents an average pooling operation.

7. The AUV fault diagnosis method based on sample playback according to claim 1, characterized in that: The global features and local features are fused and expressed as: F o =Flatten(F r )⊕Flatten(F g ), Among them, F o Represents fusion features, Flatten() represents flattening operation, ⊕ represents splicing operation, F g represents the global feature, F r Represents local features.

8. The AUV fault diagnosis method based on sample playback according to claim 1 is characterized in that: L1 regularization loss is introduced to cut the spline weight parameters of the KAN classifier.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1 to 8.

Images