AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention

CN122839241APending Publication Date: 2026-09-29ZHEJIANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610685452.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-19
Publication Date
2026-09-29

AI Technical Summary

Technical Problem

[0005]本发明要解决的技术问题是,克服现有生成式方法在处理水下机器人稀疏、低频且不平衡数据时的局限性,提出一种基于统计特征增强与多尺度注意力的水下机器人故障诊断方法

Benefits of technology

[0018]1、鲁棒性强:在极度不平衡(如1:20)的数据比例下,仍能保持较高的诊断准确率,显著优于传统的生成式方法。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122839241A_ABST
    Figure CN122839241A_ABST
Patent Text Reader

Abstract

The application relates to underwater robot technology and aims to provide an AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention. The method comprises the following steps: collecting multi-channel data of an AUV in a running state, processing the multi-channel data in a double-path parallel mode to obtain multi-dimensional statistical feature representation and original state feature representation of the original data; and then performing adaptive fusion to obtain a final diagnosis result after classifying the generated fusion features. The application has strong robustness and can still maintain a high diagnosis accuracy rate under an extremely unbalanced data ratio; the application combines the sensitivity of statistical features to distribution changes and the extraction ability of deep networks to space-time patterns, solves the problem that a single feature is difficult to represent complex underwater robot faults, and adopts deep separable convolution to greatly reduce the parameter quantity and calculation load, so that the application is suitable for the on-board or real-time diagnosis requirements of underwater robots.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of underwater robot technology, and more specifically to an online mechanical fault diagnosis method for underwater robots in unbalanced and sparse data environments. Background Technology

[0002] Autonomous Underwater Vehicles (AUVs) play a crucial role in marine monitoring, resource exploration, and other fields. Because AUVs operate autonomously underwater for extended periods, their internal states and operating conditions are opaque, making fault diagnosis paramount. However, two major challenges exist in practical applications: First, the data is extremely imbalanced, with fault samples being far scarcer than normal samples. Second, the data is sparse and low-frequency; AUV telemetry data typically has a low sampling rate (around 100Hz) and contains a large number of sparse state variables (such as depth and voltage), unlike the high-frequency, dense vibration signals commonly found in industrial scenarios.

[0003] Current mainstream solutions mainly rely on generative data augmentation techniques such as generative adversarial networks. While these methods are effective in processing high-frequency vibration signals, they often struggle to generate realistic synthetic samples when faced with the low sampling rate and sparse state data of underwater robots. In some cases, they may even cause pattern collapse or feature redundancy, leading to a decline in diagnostic performance.

[0004] Therefore, it is necessary to propose new solutions to meet the needs of actual production operations. Summary of the Invention

[0005] The technical problem this invention aims to solve is to overcome the limitations of existing generative methods in processing sparse, low-frequency, and imbalanced data from underwater robots. It proposes an underwater robot fault diagnosis method based on statistical feature enhancement and multi-scale attention. This method does not rely on generating spurious samples; instead, it addresses the imbalance problem by mining the statistical properties and multi-scale spatiotemporal features of existing data.

[0006] To solve the technical problem, the solution of the present invention is:

[0007] A fault diagnosis method for AUVs based on statistical feature enhancement and multi-scale attention is provided, including:

[0008] Multi-channel data from an underwater robot (AUV) during operation is collected and processed in a dual-path parallel manner to obtain two feature representations of the raw data; among them,

[0009] Path 1 is as follows: Perform sliding window processing on the data of each channel, calculate multiple statistical feature descriptors to form a statistical feature vector; after weighting and filtering the statistical feature vector using an attention mechanism, extract multidimensional statistical feature representations through multi-layer residual sensing;

[0010] Path 2 is as follows: Extract multi-scale spatiotemporal features of the data using multiple parallel branches; each branch adopts a combination of depthwise separable convolution and convolutional block attention mechanism, and each branch has a different convolutional kernel size; the multi-receptive field spatiotemporal features output by each branch are concatenated to obtain the original state feature representation;

[0011] Adaptive fusion is performed on the multidimensional statistical feature representations and original state feature representations output by the two paths, and the generated fused features are classified to obtain the final diagnostic result.

[0012] Description of the invention principle:

[0013] The core principle of the AUV fault diagnosis method proposed in this invention lies in its "dual-path" feature fusion mechanism, which addresses the problem that traditional generative data augmentation methods struggle to generate realistic samples when underwater robots face conditions of extremely scarce fault data and highly sparse state data (low sampling rate). The specific principle is as follows:

[0014] The applicant noted in long-term research that underwater robots exhibit sparse and low-frequency telemetry data, and that rare fault modes often show subtle shifts in the statistical distribution of the data (such as higher-order moments and entropy values) rather than obvious textural changes in the original waveform. Therefore, this invention proposes a novel statistical feature extraction path that explicitly calculates multidimensional statistical descriptors (such as skewness, kurtosis, and Shannon entropy). Through a feature attention network, these statistical features are adaptively filtered and weighted using an attention mechanism, enabling the network to "explicitly" focus on distribution changes that characterize rare faults, thereby enhancing sensitivity to minority class samples without relying on fabricated data.

[0015] Because different state variables of underwater robots (such as control signals and depth changes) have different temporal response scales, traditional single-scale convolution techniques struggle to simultaneously capture transient impacts and gradual trends. Therefore, this invention constructs a multi-scale attention module that uses convolutional kernels of different sizes to process the raw data in parallel, obtaining temporal features from multiple receptive fields. Simultaneously, by combining depthwise separable convolution with a convolutional block attention module, computational complexity is reduced while features are recalibrated from both "channel" and "spatial" dimensions. This automatically suppresses noise interference from irrelevant sensors, focusing on key time segments and channels containing fault information.

[0016] Through extensive experimental verification, the applicant discovered that the statistical features (global distribution information) and original depth features (local spatiotemporal information) in underwater robot telemetry data are highly complementary. Therefore, this invention introduces gated linear units as a fusion layer, enabling the fusion network to dynamically learn the fusion weights of the two feature paths based on the characteristics of the input data. This mechanism allows the model to adaptively balance its dependence on prior statistical knowledge and data-driven features when facing extremely imbalanced data, thereby constructing a discriminative feature space with larger inter-class distances and more compact intra-class features, achieving high-precision diagnosis of underwater robot faults.

[0017] Compared with the prior art, the present invention has the following beneficial effects and advantages:

[0018] 1. Strong robustness: It can maintain a high diagnostic accuracy even with extremely unbalanced data ratios (such as 1:20), which is significantly better than traditional generative methods.

[0019] 2. Complementary features: It combines the sensitivity of statistical features to distribution changes with the ability of deep networks to extract spatiotemporal patterns, thus solving the problem that single features are difficult to characterize complex underwater robot faults.

[0020] 3. High computational efficiency: It adopts depthwise separable convolution, which greatly reduces the number of parameters and computational load, making it suitable for the needs of underwater robot onboard or real-time diagnosis. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall structure of the statistical feature enhancement data enhancement network framework proposed in this invention.

[0022] Figure 2 This is a schematic diagram of the multi-scale attention module in this invention, illustrating the parallel processing path with different convolutional kernel sizes.

[0023] Figure 3 This is a schematic diagram of the feature attention network in this invention, illustrating the attention mask generation and residual connection process. Detailed Implementation

[0024] First, it should be noted that this invention relates to fault diagnosis technology and signal processing technology. The implementation of this invention may involve the principles of several fundamental algorithms. The applicant believes that, after carefully reading the application documents and accurately understanding the implementation principles and objectives of this invention, and in conjunction with existing known technologies, those skilled in the art can fully utilize their algorithm-writing capabilities to implement this invention. All references in this application document fall within this scope, and the applicant will not list them all.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] I. AUV Fault Diagnosis Methods

[0027] The AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention described in this invention includes the following steps:

[0028] 1. Collect multi-channel data of the underwater robot (AUV) during operation.

[0029] Data acquisition is performed by multiple sensors mounted on the AUV, collecting signals from multiple sources, including motion status and environmental parameters. The control processing board receives these sensor signals, performs analog-to-digital conversion, and obtains the underwater robot's status data. ,in The time window length, This refers to the number of channels. This status data is transmitted to a host computer running on the AUV or a cloud server located at the shore-based monitoring station for subsequent calculations and diagnostic processing using built-in software.

[0030] The AUV's operating status is at least one of the following: normal operation, minor propeller damage, severe propeller damage, excessive load, or depth sensor malfunction. When collecting data, the AUV's status variables are recorded at set time intervals, including at least: control signals, pressure, depth, roll attitude angle, pitch attitude angle, yaw attitude angle, angular velocity, acceleration, and battery voltage.

[0031] 2. Data is processed in a dual-path parallel manner to obtain two feature representations.

[0032] This invention employs a parallel setup of a statistical feature enhancement module and a multi-scale attention module to perform computational processing on the same original data using a dual-path parallel approach. The former is used to acquire multidimensional statistical features (such as mean, skewness, kurtosis, entropy, etc.) that are highly sensitive to distribution shifts caused by rare faults, while the latter is used to extract intrinsic features containing rich spatiotemporal information.

[0033] (2.1) Taking the feature enhancement module as path one, it specifically includes:

[0034] The statistical feature extraction unit extracts the state signal data within each channel. A sliding window process is performed to calculate multiple statistical feature descriptors for each channel, which are then used to construct a statistical feature vector. The statistical feature descriptor refers to at least one of the following: minimum value, maximum value, and mean value. Median, peak-to-peak value, standard deviation Zero crossing rate, kurtosis skewness Root mean square First-order difference, Shannon entropy and peak factor .

[0035] Then, the attention mechanism is used in the feature attention network to apply attention to the statistical feature vectors. Weighting and filtering are performed, and multidimensional statistical features are extracted through multi-layer residual perception to show that these features are highly sensitive to distribution shifts caused by rare faults.

[0036] When performing weighted processing, first calculate the attention mask using the following formula. :

[0037]

[0038] Then, the weighted features are calculated using the following formula. :

[0039]

[0040] in, Use the Sigmoid activation function; All are learnable weights; To modify the activation function of the linear unit; This is the feature vector of the batch mean; This indicates element-wise multiplication;

[0041] By using an attention mechanism to adaptively weight features, irrelevant features can be suppressed and the expression of key features can be enhanced.

[0042] After being processed through stacked residual layers, it is finally mapped to statistical feature embedding.

[0043] (2.2) Using the multi-scale attention module as path two, specifically including:

[0044] Extracting state data using multiple parallel branches Multi-scale spatiotemporal features are extracted; each branch employs a combination of depthwise separable convolution and convolutional block attention mechanisms, enabling feature refinement in both channel and spatial dimensions to extract intrinsic features rich in spatiotemporal information. Each branch has a different convolutional kernel size, allowing simultaneous capture of local mutations and long-term dependencies. The multi-receptive-field spatiotemporal features output from each branch are concatenated to obtain the original state feature representation.

[0045] As an example, three parallel branches are used to extract multi-scale spatiotemporal features; each branch includes depthwise separable convolutional units, a convolutional block attention mechanism module, and depthwise separable convolutional units arranged sequentially; each branch corresponds to a different convolutional kernel size. , .

[0046] The depthwise separable convolution includes depthwise convolution and pointwise convolution to reduce the number of parameters; the convolutional block attention mechanism includes sequentially connected channel attention modules and spatial attention modules; wherein,

[0047] Channel attention The calculation formula is:

[0048]

[0049] Spatial attention The calculation formula is:

[0050]

[0051] Feature map output by convolutional block attention mechanism The calculation formula is:

[0052]

[0053] in, Use the Sigmoid activation function; All are learnable weights; To modify the activation function of the linear unit; Indicates average pooling; Indicates max pooling; Input feature map; The feature map after channel attention processing; This represents one-dimensional convolution; This indicates taking the average value; This indicates taking the maximum value; This indicates element-wise multiplication.

[0054] Finally, the outputs of the three paths are concatenated, and the original state feature representation is obtained through global average pooling and linear projection. .

[0055] 3. Combine the two feature representations and output the classification result.

[0056] The fusion network module simultaneously receives outputs from both the statistical feature enhancement module and the multi-scale attention module. A gated linear unit is introduced as the fusion layer, using a gating mechanism to control the information flow between statistical and original features, enabling adaptive dynamic fusion to generate discriminative feature representations. This leverages the complementarity of the two to improve the clarity of classification boundaries in data-scarce situations. Then, a classifier is used to classify the generated fused features, outputting the final diagnostic result.

[0057] Using the above method, the present invention can diagnose typical faults such as occasional propeller damage (minor / serious), increased load, and depth sensor failure based on real underwater robot platform data.

[0058] II. AUV Fault Diagnosis System

[0059] The present invention further provides an AUV fault diagnosis system for implementing the aforementioned diagnostic method, specifically comprising:

[0060] 1. Two functional units in two hardware forms:

[0061] The data acquisition module includes multiple sensors mounted on the AUV for acquiring multi-source sensor signals, including motion status and environmental parameters.

[0062] The control processing board, mounted on the AUV, is used to receive sensor signals and output underwater robot status data.

[0063] 2. This invention further includes a statistical feature enhancement module, a multi-scale attention module, and a fusion network module. These three functional units run as software programs on a host computer on the AUV, or on a cloud server of a shore-based monitoring station. Figure 1 As shown, the statistical feature enhancement module and the multi-scale attention module are set up in parallel and run in a dual-path parallel manner; the fusion network module is placed last and receives the outputs of the two parallel modules at the same time.

[0064] 3. The structure of the multi-scale attention module is as follows: Figure 2 As shown, multiple parallel branches are used to extract multi-scale channel and temporal features from the data; in each branch, depthwise separable convolutional units, convolutional block attention modules, and depthwise separable convolutional units are arranged in sequence, and each branch has a different convolutional kernel size; the concatenation unit is connected to each branch to concatenate the output multi-receptive field spatiotemporal features to obtain the original state feature representation.

[0065] The workflow of the multi-scale attention module is as follows:

[0066] (1) Input distribution: The input feature map is simultaneously fed into three parallel multi-scale depth separable convolution branches, using three different sizes of convolution kernels: 5×5, 11×11, and 23×23, to achieve multi-scale feature extraction.

[0067] (2) Multi-scale feature extraction: Each branch performs depthwise separable convolution: First, perform depthwise convolution (channel-wise convolution) on the input feature map, and then fuse channel information through pointwise convolution (1×1 convolution) to output the feature map of the corresponding scale.

[0068] (3) Attention Mechanism Weighting: For each scale feature map, the channel dimension weight is first learned through the channel attention submodule, and multiplied element by element with the original feature map to obtain the channel weighted feature; then the channel weighted feature is sent to the spatial attention submodule to learn the spatial dimension weight, and multiplied element by element again to obtain multi-scale features that have both channel and spatial attention.

[0069] (4) Secondary multi-scale feature extraction: The three sets of multi-scale features after attention weighting are fed into a set of multi-scale depth separable convolutions with the same structure to further enhance the feature expression at different scales.

[0070] (5) Stitching and fusion output: The three sets of multi-scale features extracted in the second stage are stitched together in the channel dimension, and the final multi-scale attention enhancement feature map is obtained through fusion operation, which is used as the module output.

[0071] The core design intent of the multi-scale attention module is to: capture features under different receptive fields through multi-scale convolution to adapt to the multi-scale changes of the target object; significantly reduce the amount of computation through depthwise separable convolution to ensure model efficiency; and enhance key features from the channel and spatial dimensions by combining channel and spatial attention, suppress redundant information, and improve feature representation capabilities.

[0072] 4. The statistical feature enhancement module includes a statistical feature extraction unit and a feature attention network arranged in sequence, which are used to extract multidimensional statistical feature representations from the data that are sensitive to the distribution shift of rare faults.

[0073] The statistical feature extraction unit is used to extract multidimensional statistical features of the signal, including at least: minimum, maximum, average, median, peak-to-peak value, standard deviation, zero-crossing rate, kurtosis, skewness, root mean square, first-order difference root mean square, Shannon entropy, and peak factor.

[0074] The feature attention network consists of a normalization module, a feature attention module, a residual module, and a fully connected layer arranged in sequence, as shown in the following structure: Figure 2 As shown, the specific workflow is as follows:

[0075] (1) Input and batch normalization: The original feature map first goes through the batch normalization module to complete the data standardization and stabilize the training process.

[0076] (2) Feature attention weighting: After the batch normalized features enter the feature attention module, channel / spatial information is extracted through two convolutions respectively; attention weights are generated by activation functions and then multiplied element-wise with the original features to achieve weighted enhancement of key features.

[0077] (3) Residual stacking: The weighted features are fed into the residual module. The gradient vanishing problem is alleviated by residual connection, which ensures the training stability of deep network. The figure shows that this module will be stacked 6 times.

[0078] (4) Fully connected layer output: After passing through the fully connected layer, the high-dimensional features are mapped to the final output result (such as classification probability or regression value).

[0079] 5. The fusion network module, including gated linear units and a classifier, is used to adaptively fuse and classify the outputs of the two paths and output the final diagnostic result.

[0080] 6. Based on the knowledge of those skilled in the art, the present invention also provides a computer device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions to be executed by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the aforementioned AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention.

[0081] The present invention further provides a computer-readable storage medium storing computer instructions for causing the computer to execute the aforementioned AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention.

[0082] III. Specific application examples:

[0083] This embodiment is based on the technical solution of the present invention and uses real underwater robot platform data for verification. The experiment aims to evaluate the diagnostic performance of the statistical feature enhancement data augmentation network method proposed in this invention for underwater robot propulsion system and sensor faults under conditions of extreme data imbalance and sparse telemetry data.

[0084] This embodiment specifically includes the following steps:

[0085] Step 1: Experimental Environment and Fault Setup. Experimental data was collected from an underwater robot platform with autonomous navigation capabilities. Five operating states were set: normal state, minor propeller damage, severe propeller damage, increased load, and depth sensor failure. Data was recorded via the underwater robot's main control board at a sampling interval of 10 milliseconds (approximately 100Hz). State variables included control signals, pressure, depth, attitude angles (roll, pitch, yaw), angular velocity, acceleration, and battery voltage, forming a typical low-frequency sparse telemetry dataset.

[0086] Step 2: Dataset Construction and Imbalance Setting. To simulate the scenario of scarce fault data in actual sea trials, datasets with different imbalance ratios (the ratio of normal samples to faulty samples) were constructed, namely 1:2, 1:5, 1:10, and 1:20. Among them, 1:20 represents an extremely imbalanced and severe operating condition.

[0087] Step 3: Statistical Feature Extraction and Enhancement. A sliding window process is applied to the collected state data. Thirteen statistical features are calculated for each channel, including: minimum, maximum, mean, median, peak-to-peak value, standard deviation, zero-crossing rate, kurtosis, skewness, root mean square, first difference, Shannon entropy, and peak factor. These statistical feature vectors are then input into a feature attention network. The feature attention network weights the features using an attention mask and processes them through six residual layers to extract high-level statistical feature embeddings sensitive to the distribution shift of rare faults.

[0088] Step 4: Multi-scale spatiotemporal feature extraction. The original state data is input into the multi-scale attention module. This module has three parallel branches with convolutional kernel sizes of [sizes to be filled in]. This approach aims to cover features across short, medium, and long timescales. Each branch extracts features using depthwise separable convolutions and combines a convolutional block attention mechanism module to recalibrate features in both channel and spatial dimensions, ultimately outputting a raw feature embedding rich in spatiotemporal information.

[0089] Step 5: Feature Fusion and Diagnosis. The statistical features obtained in Step 3 and the original features obtained in Step 4 are embedded into the gated linear unit (GLU). The fusion layer dimension is set to 128. The GLU adaptively adjusts the contribution of the two features, and the generated fused features are input into the classifier to obtain the final diagnostic result.

[0090] Step 6: To verify the effectiveness of the above model, it was trained and tested. The experimental results are as follows:

[0091] Overall Accuracy: With a data ratio of 1:2 (slightly imbalanced), the average accuracy of the method of this invention reaches 95.34%; with a data ratio of 1:20 (extremely imbalanced), the average accuracy remains at 72.46%. Comparative Analysis: Under the extreme condition of 1:20, the method of this invention significantly outperforms other comparative methods. Specific Fault Identification: Under the 1:20 condition, the recognition rate of this invention for the "normal" state reaches 100%, the recognition rate for "depth sensor fault" reaches 90.91%, and the recognition rate for "increased load" reaches 77.16%. This indicates that this invention can effectively overcome the bias caused by data scarcity and accurately distinguish different types of faults. This embodiment verifies that statistical feature-enhanced data augmentation networks, under conditions of sparse sensing and extreme class imbalance, have superior robustness and diagnostic accuracy compared to traditional deep learning and generative augmentation methods.

[0092] This example first preprocesses the acquired AUV multi-channel status data using a multi-dimensional statistical feature extraction algorithm, constructing a feature vector containing temporal statistics and information entropy while preserving the original time-series signal for input into the statistical feature enhancement data augmentation network. For the actual operating conditions where fault samples are scarce and the positive-to-negative sample ratio is extremely imbalanced (e.g., 1:20), only low-frequency telemetry data due to sparse sensing is used as the training set for network training. During training, parallel feature attention networks and multi-scale attention modules respectively weight and enhance sensitive statistical distribution features and deep multi-scale spatiotemporal features, and adaptive feature fusion is achieved through gated linear units. After training, diagnostic verification is performed using an imbalanced test set containing various faults. Based on the probability distribution output by the classifier, the final fault diagnosis result is determined.

[0093] Based on extensive field tests conducted by the applicant's research team on actual underwater robot platforms, this invention is highly effective for underwater robot fault diagnosis in sparse telemetry data environments. It is of great help in solving the problem of fault identification under severe class imbalance and low sampling rate conditions, and its performance is significantly better than traditional generative data augmentation methods.

[0094] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A fault diagnosis method for AUVs based on statistical feature enhancement and multi-scale attention, characterized in that, include: Multi-channel data from an underwater robot (AUV) during operation is collected and processed in a dual-path parallel manner to obtain two feature representations of the raw data; among them... Path 1 is as follows: Perform sliding window processing on the data of each channel, calculate multiple statistical feature descriptors to form a statistical feature vector; after weighting and filtering the statistical feature vector using an attention mechanism, extract multidimensional statistical feature representations through multi-layer residual sensing; Path 2 is as follows: Extract multi-scale spatiotemporal features of the data using multiple parallel branches; each branch adopts a combination of depthwise separable convolution and convolutional block attention mechanism, and each branch has a different convolutional kernel size; the multi-receptive field spatiotemporal features output by each branch are concatenated to obtain the original state feature representation; Adaptive fusion is performed on the multidimensional statistical feature representations and original state feature representations output by the two paths, and the generated fused features are classified to obtain the final diagnostic result.

2. The method according to claim 1, characterized in that, The AUV's operating state is at least one of the following: normal operation, minor propeller damage, severe propeller damage, excessive load, or depth sensor malfunction. When collecting data, the AUV's state variables are recorded at set time intervals, including at least: control signals, pressure, depth, roll attitude angle, pitch attitude angle, yaw attitude angle, angular velocity, acceleration, and battery voltage.

3. The method according to claim 1, characterized in that, The statistical feature descriptor constituting the statistical feature vector refers to at least one of the following: minimum, maximum, mean, median, peak-to-peak value, standard deviation, zero-crossing rate, kurtosis, skewness, root mean square, first difference, Shannon entropy, and peak factor.

4. The method according to claim 1, characterized in that, In path one, a feature attention network is used to process the statistical feature vector. Perform weighted calculations to obtain weighted features. : ; ; in, For attention masking; This indicates element-wise multiplication; Use the Sigmoid activation function; All are learnable weights; To modify the activation function of the linear unit; This is the feature vector of the batch mean.

5. The method according to claim 1, characterized in that, In path two, multiple parallel branches are used to extract multi-scale spatiotemporal features; each branch includes depthwise separable convolutional units, convolutional block attention mechanism modules, and depthwise separable convolutional units arranged in sequence.

6. The method according to claim 1, characterized in that, The depthwise separable convolution includes depthwise convolution and pointwise convolution; the convolutional block attention mechanism includes sequentially connected channel attention modules and spatial attention modules; wherein, Channel attention The calculation formula is: ; Spatial attention The calculation formula is: ; Feature map output by convolutional block attention mechanism The calculation formula is: ; in, Use the Sigmoid activation function; All are learnable weights; To modify the activation function of the linear unit; Indicates average pooling; Indicates max pooling; Input feature map; The feature map after channel attention processing; This represents one-dimensional convolution; This indicates taking the average value; This indicates taking the maximum value; This indicates element-wise multiplication.

7. The method according to claim 1, characterized in that, When performing adaptive fusion, a gated linear unit is introduced as a fusion layer. The information flow between statistical features and original features is controlled by a gating mechanism to generate complementary discriminative feature representations.

8. An AUV fault diagnosis system based on statistical feature enhancement and multi-scale attention, characterized in that, include: The data acquisition module includes multiple sensors mounted on the AUV for acquiring multi-source sensor signals, including motion status and environmental parameters. The control processing board is used to receive sensor signals and output underwater robot status data; The statistical feature enhancement module and the multi-scale attention module are configured in parallel and run in a dual-path parallel manner. in, The statistical feature enhancement module includes a statistical feature extraction unit and a feature attention network arranged in sequence, which are used to extract multidimensional statistical feature representations from the data that are sensitive to the distribution shift of rare faults; The feature attention network consists of a normalization module, a feature attention module, a residual module, and a fully connected layer arranged in sequence. The multi-scale attention module uses multiple parallel branches to extract multi-scale channel and temporal features from the data; in each branch, depthwise separable convolutional units, convolutional block attention modules, and depthwise separable convolutional units are arranged in sequence, and each branch has a different convolutional kernel size; the concatenation unit is connected to each branch to concatenate the output multi-receptive field spatiotemporal features to obtain the original state feature representation; The fusion network module, including gated linear units and a classifier, is used to adaptively fuse and classify the outputs of the two paths and output the final diagnostic result. The statistical feature enhancement module, multi-scale attention module, and fusion network module run on the host computer on the AUV or on the cloud server located at the shore-based monitoring station.

9. A computer device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions that are executed by the at least one processor to cause the at least one processor to perform the AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the AUV fault diagnosis method based on statistical feature enhancement and multi-scale attention as described in any one of claims 1 to 7.