An open set fault identification method for autonomous underwater robots

Abstract

The application provides an open set fault recognition method for an autonomous underwater vehicle (AUV), and relates to the technical field of AUV intelligent fault diagnosis. Firstly, a one-dimensional convolutional neural network (CNN) is used to extract deep discriminative features from the time series data of the multi-dimensional sensors of the AUV, and an uncertainty channel attention mechanism is innovatively introduced to enhance the attention of the model to stable features. Then, an independent one-class support vector machine (OC-SVM) is trained for each known fault category, and the respective decision threshold is calculated. In the test stage, the sample features are input into each OC-SVM for voting. If the number of votes is greater than or equal to 1, it is determined as a known category and further classified by Softmax, otherwise it is determined as an unknown fault. The technical problem that the traditional closed set diagnosis model misjudges the unknown fault as the known fault is solved. The experimental results on the Haizhe public data set show that the unknown fault recognition rate and the known fault classification accuracy of the method are significantly better than those of the existing methods, and the method has a good application prospect.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for autonomous underwater vehicles (AUVs), and in particular to a method for open set fault identification for autonomous underwater robots. Background Technology

[0002] Autonomous underwater vehicles (AUVs) play a vital role in deep-sea resource exploration and environmental monitoring. With increasing demands for deep-sea missions, the complexity of AUV systems and environmental uncertainties are constantly rising, leading to increasingly diverse failure modes. In actual operation, AUVs may encounter novel and unknown faults not encountered during the training phase. Traditional closed-set fault diagnosis methods can only identify known fault categories, easily misclassifying unknown faults as known ones, resulting in incorrect response decisions, mission interruptions, data loss, or even equipment damage.

[0003] Existing open-set fault identification methods are mostly designed for faults in rotating machinery such as bearings, while research on open-set fault diagnosis specifically for AUVs remains limited. Some methods rely on complete fault attribute descriptions, making it difficult to effectively synthesize unknown fault samples for training under real-world conditions. Therefore, there is an urgent need for an intelligent diagnostic method capable of simultaneously identifying both known and unknown faults during AUV operation and accurately classifying known faults. Summary of the Invention

[0004] This invention proposes an open set fault identification method for autonomous underwater vehicles (AUVs), aiming to solve the problem that existing fault diagnosis methods for autonomous underwater vehicles (AUVs) cannot effectively identify unknown faults.

[0005] This invention provides a method for open set fault identification in autonomous underwater vehicles, the method comprising the following steps:

[0006] Step 1: Acquire multi-dimensional sensor time-series data of the autonomous underwater vehicle (AUV) and perform preprocessing;

[0007] Step 2: Construct a one-dimensional convolutional neural network as a feature extractor to extract deep discriminative features from the preprocessed data;

[0008] Step 3: For each known fault category, train an independent single-class support vector machine OC-SVM, and set the corresponding decision threshold based on the decision score distribution of the training samples of that class.

[0009] Step 4: During the testing phase, the test sample is input into the trained one-dimensional convolutional neural network to obtain a deep discriminative feature vector; the deep discriminative feature vector is then input into the single-class support vector machine (OC-SVM) corresponding to all known fault categories. If the decision score output by a certain single-class support vector machine (OC-SVM) is greater than or equal to the decision threshold corresponding to that class, then the test sample is considered to have obtained a vote for that class.

[0010] Step 5: Count the total number of votes for all single-class support vector machines (OC-SVMs): If the total number of votes is greater than or equal to 1, the test sample is determined to belong to the known fault category, and the deep discriminative feature vector is input into the Softmax classifier to output the specific fault category; if the total number of votes is equal to 0, the test sample is determined to belong to the unknown fault category.

[0011] Furthermore, the method for constructing a one-dimensional convolutional neural network as a feature extractor in step 2 to extract deep discriminative features from the preprocessed data is as follows:

[0012] An uncertain channel attention mechanism is introduced after the convolutional layer, and its computation process is as follows:

[0013] Calculate the global average pooling value for each feature channel. and global maximum pool value ;

[0014] Calculate the channel variation range ;

[0015] Channel-based global average pooling value Generate deterministic weights ;

[0016] Based on the channel change amplitude Generate Uncertain Weights and uncertainty score ;

[0017] Calculate the penalty item:

[0018]

[0019] in, As a penalty item, This is an adjustable hyperparameter used to control the strength of the penalty imposed by uncertain attention on the final attention weights. The normalized uncertainty score;

[0020] The final attention weights are obtained by fusion: ;

[0021] Wherein, the uncertainty hyperparameter The value range is [0,1], and its optimal value is determined by grid search;

[0022] The final output is the deep discriminative feature vector after attention weighting. .

[0023] Furthermore, the one-dimensional convolutional neural network in step 2 sequentially includes: a one-dimensional convolutional layer, a batch normalization layer, a max pooling layer, a dropout layer, an uncertain channel attention module, a fully connected layer, and an L2 normalization layer; the kernel size of the one-dimensional convolutional layer is 3 or 6, and the number of output channels is 64 or 128.

[0024] Furthermore, in step 3, for each known fault category, an independent single-class support vector machine (OC-SVM) is trained, and the corresponding decision threshold is set based on the decision score distribution of the training samples for that class. The specific method is as follows:

[0025] Assuming there are K known fault categories, use the feature vectors of each type of training sample extracted in step 2 to train an independent single-class support vector machine OC-SVM for each known fault category.

[0026] Each single-class support vector machine OC-SVM uses a radial basis kernel function to find a hypersphere with the smallest volume in the feature space that contains the vast majority of samples of that class.

[0027] For each single-class support vector machine (OC-SVM), calculate the distribution of its training sample decision scores and set a class-specific decision threshold. .

[0028] Furthermore, the multi-dimensional sensor timing data of the autonomous underwater vehicle (AUV) in step 1 specifically includes: depth, pressure, voltage, roll angle, pitch angle, yaw angle, three-axis acceleration, and three-axis angular velocity, totaling 12 sensing channels.

[0029] The preprocessing includes: unifying the time series length of each sample to 196 time steps by truncating or padding with tail values, and performing Z-score standardization on the data of each sensing channel so that the data distribution of each channel follows a standard normal distribution with a mean of 0 and a variance of 1.

[0030] Further, in step 4, during the testing phase, the test sample is input into the trained one-dimensional convolutional neural network to obtain a deep discriminative feature vector. This deep discriminative feature vector is then input into the single-class support vector machine (OC-SVM) corresponding to all known fault categories. If the decision score output by a certain OC-SVM is greater than or equal to the decision threshold corresponding to that class, then the test sample is considered to have obtained a vote for that class. The specific method is as follows:

[0031] Input the test samples into the trained CNN model to obtain feature vectors. ;

[0032] Input the feature vector respectively In the nth single-class support vector machine OC-SVM, the nth Each single-class support vector machine (OC-SVM) outputs a decision score. ;

[0033] like If a class casts a vote for a test sample, it is denoted as . ;otherwise .

[0034] Further, in step 5, the total number of votes for all single-class support vector machines (OC-SVMs) is counted: if the total number of votes is greater than or equal to 1, the test sample is determined to belong to the known fault category, and the deep discriminative feature vector is input into the Softmax classifier to output the specific fault category; if the total number of votes is equal to 0, the method for determining that the test sample belongs to the unknown fault is as follows:

[0035] Calculate the total number of votes:

[0036]

[0037] like If the sample is classified as a known category, its feature vector is fed into the Softmax classifier, and the category with the highest probability is output as the final diagnosis result.

[0038] like If the sample is identified as an unknown fault, the system will trigger an unknown fault warning, prompting manual review or data recording.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. Strong open set recognition capability: Through the CNN-OC-SVM combined architecture, it can effectively distinguish between known faults and unknown faults, avoiding the misjudgment of unknown faults by traditional closed set methods.

[0041] 2. More robust feature representation: The introduction of an uncertain channel attention mechanism enables the model to adaptively suppress unstable feature channels, enhancing its robustness to noise and interference.

[0042] 3. No need for unknown fault samples: OC-SVM can be trained with only known class samples, making it suitable for scenarios where unknown fault samples are difficult to obtain in practice.

[0043] 4. High diagnostic accuracy: Experiments on the "Haizhe" AUV public dataset show that the method of this invention is significantly better than the comparative algorithms such as Softmax and OpenMax in terms of weighted accuracy and F1 score.

[0044] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0045] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0046] Figure 1 This is a flowchart of the overall method of the present invention;

[0047] Figure 2 This is a schematic diagram of the CNN feature extractor structure in this invention;

[0048] Figure 3 This is a flowchart of the training phase of the present invention;

[0049] Figure 4 This is a flowchart of the testing phase of the present invention;

[0050] Figure 5 This is a schematic diagram of the overall structure of the model. Detailed Implementation

[0051] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0052] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0053] This invention provides a method for open set fault identification in autonomous underwater vehicles, the method comprising the following steps:

[0054] Step 1: Acquire multi-dimensional sensor time-series data of the autonomous underwater vehicle (AUV) and perform preprocessing;

[0055] Step 2: Construct a one-dimensional convolutional neural network as a feature extractor to extract deep discriminative features from the preprocessed data;

[0056] Step 3: For each known fault category, train an independent single-class support vector machine OC-SVM, and set the corresponding decision threshold based on the decision score distribution of the training samples of that class.

[0057] Step 4: During the testing phase, the test sample is input into the trained one-dimensional convolutional neural network to obtain a feature vector; the feature vector is then input into the single-class support vector machine (OC-SVM) corresponding to all known fault categories. If the decision score output by a certain single-class support vector machine (OC-SVM) is greater than or equal to the decision threshold corresponding to that class, then the test sample is considered to have obtained a vote for that class.

[0058] Step 5: Count the total number of votes for all single-class support vector machines (OC-SVMs): If the total number of votes is greater than or equal to 1, the test sample is determined to belong to the known fault category, and the deep discriminative feature vector is input into the Softmax classifier to output the specific fault category; if the total number of votes is equal to 0, the test sample is determined to belong to the unknown fault category.

[0059] Optionally, the method for constructing a one-dimensional convolutional neural network as a feature extractor in step 2 to extract deep discriminative features from the preprocessed data is as follows:

[0060] The calculation process is as follows:

[0061] Calculate the global average pooling value for each feature channel. and global maximum pool value ;

[0062] Calculate the channel variation range ;

[0063] Channel-based global average pooling value Generate deterministic weights ;

[0064] Based on the channel change amplitude Generate Uncertain Weights and uncertainty score ;

[0065] Calculate the penalty item:

[0066]

[0067] in, As a penalty item, This is an adjustable hyperparameter used to control the strength of the penalty imposed by uncertain attention on the final attention weights. The normalized uncertainty score;

[0068] The final attention weights are obtained by fusion: ;

[0069] Wherein, the uncertainty hyperparameter The value range is [0,1], and its optimal value is determined by grid search;

[0070] The final output is the deep discriminative feature vector after attention weighting. .

[0071] In this way, an uncertain channel attention mechanism is introduced after the global max pooling layer. By penalizing feature channels with large fluctuations, the model focuses more on stable and reliable failure modes. This enables the model to accurately identify known failures and effectively reject unknown failures in an open environment, thereby improving the robustness and safety of the model.

[0072] Optionally, the one-dimensional convolutional neural network in step 2 includes, in sequence: a one-dimensional convolutional layer, a batch normalization layer, a max pooling layer, a dropout layer, an uncertain channel attention module, a fully connected layer, and an L2 normalization layer; the kernel size of the one-dimensional convolutional layer is 3 or 6, and the number of output channels is 64 or 128.

[0073] Optionally, in step 3, for each known fault category, an independent single-class support vector machine (OC-SVM) is trained, and a corresponding decision threshold is set based on the decision score distribution of the training samples for that class; the specific method is as follows:

[0074] Assuming there are K known fault categories, use the feature vectors of each type of training sample extracted in step 2 to train an independent single-class support vector machine OC-SVM for each known fault category.

[0075] Each single-class support vector machine OC-SVM uses a radial basis kernel function to find a hypersphere with the smallest volume in the feature space that contains the vast majority of samples of that class.

[0076] For each single-class support vector machine (OC-SVM), calculate the distribution of its training sample decision scores and set a class-specific decision threshold. .

[0077] Each class-trained OC-SVM does not learn the inter-class discrimination boundary. Instead, it constructs a compact minimum hypersphere for each known class to describe the distribution area of ​​normal samples of that class in the feature space. Samples not in this area will be judged as "unknown", thus avoiding the misclassification problem of traditional multi-classification models forcibly classifying unknown samples into a certain known class.

[0078] Optionally, the multi-dimensional sensor timing data of the autonomous underwater vehicle (AUV) in step 1 specifically includes: depth, pressure, voltage, roll angle, pitch angle, yaw angle, three-axis acceleration, and three-axis angular velocity, totaling 12 sensing channels.

[0079] The preprocessing includes: unifying the time series length of each sample to 196 time steps by truncating or padding with tail values, and performing Z-score standardization on the data of each sensing channel so that the data distribution of each channel follows a standard normal distribution with a mean of 0 and a variance of 1.

[0080] Optionally, in step 4, during the testing phase, the test sample is input into the trained one-dimensional convolutional neural network to obtain a feature vector; this feature vector is then input into the single-class support vector machine (OC-SVM) corresponding to all known fault categories. If the decision score output by a certain single-class support vector machine (OC-SVM) is greater than or equal to the decision threshold corresponding to that class, then the test sample is considered to have obtained a vote for that class. The specific method is as follows:

[0081] Input the test samples into the trained CNN model to obtain feature vectors. ;

[0082] Input the feature vector respectively In the nth single-class support vector machine OC-SVM, the nth Each single-class support vector machine (OC-SVM) outputs a decision score. ;

[0083] like If a class casts a vote for a test sample, it is denoted as . ;otherwise .

[0084] Optionally, in step 5, the total number of votes for all single-class support vector machines (OC-SVMs) is counted: if the total number of votes is greater than or equal to 1, the test sample is determined to belong to the known fault category, and the deep discriminative feature vector is input into the Softmax classifier to output the specific fault category; if the total number of votes is equal to 0, the method for determining that the test sample belongs to the unknown fault is as follows:

[0085] Calculate the total number of votes:

[0086]

[0087] like If the sample is classified as a known category, its feature vector is fed into the Softmax classifier, and the category with the highest probability is output as the final diagnosis result.

[0088] like If the sample is identified as an unknown fault, the system will trigger an unknown fault warning, prompting manual review or data recording.

[0089] Figure 1 The complete flowchart of the present invention is shown, including core steps such as dataset loading and processing, training and testing;

[0090] Figure 2 The diagram shows the CNN network structure used in this model, which mainly includes convolutional layers, max pooling layers, global pooling layers, and fully connected layers. Innovatively, an attention layer is introduced after the global pooling layer to improve the robustness and security of the system.

[0091] Figure 3 The workflow of the training phase is demonstrated. First, the convolutional neural network is trained end-to-end using sample data of known categories, enabling it to extract highly discriminative feature vectors from the raw sensor data. Then, the extracted features are categorized, and a separate OC-SVM is trained for each known fault category. Each OC-SVM independently learns the distribution boundary of that category of samples in the feature space and calculates a category-specific decision threshold based on the decision score distribution of that category of samples in the training set. Finally, the training phase outputs a trained convolutional neural network feature extractor, multiple OC-SVM models, and the corresponding discriminative threshold for each category, for use in the subsequent testing phase.

[0092] Figure 4 The workflow during the testing phase is demonstrated. First, test samples are input into a pre-trained CNN to extract their corresponding feature vectors. Then, these feature vectors are sequentially input into each known-class OC-SVM. Each OC-SVM outputs a decision score, representing the signed distance of the sample to its hypersphere boundary. Each decision score is compared to a preset threshold for its corresponding class: if a class's decision score is greater than or equal to its threshold, the class is considered to accept the sample and cast a vote; otherwise, no vote is cast. If all classes do not vote (i.e., the number of votes is 0), the sample is classified as an unknown fault; if at least one class votes, it is classified as a known fault, and the class with the highest decision score is selected as the final fault type output.

[0093] Figure 5 The complete process of the algorithm of this invention is demonstrated, and the core work of each step is introduced.

[0094] Example

[0095] This embodiment uses the "Haizhe" AUV public dataset from Zhejiang University for verification. This dataset contains five operating states: normal, depth sensor constant bias fault, minor propeller damage, severe propeller damage, and increased load. Each state is treated as an "unknown fault" and verified using the leave-one-out method.

[0096] 1. Data preprocessing;

[0097] Twelve sensor channels were selected (excluding timestamps and PWM control signals), and each sample was truncated or padded to 196 time steps. Z-score normalization was used to ensure that the data for each channel conformed to a standard normal distribution. The training set contained all samples from four known classes, and the test set contained 20% of the samples from the four known classes (excluding those from the training set) and 50% of the samples from one unknown class.

[0098] 2. CNN feature extractor parameter settings;

[0099] Convolutional layer: 1 convolutional layer, kernel size = 3, number of output channels = 64, activation function = ReLU.

[0100] Batch normalization layer: used to accelerate convergence.

[0101] Max pooling layer: pooling window = 2, stride = 2.

[0102] Dropout layer: Drop rate = 0.5.

[0103] Attention module: Implemented according to the uncertainty channel attention mechanism described above, with hyperparameter α determined to be 0.6 through grid search.

[0104] Fully connected layer: Maps the pooled features to a 128-dimensional discriminative space, and then performs Dropout and L2 normalization again.

[0105] 3. Multi-path OC-SVM training;

[0106] OC-SVMs were trained for each of the four known categories using the RBF kernel function, with the kernel parameter γ determined through cross-validation. The decision threshold for each OC-SVM was set to the 5th percentile of the decision scores from the training samples.

[0107] 4. Compare the experimental results;

[0108] Compare this method (Model 1) with two baseline methods:

[0109] CNN-Softmax (Model 2): ​​No OC-SVM voting, direct Softmax classification;

[0110] CNN-OpenMax (Model 3): Uses extremum theory to correct the Softmax output and introduces unknown class probabilities.

[0111] In the five leave-one-out experiments, the average weighted accuracy of our method (CNN-OC-SVM) reached 0.84, and the average F1 score reached 0.77; while CNN-Softmax achieved 0.63 and 0, and CNN-OpenMax achieved 0.73 and 0.49, as shown in Table 1.

[0112] Table 1 shows the performance comparison of each model in leave-one-out validation.

[0113]

[0114] 5. Ablation test results;

[0115] After removing the attention mechanism, the model's weighted accuracy decreased by approximately 5 to 8 percentage points;

[0116] After replacing CNN with BiLSTM or LSTM, the unknown fault recognition rate dropped significantly (from 80% to 40%~20%), proving that CNN is more advantageous in extracting AUV temporal features.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for open-set fault identification in autonomous underwater vehicles, characterized in that, The method includes the following steps: Step 1: Acquire multi-dimensional sensor time-series data of the autonomous underwater vehicle (AUV) and perform preprocessing; Step 2: Construct a one-dimensional convolutional neural network as a feature extractor to extract deep discriminative features from the preprocessed data; Step 3: For each known fault category, train an independent single-class support vector machine OC-SVM, and set the corresponding decision threshold based on the decision score distribution of the training samples of that class. Step 4: During the testing phase, the test sample is input into the trained one-dimensional convolutional neural network to obtain a deep discriminative feature vector; the deep discriminative feature vector is then input into the single-class support vector machine (OC-SVM) corresponding to all known fault categories. If the decision score output by a certain single-class support vector machine (OC-SVM) is greater than or equal to the decision threshold corresponding to that class, then the test sample is considered to have obtained a vote for that class. Step 5: Count the total number of votes for all single-class support vector machines (OC-SVMs): If the total number of votes is greater than or equal to 1, the test sample is determined to belong to the known fault category, and the deep discriminative feature vector is input into the Softmax classifier to output the specific fault category; if the total number of votes is equal to 0, the test sample is determined to belong to the unknown fault category.

2. The method for open set fault identification of autonomous underwater robots according to claim 1, characterized in that, In step 2, a one-dimensional convolutional neural network is constructed as a feature extractor to extract deep discriminative features from the preprocessed data. The specific method is as follows: An uncertain channel attention mechanism is introduced after the convolutional layer, and its computation process is as follows: Calculate the global average pooling value for each feature channel. and global maximum pool value ; Calculate the channel variation range ; Channel-based global average pooling value Generate deterministic weights ; Based on channel change amplitude Generate Uncertainty Weights and uncertainty score ; Calculate the penalty item: in, As a penalty item, This is an adjustable hyperparameter used to control the strength of the penalty imposed by uncertain attention on the final attention weights. The normalized uncertainty score; The final attention weights are obtained by fusion: ; Wherein, the uncertainty hyperparameter The value range is [0,1], and its optimal value is determined by grid search; The final output is the deep discriminative feature vector after attention weighting. .

3. The method for open set fault identification of autonomous underwater robots according to claim 1, characterized in that, The one-dimensional convolutional neural network in step 2 includes, in sequence: a one-dimensional convolutional layer, a batch normalization layer, a max pooling layer, a dropout layer, an uncertain channel attention module, a fully connected layer, and an L2 normalization layer; the kernel size of the one-dimensional convolutional layer is 3 or 6, and the number of output channels is 64 or 128.

4. The method for open set fault identification of autonomous underwater robots according to claim 1, characterized in that, In step 3, for each known fault category, an independent single-class support vector machine (OC-SVM) is trained, and the corresponding decision threshold is set based on the decision score distribution of the training samples for that class. The specific method is as follows: Assuming there are K known fault categories, use the feature vectors of each type of training sample extracted in step 2 to train an independent single-class support vector machine OC-SVM for each known fault category. Each single-class support vector machine OC-SVM uses a radial basis kernel function to find a hypersphere with the smallest volume in the feature space that contains most of the samples of that class. For each single-class support vector machine (OC-SVM), calculate the distribution of its training sample decision scores and set a class-specific decision threshold. .

5. The method for open set fault identification of autonomous underwater robots according to claim 1, characterized in that, The multi-dimensional sensor timing data of the autonomous underwater vehicle (AUV) in step 1 specifically includes: depth, pressure, voltage, roll angle, pitch angle, yaw angle, three-axis acceleration, and three-axis angular velocity, totaling 12 sensing channels. The preprocessing includes: unifying the time series length of each sample to 196 time steps by truncating or padding with tail values, and performing Z-score standardization on the data of each sensing channel so that the data distribution of each channel follows a standard normal distribution with a mean of 0 and a variance of 1.

6. The method for open set fault identification of autonomous underwater robots according to claim 1, characterized in that, In step 4, during the testing phase, the test sample is input into the trained one-dimensional convolutional neural network to obtain a deep discriminative feature vector. This feature vector is then input into the single-class support vector machine (OC-SVM) corresponding to each known fault category. If the decision score output by a certain OC-SVM is greater than or equal to the decision threshold corresponding to that class, the test sample is considered to have received a vote from that class. The specific method is as follows: Input the test samples into the trained CNN model to obtain feature vectors. ; Input the feature vector respectively In the nth single-class support vector machine OC-SVM, the nth Each single-class support vector machine (OC-SVM) outputs a decision score. ; like If a class casts a vote for a test sample, it is denoted as . ;otherwise .

7. The method for open set fault identification of autonomous underwater robots according to claim 1, characterized in that, In step 5, the total number of votes for all single-class support vector machines (OC-SVMs) is counted: if the total number of votes is greater than or equal to 1, the test sample is determined to belong to the known fault category, and the deep discriminative feature vector is input into the Softmax classifier to output the specific fault category; if the total number of votes is equal to 0, the test sample is determined to belong to the unknown fault category. Calculate the total number of votes: like If the sample is classified as a known category, its feature vector is fed into the Softmax classifier, and the category with the highest probability is output as the final diagnosis result. like If the sample is identified as an unknown fault, the system will trigger an unknown fault warning, prompting manual review or data recording.

Images