A fault diagnosis method for underwater autonomous underwater vehicles based on search space

Abstract

The present invention discloses a method for fault diagnosis of an underwater autonomous submersible based on a search space. First, an oscillation signal of the underwater autonomous submersible is obtained in an underwater experiment and used as the input of a search network model for forward propagation. During this process, the activation state of the ReLU layer of the network model is recorded, and then the activation matrix, the kernel matrix and the final score of the network model are calculated, and the network model with the highest score is recorded as the fault diagnosis model; finally, the fault diagnosis model is used to perform real-time fault diagnosis on the underwater autonomous submersible.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and more specifically, relates to a fault diagnosis method for an underwater autonomous submersible based on a search space. Background Art

[0002] Fault diagnosis involves extracting features from information captured by device sensors to determine the fault type. Autonomous underwater vehicles (AUVs) are intelligent devices capable of independent underwater operations, reducing the cost and risk of marine operations. However, AUVs carry a certain risk of failure during their missions, necessitating the implementation of fault diagnosis solutions for these intelligent devices. Due to the complexity of sensor data and the difficulty in extracting relevant features, researchers often opt for diagnostic solutions based on machine learning or deep learning.

[0003] Machine learning has introduced diagnostic solutions such as hierarchical fault tree models and neural network models. These machine learning-based diagnostic methods are interpretable, and some solutions can effectively extract feature information from sensor oscillation signals, achieving good diagnostic rates in experiments with certain application scenarios. However, traditional machine learning-based fault diagnosis relies too heavily on the accuracy of extracted features, resulting in insufficient accuracy in the complex scenarios of actual underwater applications.

[0004] Deep learning, a specialized machine learning technique, possesses powerful nonlinear fitting capabilities and can often extract higher-level features from data. Numerous deep learning methods have been applied to fault diagnosis of autonomous underwater vehicles. These methods, thanks to sophisticated network design, have achieved excellent diagnostic results. These diagnostic solutions often exhibit enhanced robustness, with some neural network architectures even performing well in noisy environments or with partial loss of sensor data. However, most neural network architectures require manual design based on the researcher's requirements. The performance of the network is evaluated through training, and poorly performing architectures are discarded. Therefore, manually designing sophisticated network architectures is difficult and costly.

[0005] Neural architecture search (NAS) can automatically design network structures. Using specific search strategies, this technology can find a network structure that meets accuracy requirements within a defined search space. While NAS technology achieves automatic network structure design and achieves high accuracy in fault diagnosis, the enormous computational effort, time cost, GPU usage, and inherent complexity of the designed network structure limit its further application in fault diagnosis.

[0006] The above analysis demonstrates that existing machine learning-based diagnostic solutions struggle to process complex sensor information, while deep learning-based solutions often struggle to avoid the significant cost of manually designing diagnostic networks. While existing neural architecture search methods can automatically design sophisticated diagnostic network models, their computational and time costs remain high. Therefore, further research on the automated design of underwater vehicle fault diagnosis models using neural network architecture search is essential. Summary of the Invention

[0007] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide an underwater autonomous vehicle fault diagnosis method based on search space, which characterizes the performance of untrained candidate networks through a scoring algorithm. It has the advantages of fast search speed, good model diagnosis performance, low computational complexity, and small GPU resource occupation.

[0008] To achieve the above-mentioned object of the invention, the present invention provides a fault diagnosis method for an underwater autonomous underwater vehicle based on a search space, characterized by comprising the following steps:

[0009] (1) Construct a training dataset;

[0010] (1.1) Under different fault conditions, the oscillation signal of the underwater autonomous underwater vehicle at each sampling time is collected. The oscillation signal collected at the t-th sampling time under the k-th fault condition is recorded as x k,t , k=1,2,…,K, K represents the number of fault categories of the underwater autonomous underwater vehicle, t=1,2,…,T, T is the total number of samples;

[0011] (1.2), the collected oscillation signal x k,t Add fault tag y k , fault label value pair fault category number k;

[0012] (1.3), the oscillation signal x at a single moment k,t and the corresponding fault label y k As a set of training data, thus establishing a training data set X;

[0013] X={X1,X2,…,X k ,…,X K}

[0014] X k ={(x k,1 ,y k ),(x k,2 ,y k ),…,(x k,t ,y k ),…,(x k,T ,y k )}

[0015] (2) Neural Network Architecture Search

[0016] (2.1), select NAS-Bench-201 as the search space, denoted as β;

[0017] (2.2),Search candidate networks;

[0018] Get several candidate networks from the search space:

[0019] B~π(β)

[0020] Among them, β is the search space, π is the search strategy, and B represents the set of candidate networks sampled using the search strategy π;

[0021] (2.3),Evaluate the performance of each candidate network;

[0022] (2.3.1) Set the sample batch size for training the candidate network to b; shuffle the order of the data samples in the training dataset X, and then divide the training dataset X into N batches:

[0023] X={X1,X2,…,X λ ,…,X N}

[0024] (2.3.2), take X1 as the input of the first candidate network, X2,…,X λ ,…,X N In turn, they are used as inputs for the candidate networks obtained subsequently;

[0025] (2.3.2) Calculate the output feature map of each ReLU function layer in a candidate network, where the output feature map of the i-th ReLU function layer is denoted as O i , i=1,2,…,n, n represents the number of ReLU function layers contained in the candidate network, O i It is a 4-dimensional tensor, the 4 dimensions include: batch size b, feature map depth c, feature map length and width (w, h);

[0026] (2.3.3), construct activation matrix;

[0027] Each ReLU function layer outputs the feature map O i Transformed into a two-dimensional matrix Q of size b×cwh i , where O i Middle pixel o i [j,1,1,1] to o i [j,c,w,h] and Q i Element q i (j,1) to q iThe conversion relationship of (j,cwh) is as follows:

[0028] Set variable j, 1≤j≤b;

[0029] Pixel o i [j,1,1,1] to o i [j,1,1,h] corresponds to element q i (j,1) to q i (j,h);

[0030] Pixel o i [j,1,2,1] to o i [j,1,2,h] corresponds to element q i (j,h+1) to q i (j,2h);

[0031] Pixel o i [j,1,w,1] to o i [j,1,w,h] corresponds to element q i (j,(w-1)h+1) to q i (j,wh);

[0032] Pixel o i [j,2,w,1] to o i [j,2,w,h] corresponds to element q i (j,wh+(w-1)h+1) to q i (j,2wh);

[0033] Pixel o i [j,c,w,1] to o i [j,c,w,h] corresponds to element q i (j,(c-1)wh+(w-1)h+1) to q i (j,cwh);

[0034] Construct a two-dimensional matrix Q i The activation matrix A of the same size i , for Q i For any element in A, if the element of A is greater than 0, then i The element at the corresponding position in is set to 1, otherwise, it is set to 0;

[0035] (2.3.4) Calculate the kernel matrix K;

[0036] According to the activation matrix A i Calculate the matrix

[0037]

[0038] Among them, a i (j,cwh) represents A i The element value in row j and column cwh;

[0039] According to A i and Calculate the matrix K i :

[0040]

[0041] Wherein, the superscript T indicates transposition;

[0042] Calculate the kernel matrix K:

[0043]

[0044] (2.3.5) Calculate the performance indicators of each candidate network;

[0045] S τ =lg|K|,τ∈B

[0046] (2.4) Select the fault diagnosis model;

[0047] Set the number of searches M; repeat steps (2.2) to (2.3), perform performance evaluation on the candidate networks in each round of search, and then select the maximum value based on the performance indicators of all candidate networks. Record the network model corresponding to the maximum value as the fault diagnosis model;

[0048] (3) Training fault diagnosis model;

[0049] The shuffled training data set X is input into the fault diagnosis model in batches, and the fault diagnosis model is trained repeatedly until it converges.

[0050] (4) Real-time fault diagnosis;

[0051] The oscillation signal of the underwater autonomous underwater vehicle at the current moment is collected in real time, and then the oscillation signal is input into the fault diagnosis model to output the fault type in real time.

[0052] The object of the invention of the present invention is achieved like this:

[0053] The present invention is a fault diagnosis method for an underwater autonomous submersible based on a search space. First, an oscillation signal of the underwater autonomous submersible is obtained in an underwater experiment and used as the input of a search network model for forward propagation. During this process, the activation state of the ReLU layer of the network model is recorded, and then the activation matrix, the kernel matrix and the final score of the network model are calculated, and the network model with the highest score is recorded as the fault diagnosis model; finally, the fault diagnosis model is used to perform real-time fault diagnosis on the underwater autonomous submersible.

[0054] At the same time, the underwater autonomous submersible fault diagnosis method based on search space of the present invention also has the following beneficial effects:

[0055] (1) The present invention can realize the automated design of neural network diagnostic models, overcoming the difficulty of manually designing sophisticated diagnostic networks;

[0056] (2) The present invention is experimentally verified on a real underwater vehicle dataset. The experimental results show that the diagnostic architecture automatically designed by the present invention performs better than the manually designed diagnostic architecture;

[0057] (3) The special scoring mechanism adopted by the present invention can replace the expensive inner loop training step in the traditional neural architecture search, greatly reducing the cost of fault diagnosis architecture search, so that a good-performance underwater submersible fault diagnosis architecture can be automatically designed in just a few seconds on a single GPU. Compared with the traditional neural network architecture search method, the present invention takes much less time to search for a good-performance fault diagnosis architecture under the same GPU resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a fault diagnosis method for an underwater autonomous submersible vehicle according to the present invention;

[0059] Figure 2 The embodiment uses an underwater submersible structure;

[0060] Figure 3 It is the health status of the submersible that is set;

[0061] Figure 4 is a candidate architecture unit structure diagram;

[0062] Figure 5 It is a bar chart of diagnostic effect;

[0063] Figure 6 This is a comparison chart of diagnostic effects. DETAILED DESCRIPTION

[0064] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0065] Example

[0066] In this embodiment, the neural network architecture search, network training and testing are all performed in the PyTorch environment, using the NVIDIA GTX3070Ti GPU and Intel i9-12900H CPU. Figure 1 As shown, the present invention provides a method for fault diagnosis of an autonomous underwater vehicle based on a search space, comprising the following steps:

[0067] S1. Build training data set and test set;

[0068] S1.1. Collect the oscillation signal of the underwater autonomous underwater vehicle at each sampling time under different fault conditions. The oscillation signal collected at the t-th sampling time under the k-th fault condition is recorded as x k,t , k=1,2,…,K, K represents the number of fault categories of the underwater autonomous underwater vehicle, t=1,2,…,T, T is the total number of samples;

[0069] S1.2, the collected oscillation signal x k,t Add fault tag y k , fault label value pair fault category number k;

[0070] S1.3, the oscillation signal x at a single moment k,t and the corresponding fault label y k As a set of training data, thus establishing a training data set X;

[0071] X={X1,X2,…,X k ,…,X K}

[0072] X k ={(x k,1 ,y k ),(x k,2 ,y k ),…,(x k,t ,y k ),…,(x k,T ,y k )}

[0073] In this embodiment, a small four-rotor underwater submersible is selected. The structure of the underwater submersible is as follows: Figure 2 As shown, it is equipped with 4 motors, 4 propellers, 4 electronic speed controllers, 1 depth sensor and other structures.

[0074] like Figure 3As shown, five states are set for the underwater submersible, including normal state, abnormal load fault, sensor fault, severe propeller fault, and minor propeller fault. Simulation experiments are conducted under different health states, and the oscillation signals obtained by the sensor and the corresponding fault labels are recorded. The simulation experiments are repeated multiple times for each health state, and the data collection for each simulation experiment lasts 10-20 seconds. A total of 1225 data samples are collected in the dataset, with approximately 250 data samples for each operating state.

[0075] After the dataset is successfully collected, random stratified sampling is performed, and the sample space of the selected test set is 20% of the total dataset. The test set will be used to verify the performance of the diagnostic model, and the remaining samples will be used as the training set to train the neural network.

[0076] S2. Neural Network Architecture Search

[0077] S2.1. Select a search space that meets the requirements. In this embodiment, NAS-Bench-201 is used as the search space, denoted as β.

[0078] The macro skeleton structure of NAS-Bench-201 includes input layer, convolution layer, neural unit module, residual module, pooling layer, and fully connected layer. The neural unit module structure is as follows: Figure 4 As shown, the four nodes A, B, C, and D are connected by six operations. Each connection can be any of the five operations: zeroing, skip connection, 1×1 convolution, 3×3 convolution, and average pooling. Therefore, there are a total of 15,625 candidate neural unit modules.

[0079] S2.2, search candidate networks;

[0080] Get several candidate networks from the search space:

[0081] B~π(β)

[0082] Among them, β is the search space, π is the search strategy, and B represents the set of candidate networks sampled using the search strategy π;

[0083] In this embodiment, 100 candidate networks are extracted from the search space each time, and the number of searches is set to 500;

[0084] S2.3. Evaluate the performance of each candidate network;

[0085] S2.3.1. Set the sample batch size for training the candidate network to b; shuffle the order of the data samples in the training dataset X, and then divide the training dataset X into N batches:

[0086] X={X1,X2,…,X λ ,…,X N}

[0087] In this embodiment, the sample batch size b is set to 8;

[0088] S2.3.2, take X1 as the input of the first candidate network, X2,…,X λ ,…,X N In turn, they are used as inputs for the candidate networks obtained subsequently;

[0089] In this embodiment, X1 is used as the input of the first candidate network, X2 is used as the input of the second candidate network, and so on. When the batch number N is less than the number of candidate networks, the batches will be re-based on X1, X2, ..., X λ ,…,X N The order continues to provide input for subsequent candidate networks;

[0090] S2.3.2. Calculate the output feature map of each ReLU function layer in a candidate network, where the output feature map of the i-th ReLU function layer is denoted as O i , i=1,2,…,n, n represents the number of ReLU function layers contained in the candidate network, O i It is a 4-dimensional tensor, the 4 dimensions include: batch size b, feature map depth c, feature map length and width (w, h);

[0091] S2.3.3. Construct activation matrix;

[0092] Each ReLU function layer outputs the feature map O i Transformed into a two-dimensional matrix Q of size b×cwh i , where O i Middle pixel o i [j,1,1,1] to o i [j,c,w,h] and Q i Element q i (j,1) to q i The conversion relationship of (j,cwh) is as follows:

[0093] Set variable j, 1≤j≤b;

[0094] Pixel o i [j,1,1,1] to o i [j,1,1,h] corresponds to element q i (j,1) to q i (j,h);

[0095] Pixel o i [j,1,2,1] to o i [j,1,2,h] corresponds to element q i (j,h+1) to qi (j,2h);

[0096] Pixel o i [j,1,w,1] to o i [j,1,w,h] corresponds to element q i (j,(w-1)h+1) to q i (j,wh);

[0097] Pixel o i [j,2,w,1] to o i [j,2,w,h] corresponds to element q i (j,wh+(w-1)h+1) to q i (j,2wh);

[0098] Pixel o i [j,c,w,1] to o i [j,c,w,h] corresponds to element q i (j,(c-1)wh+(w-1)h+1) to q i (j,cwh);

[0099] Construct a two-dimensional matrix Q i The activation matrix A of the same size i , for Q i For any element in A, if the element of A is greater than 0, then i The element at the corresponding position in is set to 1, otherwise, it is set to 0;

[0100] S2.3.4. Calculate the kernel matrix K;

[0101] According to the activation matrix A i Calculate the matrix

[0102]

[0103] Among them, a i (j,cwh) represents A i The element value in row j and column cwh;

[0104] According to A i and Calculate the matrix K i :

[0105]

[0106] Wherein, the superscript T indicates transposition;

[0107] Calculate the kernel matrix K:

[0108]

[0109] S2.3.5. Calculate the performance indicators of each candidate network;

[0110] S τ =lg|K|,τ∈B

[0111] S2.4, select the fault diagnosis model;

[0112] Repeat steps S2.2 to S2.3 to evaluate the performance of candidate networks in each round of search. When the number of searches M reaches 200, select the maximum value based on the performance indicators of all candidate networks and record the network model corresponding to the maximum value as the fault diagnosis model.

[0113] S3, training fault diagnosis model;

[0114] The shuffled training data set X is input into the fault diagnosis model in batches, and the fault diagnosis model is trained repeatedly until it converges.

[0115] S4. Verify the diagnostic effect of the fault diagnosis model;

[0116] The sample data in the test set is input into the fault diagnosis model to output the fault type.

[0117] In this example, the performance of the diagnostic model was evaluated using accuracy (ACC), F1-Score (F1-Score), and area under the curve (AUC). Figure 5 The diagnostic performance of the proposed method is demonstrated, with a model diagnostic accuracy of 98.37%, an F1-Score of 98.13%, and an AUC of 99.93%. This demonstrates that the proposed automated diagnostic model search framework can obtain effective diagnostic models from the search space. Furthermore, searching for the optimal architecture takes only 25 seconds per epoch, and the entire NAS process takes only 3.47 hours, improving time efficiency.

[0118] In addition, the automatically designed fault diagnosis model of the present invention is compared with the manually designed fault diagnosis model. Figure 6This is a radar chart comparing the present invention with other diagnostic methods. TENAS-FD represents the present invention's method, while the other methods are manually designed diagnostic models. The artificial neural network fault diagnosis models involved in the comparison include Multi-ANN, Wen-CNN, Sequential Convolutional Neural Network (Seq-CNN), Spiking Recurrent Attention Network (SRA-SNN), Multi-GRU, Genetic Algorithm-Enabled Ensemble Learning (GASEL), and Multi-LSTM. TENAS-FD outperformed seven manually designed experimental network architectures. TENAS-FD achieved an ACC of 98.37%, an F1-Score of 98.13%, and an AUC of 99.93%, respectively. Multi-ANN achieved an ACC of 87.74%, an F1-Score of 84.21%, and an AUC of 94.25%, respectively. Wen-CNN achieved an ACC of 95.14%, an F1-Score of 95.78%, and an AUC of 99.21%, respectively. Seq-CNN achieved 97.01% ACC, 97.57% F1-Score, and 99.48% AUC, respectively. SRA-SNN achieved 97.41% ACC, 97.84% F1-Score, and 99.75% AUC, respectively. Multi-GRU achieved 94.12% ACC, 92.71% F1-Score, and 98.15% AUC, respectively. GASEL achieved 97.96% ACC, 98.57% F1-Score, and 99.34% AUC, respectively. Multi-LSTM achieved 93.17% ACC, 91.56% F1-Score, and 97.14% AUC, respectively. This shows that the present invention can effectively obtain a good-performing diagnostic network in the search space and automatically design a diagnostic model for a submarine. This invention not only automates the design of diagnostic models for underwater autonomous vehicles, but also produces better diagnostic models than manual designs, and the network design process is much less time-consuming. The proposed TENAS-FD significantly reduces the cost and difficulty of designing high-performance diagnostic models and has the potential to be applied to the design of diagnostic networks for underwater vehicles.

[0119] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A fault diagnosis method for an underwater autonomous underwater vehicle based on a search space, characterized in that: The following steps are involved: (1) Construct a training dataset; (1.1) Under different fault conditions, the oscillation signal of the underwater autonomous underwater vehicle at each sampling time is collected. The oscillation signal collected at the t-th sampling time under the k-th fault condition is recorded as x k,t , k=1,2,…,K, K represents the number of fault categories of the underwater autonomous underwater vehicle, t=1,2,…,T, T is the total number of samples; (1.2), the collected oscillation signal x k,t Add fault label y k , fault label value pair fault category number k; (1.3), the oscillation signal x at a single moment k,t and the corresponding fault label y k As a set of training data, thus establishing a training data set X; X={X1,X2,…,X k ,…,X K } X k ={(x k,1 ,and k ),(x k,2 ,and k ),…,(x k,t ,and k ),…,(x k,T ,and k )} (2) Neural Network Architecture Search (2.1), select NAS-Bench-201 as the search space, denoted as β; (2.2),Search candidate networks; Get several candidate networks from the search space: B～π(β) Among them, β is the search space, π is the search strategy, and B represents the set of candidate networks sampled using the search strategy π; (2.3),Evaluate the performance of each candidate network; (2.3.1) Set the sample batch size for training the candidate network to b; shuffle the order of the data samples in the training dataset X, and then divide the training dataset X into N batches: X={X1,X2,…,X λ ,…,X N } (2.3.2), take X1 as the input of the first candidate network, X2,…,X λ ,…,X N In turn, they are used as inputs for the candidate networks obtained subsequently; (2.3.2) Calculate the output feature map of each ReLU function layer in a candidate network, where the output feature map of the i-th ReLU function layer is denoted as O i , i=1,2,…,n, n represents the number of ReLU function layers contained in the candidate network, O i It is a 4-dimensional tensor, the 4 dimensions include: batch size b, feature map depth c, feature map length and width (w, h); (2.3.3), construct activation matrix; Each ReLU function layer outputs the feature map O i Transformed into a two-dimensional matrix Q of size b×cwh i , where O i Middle pixel o i [j,1,1,1] to o i [j,c,w,h] and Q i Element q i (j,1) to q i The conversion relationship of (j,cwh) is as follows: Set variable j, 1≤j≤b; Pixel o i [j,1,1,1] to o i [j,1,1,h] corresponds to element q i (j,1) to q i (j,h); Pixel o i [j,1,2,1] to o i [j,1,2,h] corresponds to element q i (j,h+1) to q i (j,2h); Pixel o i [j,1,w,1] to o i [j,1,w,h] corresponds to element q i (j,(w-1)h+1) to q i (j,wh); Pixel o i [j,2,w,1] to o i [j,2,w,h] corresponds to element q i (j,wh+(w-1)h+1) to q i (j,2wh); Pixel o i [j,c,w,1] to o i [j,c,w,h] corresponds to element q i (j,(c-1)wh+(w-1)h+1) to q i (j,cwh); Construct a two-dimensional matrix Q i The activation matrix A of the same size i , for Q i For any element in A, if the element of A is greater than 0, then i The element at the corresponding position in is set to 1, otherwise, it is set to 0; (2.3.4) Calculate the kernel matrix K; According to the activation matrix A i Calculate the matrix Among them, a i (j,cwh) represents A i The element value in row j and column cwh; According to A i and Calculate the matrix K i : Wherein, the superscript T indicates transposition; Calculate the kernel matrix K: (2.3.5) Calculate the performance indicators of each candidate network; S τ =lg|K|,τ∈B (2.4) Select the fault diagnosis model; Set the number of searches M; repeat steps (2.2) to (2.3), perform performance evaluation on the candidate networks in each round of search, and then select the maximum value based on the performance indicators of all candidate networks. Record the network model corresponding to the maximum value as the fault diagnosis model; (3) Training fault diagnosis model; The shuffled training data set X is input into the fault diagnosis model in batches, and the fault diagnosis model is trained repeatedly until it converges. (4) Real-time fault diagnosis; The oscillation signal of the underwater autonomous underwater vehicle at the current moment is collected in real time, and then the oscillation signal is input into the fault diagnosis model to output the fault type in real time.

Images