Improved HGCN-based dual-channel signal fusion fault diagnosis method for AUV (Autonomous Underwater Vehicle) propeller

Abstract

The invention provides an improved HGCN-based dual-channel signal fusion fault diagnosis method for an AUV (Autonomous Underwater Vehicle) propeller. The method comprises the following steps: firstly, forming a three-channel signal by two vibration signals of an autonomous underwater vehicle propeller and a random noise signal with the same length, and performing noise-assisted multivariate empirical mode decomposition; after decomposition is completed, IMFs components corresponding to noise channels are removed, and then the reserved IMFs components are fused by adopting a mean value fusion algorithm to obtain a new one-dimensional signal. An improved HGCN model is constructed by adding a graph attention mechanism and a gating loop unit module after a convolutional layer of a hypergraph convolutional network. Continuously sampling from the one-dimensional fusion signal to obtain samples, and dividing the samples into a training set and a test set according to the proportion of 1: 3; and inputting the training set into an improved HGCN model for model training, inputting the test set into the trained model for autonomous feature extraction and intelligent fault diagnosis, and finally outputting a classification result of a test sample.

Description

Technical Field

[0001] The invention relates to an AUV thruster dual-channel signal fusion fault diagnosis method based on improved HGCN, and relates to the technical field of AUV fault diagnosis. Background Art

[0002] Autonomous Underwater Vehicle (AUV) is a key tool for ocean exploration and resource development. Its operational safety is particularly important in unmanned and cable-free marine environments. Under high-pressure working conditions, AUVs are prone to entanglement, loose chute, and imbalance of single and double blades. Therefore, ensuring its safety has become an important task in the research and use process. Fault diagnosis, as a key means to ensure the safety of underwater vehicles, has far-reaching significance for the application of AUVs. When performing complex tasks, there is a complex nonlinear relationship between the AUV fault type and fault characteristics. However, the feature information provided by the single-channel signal is limited and the fault type cannot be accurately identified. Therefore, it is necessary to fuse the information of the original multi-channel signal. In addition, in order to deeply explore and explain the more complex high-order relationships between samples, a hypergraph neural network is introduced for feature extraction, thereby improving the accuracy and reliability of fault diagnosis.

[0003] In current research, for the problem of multi-channel data decomposition and fusion, the empirical mode decomposition (EMD) algorithm can decompose the signal into a series of intrinsic mode functions (IMFs), each of which corresponds to a relatively constant local frequency, but it is only applicable to single-channel signals and cannot be applied to multi-channel signal analysis. Therefore, the multivariate empirical mode decomposition (MEMD) algorithm, a multi-channel extended version based on EMD, came into being. However, through a series of experiments, it was found that the MEMD algorithm is prone to mode aliasing and endpoint effects during the signal decomposition process. To this end, an improved method is proposed, called the noise-assisted multivariate empirical mode decomposition (NA-MEMD) algorithm. Multiple experimental results show that the improved method can effectively solve the above problems and show significant improvement effects.

[0004] In the field of fault diagnosis, fault classification was initially performed using traditional machine learning methods. However, these methods are difficult to effectively capture the deep features of samples, especially in actual working conditions, where most of the signals we face are nonlinear. Graph convolution methods can better handle this problem, among which the graph convolutional network (GCN) improves the prediction ability of unknown sample labels by considering the relationship between sample nodes. However, ordinary graph convolutional networks only consider the pairwise relationship between nodes, which may lose fault information and cause some sample recognition errors. To solve this problem, a hypergraph convolutional neural network (HGCN) is proposed. Compared with ordinary GCN, HGCN supports many-to-many node connections, thereby mining more complex nonlinear relationships between AUV thruster samples and improving fault diagnosis accuracy. In order to make full use of the advantages of the HGCN model, the present invention improves it by adding a graph attention mechanism and a gated recurrent unit module. Summary of the invention

[0005] In view of the limited feature information contained in the single-channel signal and the strong nonlinear relationship between the fault characteristics and fault types of the AUV thruster when the AUV performs complex tasks, it is impossible to accurately identify the fault type of the AUV thruster. This paper proposes a dual-channel signal fusion fault diagnosis method for AUV thrusters based on improved HGCN. The method includes the following steps:

[0006] S1: Collect dual-channel vibration signals of AUV thrusters;

[0007] S2: Signal processing using NA-MEMD method;

[0008] S3: Get samples and define labels;

[0009] S4: Build an improved HGCN network model;

[0010] S5: Diagnose AUV thruster failure.

[0011] Furthermore, step S1 is specifically as follows:

[0012] The LMS data acquisition system and vibration acceleration sensor are used to collect the bidirectional vibration signals of the AUV thruster in the axial and radial directions. The speed of the AUV is constant at 3000r / min, the sampling frequency is set to 12.8kHz, and the sampling time is 10s. The vibration signals of five different fault types of the thruster are collected, namely normal, loose chute, winding, single blade imbalance, and double blade imbalance.

[0013] Furthermore, step S2 is specifically as follows:

[0014] S2-1: Adding a noise channel

[0015] Add a noise channel to ensure that the length of the noise signal is consistent with the length of the vibration signal. The three-channel signal is composed of two vibration signals and one noise signal, and the multivariate empirical mode decomposition (MEMD) is upgraded to noise-assisted multivariate empirical mode decomposition (NA-MEMD).

[0016] S2-2: Multivariate Empirical Mode Decomposition Using Three-Channel Signals

[0017] Vibration signals s1(t), s2(t) and noise signal s n (t) forms a three-dimensional signal, which is decomposed through the following steps:

[0018] (1) Using the Hmmersley sequence to select K evenly distributed sampling point sets on the three-dimensional sphere, we can obtain K direction vectors in the three-dimensional space

[0019] (2) Calculate the direction vector set of the three-dimensional signal The projection in each direction is denoted as Forming a Projection Set

[0020]

[0021] (3) Finding a projection set The maximum point of each projection in the time scale corresponding to the maximum point is

[0022]

[0023] (4) For the obtained extreme points Perform interpolation calculations to obtain K multivariate envelopes

[0024] (5) Calculate the K envelope means m(t) as: Extract the intrinsic mode function through h(t) = s(t) - m(t). If h(t) meets the stopping criteria of IMFs, then use s(t) = s(t) - h(t) as the new input signal and repeat the above steps. Otherwise, use h(t) as the new input signal and repeat the above steps. After multiple decompositions, the original signal is decomposed into a series of The sum of the residual r(t): 3×N×128000 components are obtained, where N represents the number of components obtained for each channel.

[0025] (6) The function of the noise channel is to suppress the modal aliasing and endpoint effect when the signal is decomposed, and the components of the noise channel do not participate in the subsequent processing. Therefore, the IMFs components corresponding to the noise channel are removed from the 3×N×128000 components obtained above, and only the IMFs components corresponding to the two vibration signals are retained.

[0026] S2-3: Use the mean fusion algorithm to fuse the IMFs of the two channels

[0027] After obtaining the components corresponding to the two channels, the IMFs corresponding to the two channels are fused using the mean fusion algorithm. The specific steps are as follows:

[0028] (1) For each channel, all components are averaged. That is, the N×128000 components of each channel are averaged according to the dimension of the components to obtain a one-dimensional signal of 1×128000. The two one-dimensional signals are combined into a two-dimensional matrix of 2×128000.

[0029] (2) After obtaining the average signal of each channel, perform a mean operation on the channel dimension based on the two-dimensional matrix obtained in the previous step.

[0030] (3) After two steps of averaging operation, a one-dimensional signal of 1×128000 is finally obtained after the dual-channel signal fusion.

[0031] Furthermore, step S3 is specifically as follows:

[0032] Samples are continuously sampled from the obtained one-dimensional fusion signal, each sample length is 1024, 200 samples for each fault type, and 50 samples are selected as training samples, and the remaining 150 are used as test samples. Therefore, there are 1000 samples in total, the training set contains 250 samples, and the test set contains 750 samples. The samples of the five different fault types of thrusters, normal, loose chute, winding, single-leaf imbalance, and double-leaf imbalance, are defined as labels 0, 1, 2, 3, and 4 respectively.

[0033] Further, step S4 is specifically as follows:

[0034] The improved HGCN network model is mainly composed of four parts: input layer, feature extraction layer, fully connected layer, and output layer. The core feature extraction layer uses three layers of hypergraph convolution (HypergraphConv), and a normalization layer (BatchNorm) is used after each convolution layer. Then the graph attention mechanism (GATConv) module is introduced, and a normalization layer is also used for data normalization. The output is input into the gated recurrent unit (GRU), and the data is also normalized through a normalization layer. The fully connected layer compresses the input 1024-dimensional features into 512 dimensions, and uses the Leaky_Relu activation function to avoid problems such as gradient disappearance and gradient explosion. The Dropout layer is added to prevent overfitting during training. The output layer outputs the 512-dimensional features as 5 classification categories.

[0035] When constructing the hypergraph convolution layer, the input of each hypergraph convolution layer is:

[0036] X (l+1) =D v -1 / 2 HWD e -1 / 2 X(l)P

[0037] Among them, D v is the node degree matrix, D e is the hyper-edge degree matrix, H is the incidence matrix, W is the edge weight matrix, and P is the edge weight matrix of the lth layer and the l+1th layer.

[0038] When constructing the graph attention layer, the input of the graph attention layer is the feature set h of each node The output is the feature set h' after learning

[0039]

[0040] The specific calculation process can be divided into the following steps:

[0041] (1) Calculate the attention coefficient

[0042] Calculate the similarity between it and its adjacent nodes:

[0043]

[0044] Among them, e ij represents the influence of node j on node i, a represents the function of calculating the correlation between two nodes, and the dimension of W is F×F'.

[0045] For ij Normalize to get α ij :

[0046]

[0047] Among them, N i is the set of adjacent nodes of node i.

[0048] (2) Calculate the output h' of the graph attention layer

[0049]

[0050] Before building the fully connected layer, we must first define the gated recurrent unit. There are two core gates in the gated recurrent unit, namely R t (Reset Gate) and Z t (Update gate). Both gates are composed of the last hidden H t-1 and the current input vector X t The common composition and formulas are:

[0051] R t =σ(X t W xr +H t-1 W hr +b r )

[0052] Z t =σ(X t W xz +H t-1 W hz +b z )

[0053] The activation function is Leaky_Relu, W represents the weight matrix at the corresponding position, and b represents the bias value at the corresponding moment.

[0054] When constructing the fully connected layer, the output of the gated recurrent unit is compressed from 1024 dimensions to 512 dimensions, while retaining useful information. The fully connected layer formula is defined as:

[0055] y=σ(ω·x+b)

[0056] Among them, σ is the activation function, ω is the weight matrix, and b is the bias vector.

[0057] Furthermore, step S5 is specifically as follows:

[0058] The training set is input into the improved HGCN model for model training, and then the test set is input into the trained model for autonomous feature extraction and intelligent fault diagnosis, and finally the classification results of the test samples are output.

[0059] Compared with the prior art, the present invention has the following effects:

[0060] The purpose of the present invention is to provide an AUV thruster dual-channel signal fusion fault diagnosis method based on improved HGCN. First, the original dual-channel vibration signal of the AUV thruster is subjected to noise-assisted multivariate empirical mode decomposition to obtain 3×N×128000 multi-channel components, and then the components corresponding to the noise channel are eliminated, and then the retained components are fused using the mean fusion algorithm to obtain a new signal. The obtained one-dimensional fusion signal is divided into a training set and a test set in a ratio of 1:3 and input into the improved HGCN model for autonomous feature extraction and intelligent fault diagnosis, and finally the classification results of the samples are output. This invention mainly solves the problem that the single-channel signal contains incomplete information and the HGCN model cannot fully consider the complex relationship between each sample node when diagnosing the AUV thruster fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Typical failures of AUV entities and their thrusters;

[0062] Figure 2 The collected axial and radial time domain vibration signals;

[0063] Figure 3 One-dimensional fusion signal obtained after NA-MEMD signal processing;

[0064] Figure 4 Schematic diagram of hypergraph convolution;

[0065] Figure 5 Overall flow chart of AUV thruster fault diagnosis;

[0066] Figure 6 AUV thruster fault diagnosis results diagram. DETAILED DESCRIPTION

[0067] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0068] A dual-channel signal fusion fault diagnosis method for AUV thrusters based on improved HGCN, the method comprising the following steps:

[0069] S1: Collect dual-channel vibration signals of AUV thrusters

[0070] The LMS data acquisition system and vibration acceleration sensor are used to collect the bidirectional vibration signals of the AUV thruster in the axial and radial directions. The speed of the AUV is constant at 3000r / min, the sampling frequency is set to 12.8kHz, and the sampling time is 10s. The vibration signals of five different fault types of the thruster are collected, namely normal, loose chute, winding, single blade imbalance, and double blade imbalance.

[0071] S2: Signal processing using NA-MEMD method

[0072] S2-1: Adding a noise channel

[0073] Add a noise channel to ensure that the length of the noise signal is consistent with the length of the vibration signal. A three-channel signal is formed by two vibration signals and one noise signal, and the multivariate empirical mode decomposition (MEMD) is upgraded to noise-assisted multivariate empirical mode decomposition (NA-MEMD). The noise channel can effectively suppress the modal aliasing and endpoint effects that appear in the MEMD decomposition, so that the characteristic time scales contained in the IMFs after the signal decomposition are not much different, and the characteristic scale phenomenon of the same frequency does not appear in adjacent components.

[0074] S2-2: Multivariate Empirical Mode Decomposition Using Three-Channel Signals

[0075] Vibration signals s1(t), s2(t) and noise signal s n (t) forms a three-dimensional signal, which is decomposed through the following steps:

[0076] (1) Using the Hmmersley sequence to select K evenly distributed sampling point sets on the three-dimensional sphere, we can obtain K direction vectors x in the three-dimensional space. θk .

[0077] (2) Calculate the direction vector set x of the three-dimensional signal θk The projection in each direction is denoted as Forming a Projection Set

[0078]

[0079] (3) Finding a projection set The maximum point of each projection in the time scale corresponding to the maximum point is

[0080]

[0081] (4) For the obtained extreme points Perform interpolation calculations to obtain K multivariate envelopes

[0082] (5) Calculate the K envelope means m(t) as: Extract the intrinsic mode function through h(t) = s(t) - m(t). If h(t) meets the stopping criteria of IMFs, then use s(t) = s(t) - h(t) as the new input signal and repeat the above steps. Otherwise, use h(t) as the new input signal and repeat the above steps. After multiple decompositions, the original signal is decomposed into a series of The sum of the residual r(t): 3×N×128000 components are obtained, where N represents the number of components obtained for each channel.

[0083] (6) The function of the noise channel is to suppress the modal aliasing and endpoint effect when the signal is decomposed, and the components of the noise channel do not participate in the subsequent processing. Therefore, the IMFs components corresponding to the noise channel are removed from the 3×N×128000 components obtained above, and only the IMFs components corresponding to the two vibration signals are retained.

[0084] S2-3: Use the mean fusion algorithm to fuse the IMFs of the two channels

[0085] After obtaining the components corresponding to the two channels, the IMFs corresponding to the two channels are fused using the mean fusion algorithm. The specific steps are as follows:

[0086] (1) For each channel, all components are averaged. That is, the N×128000 components of each channel are averaged according to the dimension of the components to obtain a one-dimensional signal of 1×128000. The two one-dimensional signals are combined into a two-dimensional matrix of 2×128000.

[0087] (2) After obtaining the average signal of each channel, perform a mean operation on the channel dimension based on the two-dimensional matrix obtained in the previous step.

[0088] (3) After two steps of averaging operation, a one-dimensional signal of 1×128000 is finally obtained after the dual-channel signal fusion.

[0089] S3: Obtaining samples and defining labels

[0090] Samples are continuously sampled from the obtained one-dimensional fusion signal, each sample length is 1024, 200 samples for each fault type, and 50 samples are selected as training samples, and the remaining 150 are used as test samples. Therefore, there are 1000 samples in total, the training set contains 250 samples, and the test set contains 750 samples. The samples of the five different fault types of thrusters, normal, loose chute, winding, single-leaf imbalance, and double-leaf imbalance, are defined as labels 0, 1, 2, 3, and 4 respectively.

[0091] S4: Building an improved HGCN network model

[0092] The improved HGCN network model is mainly composed of four parts: input layer, feature extraction layer, fully connected layer, and output layer. The core feature extraction layer uses three layers of hypergraph convolution (HypergraphConv), and a normalization layer (BatchNorm) is used after each convolution layer. Then the graph attention mechanism (GATConv) module is introduced, and a normalization layer is also used for data normalization. The output is input into the gated recurrent unit (GRU), and the data is also normalized through a normalization layer. The fully connected layer compresses the input 1024-dimensional features into 512 dimensions, and uses the Leaky_Relu activation function to avoid problems such as gradient disappearance and gradient explosion. The Dropout layer is added to prevent overfitting during training. The output layer outputs the 512-dimensional features as 5 classification categories.

[0093] The improved HGCN model architecture is as follows:

[0094]

[0095]

[0096] When constructing the hypergraph convolution layer, the input of each hypergraph convolution layer is:

[0097] X (l+1) =D v -1 / 2 HWD e -1 / 2 X(l)P

[0098] Among them, D v is the node degree matrix, D e is the hyper-edge degree matrix, H is the incidence matrix, W is the edge weight matrix, and P is the edge weight matrix of the lth layer and the l+1th layer.

[0099] When constructing the graph attention layer, the input of the graph attention layer is the feature set h of each node The output is the feature set h' after learning

[0100]

[0101] The specific calculation process can be divided into the following steps:

[0102] (1) Calculate the attention coefficient

[0103] Calculate the similarity between it and its adjacent nodes:

[0104]

[0105] Among them, e ijrepresents the influence of node j on node i, a represents the function of calculating the correlation between two nodes, and the dimension of W is F×F'.

[0106] For ij Normalize to get α ij :

[0107]

[0108] Among them, N i is the set of adjacent nodes of node i.

[0109] (2) Calculate the output h' of the graph attention layer

[0110]

[0111] Before building the fully connected layer, we must first define the gated recurrent unit. There are two core gates in the gated recurrent unit, namely R t (Reset Gate) and Z t (Update gate). Both gates are composed of the last hidden H t-1 and the current input vector X t The common composition and formulas are:

[0112] R t =σ(X t W xr +H t-1 W hr +b r )

[0113] Z t =σ(X t W xz +H t-1 W hz +b z )

[0114] The activation function is Leaky_Relu, W represents the weight matrix at the corresponding position, and b represents the bias value at the corresponding moment.

[0115] When constructing the fully connected layer, the output of the gated recurrent unit is compressed from 1024 dimensions to 512 dimensions, while retaining useful information. The fully connected layer formula is defined as:

[0116] y=σ(ω·x+b)

[0117] Among them, σ is the activation function, ω is the weight matrix, and b is the bias vector.

[0118] S5: Diagnosing AUV Thruster Failures

[0119] The training set is input into the improved HGCN model for model training, and then the test set is input into the trained model for autonomous feature extraction and intelligent fault diagnosis, and finally the classification results of the test samples are output.

[0120] The above are preferred embodiments of the present invention. A person skilled in the art may make various changes or improvements on this basis. Without departing from the general concept of the present invention, these changes or improvements should fall within the scope of protection required by the present invention.

Claims

1. A dual-channel signal fusion fault diagnosis method for AUV thrusters based on improved HGCN, characterized in that: The following steps are involved: S1: Collect dual-channel vibration signals of AUV thrusters; S2: Signal processing using NA-MEMD method; S3: Get samples and define labels; S4: Build an improved HGCN network model; S5: Diagnose AUV thruster failure.

Images