Underwater propeller fault identification method and identification system based on channel expansion and sequence fusion

Abstract

This invention discloses a fault identification method for underwater thrusters based on channel expansion and sequence fusion. The method acquires time-domain sequences of several dynamic signals from an underwater robot. Each time-domain sequence is sequentially processed using mean removal, wavelet decomposition, modified Bayesian algorithm, and evidence theory fusion. Then, each processed sequence undergoes a Fourier transform to obtain corresponding frequency sequences. These frequency sequences are arranged and fused to obtain a two-dimensional matrix. Using this two-dimensional matrix as input, a fault diagnosis model is used to classify fault levels and determine the degree of thruster fault. The fault diagnosis model is established through underwater thruster fault experiments. The time-series channels are expanded, and the diagnostic effects of the expanded frequency sequences are ranked and fused into a two-dimensional matrix for hyperparameter learning, significantly improving the convergence speed and accuracy of fault degree identification.

Description

Technical Field

[0001] This invention relates to underwater robot thruster fault detection, specifically to an underwater thruster fault identification method and system based on channel expansion and sequence fusion. Background Technology

[0002] With the increasing exploration and development of the ocean, underwater robots can be found in the Arctic, the Mariana Trench, and coastal ports, demonstrating their crucial role in marine exploration. However, facing the complex marine environment, the propulsion systems of underwater robots need to operate continuously under harsh, high-intensity conditions for extended periods. If a propulsion system malfunctions during engineering practice, the underwater robot will be unable to complete its intended tasks, potentially leading to even greater losses. Therefore, it is necessary to identify faults in underwater propulsion systems to effectively address the challenges posed by the complex underwater environment.

[0003] In the prior art, Chinese patent application number 202010030259.6 discloses a method for diagnosing faults in an underwater robot propulsion system. This method normalizes the raw data and directly diagnoses the degree of propulsion fault based on monitoring signals such as voltage using a convolutional neural network. However, this fault diagnosis model uses a limited number of fault data channels, resulting in less fault information in the data and limited identification effectiveness.

[0004] For example, Chinese patent application CN111275164A discloses a method for diagnosing faults in an underwater robot propulsion system. This method normalizes the collected voltage, current, output speed, and tension / compression data of the underwater robot propulsion system and uses a convolutional neural network to locate and classify faults. This approach integrates fault feature extraction with fault diagnosis modeling, directly diagnosing the degree of propulsion fault based on monitoring signals such as voltage. However, the accuracy of fault diagnosis can sometimes decrease when the position of the fault information in the monitoring signal sequence changes.

[0005] Among existing fault diagnosis technologies, the paper "Fault Diagnosis of Roller Bearings in Spinning Frames Based on FFT-1D-CNN" published in the Cotton Textile Technology Journal proposes a fault diagnosis method for roller bearings in variable speed spinning frames based on Fast Fourier Transform (FFT) and one-dimensional convolutional neural network (1D-CNN). This method uses a one-dimensional convolutional neural network to diagnose faults using vibration acceleration signals from FFT, thus meeting the end-to-end diagnostic needs of enterprises. However, this method only learns information from a single sequence of fault channels, resulting in limited fault identification accuracy.

[0006] Chinese patent application No. 202110020601.9 discloses a bearing fault diagnosis method based on multi-channel CNN multi-information fusion. In this method, the collected time-domain vibration signals from both sides of the bearing are input into a multi-information fusion convolutional neural network with multiple dual-channel convolutional kernels, which reduces the computation process of the network model. However, this method does not take into account the influence of the order of single sequence arrangement on the neural network model during signal fusion, resulting in poor recognition performance. Summary of the Invention

[0007] Purpose of the invention: To address the above-mentioned shortcomings, this invention provides a fault identification method for underwater thrusters based on channel expansion and sequence fusion, which improves the convergence speed and accuracy of fault identification.

[0008] The present invention also provides an underwater thruster fault identification system based on channel expansion and sequence fusion.

[0009] Technical Solution: To solve the above problems, this invention employs a fault identification method for underwater thrusters based on channel expansion and sequence fusion, comprising the following steps:

[0010] (1) Collect several dynamic signals of the underwater robot and obtain the time-domain sequence of several dynamic signals;

[0011] (2) The obtained time-domain sequences are subjected to mean removal, wavelet decomposition, modified Bayesian algorithm and evidence theory fusion processing in sequence to realize sequence channel expansion;

[0012] (3) Perform Fourier transform on each sequence after mean removal, each sequence after wavelet decomposition, each sequence after modified Bayes algorithm, and the fused sequence after evidence theory fusion to obtain the corresponding frequency sequences.

[0013] (5) Arrange and fuse the frequency sequences to obtain a two-dimensional matrix;

[0014] (6) Using a two-dimensional matrix as input, the fault level is classified through a fault diagnosis model to obtain the degree of thruster failure. The fault diagnosis model is established through underwater thruster failure tests.

[0015] Furthermore, in step (5), each frequency sequence is arranged according to the channel expansion processing order, that is, each column of the fused two-dimensional matrix is, in turn, the frequency sequence corresponding to each sequence after mean removal processing, the frequency sequence corresponding to each sequence after wavelet decomposition processing, the frequency sequence corresponding to each sequence after modified Bayes algorithm processing, and the frequency sequence of the fused sequence.

[0016] Furthermore, in step (5), the frequency sequences are arranged in order of signal type, that is, each column of the fused two-dimensional matrix is, in order, the frequency sequence corresponding to the first type of dynamic signal after mean-reduction processing, the frequency sequence corresponding to the wavelet decomposition processing, the frequency sequence corresponding to the modified Bayes algorithm processing, the frequency sequence corresponding to the second type of dynamic signal after mean-reduction processing, the frequency sequence corresponding to the wavelet decomposition processing, the frequency sequence corresponding to the modified Bayes algorithm processing, ..., the frequency sequence corresponding to the nth type of dynamic signal after mean-reduction processing, the frequency sequence corresponding to the wavelet decomposition processing, the frequency sequence corresponding to the modified Bayes algorithm processing, and the frequency sequence of the fused sequence.

[0017] Furthermore, in step (5), each frequency sequence is arranged in order from best to worst according to its convolutional neural network training and testing results; a one-dimensional convolutional neural network is constructed, with the frequency sequence as the input of the fault sample and the thruster fault level as the output, a cross-entropy loss function is constructed, the weights of the convolutional neural network are updated based on the gradient descent method, the convergence speed and recognition accuracy of the convolutional neural network training are recorded, and the training and testing results of the convolutional neural network are obtained.

[0018] Furthermore, in step (5), each frequency sequence is arranged in order from worst to best according to the training and testing results of its convolutional neural network; a one-dimensional convolutional neural network is constructed, with the frequency sequence as the input of the fault sample and the thruster fault level as the output, a cross-entropy loss function is constructed, the weights of the convolutional neural network are updated based on the gradient descent method, the convergence speed and recognition accuracy of the training of the convolutional neural network are recorded, and the training and testing results of the convolutional neural network are obtained.

[0019] Furthermore, the specific content of the evidence theory fusion process in step (2) is as follows:

[0020] (2.4.1) Determine the identification framework Θ for thruster faults;

[0021] (2.4.2) Based on the confidence assignment function of the fault evidence of each sequence after processing by the modified Bayesian algorithm, calculate the confidence assignment function of the fault evidence of the fused sequence:

[0022] (2.4.3) Calculate the support of the confidence assignment function of the fused sequence;

[0023] (2.4.4) Based on the support calculated in step (2.4.3), the confidence value time-domain sequence C(A) is calculated. k ), and C(A) k This serves as the final fusion sequence.

[0024] The present invention also provides an underwater thruster fault identification system based on channel expansion and sequence fusion, including a data acquisition module for acquiring several dynamic signals of an underwater robot and obtaining the time-domain sequence of the several dynamic signals;

[0025] Signal processing module: used to perform mean removal, wavelet decomposition, modified Bayesian algorithm, and evidence theory fusion processing on the obtained time-domain sequences in sequence to realize sequence channel expansion;

[0026] The signal fusion module is used to perform Fourier transforms on the sequences after mean removal processing, wavelet decomposition processing, modified Bayesian algorithm processing, and evidence theory fusion processing to obtain the corresponding frequency sequences; and to arrange and fuse the frequency sequences to obtain a two-dimensional matrix.

[0027] The fault identification module is used to take a two-dimensional matrix as input, classify the fault level through a fault diagnosis model, and obtain the degree of thruster fault. The fault diagnosis model is established through underwater thruster fault tests.

[0028] Beneficial effects: Compared with the prior art, the significant advantage of this invention is that it expands the time series channel, sorts the diagnostic effects of the expanded frequency series, and merges them into a two-dimensional matrix for hyperparameter learning, which greatly improves the convergence speed and accuracy of fault degree identification. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the identification method in this invention, in which the frequency sequence is arranged according to the channel expansion processing order.

[0030] Figure 2 This is a schematic diagram of the identification method in this invention, in which frequency sequences are arranged in order of signal type.

[0031] Figure 3 This is a schematic diagram of the identification method in this invention, in which frequency sequences are arranged in descending order of their convolutional neural network training and testing results.

[0032] Figure 4 This is a schematic diagram of the identification method in this invention, in which frequency sequences are arranged in order from worst to best according to the training and testing results of their convolutional neural networks.

[0033] Figure 5 This is a waveform diagram of the thruster control voltage change rate signal collected by this invention.

[0034] Figure 6 This is a waveform diagram of the longitudinal velocity signal of the underwater robot collected by this invention.

[0035] Figure 7This is a schematic diagram of the process of channel expansion and sequence fusion of the acquired signal in this invention.

[0036] Figure 8 This is a comparison chart of entropy curves in this invention.

[0037] Figure 9 This is a comparison chart of accuracy curves in this invention. Detailed Implementation

[0038] Example 1:

[0039] like Figure 1 As shown in this embodiment, a fault identification method for underwater thrusters based on channel expansion and sequence fusion includes the following steps:

[0040] (1) A sliding time window of length L is used to capture different types of dynamic signals of the underwater robot. In this embodiment, the dynamic signals of the thruster control voltage change rate and the longitudinal velocity of the underwater robot are used as examples to obtain the time domain sequences of the two dynamic signals: {T1(k)} = [T1(1) T1(2) … T1(L)] and {T2(k)} = [T2(1) T2(2) … T2(L)].

[0041] (2) Process the obtained time-domain sequence to achieve time-domain sequence channel expansion, specifically:

[0042] (2.1) The time-domain sequences {T1(k)} and {T2(k)} of the dynamic signal obtained in the first step are subjected to mean-removal processing to obtain the mean-removed time-domain sequence {T 1R (k)}、{T 2R (k)}, to reduce the influence of the constant components of the signal;

[0043] (2.2) For the mean-removed time-domain sequence {T} 1R (k)}、{T 2R (k)} is decomposed into wavelet scale component sequence {T}. 1RW (k)}、{T 2RW (k)}, to reduce the impact of high-frequency noise;

[0044] (2.3) For wavelet scale component sequences {T 1RW (k)}、{T 2RW (k)} is processed using the modified Bayesian algorithm to obtain the enhanced sequence {T}. 1RWMB (k)}、{T 2RWMB (k)}, to enhance the time-domain amplitude of fault characteristics;

[0045] (2.4) Enhance the sequence {T} 1RWMB (k)}、{T 2RWMB(k)} is processed based on evidence theory, specifically as follows:

[0046] (2.4.1) Let the total number of faulty focal elements be L, and construct the thruster fault identification framework as Θ, Θ={A1,A2,…,A L};

[0047] (2.4.2) Calculate the enhancement sequence {T} 1RWMB (k)}、{T 2RWMB The credibility assignment function F(A) of the fault evidence in (k)} m ):

[0048]

[0049] In equation (1), n ​​= 1, 2, ..., L, where L is the length of the sliding time window; S MB (m) represents {T 1RWMB (k)} or {T} 2RWMB (k)};

[0050] (2.4.3) Calculate the confidence assignment function F of the fault evidence in the fused sequence. ij (A k ):

[0051]

[0052] In equation (2), i ≠ j, i and j are the focal element indices, ranging from 1 to L, and are positive integers; F ij (A k ) is {T 1RWMB The credibility assignment function for fault evidence in (k)}, F ij (A k ) is {T 2RWMB (k)} The credibility assignment function for fault evidence. From formula (2), it can be seen that if F i (A k ) and F j (A k The larger the difference between F, the smaller their similarity; when F ij (A k When ) = 1, it means that the two have the same similarity.

[0053] (2.4.4) Calculate the reliability assignment function F of the fused sequence. ij (A k Support level Q i (A k ):

[0054]

[0055] (2.4.5) The confidence value time-domain sequence C(A) is calculated. k ), and C(A) k As the final fusion sequence:

[0056]

[0057] C(A k ) is represented as F ij (A k The credibility value of ).

[0058] (3) The time series {T 1R (k)}、{T 2R (k)}、{T 1RW (k)}、{T 2RW (k)}、{T 1RWMB (k)}、{T 2RWMB (k)}、{C(A k Perform Fourier transforms on each of the} to obtain the frequency sequence {f}. 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)};

[0059] The Fourier transform process is as follows: x(n) is {T} 1R (k)}、{T 2R (k)}、{T 1RW (k)}、{T 2RW (k)}、{T 1RWMB (k)}、{T 2RWMB (k)}、{C(A k If one of them is P, then its P f The point discrete Fourier transform is shown in equation (5):

[0060]

[0061]

[0062] In the formula, p = 0, 1, ..., P f -1; j is an imaginary number; a p b p Let |X(p)| be the real and imaginary parts of the complex number X(p), respectively. Take the first half of the sequence of |X(p)| in formula (6), i.e., p = 0, 1, ..., P. f / 2-1, as a frequency sequence;

[0063] (4) The single frequency sequence {f 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} are arranged according to the channel expansion processing order and fused into a two-dimensional matrix. That is, the leftmost two columns of the two-dimensional matrix are the frequency sequences {f} corresponding to the mean removal processing. 1R (p)}、{f 2R (p)}, the two columns to the right are the frequency sequences {f} corresponding to the wavelet decomposition process. 1RW (p)}、{f 2RW (p)}, the next two columns to the right are the frequency sequences {f} corresponding to the modified Bayesian algorithm processing. 1RWMB (p)}、{f 2RWMB (p)}, the next column to the right is the frequency sequence {f} corresponding to the fusion of evidence theories. C (p)}, the result is shown in formula (7).

[0064]

[0065] (5) Using a two-dimensional matrix as input, the fault level is classified through a fault diagnosis model to obtain the degree of thruster fault. The fault diagnosis model is established through underwater thruster fault tests, specifically: conducting fault tests on the underwater thruster with known fault levels to obtain several dynamic signals of the underwater robot corresponding to the fault level; performing time-domain sequence channel expansion and sequence fusion on the obtained dynamic signals, using the fused two-dimensional matrix as fault samples as input to a two-dimensional convolutional neural network, using the fault level as the output of the two-dimensional convolutional neural network, training the parameters of the two-dimensional convolutional neural network using fault samples, and obtaining the fault diagnosis model of the two-dimensional convolutional neural network after training.

[0066] Example 2

[0067] like Figure 2 As shown, this embodiment of the underwater thruster fault identification method based on channel expansion and sequence fusion includes steps (1) to (3) and step (5) in embodiment 1, which will not be repeated here. Based on steps (1) to (3) and step (5) in embodiment 1, step (4) in this embodiment is as follows:

[0068] (4) The single frequency sequence {f 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW(p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} are arranged according to signal type and fused into a two-dimensional matrix. That is, the leftmost three columns of the two-dimensional matrix are the frequency sequences {f} corresponding to the dynamic signal {T1(k)} after mean removal, wavelet decomposition, and modified Bayesian algorithm processing. 1R (p)}、{f 1RW (p)}、{f 1RWMB (p)}, the three columns to the right are the frequency sequences {f} corresponding to the dynamic signal {T2(k)} after mean subtraction, wavelet decomposition, and modified Bayesian algorithm processing. 2R (p)}、{f 2RW (p)}、{f 2RWMB (p)}, the next column to the right is the frequency sequence {f} corresponding to the fusion of evidence theories. C (p)}, the result is shown in formula (8).

[0069]

[0070] Example 3

[0071] like Figure 3 As shown, this embodiment of the underwater thruster fault identification method based on channel expansion and sequence fusion includes steps (1) to (3) and step (5) in embodiment 1, which will not be repeated here. Based on steps (1) to (3) and step (5) in embodiment 1, step (4) in this embodiment is as follows:

[0072] (4) The single frequency sequence {f 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} are fused into a two-dimensional matrix according to the order of the training and testing results of the one-dimensional convolutional neural network from best to worst.

[0073] The construction of the aforementioned one-dimensional convolutional neural network is as follows: using the frequency sequence as fault samples as the input of the one-dimensional convolutional neural network, and the thruster fault level as the output of the one-dimensional convolutional neural network, a cross-entropy loss function is constructed, the weights of the convolutional neural network are updated based on the gradient descent method, and the convergence speed and recognition accuracy of the convolutional neural network training are recorded. The single frequency sequence {f} is used as the input. 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW(p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} represents fault samples. We train the weights of a convolutional neural network and record the convergence speed and recognition accuracy of the convolutional neural network training.

[0074] Assuming the single frequency sequences are ordered from best to worst based on the training and testing results of the convolutional neural network, the order is {f}. 1R (p)}、{f 2R (p)}、{f 2RWMB (p)}、{f C (p)}、{f 2RW (p)}、{f 1RW (p)}、{f 1RWMB (p)}. Then the frequency sequence arrangement in the two-dimensional matrix is: the leftmost column of the two-dimensional matrix is ​​{f 1R (p)}, and to the right of this are {f} 2R (p)}、{f 2RWMB (p)}、{f C (p)}、{f 2RW (p)}、{f 1RW (p)}、{f 1RWMB (p)}, the result is shown in formula (9).

[0075]

[0076] Example 4:

[0077] like Figure 4 As shown, this embodiment of the underwater thruster fault identification method based on channel expansion and sequence fusion includes steps (1) to (3) and step (5) in embodiment 1, which will not be repeated here. Based on steps (1) to (3) and step (5) in embodiment 1, step (4) in this embodiment is as follows:

[0078] (4) The single frequency sequence {f 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} are fused into a two-dimensional matrix according to the order of poor to excellent training and testing results of the one-dimensional convolutional neural network.

[0079] The construction process of the above-described one-dimensional convolutional neural network is the same as that in Example 3, and will not be repeated here. Using a single frequency sequence {f}...1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} represents fault samples. We train the weights of a convolutional neural network and record the convergence speed and recognition accuracy of the convolutional neural network training.

[0080] Assuming the single frequency sequences are ordered from worst to best based on the training and testing results of the convolutional neural network, the order is {f}. 1RWMB (p)}、{f 1RW (p)}、{f 2RW (p)}、{f C (p)}、{f 2RWMB (p)}、{f 2R (p)}、{f 1R (p)}. Then the frequency sequence arrangement in the two-dimensional matrix is: the leftmost column of the two-dimensional matrix is ​​{f 1RWMB (p)}, and to the right of this are {f} 1RW (p)}、{f 2RW (p)}、{f C (p)}、{f 2RWMB (p)}、{f 2R (p)}、{f 1R (p)}, the result is shown in formula (10).

[0081]

[0082] Example 5:

[0083] This embodiment is based on an underwater thruster fault identification method based on channel expansion and sequence fusion from Embodiments 1 to 4. Fault experiments are conducted on an underwater robot thruster, and a two-dimensional convolutional neural network fault diagnosis model is constructed. Five fault severity levels are set: 0%, 10%, 20%, 30%, and 40%, where 0% indicates no thruster fault. The specific steps are as follows:

[0084] (1) As Figure 5 , 6 As shown, the collected experimental data mainly include the thruster control voltage change rate and the longitudinal velocity of the underwater robot. The experimental data were extracted using a time window of length L = 400 time steps, and the extracted data were used as fault samples. The time window of length L = 400 time steps was slid to the right, and a fault sample was constructed for each time step. Ultimately, each fault level contained 100 fault samples. Half of the fault samples were selected as training samples, and the other half as test samples.

[0085] like Figure 7 As shown, taking a 30% thruster failure level as an example, the speed signal and control voltage change rate of the first to 400 time cycles are used to obtain the time-domain sequences {T1(k)} and {T2(k)} of the two dynamic signals.

[0086] (2) Process the obtained time-domain sequence to achieve time-domain sequence channel expansion, specifically:

[0087] (2.1) The original time series {T1(k)} and {T2(k)} of the fault samples are subjected to mean-removing processing to obtain the mean-removed time-domain sequence {T1(k)}. 1R (k)}、{T 2R (k)};

[0088] (2.2) For the mean-removed time-domain sequence {T} 1R (k)}、{T 2R (k)} is decomposed into wavelet scale component sequence {T}. 1RW (k)}、{T 2RW (k)};

[0089] (2.3) For wavelet scale component sequences {T 1RW (k)}、{T 2RW (k)} is processed using the modified Bayesian algorithm to obtain the enhanced sequence {T}. 1RWMB (k)}、{T 2RWMB (k)};

[0090] (2.4) Enhance the sequence {T} 1RWMB (k)}、{T 2RWMB (k)} Based on evidence theory, a fusion sequence C(A) is obtained through fusion processing. k );

[0091] (3) The time series {T 1R (k)}、{T 2R (k)}、{T 1RW (k)}、{T 2RW (k)}、{T 1RWMB (k)}、{T 2RWMB (k)}、{C(A k Perform Fourier transforms on each of the} to obtain the frequency sequence {f}. 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)};

[0092] (4) Based on the frequency sequence arrangement order of step (4) in Examples 1 to 4, four different two-dimensional matrices are respectively fused.

[0093] (5) Using four different two-dimensional matrices of the training samples as input to the convolutional neural network fault diagnosis model, and the fault severity level as the output of the two-dimensional convolutional neural network, the parameters of the two-dimensional convolutional neural network are trained. Based on the trained fault diagnosis model, the test samples are classified by fault level and the fault severity is identified. The convolutional neural network structure parameters are set as shown in Table 1.

[0094] Table 1 Convolutional Neural Network Structure Parameters

[0095]

[0096] Diagnostic results of Examples 1 to 4 are as follows Figure 8 , 9 As shown.

[0097] like Figure 8 (a) shows a comparison of the entropy curves of the seven obtained single-frequency sequences, as follows: Figure 9 (a) shows a comparison of the accuracy curves of the seven obtained single-frequency sequences, with the single-frequency sequence {f 1R (p)}、{f 2R (p)}、{f 1RW (p)}、{f 2RW (p)}、{f 1RWMB (p)}、{f 2RWMB (p)}、{f C (p)} were used as fault samples to train and test the convolutional neural network fault diagnosis model. The convergence time and recognition accuracy of the convolutional neural network are shown in Table 2.

[0098] Table 2. Ranking of Diagnostic Effectiveness of Different Single Frequency Sequences

[0099]

[0100]

[0101] According to Table 2, the order of each single frequency sequence in step (4) of implementation 1 to 4 is shown in Table 3.

[0102] Table 3 shows the order of each individual frequency sequence.

[0103]

[0104] like Figure 8(b) shows a comparison of the entropy curves of the four two-dimensional matrices obtained in Examples 1 to 4, as follows: Figure 9 (b) shows a comparison of the accuracy curves of the four two-dimensional matrices obtained in Examples 1 to 4. The four fused two-dimensional matrices were used as fault samples for training and diagnosis of the convolutional neural network fault diagnosis model. The results show that the convergence curve of the single frequency sequence exhibits large fluctuations and poor identification performance. Compared to the single frequency sequence, the convergence speed of the two-dimensional matrices in Examples 1 to 4 is significantly faster, and the convergence curves are more stable.

[0105] like Figure 8 (c) shows the {T} 1R (k)} and {T 2R The fusion matrix of (k)} is used as a comparison of the entropy curve of the known method 1 for fault samples, as shown in the figure. Figure 9 (c) shows the {T} 1R (k)} and {T 2R A comparison of the accuracy curves of the known method 1, which uses the fusion matrix of (k)} as fault samples.

[0106] Furthermore, the training and diagnostic results of single-frequency sequence convolutional neural networks with kernel sizes of 21*1, 31*1, 41*1, 51*1, and 61*1 are shown in Table 3. The training and diagnostic results of two-dimensional matrix convolutional neural networks with kernel sizes of 21*2, 31*2, 41*2, 51*2, 61*2, 21*7, 31*7, 41*7, 51*7, and 61*7 are shown in Table 4.

[0107] Table 3 shows the convergence speed of fault identification with different kernel sizes under different single frequency sequences.

[0108]

[0109] Table 4 shows the convergence speed of fault identification under different convolution kernel sizes for different two-dimensional matrices.

[0110]

[0111] As shown in Tables 3 and 4, taking convolution kernel sizes of 51*2, 51*1, and 51*7 as examples, the known method uses {T 1R (k)} and {T 2RThe fusion matrix of (k)} used as a fault sample for fault identification exhibited divergence. With a convolution kernel size of 51*2, its convergence time was divergent, with a highest accuracy of 87%, indicating poor convergence performance. The convergence time for a single frequency sequence was 200–1520 time ticks, with a convergence accuracy of 100%, which was slow. Compared to known methods and fault identification using a single frequency sequence, the convergence times for Examples 1, 2, 3, and 4 were 250, 160, 140, and 140 time ticks respectively, with 100% convergence accuracy. The method in these examples showed faster convergence and better identification results compared to known methods and single-frequency-sequence fault identification methods. Compared to directly using time-domain sequences or single-frequency sequences as fault samples for fault diagnosis, this method effectively improves convergence speed and accuracy. It is effective for time-series channel expansion and frequency-sequence fusion.

Claims

1. A method for underwater thruster fault identification based on channel expansion and sequence fusion, characterized in that, The method comprises the following steps: (1) collecting a plurality of dynamic signals of the underwater robot, and obtaining time domain sequences of the plurality of dynamic signals; (2) sequentially performing, on each time domain sequence obtained, the following processing in sequence: de-meaning, wavelet decomposition, modified Bayesian algorithm and evidence theory fusion, so as to realize sequence channel expansion; the specific content of the evidence theory fusion processing is as follows: (2.4.1) Determining a framework of identification of propeller faults ; (2.4.2) The belief assignment function of the fault evidence of the fusion sequence is calculated according to the belief assignment function of the fault evidence of each sequence processed according to the revised Bayesian algorithm : ； wherein , assigning a credibility function to the different sequences of failure evidence; (2.4.3) Computing the support of the fusion sequence credibility assignment function : ； wherein i and j are focus element sequence numbers, and the value range is 1~L; (2.4.4) Calculate the confidence value time series from the support values calculated in step (2.4.3) and as the final fused sequence: ； (3) performing Fourier transform on each sequence after de-meaning processing, each sequence after wavelet decomposition processing, each sequence after modified Bayesian algorithm processing and the fusion sequence after evidence theory fusion processing, respectively, to obtain corresponding frequency sequences; (5) arranging and fusing the frequency sequences to obtain a two-dimensional matrix; the arrangement order of the frequency sequences includes the order of channel expansion processing, the order of signal types, the order of convolutional neural network training and test results from good to bad or the order of convolutional neural network training and test results from bad to good; (6) taking the two-dimensional matrix as input, performing fault level classification through a fault diagnosis model to obtain the propeller fault degree, and the fault diagnosis model is established through underwater propeller fault test.

2. The underwater propulsor fault identification method of claim 1, wherein, In the step (5), the frequency sequences are arranged according to the order of channel expansion processing, that is, each column of the two-dimensional matrix fused is in turn the frequency sequence corresponding to each sequence after de-meaning processing, the frequency sequence corresponding to each sequence after wavelet decomposition processing, the frequency sequence corresponding to each sequence after modified Bayesian algorithm processing and the frequency sequence of the fusion sequence.

3. The underwater thruster fault identification method of claim 1, wherein, In the step (5), the frequency sequences are arranged according to the order of signal types, that is, each column of the two-dimensional matrix fused is in turn the frequency sequence corresponding to the sequence after de-meaning processing of the first dynamic signal, the frequency sequence corresponding to the sequence after wavelet decomposition processing, the frequency sequence corresponding to the sequence after modified Bayesian algorithm processing, the frequency sequence corresponding to the sequence after de-meaning processing of the second dynamic signal, the frequency sequence corresponding to the sequence after wavelet decomposition processing, the frequency sequence corresponding to the sequence after modified Bayesian algorithm processing, …, the frequency sequence corresponding to the sequence after de-meaning processing of the nth dynamic signal, the frequency sequence corresponding to the sequence after wavelet decomposition processing, the frequency sequence corresponding to the sequence after modified Bayesian algorithm processing and the frequency sequence of the fusion sequence.

4. The underwater thruster fault identification method of claim 1, wherein, In the step (5), the frequency sequences are arranged according to the order of convolutional neural network training and test results from good to bad; a one-dimensional convolutional neural network is constructed, the frequency sequence is taken as a fault sample input, the propeller fault level is taken as an output, a cross-entropy loss function is constructed, the convolutional neural network weight value is updated based on the gradient descent method, the convergence speed and recognition accuracy of the convolutional neural network training are recorded, and the convolutional neural network training and test results are obtained.

5. The underwater thruster fault identification method of claim 1, wherein, In the step (5), the frequency sequences are arranged according to the order of convolutional neural network training and test results from bad to good; a one-dimensional convolutional neural network is constructed, the frequency sequence is taken as a fault sample input, the propeller fault level is taken as an output, a cross-entropy loss function is constructed, the convolutional neural network weight value is updated based on the gradient descent method, the convergence speed and recognition accuracy of the convolutional neural network training are recorded, and the convolutional neural network training and test results are obtained.

6. The method according to any one of claims 1 to 5, characterized in that The specific content of the fault diagnosis model is: performing a fault test on the underwater thruster with a known fault degree level, obtaining a plurality of dynamic signals of the underwater robot corresponding to the fault degree level; performing time domain sequence channel expansion and sequence fusion on the obtained plurality of dynamic signals, taking the fused two-dimensional matrix as a fault sample, taking the fault degree level as the output of the two-dimensional convolutional neural network, training the two-dimensional convolutional neural network parameters using the fault sample, and obtaining the fault diagnosis model of the two-dimensional convolutional neural network after the training is completed.

7. A recognition system for the method of diagnosing a failure of the underwater propeller according to any one of claims 1 to 6, characterized by, The method comprises a collection module configured to collect a plurality of dynamic signals of the underwater robot and obtain time domain sequences of the plurality of dynamic signals; The signal processing module is configured to perform, in sequence, mean removal, wavelet decomposition, modified Bayesian algorithm processing, and evidence theory fusion processing on each time domain sequence to realize sequence channel expansion; The signal fusion module is configured to perform Fourier transform on each sequence after the mean removal processing, each sequence after the wavelet decomposition processing, each sequence after the modified Bayesian algorithm processing, and each fusion sequence after the evidence theory fusion processing to obtain corresponding frequency sequences; and arrange and fuse the frequency sequences to obtain a two-dimensional matrix. The fault recognition module is configured to take the two-dimensional matrix as input, perform fault level classification through the fault diagnosis model, and obtain a thruster fault degree, wherein the fault diagnosis model is established through an underwater thruster fault test.

8. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 6.

Images