Underwater propeller fault diagnosis method and diagnosis system based on time-frequency power spectrum reconstruction

Abstract

The application discloses a kind of underwater propeller fault diagnosis methods based on time-frequency power spectrum reconstruction, obtain the time series of underwater robot dynamic signal;The time-frequency power spectrum of time series is calculated;Instantaneous spectrum is constructed for probability density function;Instantaneous shannon entropy is calculated, and instantaneous shannon entropy curve is formed;The position where minimum in instantaneous shannon entropy curve is determined;Then time-frequency power spectrum is reconstructed;With reconstructed time-frequency power spectrum as input, fault diagnosis model is carried out to fault grade classification, and propeller fault degree is obtained, and the fault diagnosis model is established by underwater propeller fault test.It is reduced that the influence that the position of fault information in monitoring signal sequence changes occurs through the time-frequency power spectrum of reconstructed propeller dynamic signal, iteration number is effectively reduced, and propeller fault degree identification classification accuracy is improved.

Description

Technical Field

[0001] This invention relates to underwater robot thruster fault detection, specifically to an underwater thruster fault diagnosis method and system based on time-frequency power spectrum reconstruction. Background Technology

[0002] For exploring the vast and deep oceans or lakes, intelligent and automated underwater robots are the primary choice for underwater environment detection tools. The proper functioning of the underwater robot's thrusters is fundamental to the AUV's ability to complete its missions. Therefore, thruster fault diagnosis has become an essential technical means to ensure the successful, safe, stable, and long-term execution of underwater robots.

[0003] In the prior art, Chinese patent application CN109977920A discloses a fault diagnosis method for hydro-turbine units based on time-frequency graphs and convolutional neural networks. This method uses the vibration signal of the hydro-turbine unit as the input of the CNN to identify the fault type, effectively characterizing the time-frequency features of the vibration signal and realizing intelligent fault diagnosis. However, the processing of time-frequency features is relatively cumbersome, making it difficult to achieve efficient fault diagnosis. Furthermore, it only analyzes two-dimensional time-frequency features, resulting in limited convergence speed and recognition accuracy of the convolutional neural network.

[0004] For example, Chinese patent application CN113916535A discloses a bearing diagnosis method, system, device and medium based on time frequency and CNN. This method uses convolutional neural networks to perform deep learning on the signals of bearings with variable speed, which effectively realizes the classification of different fault types and avoids the complex signal processing steps. However, the fault feature extraction of this method is relatively cumbersome, and it does not consider the influence of hyperparameters on the convolutional neural network model, resulting in limited convergence speed and fault identification and classification accuracy during the model training process. Summary of the Invention

[0005] Purpose of the invention: To address the above-mentioned shortcomings, this invention provides a fault diagnosis method for underwater thrusters based on time-frequency power spectrum reconstruction, which improves the efficiency and accuracy of fault identification.

[0006] The present invention also provides an underwater thruster fault diagnosis system based on time-frequency power spectrum reconstruction.

[0007] Technical Solution: To solve the above problems, this invention employs a fault diagnosis method for underwater thrusters based on time-frequency power spectrum reconstruction, comprising the following steps:

[0008] (1) Collect dynamic signals of the underwater robot and obtain the time series of the dynamic signals;

[0009] (2) The time-frequency power spectrum of the time series is calculated based on the smooth pseudo-Wigner-Willi distribution algorithm;

[0010] (3) Construct the probability density function for the instantaneous spectrum of each time beat in the time-frequency power spectrum;

[0011] (4) Calculate the instantaneous Shannon entropy for each time step according to the constructed probability density function, connect the instantaneous Shannon entropy of all time steps together to form the instantaneous Shannon entropy curve; and determine the location of the minimum value in the instantaneous Shannon entropy curve.

[0012] (5) Reconstruct the time-frequency power spectrum based on the location of the minimum value in the instantaneous Shannon entropy curve: If the location of the minimum value in the instantaneous Shannon entropy curve is located in the latter half of the time-frequency power spectrum, then a portion of the sequence in the first half of the time-frequency power spectrum is extracted, and the extracted portion of the sequence is placed at the end of the extracted time-frequency power spectrum to form the reconstructed time-frequency power spectrum; If the location of the minimum value in the instantaneous Shannon entropy curve is located in the first half of the time-frequency power spectrum, then a portion of the sequence in the latter half of the time-frequency power spectrum is extracted, and the extracted portion of the sequence is placed at the beginning of the extracted time-frequency power spectrum to form the reconstructed time-frequency power spectrum;

[0013] (6) Using the reconstructed time-frequency power spectrum as input, the fault level is classified through the fault diagnosis model to obtain the degree of thruster fault. The fault diagnosis model is established through underwater thruster fault test.

[0014] Furthermore, at least two dynamic signals from the underwater robot are collected, and the time-frequency power spectrum of each dynamic signal is reconstructed to obtain multiple reconstructed time-frequency power spectra. These reconstructed time-frequency power spectra are then fused to obtain a three-dimensional time-frequency power spectrum matrix. Using the three-dimensional time-frequency power spectrum matrix as input, a three-dimensional convolutional neural network fault diagnosis model is used to classify the fault level and obtain the degree of thruster fault. The three-dimensional convolutional neural network fault diagnosis model is established through underwater thruster fault experiments.

[0015] Furthermore, multiple reconstructed time-frequency power spectra were obtained. , … ,by For the first layer, For the second layer, ... The Q-th layer is fused into a three-dimensional time-frequency power spectrum matrix. .

[0016] Furthermore, the characteristic is that the time-frequency power spectrum in step (2) The calculation formula is:

[0017]

[0018] Where z(n) is a time series The analytical signal, z*(n) is the complex conjugate of z(n), |·| is the absolute value function, n is the time series index, m is the frequency series index, h(k) and g(l) represent the smoothing window functions in the frequency and time directions, respectively, N1 represents the number of bins in the frequency series, and j represents the imaginary number. , M can be an integer.

[0019] The probability density in step (3) The construction formula is:

[0020]

[0021] In step (4), the instantaneous Shannon entropy The calculation formula is:

[0022]

[0023] Furthermore, the specific content of reconstructing the time-frequency power spectrum in step (4) is as follows: determining the minimum value in the instantaneous Shannon entropy curve. Location According to location Reconstruct the time-frequency power spectrum TF(m, n);

[0024] If position Meet the conditions Then, extract the first column to the second column of the time-frequency power spectrum TF(m, n). Time-frequency power spectrum of the column The truncated time-frequency power spectrum is then placed at the end of the truncated time-frequency power spectrum to form a reconstructed time-frequency power spectrum:

[0025] ;

[0026] If position The conditions are not met. Then, extract the nth element from the time-frequency power spectrum TF(m, n). Time-frequency power spectrum from column L to column L Then, the truncated time-frequency power spectrum is placed at the front of the truncated time-frequency power spectrum to form a reconstructed time-frequency power spectrum:

[0027] .

[0028] Furthermore, the specific content of establishing the fault diagnosis model is as follows: conduct fault tests on the underwater thruster with known fault severity levels to obtain dynamic signals of the underwater robot corresponding to the fault severity level; reconstruct the time-frequency power spectrum of the obtained dynamic signals, use the reconstructed time-frequency power spectrum as fault samples, use it as input to the two-dimensional convolutional neural network, use the fault severity level as output of the two-dimensional convolutional neural network, train the parameters of the two-dimensional convolutional neural network using fault samples, and obtain the fault diagnosis model of the two-dimensional convolutional neural network after training.

[0029] Furthermore, the specific content of establishing the three-dimensional convolutional neural network fault diagnosis model is as follows: a fault test with a known fault level is conducted on the underwater thruster to obtain at least two dynamic signals of the underwater robot corresponding to the fault level; the obtained dynamic signals are reconstructed and fused using time-frequency power spectrum to obtain a three-dimensional time-frequency power spectrum matrix; the three-dimensional time-frequency power spectrum matrix is ​​used as a fault sample as the input of the three-dimensional convolutional neural network, and the fault level is used as the output of the three-dimensional convolutional neural network; the parameters of the three-dimensional convolutional neural network are trained using the fault samples; after training, the fault diagnosis model of the three-dimensional convolutional neural network is obtained.

[0030] The present invention also employs an underwater thruster fault diagnosis system based on time-frequency power spectrum reconstruction, including an acquisition module for acquiring dynamic signals of the underwater robot and obtaining the time-domain sequence of the dynamic signals;

[0031] Signal processing module: This module calculates the time-frequency power spectrum of a time series based on a smoothed pseudo-Wigner-Willy distribution algorithm; constructs a probability density function for the instantaneous spectrum of each time step in the time-frequency power spectrum; calculates the instantaneous Shannon entropy of each time step based on the constructed probability density function; connects the instantaneous Shannon entropies of all time steps to form an instantaneous Shannon entropy curve; and determines the location of the minimum value in the instantaneous Shannon entropy curve.

[0032] The reconstruction module is used to reconstruct the time-frequency power spectrum based on the location of the minimum value in the instantaneous Shannon entropy curve: if the location of the minimum value in the instantaneous Shannon entropy curve is located in the latter half of the time-frequency power spectrum, a portion of the sequence in the first half of the time-frequency power spectrum is extracted, and the extracted portion of the sequence is placed at the end of the extracted time-frequency power spectrum to form the reconstructed time-frequency power spectrum; if the location of the minimum value in the instantaneous Shannon entropy curve is located in the first half of the time-frequency power spectrum, a portion of the sequence in the latter half of the time-frequency power spectrum is extracted, and the extracted portion of the sequence is placed at the beginning of the extracted time-frequency power spectrum to form the reconstructed time-frequency power spectrum.

[0033] The fault identification module is used to classify the fault level by taking the reconstructed time-frequency power spectrum as input and using the fault diagnosis model to obtain the degree of thruster fault. The fault diagnosis model is established through underwater thruster fault tests.

[0034] Beneficial effects: Compared with the prior art, the significant advantage of this invention is that by reconstructing the time-frequency power spectrum of the thruster dynamic signal, the impact of changes in the position of fault information in the monitoring signal sequence is reduced, the number of iterations is effectively reduced, and the accuracy of thruster fault degree identification and classification is improved. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the diagnostic method for time-frequency power spectrum reconstruction in this invention.

[0036] Figure 2 This is a schematic diagram of the diagnostic method for time-frequency power spectrum reconstruction and fusion in this invention.

[0037] Figure 3 This is a waveform diagram of the thruster control voltage change rate signal collected by this invention.

[0038] Figure 4 This is a waveform diagram of the longitudinal velocity signal of the underwater robot collected by this invention.

[0039] Figure 5 This is a schematic diagram of the process of reconstructing and fusing the time-frequency power spectrum of the acquired signal in this invention.

[0040] Figure 6 This is a schematic diagram comparing the fault degree of the fault degree diagnosis results for the reconstruction, reconstruction and fusion of the time-frequency power spectrum of velocity signals in this invention.

[0041] Figure 7 This is a schematic diagram comparing the fault degree of the fault degree diagnosis results for the reconstruction, reconstruction and fusion of the time-frequency power spectrum of voltage signals in this invention. Detailed Implementation

[0042] like Figure 1 As shown in the figure, this embodiment of the underwater thruster fault diagnosis method based on time-frequency power spectrum reconstruction includes the following steps:

[0043] The first step involves using a sliding time window of length L to capture the dynamic signals of the underwater robot, and then performing mean-removing processing on the acquired dynamic signals to obtain a time series. Dynamic signals include the rate of change of thruster control voltage and the longitudinal speed of the underwater robot.

[0044] The second step involves calculating the time series based on a smoothed pseudo-Wigner-Willi distribution algorithm. The time-frequency power spectrum TF(m, n) is calculated using the following formula:

[0045] (1)

[0046] In the formula, z(n) is a time series. The analytic signal is given by z*(n), which is the complex conjugate of z(n), |·| is the absolute value function, n is the time series index, m is the frequency series index, h(k) and g(l) represent the smoothing window functions in the frequency and time directions, respectively, N1 represents the number of bins in the frequency series, and j represents the imaginary number. , M can be an integer.

[0047] The third step is to construct the instantaneous probability density function, that is, to construct the probability density function using the instantaneous spectrum of each time beat in the time-frequency power spectrum. The construction formula is as follows:

[0048] (2)

[0049] In the formula, n is the time series number and m is the frequency series number;

[0050] The fourth step is to calculate the instantaneous Shannon entropy. Calculate the instantaneous Shannon entropy for each time beat according to formula (3), connect the Shannon entropies of all time beats together to form the instantaneous Shannon entropy curve, and determine the location of the minimum value in the instantaneous Shannon entropy curve.

[0051] (3)

[0052] Fifth, determine the minimum value in the instantaneous Shannon entropy curve based on the results of step four. Location According to location For the original time-frequency power spectrum Refactor.

[0053] If position Meet the conditions Then the original time-frequency power spectrum It can be represented as:

[0054] Extract the first column of the time-frequency power spectrum to Column data Then extract the data The reconstructed time-frequency power spectrum is formed by placing it at the very end of the truncated time-frequency power spectrum. :

[0055] ;

[0056] If position Meet the conditions Then the original time-frequency power spectrum It can be represented as:

[0057] ,

[0058] Extract the original time-frequency power spectrum The Data from column L Then extract the data The reconstructed time-frequency power spectrum is formed by placing it at the very front of the truncated time-frequency power spectrum. :

[0059] .

[0060] Step 7: Reconstruct the time-frequency power spectrum As input, the fault level is classified through the fault diagnosis model to obtain the degree of thruster fault.

[0061] The fault diagnosis model was established through underwater thruster fault tests. Specifically, the following steps were taken: fault tests were conducted on the underwater thruster at known fault levels to obtain dynamic signals of the underwater robot at the corresponding fault levels; the time-frequency power spectrum of the obtained dynamic signals was reconstructed, and the reconstructed time-frequency power spectrum was used as a fault sample as the input of a two-dimensional convolutional neural network (2D convolutional neural network), with the fault level as the output of the 2D convolutional neural network. The parameters of the 2D convolutional neural network were trained using the fault samples, and after training, the fault diagnosis model of the 2D convolutional neural network was obtained.

[0062] like Figure 2 As shown in this embodiment, a deep learning-based fault diagnosis method for underwater thrusters based on time-frequency power spectrum reconstruction and fusion is described. Based on the diagnostic steps of time-frequency power spectrum reconstruction, at least two dynamic signals of the underwater robot are collected, and the time-frequency power spectrum of each dynamic signal is reconstructed to obtain multiple reconstructed two-dimensional time-frequency power spectra. , … Multiple reconstructed two-dimensional time-frequency power spectra are fused to obtain a three-dimensional time-frequency power spectrum matrix. Using the three-dimensional time-frequency power spectrum matrix as input, a three-dimensional convolutional neural network fault diagnosis model is used to classify the fault level and obtain the degree of thruster fault.

[0063] The two-dimensional time-frequency power spectrum is fused into a three-dimensional time-frequency power spectrum matrix. ,in For the first layer, For the second layer, ... For the Q-th layer, that is,

[0064] ,

[0065] ,

[0066] ...

[0067] .

[0068] The specific content of establishing the three-dimensional convolutional neural network fault diagnosis model is as follows: A fault test with a known fault level is conducted on the underwater thruster to obtain at least two dynamic signals of the underwater robot corresponding to the fault level; the obtained dynamic signals are reconstructed and fused using time-frequency power spectra to obtain a three-dimensional time-frequency power spectrum matrix; the three-dimensional time-frequency power spectrum matrix is ​​used as a fault sample as the input of the three-dimensional convolutional neural network, and the fault level is used as the output of the three-dimensional convolutional neural network; the parameters of the three-dimensional convolutional neural network are trained using the fault samples; after training, the fault diagnosis model of the three-dimensional convolutional neural network is obtained.

[0069] like Figure 3 and Figure 4 As shown, a fault experiment with known fault levels was conducted on the underwater thruster, obtaining experimental data with a sampling length of 500 time ticks. The experimental data mainly includes dynamic signals such as the thruster control voltage and the longitudinal velocity of the underwater robot. A time window of 400 seconds was set to truncate the experimental data, and the truncated data was used as fault samples. For each time tick to the right of this time window, a fault sample was constructed. Finally, 100 fault samples were constructed for each fault level. 50% of the fault samples were randomly selected as training samples, and the remaining 50% were used as test samples.

[0070] like Figure 5 As shown in (a), taking the longitudinal velocity signal from time 1 to 400 of a thruster with a failure degree of 30% as an example, the time series is obtained by removing the mean from the failure sample. .like Figure 5 As shown in (b), the time series is calculated based on the smoothed pseudo-Wigner-Willi distribution algorithm. The time-frequency power spectrum TF(m,n).

[0071] like Figure 5 As shown in (c), the instantaneous probability density function is constructed using the time-frequency power spectrum, and the instantaneous Shannon entropy is calculated to obtain the instantaneous Shannon entropy curve. In this figure, the minimum value of the instantaneous Shannon entropy curve is shown. The value is 4.009, and the time signature is... =172.

[0072] like Figure 5 As shown in (d), the first element in the original time-frequency power spectrum TF(m,n) is... The data from beat 372 to beat L=400 is truncated and placed at the beginning of the original time-frequency power spectrum TF(m,n) to obtain the reconstructed time-frequency power spectrum. time-frequency power spectrum As fault samples, they are input into a two-dimensional convolutional neural network to classify fault levels and identify the degree of fault.

[0073] like Figure 5 As shown in (e), the two-dimensional time-frequency power spectra of different types of dynamic signals are fused together to form a three-dimensional time-frequency power spectrum matrix. The three-dimensional time-frequency power spectrum matrix is ​​used as a fault sample and input into a three-dimensional convolutional neural network fault diagnosis model to classify the fault level and identify the degree of fault.

[0074] The parameters of the convolutional neural network structure are shown in Table 1.

[0075] Table 1 Convolutional Neural Network Structure Parameters

[0076]

[0077] In this embodiment Figure 1 and Figure 2 The diagnostic results of the two methods are the same as those of direct diagnosis. Figure 5 (b) shows the diagnostic results of the fault diagnosis model trained using the time-frequency power spectrum as a fault sample. Figure 6 and Figure 7 As shown.

[0078] directly with Figure 5 (b) shows the diagnostic method of using the time-frequency power spectrum as a fault sample to train the fault diagnosis model. The original time-frequency power spectrum is directly used as the fault sample for diagnosis and training. The convergence time is 500 time ticks and the convergence accuracy is 96.4%. Figure 1 The method shown uses the reconstructed time-frequency power spectrum as fault samples for diagnosis and training, with a convergence time of 136 time ticks and a convergence accuracy of 100%. Figure 2 The method shown uses the fused three-dimensional time-frequency power spectrum matrix as a fault sample for training and diagnosis, with a convergence time of 40 time ticks and a convergence accuracy of 100%. Compared to directly using the original time-frequency power spectrum as a fault sample, using the reconstructed time-frequency power spectrum as a fault sample results in faster convergence speed and 100% accuracy. Compared to using the reconstructed time-frequency power spectrum as a fault sample, the fused three-dimensional time-frequency power spectrum matrix as a fault sample provides better fault identification, significantly faster convergence speed, and 100% accuracy, with the best fault severity identification results.

[0079] Furthermore, the training and diagnostic results of the original time-frequency power spectrum convolutional neural networks with kernel sizes of 11*11, 21*21, 31*31, 41*41, 51*51, and 61*61 are shown in Table 2. The training and diagnostic results of the three-dimensional time-frequency power spectrum convolutional neural networks with kernel sizes of 21*21*2, 31*31*2, 41*41*2, 51*51*2, and 61*61*2 are also shown in Table 2.

[0080] Table 2 shows the convergence time and accuracy of fault identification with different kernel sizes under different time-frequency power spectra.

[0081]

[0082] As shown in Table 2, the convergence time of the diagnostic method using the original time-frequency power spectrum as the fault sample is 198 to 500 time ticks, and the convergence accuracy is 90.8% to 100%. The convergence time of the diagnostic method using the reconstructed time-frequency power spectrum as the fault sample is 56 to 161 time ticks, and the convergence accuracy is 100%. The convergence time of the diagnostic method using the fused time-frequency power spectrum matrix as the fault sample is 40 to 89 time ticks, and the convergence accuracy is 100%. Furthermore, under the same convolution kernel size, the convergence speed of the diagnostic method using the fused time-frequency power spectrum matrix as the fault sample is faster than that of the method using the reconstructed time-frequency power spectrum.

[0083] The time-frequency power spectrum reconstruction and fusion shown in this embodiment is effective, significantly improving convergence speed and identification accuracy. The three-dimensional convolutional neural network can fully utilize time-domain, frequency-domain, and time-frequency-domain information, learning higher-level information that two-dimensional convolutional neural networks cannot. Furthermore, the reconstruction and fusion of the signal's time-frequency power spectrum not only effectively reduces the number of iterations but also further improves the classification accuracy of fault identification.

Claims

1. A method for fault diagnosis of underwater propeller based on time-frequency power spectrum reconstruction, characterized in that, The method comprises the following steps: (1) collecting dynamic signals of the underwater robot to obtain time series of the dynamic signals; (2) calculating time-frequency power spectrum of the time series based on a smooth pseudo-Vigener-Wille distribution algorithm; (3) constructing a probability density function for the instantaneous spectrum of each time beat in the time-frequency power spectrum; (4) calculating the instantaneous Shannon entropy of each time beat according to the constructed probability density function, connecting the instantaneous Shannon entropies of all time beats together to form an instantaneous Shannon entropy curve, and determining the position of the minimum value in the instantaneous Shannon entropy curve; (5) reconstructing the time-frequency power spectrum according to the position of the minimum value in the instantaneous Shannon entropy curve: if the position of the minimum value in the instantaneous Shannon entropy curve is located in the latter half of the time-frequency power spectrum, then part of the sequence in the former half of the time-frequency power spectrum is intercepted, and the intercepted part of the sequence is placed at the rear end of the intercepted time-frequency power spectrum to form a reconstructed time-frequency power spectrum; if the position of the minimum value in the instantaneous Shannon entropy curve is located in the former half of the time-frequency power spectrum, then part of the sequence in the latter half of the time-frequency power spectrum is intercepted, and the intercepted part of the sequence is placed at the front end of the intercepted time-frequency power spectrum to form a reconstructed time-frequency power spectrum; (6) inputting the reconstructed time-frequency power spectrum into a fault diagnosis model to perform fault level classification and obtain the fault degree of the propeller, wherein the fault diagnosis model is established through underwater propeller fault tests.

2. The underwater propeller fault diagnosis method according to claim 1, characterized by, At least two kinds of dynamic signals of the underwater robot are collected, the time-frequency power spectrum of each kind of dynamic signal is reconstructed, a plurality of reconstructed time-frequency power spectrums are obtained, and fusion is performed to obtain a three-dimensional time-frequency power spectrum matrix. The three-dimensional time-frequency power spectrum matrix is inputted into a three-dimensional convolutional neural network fault diagnosis model to perform fault level classification and obtain the fault degree of the propeller, wherein the three-dimensional convolutional neural network fault diagnosis model is established through underwater propeller fault tests.

3. The underwater propeller fault diagnosis method according to claim 2, characterized by, Multiple reconstructed time-frequency power spectra were obtained. , … ,by For the first layer, For the second layer, ... The Q-th layer is fused into a three-dimensional time-frequency power spectrum matrix. .

4. The underwater propeller fault diagnosis method according to claim 1 or 2 or 3, characterized by, The step (2) calculates the time-frequency power spectrum The formula is: where z(n) is a time series of the analytic signal, z*(n) is the complex conjugate of z(n), |·| is the absolute value function, n is the time series index, m is the frequency series index, h(k) and g(l) are smoothing window functions in the frequency and time directions, respectively, N1 is the number of frequency series bins, and j is the imaginary unit, , M is an integer.

5. The underwater propeller fault diagnosis method according to claim 4, characterized in that, The probability density in the step (3) The configuration formula is: 。 6. The underwater propeller fault diagnosis method according to claim 5, characterized by, The formula for calculating the instantaneous Shannon entropy in step (4) is: 。 7. The underwater propeller fault diagnosis method according to claim 1 or 2 or 3, characterized by, The step (4) of reconstructing the time-frequency power spectrum comprises: determining the minimum value in the instantaneous Shannon entropy curve The location According to the location The time-frequency power spectrum TF(m, n) is reconstructed; If the position satisfies the condition , then the first column to the column of the time-frequency power spectrum TF(m, n) is intercepted , and the intercepted time-frequency power spectrum is placed at the rear end of the intercepted time-frequency power spectrum to form a reconstructed time-frequency power spectrum: ； If position Meet the conditions Then, extract the nth element from the time-frequency power spectrum TF(m, n). Time-frequency power spectrum from column L to column L Then, the truncated time-frequency power spectrum is placed at the front of the truncated time-frequency power spectrum to form a reconstructed time-frequency power spectrum: 。 8. The underwater propulsor fault diagnostic method of claim 1, wherein, The specific content of establishing the fault diagnosis model is that: performing fault tests on the underwater propeller with known fault degree levels, obtaining dynamic signals of the underwater robot corresponding to the fault degree levels; reconstructing the time-frequency power spectrum of the obtained dynamic signals, taking the reconstructed time-frequency power spectrum as a fault sample, taking the fault sample as the input of a two-dimensional convolutional neural network, taking the fault degree level as the output of the two-dimensional convolutional neural network, training the parameters of the two-dimensional convolutional neural network by using the fault sample, and obtaining the fault diagnosis model of the two-dimensional convolutional neural network after the training is completed.

9. The underwater propeller fault diagnostic method according to claim 2, characterized by, The specific content of establishing the three-dimensional convolutional neural network fault diagnosis model is that: performing fault tests on the underwater propeller with known fault degree levels, obtaining at least two kinds of dynamic signals of the underwater robot corresponding to the fault degree levels; reconstructing and fusing the time-frequency power spectrum of the obtained dynamic signals to obtain a three-dimensional time-frequency power spectrum matrix, taking the three-dimensional time-frequency power spectrum matrix as a fault sample, taking the fault sample as the input of a three-dimensional convolutional neural network, taking the fault degree level as the output of the three-dimensional convolutional neural network, training the parameters of the three-dimensional convolutional neural network by using the fault sample, and obtaining the fault diagnosis model of the three-dimensional convolutional neural network after the training is completed.

10. A diagnosis system employing the method for diagnosing a failure of an underwater propeller according to any one of claims 1 to 9, characterized by, The method comprises a collecting module configured to collect dynamic signals of the underwater robot to obtain time series of the dynamic signals; The signal processing module is configured to calculate a time-frequency power spectrum of the time sequence based on a smooth pseudo-Vigener-Wille distribution algorithm; construct a probability density function for an instantaneous spectrum of each time beat in the time-frequency power spectrum; calculate an instantaneous Shannon entropy of each time beat according to the constructed probability density function; connect the instantaneous Shannons entropies of all the time beats together to form an instantaneous Shannon entropy curve; and determine a position of a minimum value in the instantaneous Shannon entropy curve; The reconstruction module is configured to reconstruct the time-frequency power spectrum according to the position of the minimum value in the instantaneous Shannon entropy curve; if the position of the minimum value in the instantaneous Shannon entropy curve is located in a latter half of the time-frequency power spectrum, then a part of sequences in a former half of the time-frequency power spectrum is intercepted, and the intercepted part of sequences is arranged at a rear end of the time-frequency power spectrum after the interception to form a reconstructed time-frequency power spectrum; if the position of the minimum value in the instantaneous Shannon entropy curve is located in the former half of the time-frequency power spectrum, then a part of sequences in a latter half of the time-frequency power spectrum is intercepted, and the intercepted part of sequences is arranged at a front end of the time-frequency power spectrum after the interception to form the reconstructed time-frequency power spectrum; The fault identification module is configured to take the reconstructed time-frequency power spectrum as an input, perform fault level classification through a fault diagnosis model to obtain a propeller fault degree, and the fault diagnosis model is established through an underwater propeller fault test.

Images