An intelligent fault diagnosis method for underwater robot propeller

Abstract

The application provides a kind of underwater robot propeller cross-domain intelligent fault diagnosis method, belongs to underwater robot fault diagnosis field. Through dynamic adversarial generative network, energy features and environmental parameters are introduced, frequency domain physical constraints are constructed, pseudo-fault data suitable for the target domain are generated, and time domain and frequency domain double discriminators are used to ensure the authenticity and rationality of the generated signal. The environmental residual adaptation network constructs a residual dynamic compensation mechanism, fuses the residual between the measured features and the pseudo-fault features of the target domain with the environment factors after encoding the environmental parameters, constructs a dynamic mapping relationship between the environmental parameters and the feature residual, adaptively generates a feature compensation amount, realizes double-domain adaptation of global distribution alignment and local residual compensation, and can effectively improve the generalization ability and diagnostic accuracy of underwater robot propeller fault diagnosis in different water environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater robot fault diagnosis, and particularly relates to a cross-domain intelligent fault diagnosis method for a propeller thruster of an underwater robot. BACKGROUND

[0002] The ocean contains rich mineral resources, marine biological resources and other energy resources, and the emergence of underwater robots provides strong protection for human exploration of complex marine environments. In the complex system architecture of underwater robots, the propeller thruster system is easily affected by ocean currents, salinity and temperature in the water area, and has become the core component with the highest failure rate. Ocean currents can change the flow field distribution around the propeller thruster, and strong ocean currents or complex flow fields can significantly increase the load of the propeller thruster, increase the working current, cause speed fluctuation or aggravate cavitation effect, and irregular ocean currents can cause the propeller thruster to bear uncertain impact load, which may cause shaft vibration to increase and bearing wear; salinity affects the density and viscosity of water, resulting in an increase in the load of the propeller thruster, baseline drift of the current signal, affecting the frequency band shift and harmonic energy distribution of the signal, and high-salinity water (such as seawater) can also accelerate the electrochemical corrosion of the metal propeller thruster, thereby increasing the surface roughness of the propeller thruster and causing abnormal vibration of the propeller thruster; different temperatures also affect the propeller thruster, and the metal material of the propeller thruster shrinks at low temperature, which can increase the friction between the propeller and the shaft and cause unbalanced vibration of the propeller thruster; and high temperature can cause the viscosity of the lubricating oil to decrease, thereby increasing the risk of mechanical wear. Moreover, the electronic components in the motor controller are easily affected by temperature, and the current sampling error is large, which can mask the real fault signal, and temperature can affect the transient impact strength and baseline noise level of the generated signal.

[0003] Current fault diagnosis results of the propeller thruster of the underwater robot are mostly obtained based on laboratory or fixed water area test environments, while in actual operation, the underwater robot needs to perform tasks in different water areas, and the data obtained in the laboratory or fixed water area is source domain data, and the data obtained in actual operation is target domain data. Due to environmental factors, the vibration spectrum characteristics and sensor data distribution of the propeller thruster of the underwater robot in the target domain data are greatly different from those in the source domain data, which presents the problems of data mismatch and feature layer offset, which makes it difficult for the model trained based on a single water area to achieve effective cross-domain fault diagnosis.

[0004] Therefore, in the fault diagnosis of the propeller thruster of the underwater robot, environmental factors and physical constraints are introduced to perform intelligent fault diagnosis with domain adaptation capability, which has become an important research direction to improve the accuracy and reliability of the fault diagnosis of the propeller thruster of the underwater robot. SUMMARY

[0005] In view of the problems existing in the prior art, the present application provides a cross-domain intelligent fault diagnosis method for a propeller thruster of an underwater robot, and specifically comprises the following steps:

[0006] Obtain source domain fault diagnosis data of the propeller thruster of the underwater robot in simulated operation in a laboratory water area, and extract source domain energy features from the source domain fault diagnosis data, wherein the source domain fault diagnosis data comprises source domain vibration data and source domain current data.

[0007] Obtain target domain fault diagnosis data and environmental parameters of the propeller thruster of the underwater robot in actual operation in a real water area, wherein the target domain fault diagnosis data comprises target domain vibration data and target domain current data, and the target domain environmental parameters comprise target domain salinity and target domain temperature.

[0008] A dynamic adversarial generative network data enhancement model is constructed, the source domain energy features, the target domain fault diagnosis data and the environmental parameters are input into the dynamic adversarial generative network data enhancement model, the dynamic adversarial generative network data enhancement model determines a frequency domain physical law constraint according to the source domain energy features and the target domain environmental parameters, and generates target domain pseudo fault data, wherein the dynamic adversarial generative network data enhancement model comprises a generator and a discriminator, the generator comprises an environmental parameter encoding module, a residual attention module and an inverse wavelet packet synthesis module, and the discriminator comprises a time domain discriminator and a frequency domain discriminator.

[0009] An environmental residual adaptation network cross-domain fault diagnosis model is constructed, the target domain pseudo fault data, the target domain fault diagnosis data and the environmental parameters are input into the environmental residual adaptation network cross-domain fault diagnosis model, the environmental residual adaptation network cross-domain fault diagnosis model extracts measured features, pseudo fault features and environmental factors of the target domain, fuses a residual error between the measured features and the pseudo fault features with the environmental factors, constructs a dynamic mapping relationship between the environmental parameters and the feature residual error, adaptively generates a feature compensation amount, performs fault diagnosis on the compensated features, and obtains a fault type, wherein the environmental residual adaptation network cross-domain fault diagnosis model comprises a feature extractor, a residual compensation module and a fault classifier, and the feature extractor comprises a shared feature extractor and a private feature extractor.

[0010] Optionally, the source domain energy features are extracted from the source domain fault diagnosis data, and specifically comprise the following steps:

[0011] Wavelet packet decomposition: the source domain fault diagnosis data is decomposed by using Daubechies wavelets to five layers of wavelet packet decomposition, and the source domain fault diagnosis data is divided into 32 frequency bands to obtain 32 frequency band wavelet packet coefficients . .

[0012] Energy feature extraction: according to the 32 frequency band wavelet packet coefficients According to the energy of each frequency band, a total of 32 frequency band source domain vibration energy feature vectors are obtained and source domain current energy feature vectors , forming a frequency band energy feature vector , as shown in formula (1):

[0013] (1).

[0014] In the formula, is the energy of the i th frequency band, is the wavelet packet coefficient of the i th frequency band, is the number of wavelet packet coefficients of the i th frequency band.

[0015] Optionally, the dynamic generative adversarial network data enhancement model determines the frequency domain physical law constraint according to the source domain energy feature and the target domain environmental parameter to generate target domain pseudo-fault data, and specifically includes the following steps:

[0016] The generator receives source domain fault diagnosis data , frequency band energy feature vectors extracted by wavelet packet decomposition , target domain salinity and temperature , and a random noise vector as input.

[0017] The environmental parameter encoding module encodes the target domain salinity and temperature , and maps the salinity and temperature to frequency band offsets to ensure that the generated signal main frequency adapts to the current water area, as shown in formula (2):

[0018] (2).

[0019] In the formula, is the frequency band index offset, is the salinity influence coefficient, is the temperature influence coefficient, and are the source domain reference environmental parameters.

[0020] The residual attention module adjusts the frequency band offset of the source domain vibration energy feature , and calculates the weight of each frequency band using the SimAM frequency domain attention mechanism to generate weighted source domain vibration energy features . ​​​​

[0021] The anti-wavelet packet synthesis module converts the weighted source domain vibration energy features into time domain signals to generate target domain pseudo-vibration data .

[0022] The random noise vector is added to the target domain pseudo-vibration data to splice an input multi-layer perceptron to generate target domain pseudo-current data .

[0023] The target domain pseudo-vibration data and the target domain pseudo-current data are synthesized into target domain pseudo-fault data .

[0024] The discriminator receives target domain fault diagnosis data and target domain pseudo-fault data.

[0025] The time domain discriminator splices the target domain pseudo-vibration data and the target domain pseudo-fault data as a first input, and the target domain fault diagnosis data as a second input, extracts time domain features through a convolutional neural network, and outputs a authenticity probability .

[0026] The frequency domain discriminator performs a fast Fourier transform on the target domain vibration data to obtain an amplitude spectrum , analyzes the target domain pseudo-vibration data and the amplitude spectrum through a convolutional neural network to obtain frequency domain distribution features, and outputs a rationality probability .

[0027] The loss function of the generator is represented as:

[0028] (3).

[0029] (4).

[0030] (5).

[0031] The loss function of the discriminator is represented as:

[0032] (6).

[0033] The objective function of the dynamic adversarial generation network data enhancement model is represented as:

[0034] ​​ (7).

[0035] wherein, is a time domain discriminator loss weight, is a frequency domain discriminator weight, is target domain fault diagnosis data, subject to a real data distribution, is a fast Fourier transform, is an expectation operation, respectively averaging the target domain fault diagnosis data and the target domain pseudo-fault data, is target domain pseudo-vibration data, is target domain pseudo-current data, is a time domain adversarial loss, is a frequency domain adversarial loss, is a physical constraint loss, is a fault energy threshold, is a frequency band weight from a SimAM frequency domain attention mechanism, is a spectrum fidelity constraint loss.

[0036] Generator by minimizing the objective function such that the generated data can deceive the discriminator into thinking that the generated data is real; the discriminator by maximizing the objective function such that can accurately distinguish between real fault data and pseudo-fault data .

[0037] Optionally, the environment residual adaptation network extracts the measured features, pseudo-fault features, and environmental factors of the target domain from the cross-domain fault diagnosis model, fuses the residual between the measured features and the pseudo-fault features of the target domain with the environmental factors, constructs a dynamic mapping relationship between the environmental parameters and the feature residuals, adaptively generates a feature compensation amount, performs fault diagnosis on the compensated features, and obtains the fault type, specifically including the following steps:

[0038] For the target domain pseudo-fault data and the target domain fault diagnosis data, the shared feature extractor uses a double-layer one-dimensional convolutional neural network, the first layer of 64 5x1 convolutional kernels extracts macro features with a step of 2, after maximum pooling compression, the second layer of 128 3x1 convolutional kernels captures microscopic fluctuations, and finally flattens and outputs a 256-dimensional target domain original fault feature vector and a target domain pseudo-fault feature vector .

[0039] For the environmental parameters, the private feature extractor adopts a fully connected network with two layers of ReLU activation to compress and encode the environmental information, and linearly maps the output to a 16-dimensional environmental factor vector .

[0040] The MMD is used to globally align the target domain original fault feature vector and the target domain pseudo fault feature vector, so as to reduce the overall distribution difference between the target domain original fault feature and the pseudo fault feature.

[0041] Residual compensation module The input is the target domain original fault feature vector , the target domain pseudo fault feature vector and the environmental factor vector.

[0042] The local feature residual remaining after the global alignment of the target domain original fault feature vector and the target domain pseudo fault feature vector is calculated, as shown in equation (8):

[0043] (8)。

[0044] The environmental factor is concatenated with the local feature residual to form a 272-dimensional vector input into the LSTM network, to establish a dynamic mapping between the environmental parameters and the feature offset , as shown in equation (9):

[0045] (9)。

[0046] In the formula, is the forget gate output vector, is the input gate output vector, is the output gate output vector, is the candidate state, is the input weight matrix, is the recurrent weight matrix, is the bias vector, is the Sigmoid function, and tanh is the hyperbolic tangent function.

[0047] The cell state is updated, as shown in equations (10) and (11):

[0048] (10)。

[0049] (11)。

[0050] In the formula, is the updated cell state at the current time, is the cell state at the previous time, is the current hidden state.

[0051] Final output gate generates 64-dimensional hidden state , Decoded by fully connected layer into 256-dimensional feature compensation .

[0052] Feature compensation The amplitude of the feature compensation is output by the LSTM network and matches the scale of the feature extractor, ensuring the direction consistency of the physical constraints and the compatibility of the feature space scale.

[0053] Apply the feature compensation to the original fault feature vector of the target domain to generate the compensated feature vector of the target domain .

[0054] The target domain feature vector generated by the residual compensation module is used as the input of the fault classifier, and the target domain feature vector is mapped to the fault class space through a fully connected layer, as shown in equation (12):

[0055] (12)。

[0056] In the formula, is the classification layer weight matrix, is the bias vector, and represents the fault probability distribution.

[0057] The feature alignment loss is represented as:

[0058] (13)。

[0059] In the formula, represents the expectation, is the target domain measured sample, is the target domain data distribution, is the minimum direct difference between the original fault features and the pseudo fault features of the target domain, MMD is the maximum mean difference, and γ is a hyperparameter.

[0060] The physical constraint loss is represented as:

[0061] (14)。

[0062] In the formula, is the regularization coefficient, is the parameter matrix of the LSTM network.

[0063] The fault classification loss is represented as:

[0064] (15)。

[0065] wherein, is the expectation of the target domain pseudo-fault data distribution, is the compensated pseudo-fault feature, is the target domain pseudo-fault data label, is the fault classification probability.

[0066] The fault classification objective function is represented as:

[0067] (16).

[0068] wherein, is the feature alignment weight; is the physical constraint weight; is the fault classification weight.

[0069] Optionally, the salinity influence coefficient and the temperature influence coefficient are calibrated through experiments, specifically including the following steps:

[0070] Step 1: In the laboratory water area, change the salinity and temperature, and set 25 groups of working conditions, wherein the salinity is set to 0%, 15%, 25%, 35%, and 45%, the temperature is set to 5℃, 10℃, 15℃, 25℃, and 35℃, and 0% salinity and 25℃ temperature are set as the source domain reference environmental parameters.

[0071] Step 2: Under each group of working conditions, control the underwater robot propeller to run at a constant speed of 1000RPM, and calculate the nominal fundamental frequency according to the formula .

[0072] Step 3: Extract the measured fundamental frequency by FFT analysis of the collected source domain vibration data.

[0073] Step 4: Calculate the fundamental frequency offset .

[0074] Step 5: Establish a multiple linear regression model .

[0075] Step 6: Fit the coefficient formula using the least squares method to obtain the salinity influence coefficient and the temperature influence coefficient .

[0076] Optionally, the fault energy threshold is calibrated through experiments, specifically including the following steps:

[0077] Step 1: In the laboratory water area, the salinity and temperature are changed, and a total of 25 groups of working conditions are set, wherein the salinity is set to 0%, 15%, 25%, 35%, and 45%, and the temperature is set to 5 DEG C, 10 DEG C, 15 DEG C, 25 DEG C, and 35 DEG C.

[0078] Step 2: 100 groups of vibration data of underwater robot propeller normal work are collected, each group of vibration data is decomposed by 5 layers of wavelet packet, each group of vibration data is divided into 32 frequency bands, the energy of each frequency band is calculated according to formula (1), and the mean value of 100 groups of energy is calculated And the standard deviation .

[0079] Step 3: the fault energy threshold value is calculated according to formula .

[0080] The application provides a cross-domain intelligent fault diagnosis method for underwater robot propeller, aiming at the problems of data distribution difference and feature layer offset in cross-domain fault diagnosis of underwater robot propeller in different salinity and temperature water environment, and proposes an underwater robot propeller field adaptive fault diagnosis method based on dynamic generative adversarial network and environment residual adaptive network double network fusion. The dynamic generative adversarial network introduces energy features and environmental parameters, constructs a frequency domain physical constraint, generates pseudo-fault data suitable for the target domain, and uses time domain and frequency domain discriminators to ensure the authenticity and rationality of the generated signal; the environment residual adaptive network constructs a residual dynamic compensation mechanism, fuses the residual between the measured features and the pseudo-fault features of the target domain, and the environment factors after the environmental parameters are encoded, constructs a dynamic mapping relationship between the environmental parameters and the feature residuals, adaptively generates a feature compensation amount, realizes double-domain adaptation of global distribution alignment and local residual compensation, and can effectively improve the generalization ability and diagnostic accuracy of underwater robot propeller fault diagnosis in different water environments. BRIEF DESCRIPTION OF DRAWINGS

[0081] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0082] Figure 1 The technical principle diagram of the application.

[0083] Figure 2 The dynamic generative adversarial network data enhancement model framework provided by the application.

[0084] ​​Figure 3 The environment residual adaptation network cross-domain fault diagnosis model provided by the application. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the application.

[0086] The underwater robot propeller fault diagnosis relies on sensor data, but the underwater robot propeller cross-domain fault diagnosis often faces two problems of data mismatch and feature layer offset, as follows:

[0087] The data mismatch problem: when the underwater robot is working, the differences in different water environments (ocean currents, salinity and temperature) will cause the physical mechanism of the propeller fault feature to change, resulting in serious mismatch between the source domain and the target domain fault data. The traditional data enhancement method (such as GAN) can solve the ocean current problem through multi-sensor mode, but cannot construct the dynamic correlation between other environmental factors (salinity and temperature) and fault data, and the generated pseudo-fault data has a “physical law fault” (such as the generated data does not contain the corrosion vibration feature caused by salinity) with the target domain measured data. When directly used for cross-domain fault diagnosis, the model is prone to overfitting, resulting in inaccurate fault diagnosis.

[0088] The feature layer offset problem: even if pseudo-fault data adapted to the target domain environment is generated, it cannot cover all environment-related physical effects, and there is still a feature layer offset between the pseudo-fault data and the target domain measured data. The interaction of salinity and temperature may cause nonlinear feature offset, and the traditional cross-domain fault diagnosis method (such as DANN) can only align the global distribution and cannot construct the mapping relationship between the environmental parameters and the feature residual, and cannot handle such residual.

[0089] To solve the above two problems, the application proposes an underwater robot propeller fault diagnosis method based on dynamic adversarial generation network (DAGAN) and environment residual adaptation network (ERAN) double network fusion, and the technical roadmap for solving the problem is as follows: Figure 1As shown, the dynamic counter-generation mechanism is generated by the environmental parameters (salinity and temperature), and the pseudo-fault data of the adaptive target domain physical law are generated by combining the frequency domain physical constraints, so as to solve the problem of data mismatch between the source domain and the target domain. In view of the feature layer offset problem of the target domain pseudo-fault data generated by the DAGAN and the target domain measured data, the environmental residual adaptation network (ERAN) method is provided, a residual compensation module is constructed, a dynamic mapping relationship between the environmental parameters (salinity and temperature) and the inter-domain feature residual is established, and the feature layer offset is accurately corrected.

[0090] The embodiment of the present application provides a kind of underwater robot propeller cross-domain intelligent fault diagnosis method, specifically comprising the following steps:

[0091] 1. Sensor data acquisition.

[0092] The imbalance of propeller blade, the local defect of propeller, and the propeller jammed by foreign matter in the underwater robot propeller system can cause the positioning error of underwater robot, the track deviation and other underwater robot operation problems, therefore, the embodiment sets four propeller fault types of propeller single blade imbalance, propeller double blade imbalance, propeller local defect and propeller jammed by foreign matter.

[0093] The four propeller fault types (single blade imbalance, double blade imbalance, local defect and foreign matter jam) of underwater robot can all cause mechanical vibration anomaly, and the fault frequency can be analyzed by vibration and current signals to judge the fault type: the imbalance of propeller blade (single blade and double blade) can cause periodic centrifugal force when propeller rotates, and centrifugal vibration of specific frequency is generated, which causes the amplitude to increase with the increase of rotating speed; the local defect of propeller blade can cause transient impact vibration when propeller rotates and interacts with water flow, which causes short-time high-amplitude impact in time domain, wideband energy increase in frequency domain, and transient pulse or intermittent fluctuation of current; the propeller jammed by foreign matter can cause mechanical jam, which causes random fluctuation of vibration amplitude in time domain, increase of low-frequency energy in frequency domain, and sudden increase of current. The vibration signal directly reflects the mechanical vibration mode, the current signal reflects the electromagnetic torque change of motor, the four fault types are reflected by vibration signal and current signal, and the fusion of the two can effectively avoid the overlap of characteristic frequency, therefore, the embodiment adopts nine-axis inertial measurement unit (IMU) and current sensor to measure the source domain vibration data and source domain current data of underwater robot propeller in laboratory water during simulation operation, and the target domain vibration data and target domain current data of underwater robot propeller during actual operation in real water environment, to diagnose the four fault types.

[0094] For the dynamic influence of ocean current, this method indirectly models through multi-sensor fusion strategy: the vibration signals collected by IMU can reflect the flow field disturbance caused by ocean current, and the motor load fluctuations captured by current sensor are related to the changes of propeller resistance caused by ocean current. The joint analysis of the two types of signals has covered the comprehensive effect of ocean current on the mechanical and electromagnetic characteristics of the propeller. Therefore, this method no longer inputs the ocean current parameters, but realizes the self-adaptation to the ocean current environment by fusing the energy features of vibration and current signals.

[0095] 2. Wavelet packet decomposition and energy feature extraction.

[0096] 2.1 Wavelet packet decomposition.

[0097] Wavelet packet decomposition (WPT) can decompose vibration and current signals into different frequency bands through multi-scale decomposition, and effectively capture the frequency band energy features of the four faults. Since the Daubechies wavelet has high low-frequency resolution, it is suitable for analyzing periodic mechanical vibration and current harmonics. The Daubechies wavelet is used to analyze the source domain fault diagnosis data Five-layer wavelet packet decomposition (2 5 = 32 frequency bands) is performed on the vibration and current signals. The vibration and current signals are divided into 32 frequency bands, and 32 frequency band wavelet packet coefficients are obtained, each frequency band is 1000 Hz / (2*32) = 15.625 Hz, covering the full frequency band of 0-500 Hz (Nyquist frequency).

[0098] 2.2 Energy feature extraction.

[0099] According to the 32 frequency band wavelet packet coefficients , the energy of each frequency band is calculated, and the source domain vibration energy feature vector and the source domain current energy feature vector of 32 frequency bands are obtained, forming the frequency domain energy feature vector , as shown in equation (1):

[0100] (1).

[0101] In the formula, is the energy of the i th frequency band, is the wavelet packet coefficient of the i th frequency band, is the number of wavelet packet coefficients of the i th frequency band.

[0102] 3. Constructing a dynamic adversarial generative network (DAGAN) data augmentation model.

[0103] The traditional data enhancement method GAN mainly relies on static adversarial training, lacks adaptability adjustment to dynamic environment or data characteristics, and the generated signal lacks physical law constraint, which easily leads to overfitting of the diagnostic model. Therefore, a dynamic adversarial generation network (DAGAN) method is proposed. Unlike traditional GAN, the dynamic adversarial generation mechanism driven by environmental parameters and frequency domain physical constraints is introduced to ensure that the generated data adapts to the specific water area conditions. The environmental parameters (salinity and temperature) are used to generate pseudo-fault data that conforms to the current water area physical law, breaking through the static data generation mode of traditional GAN, and providing high-quality data for the fault diagnosis method of the environmental residual adaptation network (ERAN) to solve the problem of mismatch between source domain and target domain data.

[0104] The source domain fault diagnosis data already contains the physical effects of the source domain salinity and temperature , but there are differences in the target domain salinity and temperature , resulting in a shift in the frequency band distribution and energy characteristics of the signal. Therefore, DAGAN needs to input the target domain environmental parameters and to construct the changes in physical laws caused by environmental differences, forcing the generated target domain pseudo-fault data to adapt to the physical constraints of the target domain.

[0105] As shown in Figure 2 , the dynamic adversarial generation network data enhancement model includes a generator and a discriminator . The generator includes an environmental parameter encoding module, a residual attention module, and an inverse wavelet packet synthesis module.

[0106] The generator receives the source domain fault diagnosis data , the frequency band energy feature vector extracted by wavelet packet decomposition , the target domain salinity and temperature , and a random noise vector as inputs.

[0107] The environmental parameter encoding module encodes the target domain salinity and temperature , mapping the salinity and temperature to frequency band offsets to ensure that the main frequency of the generated signal adapts to the current water area, as shown in equation (2):

[0108] (2)。

[0109] In the formula, is the frequency band index offset, This is the salinity influence coefficient. This is the temperature influence coefficient. and The source area reference environmental parameters are (salinity 0%, temperature 25℃).

[0110] Salinity Influence Coefficient and temperature influence coefficient The calibration is performed through experiments, specifically including the following steps:

[0111] Step 1: In the laboratory water, the salinity and temperature were varied, and a total of 25 sets of operating conditions were set. The salinity was set to 0%, 15%, 25%, 35%, and 45%, and the temperature was set to 5℃, 10℃, 15℃, 25℃, and 35℃. The 0% salinity and 25℃ temperature were set as the source area reference environmental parameters.

[0112] Step 2: Under each set of operating conditions, control the underwater robot's propeller-type thruster to operate at a constant speed of 1000 RPM, according to the formula... Determine the nominal fundamental frequency .

[0113] Step 3: Extract the measured fundamental frequency from the source domain vibration data acquired through FFT analysis. .

[0114] Step 4: Calculate the fundamental frequency offset .

[0115] Step 5: Establish a multiple linear regression model .

[0116] Step 6: Fit coefficients using the least squares method. The salinity influence coefficient was obtained. and temperature influence coefficient .

[0117] Since propeller vibration is mainly determined by fluid-structure interaction, its frequency domain energy distribution is physically constrained by environmental parameters (salinity, temperature) (Equation (2)). Therefore, when generating pseudo-fault data in the target domain, it is only necessary to base it on the vibration energy characteristics of the source domain. Frequency band shifting and weighting can be performed without introducing random noise. However, the current signal is affected by random electromagnetic interference and noise, and the uncertainty needs to be simulated by the noise vector z.

[0118] Residual attention module for source domain vibrational energy characteristics Perform frequency band offset adjustment The SimAM frequency domain attention mechanism is used to calculate the weights of each frequency band. And generate weighted source domain vibration energy characteristics .

[0119] The anti-wavelet packet synthesis module converts the weighted source domain vibration energy features into time domain signals to generate target domain pseudo-vibration data .

[0120] The random noise vector is concatenated with the target domain pseudo-vibration data to input a multi-layer perceptron to generate target domain pseudo-current data .

[0121] The target domain pseudo-vibration data and the target domain pseudo-current data are synthesized into target domain pseudo-fault data .

[0122] The discriminator includes a time domain discriminator and a frequency domain discriminator .

[0123] The discriminator receives target domain fault diagnosis data and target domain pseudo-fault data.

[0124] The time domain discriminator concatenates the target domain pseudo-vibration data and the target domain pseudo-fault data as a first input, and the target domain fault diagnosis data as a second input, extracts time domain features through a convolutional neural network, and outputs a authenticity probability .

[0125] The frequency domain discriminator performs a fast Fourier transform on the target domain vibration data to obtain an amplitude spectrum , analyzes the target domain pseudo-vibration data and the amplitude spectrum through a convolutional neural network to obtain a frequency domain distribution feature, and outputs a rationality probability .

[0126] The loss function of the generator is represented as:

[0127] (3).

[0128] (4).

[0129] (5).

[0130] The loss function of the discriminator is represented as:

[0131] ​​(6).

[0132] The objective function of the dynamic generative adversarial network data enhancement model is represented as:

[0133] (7).

[0134] In the formula, is the time domain discriminator loss weight (0.4), is the frequency domain discriminator weight (0.6), is the target domain fault diagnosis data, which is subject to the real data distribution, is the fast Fourier transform, is the expectation operation, which is used to average the target domain fault diagnosis data and the target domain pseudo-fault data, is the target domain pseudo-vibration data, is the target domain pseudo-current data, is the time domain adversarial loss, is the frequency domain adversarial loss, is the physical constraint loss, which ensures that the energy of the key fault frequency band of the generated signal is not lower than the fault energy threshold to avoid generating invalid fault features, is the frequency band weight from the SimAM frequency domain attention mechanism, is the spectrum fidelity constraint loss, which forces the Euclidean distance between the target domain fault diagnosis data spectrum and the target domain fault diagnosis data spectrum to be minimized.

[0135] The fault energy threshold is calibrated through experiments, including the following steps:

[0136] Step 1: In the laboratory water area, change the salinity and temperature, and set 25 groups of working conditions, wherein the salinity is set to 0%, 15%, 25%, 35%, and 45%, and the temperature is set to 5℃, 10℃, 15℃, 25℃, and 35℃.

[0137] Step 2: Collect 100 groups of underwater robot propeller normal working vibration data, and perform 5-layer wavelet packet decomposition on each group of vibration data. Divide each group of vibration data into 32 frequency bands, calculate the energy of each frequency band according to formula (1), and calculate the mean and standard deviation of the energy of 100 groups.

[0138] Step 3: Calculate the fault energy threshold according to formula .

[0139] Generator ​by minimizing the objective function such that the generated data is able to fool the discriminator into thinking that the generated data is real; the discriminator by maximizing the objective function such that the discriminator is able to accurately distinguish between real failure data and pseudo failure data .

[0140] 4. Constructing an environmental residual adaptation network cross-domain fault diagnosis model.

[0141] Although the pseudo failure data adapted to the environment of the target domain is generated by the DAGAN, it cannot cover all environment-related physical effects, and there is still a deviation between the pseudo failure data and the feature layer of the measured data of the target domain. The influence of temperature and salinity on the features has nonlinear dynamic characteristics, and traditional cross-domain fault diagnosis methods (such as DANN) can only align the global distribution and cannot construct the mapping relationship between the environmental parameters and the feature residuals. The residual information specific to the fault between different data features is not optimized, and it is difficult to handle extreme environmental differences. For traditional domain adaptation methods that only rely on global distribution alignment, it is difficult to handle the problem of local feature deviation caused by environmental parameters. A residual compensation module is designed to capture the residual between the measured features of the target domain and the pseudo failure features generated by the DAGAN, combined with the salinity and temperature environmental parameter encoding, dynamically predict the feature residual compensation amount strongly related to the environment, realize the global distribution alignment of the feature deviation, and compensate for the local feature deviation caused by the environmental difference, realize the local distribution alignment of the feature deviation, and provide an "environmentally sensitive residual adaptive compensation" cross-domain fault diagnosis method for underwater robot thrusters in complex water environments. Realize cross-domain adaptive diagnosis, solve the feature layer deviation problem in complex water environments, and significantly improve the cross-domain generalization ability and diagnosis accuracy.

[0142] As shown in Figure 3 , the ERAN includes a feature extractor , a residual compensation module , and a fault classifier , and the feature extractor includes a shared feature extractor and a private feature extractor.

[0143] The target domain pseudo failure data, target domain fault diagnosis data, and environmental parameters are input into the environmental residual adaptation network cross-domain fault diagnosis model.

[0144] For target domain pseudo-fault data and target domain fault diagnosis data, the shared feature extractor adopts a double-layer one-dimensional convolutional neural network. The first layer of 64 5x1 convolution kernels extracts macro features with a step of 2. After maximum pooling compression, the second layer of 128 3x1 convolution kernels captures microscopic fluctuations. Finally, the 256-dimensional target domain original fault feature vector is flattened and output through a fully connected layer and the target domain pseudo-fault feature vector .

[0145] For environmental parameters, the private feature extractor adopts a two-layer ReLU-activated fully connected network to compress and encode environmental information, and linearly maps to output a 16-dimensional environmental factor vector .

[0146] MMD is used to globally align the target domain original fault feature vector and the target domain pseudo-fault feature vector, to reduce the overall distribution difference between the target domain original fault feature and the pseudo-fault feature.

[0147] Residual compensation module The input is the target domain original fault feature vector , the target domain pseudo-fault feature vector, and the environmental factor vector.

[0148] Calculate the local feature residual left after global alignment of the target domain original fault feature vector and the target domain pseudo-fault feature vector, as shown in equation (8):

[0149] (8)。

[0150] Concatenate the environmental factor and the local feature residual into a 272-dimensional vector to input the LSTM network, to establish a dynamic mapping between the environmental parameters and the feature offset , as shown in equation (9):

[0151] (9)。

[0152] In the formula, is the forget gate output vector, which controls the retention proportion of the historical state (salinity dominant), is the input gate output vector, which controls the proportion of new information written (temperature dominant), is the output gate output vector, which controls the degree of current cell state output to the hidden state, is the candidate state, a temporary state basis vector generated by the current input, is the input weight matrix, used for linear transformation of the current input, is the recurrent weight matrix, used for linear transformation of the previous hidden state, Bias vector, Sigmoid is the sigmoid function, and tanh is the hyperbolic tangent function.

[0153] Computational process:

[0154] Concatenate input vectors: (272 dimensions).

[0155] Linear transformation: (256-dimensional vector).

[0156] Divide into four parts: .

[0157] Activate separately: ; ; ; .

[0158] Considering the long-term nature of the cumulative effect of salinity corrosion and the immediate nature of the friction effect of temperature, the LSTM network realizes differentiated modeling through the gating mechanism: the forgetting gate mainly captures the long-term effect of salinity, and its output value is positively correlated with the weight of the historical residual; the input gate mainly responds to the immediate change of temperature, and its output value is positively correlated with the temperature component in the current environmental factor; the candidate state generates a compensation basis vector through the fusion of the gating signal, realizing the dynamic fitting of the environment-sensitive residual, that is, the LSTM network controls the memory strength of the historical residual through the forgetting gate , the input gate adjusts the contribution of the current environmental factor, and the candidate state generates a compensation basis vector.

[0159] Update the cell state, as shown in equation (10):

[0160] (10).

[0161] In the formula, is the updated cell state at the current time, which stores the core memory unit of long-term environmental dependence, is the cell state at the last time, which is the cumulative state of historical environmental impact, represents the reserved part of the last cell state controlled by the forgetting gate, represents the adopted part of the candidate cell state controlled by the input gate.

[0162] Update the hidden state, as shown in equation (11):

[0163] (11).

[0164] In the formula, For the current hidden state, it represents the environment-residual dynamic correlation feature (used to generate compensation).

[0165] The output (hidden state ) of the LSTM network will generate a feature compensation through a fully connected layer, and then be used to correct the features of the target domain. The LSTM network parameters ( , , ) are learned through training, rather than manually set, realizing the innovative idea of environment-sensitive residual adaptive compensation.

[0166] In summary, formula (9) calculates three gates and candidate states, formula (10) updates the cell state, and formula (11) updates the hidden state. The entire LSTM uses a gating mechanism to dynamically adjust memory and output according to environmental parameters (salinity, temperature) and feature residuals, achieving differentiated modeling of long-term and short-term environmental effects, and thus accurately compensating for feature shifts.

[0167] The final output gate generates a 64-dimensional hidden state , which is decoded into a 256-dimensional feature compensation through a fully connected layer .

[0168] The magnitude of the feature compensation is matched with the scale of the LSTM output and the feature extractor, ensuring the consistency of the direction of physical constraints and the compatibility of the feature space scale.

[0169] The feature compensation is applied to the original fault feature vector of the target domain to generate the compensated target domain feature vector .

[0170] The target domain feature vector generated by the residual compensation module is used as the input of the fault classifier, and the target domain feature vector is mapped to the fault class space through a fully connected layer, as shown in formula (12):

[0171] (12).

[0172] In the formula, is the classification layer weight matrix, is the bias vector, represents the probability distribution of the four types of faults (single-leaf imbalance / double-leaf imbalance / local defect / foreign object jam).

[0173] The feature alignment loss is represented as:

[0174] (13).

[0175] where, denotes expectation, is the target domain real sample, is the target domain data distribution, is the direct difference between the target domain original fault feature and the pseudo fault feature, MMD is the maximum mean difference, which ensures the alignment of the feature distribution of the two types of data, and γ is a hyperparameter used to balance the two losses.

[0176] Physical meaning: The former realizes local sample alignment by minimizing the Euclidean distance of the feature vector of each sample; the latter realizes global distribution alignment by measuring and aligning the overall difference of the feature distribution, and provides a reliable feature reference for the residual compensation module.

[0177] The physical constraint loss is represented as:

[0178] (14).

[0179] where, is a regularization coefficient (0.1), which is used to control the model complexity and avoid overfitting, is the parameter matrix of the LSTM network.

[0180] Physical meaning: The cosine similarity is used to force the dominant direction of the feature compensation amount to be consistent with the linear combination of the environmental parameters , ensuring that the compensation logic conforms to the physical causal relationship.

[0181] The fault classification loss is represented as:

[0182] (15).

[0183] where, is the expectation of the target domain pseudo fault data distribution, is the compensated pseudo fault feature, is the label of the target domain pseudo fault data, is the fault classification probability.

[0184] Physical meaning: Ensure that the compensated feature can accurately classify the fault type.

[0185] The fault classification objective function is represented as:

[0186] (16).

[0187] where, is the feature alignment weight (1.0), is the physical constraint weight (0.5), For the fault classification weight (value 1.0), the setting is based on the following considerations: among them, the feature alignment is the basis of domain adaptation, and it is necessary to ensure the consistency of the feature distribution between the source domain and the target domain in priority, so the highest weight is given =1.0; and it is necessary to introduce physical law constraints on the basis of feature alignment to avoid the compensation amount deviating from the actual physical effect, but the weight is lower than the feature alignment to prevent over-constraint, so the weight =0.5; the fault classification is as important as the feature alignment, which ensures that the compensated features are not only distribution-aligned, but also accurately distinguish fault types to achieve the diagnosis target, so =1.0.

[0188] The optimization target is the feature alignment loss, the physical constraint loss and the fault classification loss, and the purpose is to minimize the fault classification error, realize domain-invariant feature learning and ensure the consistency of the physical law.

[0189] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

Claims

1. A cross-domain intelligent fault diagnosis method for underwater robot propeller-type thrusters, characterized in that, Specifically, the steps include the following: The source domain fault diagnosis data of the underwater robot propeller-type thruster was obtained in the laboratory water area during simulated operation, and the source domain energy characteristics were extracted from the source domain fault diagnosis data. The source domain fault diagnosis data includes source domain vibration data and source domain current data. Acquire target domain fault diagnosis data and environmental parameters of underwater robot propeller thrusters in real water environment. The target domain fault diagnosis data includes target domain vibration data and target domain current data, and the target domain environmental parameters include target domain salinity and target domain temperature. A dynamic adversarial generative network (PGN) data augmentation model is constructed. Source domain energy characteristics, target domain fault diagnosis data, and environmental parameters are input into the PGN data augmentation model. The PGN data augmentation model determines frequency domain physical law constraints based on source domain energy characteristics and target domain environmental parameters, and generates target domain pseudo fault data. The PGN data augmentation model includes a generator and a discriminator. The generator includes an environmental parameter encoding module, a residual attention module, and an inverse wavelet packet synthesis module. The discriminator includes a time domain discriminator and a frequency domain discriminator. A cross-domain fault diagnosis model for environmental residual adaptation network is constructed. Target domain pseudo-fault data, target domain fault diagnosis data, and environmental parameters are input into the model. The model extracts measured features, pseudo-fault features, and environmental factors from the target domain. The residuals between the measured features and pseudo-fault features are fused with the environmental factors to construct a dynamic mapping relationship between environmental parameters and feature residuals. Feature compensation quantities are adaptively generated, and fault diagnosis is performed on the compensated features to obtain the fault type. The cross-domain fault diagnosis model for environmental residual adaptation network includes a feature extractor, a residual compensation module, and a fault classifier. The feature extractor includes a shared feature extractor and a private feature extractor. The extraction of source domain energy features from source domain fault diagnosis data specifically includes the following steps: Wavelet packet decomposition: Using Daubechies wavelet to analyze source domain fault diagnosis data Perform five-level wavelet packet decomposition, and then... The frequency bands were divided into 32 bands, resulting in wavelet packet coefficients for each of the 32 bands. ; Energy feature extraction: based on wavelet packet coefficients of 32 frequency bands Based on the energy calculation for each frequency band, a total of 32 source domain vibration energy characteristic vectors were obtained. Source domain current energy eigenvector Forming frequency domain energy feature vectors As shown in formula (1): (1)； In the formula, For the first Energy of each frequency band For the first Wavelet packet coefficients for each frequency band For the first The number of wavelet packet coefficients in each frequency band; The dynamic adversarial generative network data augmentation model determines frequency domain physical constraints based on source domain energy characteristics and target domain environmental parameters, and generates target domain pseudo-fault data, specifically including the following steps: Generator Receive source domain fault diagnosis data Frequency band energy feature vector extracted by wavelet packet decomposition Target area salinity and temperature and random noise vector As input; The environmental parameter encoding module for target domain salinity and temperature Encode the salinity and temperature The frequency band offset is mapped to ensure that the main frequency of the generated signal is adapted to the current water area, as shown in formula (2): (2)； In the formula, This is the frequency band index offset. This is the salinity influence coefficient. This is the temperature influence coefficient. and These are the baseline environmental parameters for the source domain. Residual attention module for source domain vibrational energy characteristics Perform frequency band offset adjustment The SimAM frequency domain attention mechanism is used to calculate the weights of each frequency band. And generate weighted source domain vibration energy characteristics ; The inverse wavelet packet synthesis module extracts the weighted source domain vibration energy characteristics. Convert to time domain signal to generate target domain pseudo-vibration data ; random noise vector Pseudo-vibration data in the target domain By splicing the input to a multilayer perceptron, pseudo-current data in the target domain is generated. ; Target domain pseudo vibration data and target domain pseudo current data Synthesized into target domain pseudo-fault data ; Discriminator Receive target domain fault diagnosis data Pseudo-fault data in the target domain; Time-domain discriminant The target domain pseudo-vibration data and target domain pseudo-fault data are concatenated as the first input, and the target domain fault diagnosis data is used as the second input. A convolutional neural network is used to extract time-domain features and output the authenticity probability. ; Frequency domain discriminator The amplitude spectrum is obtained by performing a fast Fourier transform on the vibration data in the target domain. Analyzing pseudo-vibration data and amplitude spectrum in the target domain using convolutional neural networks Frequency domain distribution characteristics and output probability of rationality ; Generator loss function Represented as: (3)； (4)； (5)； Discriminator loss function Represented as: (6)； The objective function of the dynamic adversarial generative network data augmentation model is expressed as: (7)； In the formula, For the time-domain discriminator loss weights, For the frequency domain discriminator weights, For target domain fault diagnosis data, For Fast Fourier Transform, For the expected value calculation, the average values ​​are calculated for the target domain fault diagnosis data and the target domain pseudo-fault data, respectively. For pseudo-vibration data in the target domain, For the target domain pseudo current data, For temporal domain countermeasures losses, For frequency domain adversarial loss, For physical constraint loss, The fault energy threshold, For the frequency band weights derived from the SimAM frequency domain attention mechanism, Constraint loss for spectrum fidelity; Generator By minimizing the objective function This makes the generated data Able to deceive the discriminator ,let The discriminator assumes the generated data is real. By maximizing the objective function ,make Able to accurately distinguish real fault data and pseudo-fault data ; The cross-domain fault diagnosis model of the environmental residual adaptation network extracts measured features, pseudo-fault features, and environmental factors of the target domain. It then fuses the residuals between the measured features and pseudo-fault features of the target domain with the environmental factors to construct a dynamic mapping relationship between environmental parameters and feature residuals. Adaptively, it generates feature compensation quantities and performs fault diagnosis on the compensated features to obtain the fault type. Specifically, it includes the following steps: For both pseudo-fault data and fault diagnosis data in the target domain, the shared feature extractor employs a two-layer one-dimensional convolutional neural network. The first layer uses 64 5×1 convolutional kernels with a stride of 2 to extract macroscopic features, which are then compressed using max pooling. The second layer uses 128 3×1 convolutional kernels to capture microscopic fluctuations. Finally, the extracted features are flattened and output as a 256-dimensional original fault feature vector of the target domain through a fully connected layer. and target domain pseudo-fault feature vector ; For environmental parameters, the private feature extractor uses a two-layer ReLU activated fully connected network to compress and encode environmental information, and outputs a 16-dimensional environmental factor vector through linear mapping. ; MMD is used to globally align the original fault feature vector and the pseudo fault feature vector in the target domain, so as to reduce the overall distribution difference between the original fault features and the pseudo fault features in the target domain. Residual compensation module The input is the original fault feature vector of the target domain. Target domain pseudo-fault feature vector and environmental factor vector; The local feature residuals remaining after global alignment of the original fault feature vector in the target domain and the pseudo fault feature vector in the target domain are calculated, as shown in Equation (8): (8)； Environmental factors With local feature residuals The vectors are concatenated into a 272-dimensional vector and input into an LSTM network to establish environmental parameters. With feature offset The dynamic mapping is shown in formula (9): (9)； In the formula, Output vector for the forget gate. The input gate output vector, The output gate outputs a vector. Candidate state For the input weight matrix, This is a cyclic weight matrix. For bias vectors, Here, is the Sigmoid function, and tanh is the hyperbolic tangent function; Update the cell state as shown in formulas (10) and (11): (10)； (11)； In the formula, This represents the cell state updated at the current moment. This represents the cell state at the previous moment. Currently hidden; The final output gate generates a 64-dimensional hidden state. , Decoded into 256-dimensional feature compensation quantities by a fully connected layer ; Feature compensation amount The magnitude of the signal is matched with the scale of the feature extractor through the output of the LSTM network to ensure the consistency of the physical constraint direction and the compatibility of the feature space scale. Feature compensation amount Applied to the original fault feature vector in the target domain Generate the compensated target domain feature vector ; The target domain feature vector generated by the residual compensation module The target domain feature vector serves as the input to the fault classifier. The fault category space is mapped through a fully connected layer, as shown in Equation (12): (12)； In the formula, This is the classification layer weight matrix. For bias vectors, Represents the failure probability distribution; The feature alignment loss is expressed as: (13)； In the formula, Expressing expectations, For the measured samples in the target domain, For the distribution of data in the target domain, To minimize the direct difference between the original fault features and the pseudo-fault features in the target domain, MMD is the maximum mean difference, and γ is a hyperparameter. The physical constraint loss is expressed as: (14)； In the formula, The regularization coefficient is . This is the parameter matrix of the LSTM network; Fault classification loss is represented as: (15)； In the formula, To calculate the expectation of the distribution of pseudo-fault data in the target domain, For the compensated pseudo-fault characteristics, For pseudo-fault data labels in the target domain, Classify the probability of faults; The objective function for fault classification is expressed as: (16)； In the formula, Feature alignment weights; For physical constraint weights; The weights are used for fault classification.

2. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 1, characterized in that, The salinity influence coefficient and temperature influence coefficient The calibration is performed through experiments, specifically including the following steps: Step 1: In the laboratory water, the salinity and temperature were changed, and a total of 25 sets of working conditions were set. The salinity was set to 0%, 15%, 25%, 35%, and 45%, and the temperature was set to 5℃, 10℃, 15℃, 25℃, and 35℃. The 0% salinity and 25℃ temperature were set as the source area reference environmental parameters. Step 2: Under each set of operating conditions, control the underwater robot's propeller-type thruster to operate at a constant speed of 1000 RPM, according to the formula... Determine the nominal fundamental frequency ; Step 3: Extract the measured fundamental frequency from the source domain vibration data acquired through FFT analysis. ; Step 4: Calculate the fundamental frequency offset ; Step 5: Establish a multiple linear regression model ; Step 6: Fit coefficients using the least squares method. The salinity influence coefficient was obtained. and temperature influence coefficient .

3. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 2, characterized in that, The fault energy threshold The calibration is performed through experiments, specifically including the following steps: Step 1: In the laboratory water, the salinity and temperature were changed, and a total of 25 working conditions were set. The salinity was set to 0%, 15%, 25%, 35%, and 45%, and the temperature was set to 5℃, 10℃, 15℃, 25℃, and 35℃. Step 2: Collect 100 sets of vibration data of the underwater robot's propeller thruster during normal operation. Perform 5-level wavelet packet decomposition on each set of vibration data to divide each set of vibration data into 32 frequency bands. Calculate the energy of each frequency band according to formula (1). Calculate the average of 100 energy values. and standard deviation ; Step 3: According to the formula Calculate the fault energy threshold .

Images