Underwater robot propeller type thruster cross-domain intelligent fault diagnosis method

Abstract

The invention provides an underwater robot propeller type thruster cross-domain intelligent fault diagnosis method, and belongs to the field of underwater robot fault diagnosis. Energy characteristics and environmental parameters are introduced through a dynamic adversarial generative network, frequency domain physical constraints are constructed, pseudo fault data adaptive to a target domain are generated, and authenticity and rationality of generated signals are guaranteed by using time domain and frequency domain double discriminators. The environment residual error adaptation network fuses the residual error between the actual measurement characteristic of the target domain and the pseudo fault characteristic with the environment factor after the environment parameter coding by constructing a residual error dynamic compensation mechanism, constructs the dynamic mapping relation between the environment parameter and the characteristic residual error, adaptively generates the characteristic compensation amount, and finally performs the self-adaption on the characteristic compensation amount. Double-layer domain adaptation of global distribution alignment and local residual compensation is achieved, and the generalization ability and diagnosis accuracy of underwater robot propeller fault diagnosis in different water area environments can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater robot fault diagnosis, and in particular to a cross-domain intelligent fault diagnosis method for a propeller thruster of an underwater robot. Background Art

[0002] The ocean is rich in mineral resources, marine biological resources, and other energy resources. The emergence of underwater robots provides a strong foundation for human exploration of the complex marine environment. Within the complex system architecture of underwater robots, propeller-type propulsion systems are susceptible to the influence of ocean currents, salinity, and temperature, and have become the core component with the highest failure rate. Ocean currents alter the flow distribution around propellers. Strong or complex currents can significantly increase propeller loads, leading to increased operating current, speed fluctuations, or cavitation. Irregular currents subject propellers to uncertain impact loads, potentially causing increased shaft vibration and bearing wear. Salinity affects water density and viscosity, increasing propeller loads and baseline drift in current signals, affecting the signal's frequency band offset and harmonic energy distribution. High-salinity water (such as seawater) also accelerates electrochemical corrosion of metal propellers, increasing surface roughness and causing abnormal propeller vibration. Temperature variations also affect propellers. At low temperatures, the metal material shrinks, increasing friction between the propeller and shaft and causing unbalanced vibration. High temperatures, on the other hand, reduce lubricant viscosity, increasing the risk of mechanical wear. Furthermore, the electronic components in the motor controller are susceptible to temperature fluctuations, resulting in large current sampling errors that can mask true fault signals. Temperature also affects the transient impact strength and baseline noise level of the generated signal.

[0003] Current fault diagnosis results for propeller thrusters on underwater robots are mostly obtained in laboratory or fixed-water test environments. However, in actual operations, underwater robots must perform tasks in different waters. Data obtained in these labs or fixed-water environments is considered source domain data, while data obtained during actual operations is considered target domain data. Due to environmental factors, the vibration spectrum characteristics and sensor data distribution of target domain data for propeller thrusters on underwater robots differ significantly from those in the source domain, resulting in data mismatch and feature layer offset. This makes it difficult for models trained on a single water area to achieve effective cross-domain fault diagnosis.

[0004] Therefore, introducing environmental factors and physical constraints into the fault diagnosis of underwater robot propeller thrusters and conducting intelligent fault diagnosis with domain adaptability has become an important research direction to improve the accuracy and reliability of underwater robot propeller thruster fault diagnosis. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides a cross-domain intelligent fault diagnosis method for a propeller thruster of an underwater robot, which specifically includes the following steps: Source domain fault diagnosis data of a propeller-type thruster of an underwater robot during simulated operation in laboratory waters is obtained, and source domain energy features are extracted from the source domain fault diagnosis data, wherein the source domain fault diagnosis data includes source domain vibration data and source domain current data.

[0006] The target domain fault diagnosis data and environmental parameters of the underwater robot propeller thruster in actual operation in a real water environment are obtained, wherein 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.

[0007] A dynamic adversarial generative network data enhancement model is constructed, and the source domain energy characteristics, target domain fault diagnosis data and environmental parameters are input into the dynamic adversarial generative network data enhancement model. The dynamic adversarial generative network data enhancement model determines the frequency domain physical law constraints based on the source domain energy characteristics and target domain environmental parameters, and generates target domain pseudo-fault data. The dynamic adversarial generative 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, and the discriminator includes a time domain discriminator and a frequency domain discriminator.

[0008] An environmental residual adaptive network cross-domain fault diagnosis model is constructed, and the target domain pseudo-fault data, target domain fault diagnosis data and environmental parameters are input into the environmental residual adaptive network cross-domain fault diagnosis model. The environmental residual adaptive network cross-domain fault diagnosis model extracts the measured features, pseudo-fault features and environmental factors of the target domain, and fuses the residuals between the measured features and pseudo-fault features of the target domain with the environmental factors, constructs a dynamic mapping relationship between environmental parameters and feature residuals, adaptively generates feature compensation amounts, performs fault diagnosis on the compensated features, and obtains the fault type. The environmental residual adaptive network cross-domain fault diagnosis model 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.

[0009] Optionally, extracting 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-layer wavelet packet decomposition and Divide into 32 frequency bands and obtain 32 frequency band wavelet packet coefficients .

[0010] Energy feature extraction: based on 32 frequency band wavelet packet coefficients By calculating the energy of each frequency band, a total of 32 frequency bands of source domain vibration energy eigenvectors are obtained and the source domain current energy eigenvector , forming the frequency domain energy feature vector , as shown in formula (1): (1).

[0011] Where, For the The energy of the frequency band, For the The wavelet packet coefficients of the frequency band, For the The number of wavelet packet coefficients in a frequency band.

[0012] Optionally, the dynamic adversarial network data enhancement model determines the frequency domain physical law constraints based on the source domain energy characteristics and the target domain environmental parameters to generate 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 domain salinity and temperature and a random noise vector as input.

[0013] Environmental parameter encoding module for target domain salinity and temperature Encode the salinity and temperature Mapped to a frequency band offset to ensure that the main frequency of the generated signal is adapted to the current water area, as shown in formula (2): (2).

[0014] Where, is the frequency band index offset, is the salinity influence coefficient, is the temperature influence coefficient, and are the source domain benchmark environment parameters.

[0015] Residual attention module for source domain vibration energy features Adjust the frequency band offset , and use SimAM frequency domain attention mechanism to calculate the weight of each frequency band And generate weighted source domain vibration energy features .

[0016] The inverse wavelet packet synthesis module converts the weighted source domain vibration energy characteristics Convert to time domain signal to generate target domain pseudo vibration data .

[0017] The random noise vector Pseudo vibration data in the target domain Splicing input into the multi-layer perceptron to generate pseudo current data in the target domain .

[0018] The target domain pseudo vibration data and target domain pseudo current data Synthesize into pseudo fault data in the target domain .

[0019] Discriminator Receive target domain fault diagnosis data with pseudo fault data in the target domain.

[0020] Time Domain Discriminator The target domain pseudo vibration data and target domain pseudo fault data are spliced ​​as the first input, and the target domain fault diagnosis data is used as the second input. The convolutional neural network is used to extract the time domain features and output the authenticity probability. .

[0021] Frequency domain discriminator Perform fast Fourier transform on the target domain vibration data to obtain the amplitude spectrum , using convolutional neural networks to analyze target domain pseudo vibration data and amplitude spectra Frequency domain distribution features and output rationality probability .

[0022] Generator The loss function Expressed as: (3).

[0023] (4).

[0024] (5).

[0025] Discriminator The loss function Expressed as: (6).

[0026] The objective function of the dynamic adversarial generation network data enhancement model is expressed as: (7).

[0027] Where, is the loss weight of the time domain discriminator, is the frequency domain discriminator weight, is the target domain fault diagnosis data, which obeys the real data distribution. is the fast Fourier transform, For the expected operation, the target domain fault diagnosis data and the target domain pseudo-fault data are averaged respectively. is the pseudo vibration data in the target domain, is the target domain pseudo current data, is the time domain adversarial loss, is the frequency domain adversarial loss, is the physical constraint loss, is the fault energy threshold, is the frequency band weight from the SimAM frequency domain attention mechanism, Constrained loss for spectral fidelity.

[0028] Generator By minimizing the objective function , so that the generated data Can deceive the discriminator ,let The discriminator considers the generated data to be real; By maximizing the objective function ,make Able to accurately distinguish real fault data and pseudo-fault data .

[0029] Optionally, the cross-domain fault diagnosis model of the environmental residual adaptation network extracts the measured features, pseudo-fault features, and environmental factors of the target domain, fuses the residuals between the measured features and the pseudo-fault features of the target domain with the environmental factors, constructs a dynamic mapping relationship between environmental parameters and feature residuals, adaptively generates feature compensation amounts, performs fault diagnosis on the compensated features, and obtains the fault type, specifically comprising the following steps: For the target domain pseudo-fault data and target domain fault diagnosis data, the shared feature extractor uses a two-layer one-dimensional convolutional neural network. The first layer uses 64 5×1 convolution kernels with a stride of 2 to extract macro features. After maximum pooling compression, the second layer uses 128 3×1 convolution kernels to capture micro fluctuations. Finally, it is flattened and outputs a 256-dimensional target domain original fault feature vector through a fully connected layer. and the target domain pseudo-fault feature vector .

[0030] For environmental parameters, the private feature extractor uses a two-layer ReLU activated fully connected network to compress and encode environmental information, and linearly maps the output to a 16-dimensional environmental factor vector .

[0031] 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 features and pseudo-fault features.

[0032] 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.

[0033] Calculate the local feature residual remaining after global alignment of the original fault feature vector in the target domain and the pseudo-fault feature vector in the target domain, as shown in formula (8): (8).

[0034] Environmental factors and local feature residual Splice into 272-dimensional vector input LSTM network to establish environment parameters Offset from feature The dynamic mapping is shown in formula (9): (9).

[0035] Where, is the output vector of the forget gate, is the input gate output vector, is the output vector of the output gate, For candidate status, is the input weight matrix, is the circulant weight matrix, is the bias vector, is the Sigmoid function, and tanh is the hyperbolic tangent function.

[0036] Update the cell state as shown in formulas (10) and (11): (10).

[0037] (11).

[0038] Where, is the cell state after updating at the current moment, is the cell state at the previous moment, The current hidden state.

[0039] The final output gate generates a 64-dimensional hidden state , Decoded into 256-dimensional feature compensation by the fully connected layer .

[0040] Feature compensation The amplitude of is matched with the scale of the feature extractor through the LSTM network output, ensuring that the directional consistency of the physical constraint is compatible with the feature space scale.

[0041] The characteristic compensation Applied to the original fault feature vector in the target domain , generate the compensated target domain feature vector .

[0042] The target domain feature vector generated by the residual compensation module is As the input of the fault classifier, the target domain feature vector Mapped to the fault category space through the fully connected layer, as shown in formula (12): (12).

[0043] Where, is the classification layer weight matrix, is the bias vector, represents the failure probability distribution.

[0044] The feature alignment loss is expressed as: (13).

[0045] Where, Express expectations, is the measured sample in the target domain, is the target domain data distribution, 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.

[0046] The physical constraint loss is expressed as: (14).

[0047] Where, is the regularization coefficient, is the parameter matrix of the LSTM network.

[0048] The fault classification loss is expressed as: (15).

[0049] Where, To find the expected distribution of pseudo-fault data in the target domain, is the pseudo-fault signature after compensation, is the pseudo fault data label of the target domain, is the probability of fault classification.

[0050] The fault classification objective function is expressed as: (16).

[0051] Where, is the feature alignment weight; is the physical constraint weight; is the fault classification weight.

[0052] Optionally, the salinity influence coefficient and temperature effect coefficient Calibration is performed through experiments, which specifically includes the following steps: Step 1: In the laboratory waters, change the salinity and temperature, and set a total of 25 groups of operating conditions. Among them, the salinity is set to: 0%, 15%, 25%, 35%, 45%, and the temperature is set to: 5℃, 10℃, 15℃, 25℃, 35℃. 0% salinity and 25℃ temperature are set as the source domain baseline environmental parameters.

[0053] Step 2: Under each set of working conditions, control the propeller of the underwater robot to run at a constant speed of 1000RPM. According to the formula Obtain the nominal fundamental frequency .

[0054] Step 3: Extract the measured fundamental frequency from the collected source domain vibration data through FFT analysis .

[0055] Step 4: Calculate the fundamental frequency offset .

[0056] Step 5: Build a multiple linear regression model .

[0057] Step 6: Use the least squares method to fit the coefficient formula , and obtain the salinity influence coefficient and temperature effect coefficient .

[0058] Optionally, the fault energy threshold Calibration is performed through experiments, which specifically includes the following steps: Step 1: In the laboratory waters, change the salinity and temperature, and set a total of 25 groups of operating conditions, among which the salinity is set to: 0%, 15%, 25%, 35%, 45%, and the temperature is set to: 5℃, 10℃, 15℃, 25℃, 35℃.

[0059] Step 2: Collect 100 sets of vibration data of the propeller thrusters of underwater robots in normal operation, perform 5-layer wavelet packet decomposition on each set of vibration data, divide each set of vibration data into 32 frequency bands, calculate the energy of each frequency band according to formula (1), and calculate the energy of each frequency band. , calculate the mean of 100 groups of energy and standard deviation .

[0060] Step 3: According to the formula Calculating the fault energy threshold .

[0061] The present invention provides a cross-domain intelligent fault diagnosis method for propeller-type thrusters of underwater robots. To address the data distribution differences and feature layer offsets that exist when diagnosing cross-domain faults of underwater robot thrusters in water environments with different salinities and temperatures, an adaptive fault diagnosis method for underwater robot thrusters based on the fusion of a dynamic adversarial generative network and an environmental residual adaptation network is proposed. The dynamic adversarial generative network introduces energy features and environmental parameters, constructs frequency domain physical constraints, generates pseudo-fault data adapted to the target domain, and uses dual discriminators in the time domain and frequency domain to ensure the authenticity and rationality of the generated signal. The environmental residual adaptation network constructs a residual dynamic compensation mechanism, fuses the residuals between the measured features and pseudo-fault features in the target domain with the environmental factors encoded by the environmental parameters, constructs a dynamic mapping relationship between the environmental parameters and the feature residuals, and adaptively generates feature compensation amounts, achieving dual-domain adaptation of global distribution alignment and local residual compensation. This method can effectively improve the generalization ability and diagnostic accuracy of underwater robot thruster fault diagnosis in different water environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 This is a technical principle diagram of the present invention.

[0064] Figure 2 This is a framework diagram of the dynamic adversarial generative network data enhancement model provided by the present invention.

[0065] Figure 3 The present invention provides an environmental residual adaptation network cross-domain fault diagnosis model. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] Fault diagnosis of propeller-type thrusters on underwater robots relies on sensor data. However, cross-domain fault diagnosis of propeller-type thrusters on underwater robots often faces two major problems: data mismatch and feature layer offset. The specific problems are as follows: Data mismatch problem: When underwater robots are operating, the differences in the environment (ocean currents, salinity, and temperature) in different waters will cause changes in the physical mechanism of propeller thruster fault characteristics, resulting in serious mismatch between the source domain and target domain fault data. Traditional data enhancement methods (such as GAN) can solve the ocean current problem through multi-sensor methods, but cannot establish a dynamic association between other environmental factors (salinity and temperature) and fault data. There is a "physical law fault" between the generated pseudo-fault data and the actual measured data in the target domain (for example, the generated data does not include the corrosion vibration characteristics caused by salinity). When directly used for cross-domain fault diagnosis, the model is prone to overfitting, resulting in inaccurate fault diagnosis.

[0068] Feature layer offset problem: Even if pseudo-fault data adapted to the target domain environment is generated, it cannot cover all physical effects related to the environment. There is still an offset between the feature layers of the pseudo-fault data and the actual measured data in the target domain. The interaction between salinity and temperature may cause nonlinear feature offset. Traditional cross-domain fault diagnosis methods (such as DANN) can only align the global distribution and cannot construct a mapping relationship between environmental parameters and feature residuals, and cannot handle such residuals.

[0069] In response to the above two problems, this paper proposes a fault diagnosis method for underwater robot propeller thrusters based on the dual network fusion of dynamic adversarial generative network (DAGAN) and environmental residual adaptation network (ERAN). The technical roadmap for solving this problem is as follows: Figure 1 As shown in the figure, a dynamic adversarial generation mechanism is driven by environmental parameters (salinity and temperature). Combined with frequency-domain physical constraints, pseudo-fault data is generated that adapts to the physical laws of the target domain, resolving the data mismatch between the source and target domains. To address the feature layer offset between the pseudo-fault data in the target domain generated by DAGAN and the measured data in the target domain, the present invention proposes an Environmental Residual Adaptation Network (ERAN) method. By constructing a residual compensation module, a dynamic mapping relationship between environmental parameters (salinity and temperature) and inter-domain feature residuals is established, achieving precise correction of the feature layer offset.

[0070] An embodiment of the present invention provides a cross-domain intelligent fault diagnosis method for a propeller thruster of an underwater robot, which specifically includes the following steps: 1. Sensor data acquisition.

[0071] Unbalanced propeller blades, partial loss of propellers, and propellers stuck by foreign objects in the underwater robot propulsion system can all lead to various underwater robot operation problems such as inaccurate positioning of the underwater robot and track deviation. Therefore, this embodiment sets four types of propeller propeller failures: single-blade imbalance of propellers, double-blade imbalance of propellers, partial defect of propellers, and foreign object jamming of propellers.

[0072] The four types of propeller propeller failures of underwater robots (single-blade imbalance, double-blade imbalance, local defect, and foreign object jamming) will all cause abnormal mechanical vibrations. The fault frequency can be analyzed through vibration and current signals to determine the fault type: propeller propeller blade imbalance (single-blade and double-blade) will cause the propeller to generate periodic centrifugal force when rotating, generating centrifugal force vibrations of a specific frequency, causing the amplitude to increase with increasing speed; local defects in propeller propeller blades will cause transient impact vibrations when the propeller interacts with the water flow during rotation, resulting in short-term high-amplitude impacts in the time domain, increased wide-band energy in the frequency domain, and transient pulses or intermittent fluctuations in the current; foreign object jamming of the propeller propeller will cause mechanical obstruction, causing random fluctuations in the vibration amplitude in the time domain, increased energy in the low-frequency band of the frequency domain, and a sudden increase in current. The vibration signal directly reflects the mechanical vibration mode, and the current signal reflects the change in the electromagnetic torque of the motor. The four fault types will all be manifested through the vibration signal and the current signal, and the fusion of the two can effectively avoid the overlap of characteristic frequencies. Therefore, this embodiment uses a nine-axis inertial measurement unit (IMU) and a current sensor to measure the source domain vibration data and source domain current data of the underwater robot propeller thruster in simulated operation in laboratory waters, as well as the target domain vibration data and target domain current data of the underwater robot propeller thruster in actual operation in real water environments, in order to diagnose the four fault types.

[0073] To address the dynamic impact of ocean currents, this method indirectly models the impact through a multi-sensor fusion strategy: the vibration signal collected by the IMU can reflect the flow field disturbance caused by the ocean current, and the motor load fluctuation captured by the current sensor is related to the change in propeller resistance caused by the ocean current. The joint analysis of the two types of signals covers the combined effect of the ocean current on the mechanical and electromagnetic properties of the propeller. Therefore, this method no longer inputs ocean current parameters, but instead achieves adaptation to the ocean current environment by fusing the energy characteristics of the vibration and current signals.

[0074] 2. Wavelet packet decomposition and energy feature extraction.

[0075] 2.1 Wavelet packet decomposition.

[0076] 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 characteristics of the four faults. Due to the high low-frequency resolution of Daubechies wavelet, it is suitable for analyzing periodic mechanical vibration and current harmonics. Daubechies wavelet is used to analyze source domain fault diagnosis data. Perform five-layer wavelet packet decomposition (2 5 = 32 frequency bands), Divide into 32 frequency bands and obtain 32 frequency band wavelet packet coefficients , each frequency band is 1000Hz / (2×32)=15.625Hz, covering the full frequency band of 0-500Hz (Nyquist frequency).

[0077] 2.2 Energy feature extraction.

[0078] According to the 32 frequency band wavelet packet coefficients By calculating the energy of each frequency band, a total of 32 frequency bands of source domain vibration energy eigenvectors are obtained and the source domain current energy eigenvector , forming the frequency domain energy feature vector , as shown in formula (1): (1).

[0079] Where, For the The energy of the frequency band, For the The wavelet packet coefficients of the frequency band, For the The number of wavelet packet coefficients in a frequency band.

[0080] 3. Build a dynamic adversarial generative network (DAGAN) data augmentation model.

[0081] Traditional data enhancement methods such as Generative Adversarial Networks (GANs) mainly rely on static adversarial training and lack adaptive adjustment to dynamic environments or data characteristics. The generated signals lack physical constraints, which easily leads to overfitting of the diagnostic model. Therefore, a dynamic adversarial generative network (DAGAN) method is proposed. Unlike traditional GANs, a dynamic adversarial generation mechanism driven by environmental parameters and frequency domain physical constraints is introduced to ensure that the generated data adapts to specific water conditions. Pseudo-fault data that conforms to the physical laws of the current water area is generated by driving environmental parameters (salinity, temperature). This breaks through the static data generation mode of traditional GANs and provides high-quality data for the fault diagnosis method of the Environmental Residual Adaptation Network (ERAN) to solve the problem of data mismatch between the source and target domains.

[0082] Source domain fault diagnosis data already contains source domain salinity and temperature physical effects under the target domain salinity and temperature There are differences, which leads to the shift of the frequency band distribution and energy characteristics of the signal. Therefore, DAGAN needs to input the target domain environment parameters and ,The physical law changes caused by the construction environment differences force the generated target domain pseudo-fault data to adapt to the physical constraints of the target domain.

[0083] like Figure 2 As shown, the dynamic adversarial network data enhancement model includes a generator and the discriminator , generator It includes environment parameter encoding module, residual attention module and inverse wavelet packet synthesis module.

[0084] Generator Receive source domain fault diagnosis data Frequency band energy feature vector extracted by wavelet packet decomposition , target domain salinity and temperature and a random noise vector as input.

[0085] Environmental parameter encoding module for target domain salinity and temperature Encode the salinity and temperature Mapped to a frequency band offset to ensure that the main frequency of the generated signal is adapted to the current water area, as shown in formula (2): (2).

[0086] Where, is the frequency band index offset, is the salinity influence coefficient, is the temperature influence coefficient, and are the source domain benchmark environmental parameters (salinity 0%, temperature 25°C).

[0087] Salinity influence coefficient and temperature effect coefficient Calibration is performed through experiments, which specifically includes the following steps: Step 1: In the laboratory waters, change the salinity and temperature, and set a total of 25 groups of operating conditions. Among them, the salinity is set to: 0%, 15%, 25%, 35%, 45%, and the temperature is set to: 5℃, 10℃, 15℃, 25℃, 35℃. 0% salinity and 25℃ temperature are set as the source domain baseline environmental parameters.

[0088] Step 2: Under each set of working conditions, control the propeller of the underwater robot to run at a constant speed of 1000RPM. According to the formula Obtain the nominal fundamental frequency .

[0089] Step 3: Extract the measured fundamental frequency from the collected source domain vibration data through FFT analysis .

[0090] Step 4: Calculate the fundamental frequency offset .

[0091] Step 5: Build a multiple linear regression model .

[0092] Step 6: Use the least squares method to fit the coefficient formula , and obtain the salinity influence coefficient and temperature effect coefficient .

[0093] Since propeller vibration is mainly determined by fluid-structure interaction, its frequency domain energy distribution is subject to physical constraints of environmental parameters (salinity, temperature) (Formula (2)). Therefore, when generating target domain pseudo-fault data, it is only necessary to base the energy characteristics of the source domain vibration on the energy characteristics of the source domain. Perform frequency band shifting and weighting without introducing random noise , and the current signal is affected by random electromagnetic interference and noise, so the uncertainty needs to be simulated by the noise vector z.

[0094] Residual attention module for source domain vibration energy features Adjust the frequency band offset , and use SimAM frequency domain attention mechanism to calculate the weight of each frequency band And generate weighted source domain vibration energy features .

[0095] The inverse wavelet packet synthesis module converts the weighted source domain vibration energy characteristics Convert to time domain signal to generate target domain pseudo vibration data .

[0096] The random noise vector Pseudo vibration data in the target domain Splicing input into the multi-layer perceptron to generate pseudo current data in the target domain .

[0097] The target domain pseudo vibration data and target domain pseudo current data Synthesize into pseudo fault data in the target domain .

[0098] Discriminator Including time domain discriminator and frequency domain discriminator .

[0099] Discriminator Receive target domain fault diagnosis data with pseudo fault data in the target domain.

[0100] Time Domain Discriminator The target domain pseudo vibration data and target domain pseudo fault data are spliced ​​as the first input, and the target domain fault diagnosis data is used as the second input. The convolutional neural network is used to extract the time domain features and output the authenticity probability. .

[0101] Frequency Domain Discriminator Perform fast Fourier transform on the target domain vibration data to obtain the amplitude spectrum , using convolutional neural networks to analyze target domain pseudo vibration data and amplitude spectra Frequency domain distribution features and output rationality probability .

[0102] Generator The loss function Expressed as: (3).

[0103] (4).

[0104] (5).

[0105] Discriminator The loss function Expressed as: (6).

[0106] The objective function of the dynamic adversarial generation network data enhancement model is expressed as: (7).

[0107] Where, is the time domain discriminator loss weight (value is 0.4), is the frequency domain discriminator weight (value is 0.6), is the target domain fault diagnosis data, which obeys the real data distribution. is the fast Fourier transform, For the expected operation, the target domain fault diagnosis data and the target domain pseudo-fault data are averaged respectively. is the pseudo vibration data in the target domain, is the target domain pseudo current data, is the time domain adversarial loss, is the frequency domain adversarial loss, To ensure the energy of the key fault frequency band of the generated signal for physical constraint loss Not less than the fault energy threshold , to avoid generating invalid fault signatures, is the frequency band weight from the SimAM frequency domain attention mechanism, For the spectral fidelity constraint loss, the Euclidean distance between the spectrum of the fault diagnosis data in the target domain and the spectrum of the fault diagnosis data in the target domain is forced to be minimized.

[0108] Fault energy threshold Calibration is performed through experiments, which specifically includes the following steps: Step 1: In the laboratory waters, change the salinity and temperature, and set a total of 25 groups of operating conditions, among which the salinity is set to: 0%, 15%, 25%, 35%, 45%, and the temperature is set to: 5℃, 10℃, 15℃, 25℃, 35℃.

[0109] Step 2: Collect 100 sets of vibration data of the propeller thrusters of underwater robots in normal operation, perform 5-layer wavelet packet decomposition on each set of vibration data, divide each set of vibration data into 32 frequency bands, calculate the energy of each frequency band according to formula (1), and calculate the energy of each frequency band. , calculate the mean of 100 groups of energy and standard deviation .

[0110] Step 3: According to the formula Calculating the fault energy threshold .

[0111] Generator By minimizing the objective function , so that the generated data Can deceive the discriminator ,let The discriminator considers the generated data to be real; By maximizing the objective function ,make Able to accurately distinguish real fault data and pseudo-fault data .

[0112] 4. Construct a cross-domain fault diagnosis model of the environmental residual adaptation network.

[0113] Although DAGAN generates pseudo-fault data tailored to the target domain environment, it cannot capture all environmentally relevant physical effects, and there is still an offset between the feature layers of the pseudo-fault data and the measured data in the target domain. The effects of temperature and salinity on features exhibit nonlinear dynamic characteristics. Traditional cross-domain fault diagnosis methods (such as DANN) can only align global distributions and cannot establish a mapping relationship between environmental parameters and feature residuals. They also fail to optimize fault-specific residual information between different data features, making them difficult to handle extreme environmental differences. Traditional domain adaptation methods only rely on global distribution alignment and are difficult to deal with the problem of local feature offset caused by environmental parameters. Therefore, a residual compensation module is designed. By capturing the measured features of the target domain and the pseudo-fault feature residuals generated by DAGAN, combined with the salinity and temperature environmental parameter encoding, the feature residual compensation amount that is strongly related to the environment is dynamically predicted to achieve global distribution alignment of feature offset, and compensate for the local feature deviation caused by environmental differences, achieving local distribution alignment of feature offset. This provides a cross-domain fault diagnosis method of "environmentally sensitive residual adaptive compensation" for underwater robot thrusters in complex water environments, realizes cross-domain adaptive diagnosis, solves the problem of feature layer offset in complex waters, and significantly improves cross-domain generalization capability and diagnostic accuracy.

[0114] like Figure 3 As shown, ERAN includes a feature extractor , residual compensation module and fault classifier ,The feature extractor includes shared feature extractor and private feature extractor.

[0115] The target domain pseudo-fault data, target domain fault diagnosis data and environmental parameters are input into the cross-domain fault diagnosis model of the environmental residual adaptation network.

[0116] For the target domain pseudo-fault data and target domain fault diagnosis data, the shared feature extractor uses a two-layer one-dimensional convolutional neural network. The first layer uses 64 5×1 convolution kernels with a stride of 2 to extract macro features. After maximum pooling compression, the second layer uses 128 3×1 convolution kernels to capture micro fluctuations. Finally, it is flattened and outputs a 256-dimensional target domain original fault feature vector through a fully connected layer. and the target domain pseudo-fault feature vector .

[0117] For environmental parameters, the private feature extractor uses a two-layer ReLU activated fully connected network to compress and encode environmental information, and linearly maps the output to a 16-dimensional environmental factor vector .

[0118] 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 features and pseudo-fault features.

[0119] 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.

[0120] Calculate the local feature residual remaining after global alignment of the original fault feature vector in the target domain and the pseudo-fault feature vector in the target domain, as shown in formula (8): (8).

[0121] Environmental factors and local feature residual Splice into 272-dimensional vector input LSTM network to establish environment parameters Offset from feature The dynamic mapping is shown in formula (9): (9).

[0122] Where, Output vector of the forget gate, controlling the proportion of historical state retention (saltness dominated), Output vector for the input gate, controlling the ratio of new information written (temperature dominated), Output vector for the output gate to control the current cell state The degree to which the output reaches the hidden state, For the candidate state, the temporary state basis vector generated by the current input, is the input weight matrix, which is used for the linear transformation of the current input. is the cyclic weight matrix used for the linear transformation of the previous hidden state, is the bias vector, is the Sigmoid function, and tanh is the hyperbolic tangent function.

[0123] Calculation process: Concatenate input vectors: (272 dimensions).

[0124] Linear transformation: (256-dimensional vector).

[0125] Divided into four parts: .

[0126] Activate separately: ; ; ; .

[0127] Considering that the corrosion cumulative effect of salinity is long-term and the friction effect of temperature is immediate, the LSTM network realizes differentiated modeling through the gating mechanism: the forget 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 of the current environmental factor; the candidate state generates a compensation basis vector by fusing the gating signal to realize the dynamic fitting of the environment-sensitive residual, that is, the LSTM network realizes the dynamic fitting of the environment-sensitive residual through the forget gate. Control the strength of historical residual memory, input gate Adjust the contribution of current environmental factors, candidate state Generate compensation basis vectors.

[0128] Update the cell state as shown in formula (10): (10).

[0129] Where, The core memory unit that stores long-term environmental dependencies is the cell state updated at the current moment. is the cell state at the previous moment, the cumulative state of historical environmental influences, Indicates that the forget gate controls the retained part of the previous cell state, Represents the adopted part of the input gate controlling the candidate cell state.

[0130] Update the hidden state as shown in formula (11): (11).

[0131] Where, is the current hidden state, which represents the dynamic correlation feature between the environment and the residual (used to generate compensation).

[0132] The output of the LSTM network (hidden state ) will generate feature compensation through a fully connected layer, which is then used to correct the characteristics of the target domain. The LSTM network parameters ( , , ) are all obtained through training and learning, rather than manual setting, realizing the innovative idea of ​​environment-sensitive residual adaptive compensation.

[0133] In summary, formula (9) calculates the 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, thereby accurately compensating for feature offsets.

[0134] The final output gate generates a 64-dimensional hidden state , Decoded into 256-dimensional feature compensation by the fully connected layer .

[0135] Feature compensation The magnitude of is matched to the scale of the feature extractor through the LSTM output, ensuring that the directional consistency of the physical constraints is compatible with the feature space scale.

[0136] The characteristic compensation Applied to the original fault feature vector in the target domain , generate the compensated target domain feature vector .

[0137] The target domain feature vector generated by the residual compensation module is As the input of the fault classifier, the target domain feature vector Mapped to the fault category space through the fully connected layer, as shown in formula (12): (12).

[0138] Where, is the classification layer weight matrix, is the bias vector, Represents the probability distribution of four types of faults (single-leaf imbalance / double-leaf imbalance / partial defect / foreign object jamming).

[0139] The feature alignment loss is expressed as: (13).

[0140] Where, Express expectations, is the measured sample in the target domain, is the target domain data distribution, In order 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 to ensure that the feature distributions of the two types of data are aligned, and γ is a hyperparameter used to balance the two losses.

[0141] Physical significance: Local sample alignment is achieved by minimizing the Euclidean distance of each sample's feature vector. Global distribution alignment is achieved by measuring and aligning the overall differences in feature distributions through MMD, providing a reliable feature benchmark for the residual compensation module.

[0142] The physical constraint loss is expressed as: (14).

[0143] Where, is the regularization coefficient (value is 0.1), which is used to control the complexity of the model and avoid overfitting. is the parameter matrix of the LSTM network.

[0144] Physical meaning: Forced feature compensation through cosine similarity The linear combination of the dominant direction and the environmental parameters Consistency ensures that compensation logic complies with physical causality.

[0145] The fault classification loss is expressed as: (15).

[0146] Where, To find the expected distribution of pseudo-fault data in the target domain, is the pseudo-fault signature after compensation, is the pseudo fault data label of the target domain, is the probability of fault classification.

[0147] Physical meaning: Ensure that the compensated features can accurately classify the fault type.

[0148] The fault classification objective function is expressed as: (16).

[0149] Where, is the feature alignment weight (value is 1.0), is the physical constraint weight (value is 0.5), is the fault classification weight (value is 1.0), which is set based on the following considerations: Feature alignment is the basis of domain adaptation, and it is necessary to prioritize the consistency of feature distribution between the source domain and the target domain, so it is given the highest weight. =1.0; and physical law constraints need to be introduced on the basis of feature alignment to prevent the compensation amount from deviating from the actual physical effect, but the weight is lower than feature alignment to prevent over-constraint, so the weight =0.5; Fault classification and feature alignment are equally important, ensuring that the compensated features are not only distributed and aligned, but also accurately distinguish the fault type and achieve the diagnosis goal, so =1.0.

[0150] The optimization objectives are feature alignment loss, physical constraint loss and fault classification loss, with the aim of minimizing fault classification error, achieving domain-invariant feature learning and ensuring consistency of physical laws.

[0151] The embodiments of the present invention have so far described the technical solutions of the present invention in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters, characterized in that: The specific steps include: Acquire source domain fault diagnosis data of an underwater robot propeller thruster during simulated operation in laboratory waters, and extract source domain energy features from the source domain fault diagnosis data, wherein the source domain fault diagnosis data includes source domain vibration data and source domain current data; Obtain target domain fault diagnosis data and environmental parameters of the underwater robot propeller thruster in actual operation in a real water environment, where 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 data enhancement model is constructed. Source domain energy characteristics, target domain fault diagnosis data, and environmental parameters are input into the dynamic adversarial generative network data enhancement model. The dynamic adversarial generative network data enhancement model determines the frequency domain physical law constraints based on the source domain energy characteristics and target domain environmental parameters, and generates pseudo-fault data in the target domain. The dynamic adversarial generative 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. The discriminator includes a time domain discriminator and a frequency domain discriminator. An environmental residual adaptive network cross-domain fault diagnosis model is constructed, and the target domain pseudo-fault data, target domain fault diagnosis data and environmental parameters are input into the environmental residual adaptive network cross-domain fault diagnosis model. The environmental residual adaptive network cross-domain fault diagnosis model extracts the measured features, pseudo-fault features and environmental factors of the target domain, and fuses the residuals between the measured features and pseudo-fault features of the target domain with the environmental factors, constructs a dynamic mapping relationship between environmental parameters and feature residuals, adaptively generates feature compensation amounts, performs fault diagnosis on the compensated features, and obtains the fault type. The environmental residual adaptive network cross-domain fault diagnosis model 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.

2. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 1 is characterized in that: The method of extracting 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-layer wavelet packet decomposition and Divide into 32 frequency bands and obtain 32 frequency band wavelet packet coefficients ; Energy feature extraction: based on 32 frequency band wavelet packet coefficients By calculating the energy of each frequency band, a total of 32 frequency bands of source domain vibration energy eigenvectors are obtained and the source domain current energy eigenvector , forming the frequency domain energy feature vector , as shown in formula (1): (1)； Where, For the The energy of the frequency band, For the The wavelet packet coefficients of the frequency band, For the The number of wavelet packet coefficients in a frequency band.

3. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 2 is characterized in that: The dynamic adversarial network data enhancement model determines the frequency domain physical law constraints based on the source domain energy characteristics and the 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 domain salinity and temperature and a random noise vector As input; Environmental parameter encoding module for target domain salinity and temperature Encode the salinity and temperature Mapped to a frequency band offset to ensure that the main frequency of the generated signal is adapted to the current water area, as shown in formula (2): (2)； Where, is the frequency band index offset, is the salinity influence coefficient, is the temperature influence coefficient, and is the source domain benchmark environmental parameter; Residual attention module for source domain vibration energy features Adjust the frequency band offset , and use SimAM frequency domain attention mechanism to calculate the weight of each frequency band And generate weighted source domain vibration energy features ; The inverse wavelet packet synthesis module converts the weighted source domain vibration energy characteristics Convert to time domain signal to generate target domain pseudo vibration data ; The random noise vector Pseudo vibration data in the target domain Splicing input into the multi-layer perceptron to generate pseudo current data in the target domain ; The target domain pseudo vibration data and target domain pseudo current data Synthesize into pseudo fault data in the target domain ; Discriminator Receive target domain fault diagnosis data and target domain pseudo-fault data; Time Domain Discriminator The target domain pseudo vibration data and target domain pseudo fault data are spliced ​​as the first input, and the target domain fault diagnosis data is used as the second input. The convolutional neural network is used to extract the time domain features and output the authenticity probability. ; Frequency domain discriminator Perform fast Fourier transform on the target domain vibration data to obtain the amplitude spectrum , using convolutional neural networks to analyze target domain pseudo vibration data and amplitude spectra Frequency domain distribution features and output rationality probability ; Generator The loss function Expressed as: (3)； (4)； (5)； Discriminator The loss function Expressed as: (6)； The objective function of the dynamic adversarial generation network data enhancement model is expressed as: (7)； Where, is the loss weight of the time domain discriminator, is the frequency domain discriminator weight, is the target domain fault diagnosis data, is the fast Fourier transform, For the expected operation, the target domain fault diagnosis data and the target domain pseudo-fault data are averaged respectively. is the pseudo vibration data in the target domain, is the target domain pseudo current data, is the time domain adversarial loss, is the frequency domain adversarial loss, is the physical constraint loss, is the fault energy threshold, is the frequency band weight from the SimAM frequency domain attention mechanism, Constraining losses for spectral fidelity; Generator By minimizing the objective function , so that the generated data Can deceive the discriminator ,let The discriminator considers the generated data to be real; By maximizing the objective function ,make Able to accurately distinguish real fault data and pseudo-fault data .

4. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 3 is characterized in that: The cross-domain fault diagnosis model of the environmental residual adaptation network extracts the measured features, pseudo-fault features, and environmental factors of the target domain, fuses the residuals between the measured features and pseudo-fault features of the target domain with the environmental factors, constructs a dynamic mapping relationship between environmental parameters and feature residuals, adaptively generates feature compensation, performs fault diagnosis on the compensated features, and obtains the fault type. Specifically, the model includes the following steps: For the target domain pseudo-fault data and target domain fault diagnosis data, the shared feature extractor uses a two-layer one-dimensional convolutional neural network. The first layer uses 64 5×1 convolution kernels with a stride of 2 to extract macro features. After maximum pooling compression, the second layer uses 128 3×1 convolution kernels to capture micro fluctuations. Finally, it is flattened and outputs a 256-dimensional target domain original fault feature vector through a fully connected layer. and the 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 linearly maps the output to a 16-dimensional environmental factor vector ; MMD is used to globally align the original fault feature vector of the target domain and the pseudo-fault feature vector of the target domain to reduce the overall distribution difference between the original fault features and the pseudo-fault features of 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; Calculate the local feature residual remaining after global alignment of the original fault feature vector in the target domain and the pseudo-fault feature vector in the target domain, as shown in formula (8): (8)； Environmental factors and local feature residual Splice into 272-dimensional vector input LSTM network to establish environment parameters Offset from feature The dynamic mapping is shown in formula (9): (9)； Where, is the output vector of the forget gate, is the input gate output vector, is the output vector of the output gate, For candidate status, is the input weight matrix, is the circulant weight matrix, is the bias vector, is the Sigmoid function, and tanh is the hyperbolic tangent function; Update the cell state as shown in formulas (10) and (11): (10)； (11)； Where, is the cell state after updating at the current moment, is the cell state at the previous moment, is the current hidden state; The final output gate generates a 64-dimensional hidden state , Decoded into 256-dimensional feature compensation by the fully connected layer ; Feature compensation The amplitude is matched with the scale of the feature extractor through the LSTM network output to ensure that the direction consistency of the physical constraint is compatible with the feature space scale; The characteristic compensation 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 is As the input of the fault classifier, the target domain feature vector Mapped to the fault category space through the fully connected layer, as shown in formula (12): (12)； Where, is the classification layer weight matrix, is the bias vector, represents the failure probability distribution; The feature alignment loss is expressed as: (13)； Where, Express expectations, is the measured sample in the target domain, is the target domain data distribution, 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)； Where, is the regularization coefficient, is the parameter matrix of the LSTM network; The fault classification loss is expressed as: (15)； Where, To find the expected distribution of pseudo-fault data in the target domain, is the pseudo-fault signature after compensation, is the pseudo fault data label of the target domain, is the probability of fault classification; The fault classification objective function is expressed as: (16)； Where, is the feature alignment weight; is the physical constraint weight; is the fault classification weight.

5. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 4 is characterized in that: The salinity influence coefficient and temperature effect coefficient Calibration is performed through experiments, which specifically includes the following steps: Step 1: In the laboratory waters, change the salinity and temperature, setting a total of 25 groups of operating conditions. Among them, the salinity is set to: 0%, 15%, 25%, 35%, 45%, and the temperature is set to: 5℃, 10℃, 15℃, 25℃, 35℃. 0% salinity and 25℃ temperature are set as the source domain baseline environmental parameters; Step 2: Under each set of working conditions, control the propeller of the underwater robot to run at a constant speed of 1000RPM. According to the formula Obtain the nominal fundamental frequency ; Step 3: Extract the measured fundamental frequency from the collected source domain vibration data through FFT analysis ; Step 4: Calculate the fundamental frequency offset ; Step 5: Build a multiple linear regression model ; Step 6: Use the least squares method to fit the coefficient formula , and obtain the salinity influence coefficient and temperature effect coefficient .

6. The cross-domain intelligent fault diagnosis method for underwater robot propeller thrusters according to claim 5 is characterized in that: The fault energy threshold Calibration is performed through experiments, which specifically includes the following steps: Step 1: In the laboratory waters, change the salinity and temperature, setting a total of 25 groups of operating conditions, where the salinity is set to 0%, 15%, 25%, 35%, and 45%, and the temperature is set to 5°C, 10°C, 15°C, 25°C, and 35°C; Step 2: Collect 100 sets of vibration data of the propeller thrusters of underwater robots in normal operation, perform 5-layer wavelet packet decomposition on each set of vibration data, divide each set of vibration data into 32 frequency bands, calculate the energy of each frequency band according to formula (1), and calculate the energy of each frequency band. , calculate the mean of 100 groups of energy and standard deviation ; Step 3: According to the formula Calculating the fault energy threshold .

Images