Fault diagnosis method for underwater vector thruster

Abstract

The invention provides a fault diagnosis method for an underwater vector propeller, and relates to the technical field of fault diagnosis. The method specifically comprises the following steps: generating a historical fault data set; preprocessing the historical fault data set; constructing a fusion model of the multi-source heterogeneous data; inputting the trained data into a data fusion model, and outputting a fusion result; inputting data from the training set into three hybrid deep learning models combined with a self-attention module for training, then taking prediction results of the three hybrid deep learning models as input of ensemble learning, and generating a final prediction result by using a weighted ensemble strategy; adding a white whale optimization algorithm of a dynamic pinhole imaging strategy to optimize hyper-parameters in the neural network model; and performing fault diagnosis. According to the technical scheme, the problems that in the prior art, an underwater vehicle is low in reliability and safety, high in maintenance cost and short in equipment service life are solved.

Description

Technical Field

[0001] The invention relates to the field of fault diagnosis, and in particular to a fault diagnosis method for an underwater vector thruster. Background Art

[0002] Underwater vector thrusters play a vital role in the field of underwater navigation. However, due to the complex and changeable underwater environment, underwater vector thrusters face many potential failure risks.

[0003] Underwater vector thrusters are usually complex systems that integrate technologies from multiple fields such as mechanics, electrical, and electronics. The mechanical parts include drive shafts, propellers, and sealing devices, which are susceptible to seawater corrosion, wear, and fatigue damage during long-term underwater operation. For example, the drive shaft may wear out due to long-term high-speed rotation, resulting in a decrease in power transmission efficiency; the propeller may be deformed or damaged by underwater foreign objects, affecting the propulsion effect. In terms of electrical systems, motors, controllers, and other components may have short circuits, open circuits, overheating, and other faults. For example, the motor winding may short-circuit due to insulation damage, resulting in excessive current and burning the motor; the cable may break due to external force pulling or corrosion, affecting power transmission. The electronic part mainly includes sensors, control systems, etc. The failure of the sensor may make it impossible to accurately obtain the operating status information of the thruster, affecting the control accuracy and stability; the failure of the control system may cause the thruster to lose control and fail to work properly. Moreover, the particularity of the underwater environment brings great difficulties to the fault diagnosis and maintenance of underwater vector thrusters. Harsh conditions such as high pressure, low temperature, and darkness underwater require fault diagnosis equipment and technicians to have a higher level of professionalism and adaptability. At the same time, underwater maintenance operations usually require special equipment and tools, and the maintenance costs are high and the risks are relatively high.

[0004] Therefore, there is a need for a fault diagnosis method for an underwater vector thruster that can improve the reliability and safety of underwater vehicles, reduce maintenance costs, and increase the service life of equipment. Summary of the invention

[0005] The main purpose of the present invention is to provide a fault diagnosis method for an underwater vector thruster to solve the problems of low reliability and low safety, high maintenance cost and short equipment service life of underwater vehicles in the prior art.

[0006] To achieve the above object, the present invention provides a fault diagnosis method for an underwater vector thruster, which specifically comprises the following steps: S1, generates a historical fault data set by obtaining historical fault data from sensors carried on the underwater vector thruster.

[0007] S2, preprocessing of historical fault data sets, including: data cleaning, noise reduction and standardization.

[0008] S3, extracts the historical fault data features of the preprocessed multi-source sensors, uses the support vector machine algorithm to obtain the optimal classification hyperplane, and constructs a fusion model for multi-source heterogeneous data; enters the trained data into the data fusion model, outputs the fusion results, and obtains the processed data set.

[0009] S4, the data set obtained after data processing is divided into a training set and a test set according to a certain ratio. In the training stage, the data from the training set is input into three hybrid deep learning models combined with self-attention modules for training. The three hybrid deep learning models combined with self-attention modules are: convolutional neural network-long short-term memory network, convolutional neural network-gated recurrent unit and convolutional neural network-gated recurrent unit-long short-term memory network. The prediction results of the three hybrid deep learning models are then used as the input of ensemble learning, and the optimal weights are obtained through the improved random grid search method, and the weighted integration strategy is used to generate the final prediction results.

[0010] S5, the Beluga optimization algorithm with dynamic pinhole imaging strategy is used to optimize the hyperparameters in the neural network model, including the number of convolution kernels M, dimension N, and the number of neurons in the long short-term memory network layer and the gated recurrent unit layer of the convolutional neural network layer.

[0011] S6, input the test set into the fault diagnosis model optimized by the improved Beluga optimization algorithm to evaluate the model fault diagnosis effect. If the diagnosis effect is not optimal, continue to use the improved Beluga optimization algorithm to optimize the hyperparameters; if the diagnosis effect is optimal, the training is completed and the final fault diagnosis model is obtained.

[0012] S7, using the sensors carried on the underwater vector thruster to collect current, acceleration, vibration frequency and pressure online, perform data preprocessing, and then input them into the final fault diagnosis model obtained in step S6 to perform fault diagnosis.

[0013] Furthermore, step S2 specifically includes the following steps: S2.1, using 3 Data cleaning is performed according to the principle, and the mean and standard deviation of the historical fault data obtained by each sensor are calculated, and the mean value is ±3 Data with a wide range are marked as outliers and are deleted or replaced with the median.

[0014] S2.2, in the noise reduction processing stage, the historical fault data obtained by the sensor is subjected to empirical mode decomposition (EMD) to decompose the data into several intrinsic mode functions (IMFs) and a residual term. According to the frequency characteristics and energy distribution of the IMFs, a part of the IMFs is selected for processing, and the remaining IMFs and residual terms are reconstructed to obtain the noise-reduced signal.

[0015] S2.3, in the data standardization stage, for each historical fault data, find the minimum value and maximum value , each historical fault data point of each sensor is standardized using the Min-Max normalization formula, and the historical fault data is mapped to the [0,1] interval.

[0016] Furthermore, step S3 specifically includes the following steps: S3.1, based on the results of data preprocessing in step S2, update the multivariate data set, and the updated data set is , the data status is , extracting multi-source heterogeneous data In the The characteristic vector of the attribute is calculated as follows: (1); Among them, the feature vector of multi-source heterogeneous data is Indicates that the data attribute is recorded during the period express.

[0017] S3.2, set the value obtained in step S3.1 is the input vector , the kernel function is , the feature vector set is , the output vector is , in order to obtain the optimal hyperplane of the data, the process is as follows: (2); In the formula, the optimal hyperplane is obtained using The weighting coefficient is expressed as Indicates; the bias term is expressed as The number of eigenvectors is represented by express; is a symbolic function; The symbolic function is defined as: (3).

[0018] S3.3, based on the obtained optimal hyperplane, the multi-source heterogeneous data is divided into h groups of data, and the data of different groups come from different sensors. The generalization parameter is set as , in order to establish a fusion model of multi-source heterogeneous data of support vector machine: (4); Among them, the fusion model of multi-source heterogeneous data is constructed using Representation; Multi-source heterogeneous data , The optimal solutions are respectively , Indicates; constraints are expressed by Indicates the number of sensors. express, , Represents multi-source heterogeneous data , The optimal solution mean, is the kernel function.

[0019] S3.4, based on the determined optimal solution of the data, establish the decision function of the data and complete the model solution. The process is shown in the following formula: (5); Among them, the decision function constructed is Represents; the mean value of multi-source heterogeneous data is expressed as ; Data optimal solution express.

[0020] Furthermore, step S4 specifically includes the following steps: S4.1, build a convolutional neural network-long short-term memory network model: The formulas for convolution calculation and maximum pooling in convolutional neural networks are as follows: (6); (7); in, is the input feature vector; represents the convolution kernel; Indicates network bias; express Layer output, express Layer input; is the pooling width; is the ReLU activation function; express The indices of the different neurons in the layer.

[0021] The update formula of the long short-term memory network is as follows: (8); (9); (10); (11); (12); In the formula, , , They are the input gate, forget gate, and output gate at time step The output value of , , At each time step Input data, unit state, hidden state; , is the previous time step The hidden state and unit state of ; among them, , , , Input To the weight matrix of the input gate, forget gate, output gate, and candidate unit state; , , , They are the hidden states of the previous moment. To the weight matrix of the input gate, forget gate, output gate, and candidate unit state; , , and They are the bias items of input gate, forget gate, output gate, and candidate unit state respectively; is the sigmoid function.

[0022] Furthermore, step S4 further includes the following steps: S4.2, build a convolutional neural network-gated recurrent unit model: The long short-term memory network is changed to a gated recurrent unit network. The updated gated recurrent unit formula is as follows: (13); (14); (15); (16); In the formula, , At time step The output value of the reset gate and update gate; , At time step Candidate hidden states, hidden states; , , They are reset gate, update gate, and weight matrix for calculating candidate hidden states; , , They are the reset gate, update gate, and bias item for calculating candidate hidden states.

[0023] S4.3, the convolutional neural network-gated recurrent unit-long short-term memory network model is constructed by adding a gated recurrent unit layer and a long short-term memory network layer in sequence after the convolutional neural network layer.

[0024] Furthermore, step S4 specifically includes the following steps: S4.4, features obtained after extraction by three models , then Perform a linear transformation, Mapped to query space , key space Sum value space ; Then use the scaled dot product and Softmax function to calculate each attention distribution, and weighted sum the attention distribution to get the corresponding output ; Finally, multiple output results are spliced ​​together using the splicing method. The formula is as follows: (17); (18); (19); (20); (twenty one); Where: , , are the linear transformation parameters of query space, key space, and value space respectively; For the query vector, For the key vectors, For the A vector of values, is a matrix consisting of the number of dimensions for each key; is the element vector in the key space; is the element vector in the value space; is the feature concatenation function; for The transpose transformation of An attention scoring function for scaling dot products; is the Softmax function, which converts a real number vector into a probability distribution vector; Indicates The output result of each head; Represents the final multi-head attention mechanism output, by splicing multiple get.

[0025] S4.5, the optimal weighted set weight is obtained according to the model search results, and the model with the largest prediction probability is selected from the weighted prediction results as the final fault diagnosis type output. The principle of weighted integration is shown in (22): (twenty two); In the formula, is the number of models, the value is 3; is the predicted probability of the model; is the weight and satisfies , .

[0026] Furthermore, step S5 specifically includes the following steps: S5.1, according to the population mechanism of the improved Beluga optimization algorithm, the Beluga is regarded as the search subject, and the movement effect in the search space is achieved by changing the position vector of the Beluga. The search agent position matrix for: (twenty three); In the formula, To search for the number of white whales, is the dimension of the design variables.

[0027] Fitness value function of Beluga whale Expressed as: (twenty four).

[0028] Balance Factor Changes in each iteration: (25); In the formula, For the current iteration, is the maximum number of iterations, is a random number in the interval (0,1). When , it corresponds to the development stage; the balance coefficient ≤0.5 corresponds to the exploration stage. As the number of iterations increases, the probability of entering the development stage increases, while the probability of entering the exploration stage decreases.

[0029] Furthermore, step S5 specifically includes the following steps: S5.2, in the exploration phase, the movement of the white whale determines the position update of the BWO algorithm in the detection phase. The position update of the white whale is as follows: (26); For the next iteration The individual in The position on the dimension, is an even number, is an odd number, and is A positive integer randomly selected from , and not equal, and Respectively represent the current iteration , The location of an individual, , is a random number in (0,1).

[0030] S5.3, in the development stage, beluga whales move and feed according to the location of nearby beluga whales. Beluga whales catch prey by sharing each other's location information using the Levy running strategy. The specific formula is: (27); (28); and Respectively represent the , The current location of each individual, For the The new position of each individual, It is the best location. and is a random number in (0,1), are random jumps, measuring the strength of Levy's run.

[0031] is a function of the Levy flight strategy and is calculated as follows: (29); (30); in, and is a random number that follows a normal distribution, is a parameter that determines the scale of the Levy distribution. is the gamma function in The value at is a constant set to 1.5.

[0032] Furthermore, step S5 specifically includes the following steps: S5.4, during the whale fall phase, beluga whales either migrate to other places or die; in order to keep the population size unchanged, the position of the beluga whale and the step size of the whale fall are used to establish the updated position, the formula is as follows: (31); in is a random number in (0,1), is the step size of the whale's descent, established as: (32); is the step factor, which is related to the probability of beluga whale decline and population size, specifically: ; and are the upper and lower bounds of the variable respectively; the probability of dropping the white whale Expressed as a linear function, the formula is as follows: (33).

[0033] S5.5, the dynamic pinhole imaging strategy is used to improve the White Whale optimization algorithm for global optimization, and the following mathematical model is obtained: (34); and are the lengths of the virtual candle at the current optimal position and the opposite position, respectively.

[0034] Set the ratio of the two candles as a variable , from which we can get: (35); in, Indicates The white whale The first iteration Dimension location.

[0035] The present invention has the following beneficial effects: The method proposed in the present invention realizes efficient and accurate diagnosis of thruster faults by comprehensively utilizing multi-source sensor data carried on vector thrusters. The core of the method lies in the weighted integration method based on the hybrid deep learning model fusion multi-head attention mechanism. Through multiple links such as data preprocessing, data fusion, model training and optimization, the accurate diagnosis of vector thruster faults is finally realized. In the data fusion stage, the present invention adopts the data fusion model of support vector machine to effectively fuse multi-source data and extract key features that can reflect the state of the thruster. This step not only improves the utilization rate of data, but also provides more accurate and comprehensive input for subsequent fault diagnosis; in the model training stage, the convolutional neural network is combined with the long short-term memory network and the gated recurrent unit to form three hybrid deep neural networks as the basis of the fault diagnosis model. The powerful feature extraction ability of the convolutional neural network can well extract the local fault features of the running sequence, and the good processing ability of the long short-term memory network and the gated recurrent unit for time series data and the ability to capture contextual information can better discover the intrinsic fault information of the thruster data. At the same time, the model integrates the multi-head attention mechanism, and through the integration of multiple attention heads, it can reduce the dependence on a single feature and improve the robustness of the model. The multi-head attention mechanism can also automatically learn to adapt to different input data distributions and improve the generalization ability of the model. Then, the fast optimization of the integrated weights of the convolutional neural network-long short-term memory network, the convolutional neural network-gated recurrent unit and the convolutional neural network-gated recurrent unit-long short-term memory network is realized by an improved random grid search algorithm. This search algorithm searches all possible parameter combinations according to the given parameter range and step size, and can balance the search speed and the demand for the global optimal solution, better play the advantages of each model, and improve the detection accuracy and generalization ability of the final weighted integrated model; in the model optimization stage, the improved Beluga optimization algorithm is used to optimize the fault diagnosis model. The Beluga optimization algorithm has strong global search ability and fast convergence characteristics. It can adaptively adjust the optimization strategy and achieve effective optimization of model parameters within a limited number of iterations. And the addition of the dynamic pinhole imaging strategy increases the exploration ability of the Beluga optimization algorithm in the search space, guides the Beluga individuals to move faster to the optimal solution area, and finally forms a better fault diagnosis model. In general, the method of the present invention plays a key role in the fault diagnosis of submarine-launched unmanned aerial vehicles and provides a strong guarantee for the safe and stable operation of vector thrusters. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. Obviously, the drawings described below are some implementations of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 A flow chart of a fault diagnosis method for an underwater vector thruster of the present invention is shown.

[0037] Figure 2 The structure and sensor arrangement diagram of the underwater vector thruster of the present invention are shown.

[0038] Figure 3 A fault diagnosis system diagram of the underwater vector thruster of the present invention is shown.

[0039] The reference numerals in the above drawings are: 1. Vibration sensor; 2. Acceleration sensor; 3. Pressure sensor; 4. Current sensor; 5. Hall angle sensor. DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be described clearly and completely below 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 creative work are within the scope of protection of the present invention.

[0041] like Figure 1 A fault diagnosis method for an underwater vector thruster shown in the figure specifically comprises the following steps: S1, generate a historical fault data set through historical fault data obtained by sensors carried on the underwater vector thruster. The sensors carried on the underwater vector thruster, such as current sensors, acceleration sensors, vibration sensors and pressure sensors, measure multi-source sensor data such as current, acceleration, vibration frequency and pressure to collect historical fault data of the vector thruster, including data on electrical system faults, power system faults and mechanical mechanism faults, and generate a historical data set.

[0042] S2, preprocessing of historical fault data sets, including: data cleaning, noise reduction and standardization.

[0043] S3, extracts the historical fault data features of the preprocessed multi-source sensors, uses the support vector machine algorithm to obtain the optimal classification hyperplane, and constructs a fusion model for multi-source heterogeneous data; enters the trained data into the data fusion model, outputs the fusion results, and obtains the processed data set.

[0044] S4, the data set obtained after data processing is divided into a training set and a test set according to a certain ratio. In the training stage, the data from the training set is input into three hybrid deep learning models combined with self-attention modules for training. The three hybrid deep learning models combined with self-attention modules are: convolutional neural network-long short-term memory network, convolutional neural network-gated recurrent unit and convolutional neural network-gated recurrent unit-long short-term memory network. The prediction results of the three hybrid deep learning models are then used as the input of ensemble learning, and the optimal weights are obtained through the improved random grid search method, and the weighted integration strategy is used to generate the final prediction results.

[0045] S5, the Beluga optimization algorithm with dynamic pinhole imaging strategy is used to optimize the hyperparameters in the neural network model, including the number of convolution kernels M, dimension N, and the number of neurons in the long short-term memory network layer and the gated recurrent unit layer of the convolutional neural network layer.

[0046] S6, input the test set into the fault diagnosis model optimized by the improved Beluga optimization algorithm to evaluate the model fault diagnosis effect. If the diagnosis effect is not optimal, continue to use the improved Beluga optimization algorithm to optimize the hyperparameters; if the diagnosis effect is optimal, the training is completed and the final fault diagnosis model is obtained.

[0047] S7, using the sensors carried on the underwater vector thruster to collect current, acceleration, vibration frequency and pressure online, perform data processing, and then input them into the final fault diagnosis model obtained in step S6 to perform fault diagnosis.

[0048] Specifically, step S2 includes the following steps: S2.1, data cleaning stage, for outlier detection and processing, 3 Principle: For data that obeys a normal distribution, if the distance between a data point and the mean exceeds three times the standard deviation, it is considered an outlier. Data cleaning is performed according to the principle, and the mean and standard deviation of the historical fault data obtained by each sensor are calculated, and the mean value is ±3 Data with a wide range are marked as outliers and are deleted or replaced with the median.

[0049] S2.2, in the noise reduction processing stage, the historical fault data obtained by the sensor is subjected to empirical mode decomposition (EMD). IMFs have different frequencies and time scales. Usually, high-frequency IMFs correspond to noise and fast-changing parts, while low-frequency IMFs contain the main trends of the data. The purpose of noise reduction can be achieved by removing high-frequency IMFs or performing threshold processing on them. For example, the acceleration sensor data is decomposed into several intrinsic mode functions (IMFs) and a residual term. According to the frequency characteristics and energy distribution of the IMFs, a part of the IMFs is selected for processing, and the remaining IMFs and residual terms are reconstructed to obtain the noise-reduced signal.

[0050] S2.3, in the data standardization stage, for each historical fault data, find the minimum value and maximum value , using the Min-Max normalization formula: Each historical fault data point of each sensor is standardized, and the historical fault data is mapped to the [0,1] interval.

[0051] Specifically, step S3 includes the following steps: S3.1, based on the results of data preprocessing in step S2, update the multivariate data set, and the updated data set is , the data status is , extracting multi-source heterogeneous data In the The characteristic vector of the attribute is calculated as follows: (1); Among them, the feature vector of multi-source heterogeneous data is Indicates that the data attribute is recorded during the period express.

[0052] S3.2, using the support vector machine algorithm to accelerate the convergence speed through the kernel function, control the dimension of the space, map the data feature nonlinear vector space to the linear separable space, obtain the optimal classification hyperplane, and build a data fusion model for multi-source heterogeneous data. Set the data obtained in step S3.1 is the input vector , the kernel function is , the feature vector set is , the output vector is , in order to obtain the optimal hyperplane of the data, the process is as follows: (2); In the formula, the optimal hyperplane is obtained using The weighting coefficient is expressed as Indicates; the bias term is expressed as The number of eigenvectors is represented by express; is a symbolic function.

[0053] The symbolic function is defined as: (3).

[0054] S3.3, based on the obtained optimal hyperplane, the multi-source heterogeneous data is divided into h groups of data, and the data of different groups come from different sensors. The generalization parameter is set as , in order to establish a fusion model of multi-source heterogeneous data of support vector machine: (4); Among them, the fusion model of multi-source heterogeneous data is constructed using Representation; Multi-source heterogeneous data , The optimal solutions are , Indicates; constraints are expressed by Indicates the number of sensors. express, , Represents multi-source heterogeneous data , The optimal solution mean, is the kernel function that maps the data into a high-dimensional space.

[0055] S3.4, based on the determined optimal solution of the data, establish the decision function of the data and complete the model solution. The process is shown in the following formula: (5); Among them, the decision function constructed is Represents; the mean value of multi-source heterogeneous data is expressed as ; Data optimal solution express.

[0056] Based on the data fusion process, the training set that has completed the training processing is entered into the data fusion model. According to the kernel function, loss function and generalization parameters, the constructed data fusion model is verified, and the fusion result is output to realize the fusion of multi-source data.

[0057] Specifically, the data set obtained after data processing is divided into training set and test set according to a certain ratio, which is set to 80% training set and 20% test set. In the training stage, the data from the training set is input into three hybrid deep learning models combined with self-attention modules for training, and then the prediction results of the three models are used as the input of ensemble learning, and the optimal weights are obtained through the improved random grid search method, and the weighted ensemble strategy is used to generate the final prediction results.

[0058] Step S4 specifically includes the following steps: S4.1, build a convolutional neural network-long short-term memory network model: The formulas for convolution calculation and maximum pooling in convolutional neural networks are as follows: (6); (7); in, is the input feature vector; represents the convolution kernel; Indicates network bias; express Layer output, express Layer input; is the pooling width; is the ReLU activation function; express The indices of the different neurons in the layer.

[0059] Each LSTM network unit has a forget gate, an input gate, and an output gate. The input gate can pass or partially pass the final output, the forget gate can control the information that needs to be updated, and the output gate is responsible for receiving the output state. The update formula of the LSTM network is as follows: (8); (9); (10); (11); (12); In the formula, , , They are the input gate, forget gate, and output gate at time step The output value of , , At each time step Input data, unit state, hidden state; , is the previous time step The hidden state and unit state of ; among them, , , , Input To the weight matrix of the input gate, forget gate, output gate, and candidate unit state; , , , They are the hidden states of the previous moment respectively To the weight matrix of the input gate, forget gate, output gate, and candidate unit state; , , and They are the bias items of input gate, forget gate, output gate, and candidate unit state respectively; is the sigmoid function.

[0060] The convolutional neural network layer parameters are composed of two multi-scale convolutional layers and two maximum pooling layers. The two convolutional layers have 64 filters, and the sizes of the convolution kernels are 5 and 3 respectively. The long short-term memory network layer can effectively capture the contextual information in the time series data, so as to better judge whether the thruster is faulty based on the forward and backward dependencies. The use of a two-layer long short-term memory network recursive neural network can achieve good performance. In summary, the local features in the convolutional neural network layer are extracted through the convolutional neural network-long short-term memory network model and input into two long short-term memory network layers with 32 and 64 hidden units respectively.

[0061] Specifically, step S4 also includes the following steps: S4.2, build a convolutional neural network-gated recurrent unit model: The LSTM network is changed to a gated recurrent unit network. The gated recurrent unit is an improved network of the LSTM network. It is simpler than the LSTM network and has only two gate recurrent units: reset gate and update gate. The reset gate determines how to combine new input information with previous memory, and the update gate determines which state information from the previous moment is discarded and which state information is updated. The updated gated recurrent unit formula is as follows: (13); (14); (15); (16); In the formula, , At time step The output value of the reset gate and update gate; , At time step Candidate hidden states, hidden states; , , They are reset gate, update gate, and weight matrix for calculating candidate hidden states; , , They are the reset gate, update gate, and bias item for calculating candidate hidden states.

[0062] In the structure of the convolutional neural network-gated recurrent unit model, the parameter settings of the convolutional neural network layer are the same as those of the convolutional neural network-long short-term memory network. For the gated recurrent unit layer, the gated recurrent unit has better parallelism and a simpler structure than the long short-term memory network. Two gated recurrent unit layers are still added after the convolutional neural network layer to further extract the fault information of the thruster. 64 hidden units are used in both gated recurrent unit layers to further improve the performance of the model.

[0063] S4.3, the convolutional neural network-gated recurrent unit-long short-term memory network model is constructed by adding a gated recurrent unit layer and a long short-term memory network layer in sequence after the convolutional neural network layer.

[0064] In order to improve the accuracy and robustness of the weighted integrated thruster fault diagnosis model, three deep learning neural networks, namely convolutional neural network, long short-term memory network and gated recurrent unit, are combined to form a convolutional neural network-gated recurrent unit-long short-term memory network hybrid deep learning thruster fault diagnosis model. The introduction of different types of fault diagnosis models can capture more features of thruster fault data from different aspects, reduce the possibility of degradation of the fault diagnosis performance of the final integrated model due to the deviation of a single model, and thus achieve the purpose of improving the robustness of the weighted integrated model.

[0065] The convolutional neural network layer of the convolutional neural network-gated recurrent unit-long short-term memory network fault diagnosis model has two convolutional layers and two pooling layers, and the parameter settings are the same as those of the convolutional neural network-long short-term memory network model. A gated recurrent unit layer and a long short-term memory network layer are added after the convolutional neural network layer, and both have 64 hidden units.

[0066] Specifically, the attention mechanism is a method that imitates human visual attention. At a specific moment, only a few features in a large amount of input data are important. The attention mechanism can make the model pay more attention to these key parts, thereby improving the overall performance of the model.

[0067] Step S4 specifically includes the following steps: S4.4, features obtained after extraction by three models , then Perform a linear transformation, Mapped to query space , key space Sum value space ; Then use the scaled dot product and Softmax function to calculate each attention distribution, and weighted sum the attention distribution to get the corresponding output ; Finally, multiple output results are spliced ​​together using the splicing method. The formula is as follows: (17); (18); (19); (20); (twenty one); Where: , , are the linear transformation parameters of query space, key space, and value space respectively; For the query vector, For the key vectors, For the A vector of values, is a matrix consisting of the number of dimensions for each key; is the element vector in the key space; is the element vector in the value space; is the feature concatenation function; for The transpose transformation of An attention scoring function for scaling dot products; is the Softmax function, which converts a real number vector into a probability distribution vector; Indicates The output result of each head; Represents the final multi-head attention mechanism output, by splicing multiple get.

[0068] The linear transformation parameters of Q, K, and V are 32, the number of heads is 8, and the linear transformation parameter after Concat is 256. The classification layer adds a ReLU activation layer after the linear transformation with a parameter of 256, and then fully connects with 16 neurons, and finally activates it through the Softmax function.

[0069] S4.5, In actual operation, the thruster may experience various forms of failures, and the frequency and time of each failure are random, so the collected fault data may show unbalanced characteristics. For the same data set, fault diagnosis models with different structures will show different characteristics, thus having different diagnostic performance for different types of faults. The weighted ensemble method can solve problems such as data imbalance by assigning different weights to the ensemble model and combining the advantages of each model, thereby improving the accuracy and robustness of the overall fault diagnosis.

[0070] The present invention adopts a weighted integration method to integrate the diagnostic results of three different hybrid deep learning models. By assigning different weights to the prediction results of each model, the advantages of each model are fully utilized to obtain the maximum accuracy and comprehensive diagnostic results. First, a random search algorithm is used to randomly search the weights of the three independent models within a given range. Then, the weight combination set with the highest accuracy is selected from the random search, and a more intensive search is performed in the parameter space near each set of weight combination sets through grid search to obtain the optimal weight parameter combination. Finally, the optimal weighted set weight is obtained according to the model search results, and the model with the largest prediction probability is selected from the weighted prediction results as the final fault diagnosis type output. The principle of weighted integration is shown in (22): (twenty two); In the formula, is the number of models, the value is 3; is the predicted probability of the model; is the weight and satisfies , .

[0071] Specifically, in the process of optimizing the improved Beluga algorithm, the parameters of the neural network prediction model, including the number M and dimension N of the convolution kernels, the number of neurons in the long short-term memory network layer, and the number of neurons in the gated recurrent unit layer, are combined to form a parameter space, and Beluga is used as a search agent to find the optimal position to optimize the parameter space. At the same time, the mean square error (MSE) is defined as the fitness function to evaluate the performance of each parameter combination in the model prediction task.

[0072] The improved White Whale optimization algorithm consists of three stages: exploration stage, development stage and whale fall stage. When it is greater than 0.5, it corresponds to the occurrence of the exploration phase; otherwise, it corresponds to the occurrence of the development phase. When the probability of whale drop is greater than the balance factor, the whale drop phase occurs, and then the fitness value is calculated, and the best parameter combination solution found in the current iteration is output. After the exploration or development phase and the whale drop phase, it is considered to have completed an iteration, and iterates continuously through the steps of initializing the population and updating the population until the maximum number of iterations is completed. Finally, the fitness values ​​calculated in each iteration are compared, and the parameter combination corresponding to the minimum fitness value is the optimal parameter combination solution.

[0073] Step S5 specifically includes the following steps: S5.1, according to the population mechanism of the improved Beluga optimization algorithm, the Beluga is regarded as the search subject, and the movement effect in the search space is achieved by changing the position vector of the Beluga. The search agent position matrix for: (twenty three); In the formula, To search for the number of white whales, is the dimension of the design variables.

[0074] Fitness value function of Beluga whale Expressed as: (twenty four);

[0075] Balance Factor Changes in each iteration: (25); In the formula, For the current iteration, is the maximum number of iterations, is a random number in the interval (0,1). When , it corresponds to the development stage; the balance coefficient ≤0.5 corresponds to the exploration stage. As the number of iterations increases, the probability of entering the development stage increases, while the probability of entering the exploration stage decreases.

[0076] Specifically, step S5 includes the following steps: S5.2, in the exploration phase, the movement of the white whale determines the position update of the BWO algorithm in the detection phase. The position update of the white whale is as follows: (26); For the next iteration The individual in The position on the dimension, is an even number, is an odd number, and is A positive integer randomly selected from , and not equal, and Respectively represent the current iteration , The location of an individual, , is a random number in (0,1).

[0077] S5.3, in the development stage, beluga whales move and feed according to the location of nearby beluga whales. Beluga whales catch prey by sharing each other's location information using the Levy running strategy. The specific formula is: (27); (28); and Respectively represent the , The current location of each individual, For the The new position of each individual, It is the best location. and is a random number in (0,1), are random jumps, measuring the strength of Levy's run.

[0078] is a function of the Levy flight strategy and is calculated as follows: (29); (30); in, and is a random number that follows a normal distribution, is a parameter that determines the scale of the Levy distribution. is the gamma function in The value at is a constant set to 1.5.

[0079] Specifically, step S5 includes the following steps: S5.4, during the whale fall phase, beluga whales either migrate to other places or die; in order to keep the population size unchanged, the position of the beluga whale and the step size of the whale fall are used to establish the updated position, the formula is as follows: (31); in is a random number in (0,1), is the step size of the whale's descent, established as: (32).

[0080] is the step factor, which is related to the probability of beluga whale decline and population size, specifically: ; and are the upper and lower bounds of the variable respectively; the probability of dropping the white whale Expressed as a linear function, the formula is as follows: (33).

[0081] S5.5, the dynamic pinhole imaging strategy is used to improve the White Whale optimization algorithm for global optimization. The dynamic pinhole imaging strategy is a general strategy proposed by imitating the pinhole imaging theory in optics. It is more accurate than the common opposition-based learning strategy and can produce more diverse corresponding points. Applying it to the entire search space, the following mathematical model is obtained: (34); and are the lengths of the virtual candle at the current optimal position and the opposite position, respectively.

[0082] In the present invention, the position of the virtual candle in the ocean is also the position of the white whale, but the point representing a single white whale has no effective length. Therefore, the ratio of the two candles is set as the variable , from which we can get: (35); in, Indicates The white whale The first iteration Dimension location.

[0083] When the lengths of the two virtual candles are the same, this strategy becomes a basic reverse learning strategy. Properly adjusting the value of K can change the position of the relative point, giving individual white whales more search opportunities. In each iteration, a new relative point is generated near the center line of the search space, and when the other party is in a better position, its position becomes the new boundary.

[0084] Specifically, in step S6, accuracy, precision and recall are used as evaluation indicators of the model diagnosis effect, and the test set data is input into the loaded model to ensure that the model can correctly receive and process the input data, run the model for prediction, and obtain the fault diagnosis result of each sample. For each sample in the test set, find its corresponding true fault category label or normal state. Traverse each sample in the test set, compare the prediction result of the model with the corresponding true label, compare whether the predicted category label is consistent with the true label, and then record the comparison result of each sample, and calculate the evaluation accuracy, precision and recall. The calculation formula is as follows: (35); (36); (37); in, (True Positive) represents the true positive, that is, the number of samples that are actually faulty and diagnosed as faulty; (True Negative) represents true negative examples, that is, the number of samples that are actually normal and diagnosed as normal; (FalsePositive) represents false positive examples, that is, the number of samples that are actually normal but diagnosed as faulty; (False Negative) represents false negative examples, that is, the number of samples that are actually faulty but diagnosed as normal.

[0085] Compare the current evaluation result with the previous optimal result. If the current result is better than the optimal result, update the optimal result and the corresponding hyperparameter combination. When a certain number of iterations is reached and the evaluation index is no longer significantly improved, the training is completed and the final fault diagnosis model is obtained; if the stopping condition is not met, continue to use the improved White Whale optimization algorithm to optimize the hyperparameters until it is met.

[0086] Specifically, in step S7, first ensure that each sensor is working properly and is in an online state, use the sensor to collect data in real time, and continuously obtain multi-source data such as fuel combustion pressure, ejection pressure, thruster attitude, thruster internal acceleration, and carrier water pressure according to the characteristics and sampling frequency of different sensors. Then, the multi-source sensor data is subjected to data preprocessing in step S2 and multi-source data fusion in step S3, and finally the fusion result of the multi-source data is obtained. The processed multi-source sensor data is input into the fault diagnosis model, the fault diagnosis model is run, the input data is predicted, and the fault diagnosis result of the vector thruster system is output according to the learned fault mode and characteristics.

[0087] like Figure 2As shown, the vector thruster system is equipped with various types of sensors to monitor and collect key data to ensure the normal operation of the vector thruster. Specifically, the vibration sensor 1 installed at the connection between the thruster housing and the transmission shaft is used to monitor the vibration of the transmission shaft during operation to determine whether the transmission shaft has wear and imbalance problems. Two acceleration sensors 2 are installed on the front side of the interior to collect the state and dynamic changes during the operation of the thruster to ensure the stability and safety of the operation. The current sensor 4 is installed on the power supply line of the motor to monitor the working current of the motor in real time to determine whether the motor has overload or short circuit faults. In addition, the inner wall of the thruster is equipped with two pairs of pressure sensors 3 to collect pressure data when the thruster is running, and can also monitor water pressure changes to promptly detect watertightness problems and ensure safety when running underwater. The Hall angle sensor 5 is installed near the transmission mechanism of the vector thruster, which can measure the deflection angle of the vector thruster in real time and accurately, and feed back the angle information to the control system, which can detect abnormal fluctuations or stuck faults, provide a basis for timely taking maintenance measures, and ensure the safe operation of the system.

[0088] like Figure 3 In the fault diagnosis system of the present invention shown, the sensor module includes: a vibration sensor, an acceleration sensor, a Hall angle sensor, a pressure sensor and a current sensor. The data acquisition unit in the data acquisition and transmission module is responsible for receiving signals from various sensors and converting them into digital signals for processing. The data acquisition unit needs to have high-speed sampling, high-precision conversion and anti-interference capabilities to ensure that the collected signals are accurate and reliable; the data transmission system in the data acquisition and transmission module transmits the collected data to the central processing unit of the fault diagnosis system, and can adopt wired transmission methods such as serial ports and Ethernet, which have the advantages of stable transmission and fast speed; the central processing unit can perform data analysis, fault diagnosis and database processing.

[0089] The central processing unit receives data from the data acquisition and transmission module and uses a variety of data analysis algorithms to process it in depth, including time domain analysis, frequency domain analysis, wavelet analysis, pattern recognition and other methods to extract fault feature information. By learning and modeling a large amount of historical data, an intelligent fault diagnosis model is established to realize automatic diagnosis of thruster faults. According to the analyzed and processed data, combined with the intelligent diagnosis model, it is accurately judged whether the thruster has a fault, and the type, location and severity of the fault are determined. The central processing unit can adopt rule-based diagnosis methods, model-based diagnosis methods or artificial intelligence-based diagnosis methods, such as neural networks, support vector machines, fuzzy logic, etc. Once the fault is determined, the central processing unit will issue a corresponding alarm signal and provide fault handling suggestions. A complete database system is established to store information such as the operation data, fault history records, and diagnosis results of the thruster. The database can be used for fault analysis and trend prediction, and provide decision support for the maintenance and management of the thruster. At the same time, the database can also realize data backup and recovery to ensure data security and reliability.

[0090] The human-computer interaction interface has the functions of display, alarm, parameter setting and reporting. The human-computer interaction interface provides an intuitive and clear human-computer interaction interface, and displays the fault diagnosis results to the user in a graphical and digital way. Including information such as fault type, location, severity, historical records, etc. At the same time, when a fault is detected, the system can issue an audible and visual alarm to remind the user to deal with it in time. Users can set the parameters of the sensor, the parameters of the diagnostic algorithm, the alarm threshold, etc. through the human-computer interaction interface. In addition, users can also debug and maintain the fault diagnosis system to ensure the normal operation of the system. The system can generate various reports according to user needs, such as fault statistics reports, operating status reports, etc. Users can choose to print reports for archiving and analysis.

[0091] The maintenance and management module has the functions of fault prediction, maintenance guidance and remote monitoring. Based on historical data and fault diagnosis results, data analysis and prediction algorithms are used to predict and analyze the trend of thruster failures. By discovering potential fault hazards in advance, a preventive maintenance plan is formulated to reduce the probability of failure and improve the reliability and availability of the thruster. When a thruster fails, the system can provide detailed maintenance guidance, including troubleshooting steps, required tools and spare parts, maintenance precautions and other information. At the same time, the system can record the maintenance process and results to provide a reference for subsequent fault analysis and maintenance. Supporting remote monitoring and management functions, users can remotely access the fault diagnosis system through the network to understand the operating status and fault conditions of the thruster in real time. The remote monitoring and management function can improve maintenance efficiency and reduce maintenance costs, especially for distributed thruster systems or thrusters for remote operations.

[0092] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A fault diagnosis method for an underwater vector thruster, characterized in that: The specific steps include: S1, generates a historical fault data set by obtaining historical fault data from sensors carried on the underwater vector thruster; S2, preprocessing of historical fault data sets, including: data cleaning, noise reduction and standardization; S3, extract the historical fault data features of the pre-processed multi-source sensors, use the support vector machine algorithm to obtain the optimal classification hyperplane, and build a fusion model for multi-source heterogeneous data; input the trained data into the data fusion model, output the fusion results, and obtain the data set after data processing; S4, the data set obtained after data processing is divided into a training set and a test set according to a certain ratio. In the training stage, the data from the training set is input into three hybrid deep learning models combined with self-attention modules for training. The three hybrid deep learning models combined with self-attention modules are: convolutional neural network-long short-term memory network, convolutional neural network-gated recurrent unit and convolutional neural network-gated recurrent unit-long short-term memory network. Then the prediction results of the three hybrid deep learning models are used as the input of ensemble learning, and the optimal weights are obtained through the improved random grid search method, and the weighted ensemble strategy is used to generate the final prediction results; S5, the White Whale optimization algorithm with dynamic pinhole imaging strategy is used to optimize the hyperparameters in the neural network model, including: the number of convolution kernels M, the dimension N, the number of neurons in the long short-term memory network layer and the gated recurrent unit layer of the convolutional neural network layer; S6, input the test set into the fault diagnosis model optimized by the improved Beluga optimization algorithm to evaluate the model fault diagnosis effect. If the diagnosis effect is not optimal, continue to use the improved Beluga optimization algorithm to optimize the hyperparameters. If the diagnosis effect is optimal, the training is completed and the final fault diagnosis model is obtained. S7, using the sensors carried on the underwater vector thruster to collect current, acceleration, vibration frequency and pressure online, perform data preprocessing, and then input them into the final fault diagnosis model obtained in step S6 to perform fault diagnosis.

2. A fault diagnosis method for an underwater vector thruster according to claim 1, characterized in that: Step S2 specifically includes the following steps: S2.1, using 3 Data cleaning is performed according to the principle, and the mean and standard deviation of the historical fault data obtained by each sensor are calculated, and the mean value is ±3 Data with a range are marked as outliers, which are deleted or replaced with the median; S2.2, in the noise reduction processing stage, the historical fault data obtained by the sensor is subjected to empirical mode decomposition (EMD), and the data is decomposed into several intrinsic mode functions (IMFs) and a residual term. According to the frequency characteristics and energy distribution of the IMFs, a part of the IMFs is selected for processing, and the remaining IMFs and residual terms are reconstructed to obtain the noise-reduced signal; S2.3, in the data standardization stage, for each historical fault data, find the minimum value and maximum value , each historical fault data point of each sensor is standardized using the Min-Max normalization formula, and the historical fault data is mapped to the [0,1] interval.

3. The fault diagnosis method of an underwater vector thruster according to claim 1, characterized in that: Step S3 specifically includes the following steps: S3.1, based on the results of data preprocessing in step S2, update the multivariate data set, and the updated data set is , the data status is , extracting multi-source heterogeneous data In the The characteristic vector of the attribute is calculated as follows: （1）； Among them, the feature vector of multi-source heterogeneous data is Indicates that the data attribute is recorded during the period express; S3.2, set the value obtained in step S3.1 is the input vector , the kernel function is , the feature vector set is , the output vector is , in order to obtain the optimal hyperplane of the data, the process is as follows: （2）； In the formula, the optimal hyperplane is obtained using The weighting coefficient is expressed as Indicates; the bias term is expressed as The number of eigenvectors is represented by express; is a symbolic function; The symbolic function is defined as: （3）； S3.3, based on the obtained optimal hyperplane, the multi-source heterogeneous data is divided into h groups of data, and the data of different groups come from different sensors. The generalization parameter is set as , in order to establish a fusion model of multi-source heterogeneous data of support vector machine: （4）； Among them, the fusion model of multi-source heterogeneous data is constructed using Representation; Multi-source heterogeneous data , The optimal solutions are , Indicates; constraints are expressed by Indicates the number of sensors. express, , Represents multi-source heterogeneous data , The optimal solution mean, is the kernel function; S3.4, based on the determined optimal solution of the data, establish the decision function of the data and complete the model solution. The process is shown in the following formula: （5）； Among them, the decision function constructed is Represents; the mean value of multi-source heterogeneous data is expressed as ; Data optimal solution express.

4. The fault diagnosis method of an underwater vector thruster according to claim 1, characterized in that: Step S4 specifically includes the following steps: S4.1, build a convolutional neural network-long short-term memory network model: The formulas for convolution calculation and maximum pooling in convolutional neural networks are as follows: （6）； （7）； in, is the input feature vector; represents the convolution kernel; Indicates network bias; express Layer output, express Layer input; is the pooling width; is the ReLU activation function; express The indices of the different neurons in the layer; The update formula of the long short-term memory network is as follows: （8）； （9）； （10）； （11）； （12）； In the formula, , , They are the input gate, forget gate, and output gate at time step The output value of , , At each time step Input data, unit state, hidden state; , is the previous time step The hidden state and unit state of ; among them, , , , Input To the weight matrix of the input gate, forget gate, output gate, and candidate unit state; , , , They are the hidden states of the previous moment respectively To the weight matrix of the input gate, forget gate, output gate, and candidate unit state; , , and They are the bias items of input gate, forget gate, output gate, and candidate unit state respectively; is the sigmoid function.

5. A fault diagnosis method for an underwater vector thruster according to claim 4, characterized in that: Step S4 also includes the following steps: S4.2, build a convolutional neural network-gated recurrent unit model: The long short-term memory network is changed to a gated recurrent unit network. The updated gated recurrent unit formula is as follows: (13)； (14)； (15)； (16)； In the formula, , At time step The output value of the reset gate and update gate; , At time step Candidate hidden states, hidden states; , , They are reset gate, update gate, and weight matrix for calculating candidate hidden states; , , They are reset gate, update gate, and bias item for calculating candidate hidden states respectively; S4.3, the convolutional neural network-gated recurrent unit-long short-term memory network model is constructed by adding a gated recurrent unit layer and a long short-term memory network layer in sequence after the convolutional neural network layer.

6. A fault diagnosis method for an underwater vector thruster according to claim 5, characterized in that: Step S4 specifically includes the following steps: S4.4, features obtained after extraction by three models , then Perform a linear transformation, Mapped to query space , key space Sum value space ; Then use the scaled dot product and Softmax function to calculate each attention distribution, and weighted sum the attention distribution to get the corresponding output ; Finally, multiple output results are spliced ​​together using the splicing method. The formula is as follows: (17)； (18)； (19)； (20)； (21)； Where: , , are the linear transformation parameters of query space, key space, and value space respectively; For the query vector, For the key vectors, For the A vector of values, is a matrix with the number of dimensions for each key; is the element vector in the key space; is the element vector in the value space; is the feature concatenation function; for The transpose transformation of An attention scoring function for scaling dot products; is the Softmax function, which converts a real number vector into a probability distribution vector; Indicates The output result of each head; Represents the final multi-head attention mechanism output, by splicing multiple get; S4.5, the optimal weighted set weight is obtained according to the model search results, and the model with the largest prediction probability is selected from the weighted prediction results as the final fault diagnosis type output. The principle of weighted integration is shown in (22): （22）； In the formula, is the number of models, the value is 3; is the predicted probability of the model; is the weight and satisfies , .

7. The fault diagnosis method of an underwater vector thruster according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1, according to the population mechanism of the improved Beluga optimization algorithm, the Beluga is regarded as the search subject, and the movement effect in the search space is achieved by changing the position vector of the Beluga. The search agent position matrix for: （23）； In the formula, To search for the number of white whales, is the dimension of the design variables; Fitness value function of Beluga whale Expressed as: （24）； Balance Factor Changes in each iteration: （25）； In the formula, For the current iteration, is the maximum number of iterations, is a random number in the interval (0,1). When , it corresponds to the development stage; the balance coefficient ≤0.5 corresponds to the exploration stage. As the number of iterations increases, the probability of entering the development stage increases, while the probability of entering the exploration stage decreases.

8. A fault diagnosis method for an underwater vector thruster according to claim 7, characterized in that: Step S5 specifically includes the following steps: S5.2, in the exploration phase, the movement of the white whale determines the position update of the BWO algorithm in the detection phase. The position update of the white whale is as follows: （26）； For the next iteration The individual in The position on the dimension, is an even number, is an odd number, and is A positive integer randomly selected from , and not equal, and Respectively represent the current iteration , The location of an individual, , is a random number in (0,1); S5.3, in the development stage, beluga whales move and feed according to the location of nearby beluga whales. Beluga whales catch prey by sharing each other's location information using the Levy running strategy. The specific formula is: （27）； （28）； and Respectively represent the , The current location of each individual, For the The new position of each individual, It is the best location. and is a random number in (0,1), is a random jump, measuring the strength of Levy's operation; is a function of the Levy flight strategy and is calculated as follows: （29）； （30）； in, and is a random number that follows a normal distribution, is a parameter that determines the scale of the Levy distribution. is the gamma function in The value at is a constant set to 1.

5.

9. A fault diagnosis method for an underwater vector thruster according to claim 8, characterized in that: Step S5 specifically includes the following steps: S5.4, during the whale fall phase, beluga whales either migrate to other places or die; in order to keep the population size unchanged, the position of the beluga whale and the step size of the whale fall are used to establish the updated position, the formula is as follows: （31）； in is a random number in (0,1), is the step size of the whale's descent, established as: （32）； is the step factor, which is related to the probability of beluga whale decline and population size, specifically: ; and are the upper and lower bounds of the variable respectively; the probability of dropping the white whale Expressed as a linear function, the formula is as follows: （33）； S5.5, the dynamic pinhole imaging strategy is used to improve the White Whale optimization algorithm for global optimization, and the following mathematical model is obtained: （34）； and are the lengths of the virtual candle at the current optimal position and the opposite position, respectively; Set the ratio of the two candles as a variable , from which we can get: （35）； in, Indicates The white whale The first iteration Dimension location.

Images