A life prediction method and system for vector thruster

Abstract

The present invention discloses a life prediction method and system for a vector thruster, which belongs to the technical field of vector thruster life prediction, and the steps are: collecting historical sensor data of the vector thruster, including current voltage, vibration signal, pressure, acceleration, temperature and sound data; preprocessing the collected historical sensor data; constructing a data cube and dividing it into a training set and a test set; building a life prediction model based on a three-branch fusion convolutional neural network, the three-branch fusion convolutional neural network model combines a spectrum attention branch module, a bidirectional space attention branch module and a multi-attention fusion branch module, and can effectively extract and fuse features from multiple angles such as spectrum and space, greatly improving the recognition ability and generalization performance of the model. The present invention can provide a more accurate vector thruster life prediction, which is helpful to discover potential faults in advance, reduce maintenance costs, and extend the service life of equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of life prediction of vector thrusters, and in particular relates to a life prediction method and system for vector thrusters. Background Art

[0002] With the increasing number of ocean exploration and utilization activities, the reliability and economy of underwater equipment have become key issues. As one of the core components of modern underwater vehicles, the performance of vector thrusters directly affects the operating efficiency and navigation safety of the entire system. They play an important role in various underwater platforms such as modern marine technology and commercial activities. However, due to the particularity of the working environment, vector thrusters face a series of challenges such as corrosion and wear, which not only limits their service life but also increases maintenance costs. Therefore, it is particularly necessary to establish an effective vector thruster life prediction method to achieve long-term stable operation of the equipment and reduce operating costs.

[0003] Vector thrusters can achieve multi-axis control of underwater vehicles by changing the direction of thrust, allowing the device to accurately locate and navigate in three-dimensional space, allowing the vehicle to more easily perform complex operations such as hovering, sideways, vertical lifting, etc. Compared with traditional fixed-direction propulsion systems, vector thrusters can adjust direction more quickly, provide timely thrust changes, and enable the device to respond quickly to changes in the external environment and improve reaction speed. In addition, vector thrusters can adjust direction while maintaining the original speed, reducing the process of deceleration and re-acceleration in traditional propulsion methods, improving overall work efficiency, and changing the state of motion by adjusting the thrust direction rather than increasing power output, which can reduce unnecessary energy loss and improve energy efficiency. In complex underwater environments, such as rugged seabed terrain and places with uncertain water flow, vector thrusters can allow equipment to cope with various situations more easily and maintain a stable posture.

[0004] Traditionally, the maintenance of vector thrusters relies on regular inspection and replacement of parts. Although this method is intuitive and simple, it has significant disadvantages: on the one hand, excessive maintenance may lead to unnecessary replacement of parts that are still in good condition, increasing maintenance costs and causing waste of resources; on the other hand, insufficient maintenance may cause safety accidents. In view of this, it is becoming increasingly important to introduce condition prediction technology based on condition monitoring. By real-time monitoring of the working status of vector thrusters and combining historical data to predict their life, it is not only possible to detect potential faults in advance, but also to optimize maintenance plans, extend the service life of equipment, reduce the need to replace new equipment, and improve overall economic benefits. At the same time, for equipment that needs to run continuously for a long time, life prediction helps to maintain the continuity of tasks, avoid delays or cancellations of tasks due to equipment failures, and arrange tasks more reasonably to ensure that the equipment is in the best working condition during critical tasks. Therefore, the service life prediction of vector thrusters is of great significance in ensuring equipment safety, optimizing maintenance strategies, controlling operating costs, and improving task execution efficiency. Summary of the invention

[0005] In view of the above problems existing in the prior art, the present invention proposes a life prediction method and system for a vector thruster, which has a reasonable design, solves the shortcomings of the prior art and has good effects.

[0006] A life prediction method for a vector thruster comprises the following steps:

[0007] Step 1: Collect historical sensor data of the vector thruster, including current and voltage, vibration signals, pressure, acceleration, temperature, and sound data, and add a remaining life label to the historical sensor data;

[0008] Step 2: Preprocess the historical sensor data collected in step 1;

[0009] Step 3: Construct a data cube and divide it into training set and test set;

[0010] Step 4: Construct a life prediction model based on a three-branch fusion convolutional neural network. The three-branch fusion convolutional neural network includes a spectrum attention branch module, a bidirectional spatial attention branch module, and a multi-attention fusion branch module; the spectrum attention branch module is used to extract useful features at the spectrum level, the bidirectional spatial attention branch module is used to dynamically adjust the weight distribution at different positions, and the multi-attention fusion branch module is used to fuse spectrum and spatial features through a multi-attention mechanism;

[0011] Step 5: Use the training set and test set to train and test the life prediction model. In actual application, the multi-source sensor data acquired in real time is input into the model to obtain the life prediction results, thereby realizing the life prediction of the vector thruster.

[0012] Furthermore, in step 2, the timestamps of the historical sensor data are first aligned to ensure data consistency; then data cleaning is performed, i.e., data deduplication, noise elimination, missing data completion, and anomaly detection; finally, the data is normalized to ensure that all sensor data are on the same scale.

[0013] Furthermore, in step 3, first, a time window size is defined every hour, and then the number of sensors, the number of time windows and the number of sampling points are determined respectively, and the number of sampling points in each time window is determined by the sampling frequency of the sensor; finally, a three-dimensional array is constructed for each time window; wherein the first dimension represents the number of time windows, the second dimension represents the number of sensors, and the third dimension represents the number of sampling points in each time window; and the three-dimensional array is divided into a training set and a test set according to a certain ratio.

[0014] Furthermore, in step 4, the original data is first processed using a 3D convolution operation to extract low-order spectrum features, and the output feature map is defined as , The expression is:

[0015] ; (1)

[0016] in, Represents a three-dimensional input data block, where is the size of the input block, is the number of spectral features, i.e., the spectral dimension; and Represent weight and bias parameters respectively; Represents a three-dimensional convolution operation, which is to convolve the input data through the convolution kernel to generate a new feature map;

[0017] The feature map Input into the spectral attention branch module, perform maximum pooling operation and average pooling operation respectively, the expression is:

[0018] ; (2)

[0019] in, is the pixel in the input block The elements at represents the output of the average pooling operation, represents the output of the maximum pooling operation;

[0020] Then and They are input into the shared selective multilayer perceptron, which is constructed by considering local neighborhoods to refine the hidden layer, thereby modeling the long-term dependencies of all spectral features. The value is set to 9, and the output vector of the hidden layer ,Depend on constitute, , The output vector The mth element in ;

[0021] The expression is:

[0022] ; (3)

[0023] in, represents the ceiling function, which rounds a given number to the nearest integer; Represents the first Elements adjacent to A collection of spectra; Every The shared parameters of Represents the first Elements adjacent to The value of each spectral band;

[0024] right After deconvolution, Leaky-ReLU activation function and batch normalization, it is finally output from the output layer;

[0025] The outputs of the two shared selective multilayer perceptrons are merged together by element-by-element addition to obtain a feature map, which is nonlinearly activated by a sigmoid function to generate an attention weight matrix , the expression is:

[0026] ; (4)

[0027] Finally, use the swish function to Perform feature enhancement and linear activation.

[0028] Furthermore, the bidirectional spatial attention branch module includes a horizontal channel and a vertical channel. First, a kernel of two spatial dimensions is used on each channel. and Perform two independent global average pooling operations to encode the feature map ,nuclear Indicates that the size in the horizontal direction is , a pooling kernel of size 1 in the vertical direction, kernel Indicates that the size in the horizontal direction is 1 and the size in the vertical direction is The pooling kernel is expressed as:

[0029] ; (5)

[0030] in, Indicates at height and Input feature pixels at channels, Indicates the width and Input feature pixels at channels, Indicates the horizontal direction The global average pooling result is Indicates the vertical direction Global average pooling results;

[0031] Will and Spliced ​​into and ,right Reshape ,Will and Perform feature fusion, which includes channel compression and feature nonlinear reconstruction, to generate feature maps , through the cascade operation, the feature map U is divided into two tensor matrices along the horizontal and vertical directions of the space and , and then input into the convolution layer with 1×1 convolution kernel, BN layer and Leaky-ReLU activation function layer, and processed by Sigmoid nonlinear activation function, and then feature fusion is performed by element-by-element multiplication to finally obtain the spatial attention weight matrix The same shape as the input data block X, is the size of the input block, expressed as:

[0032] ; (6)

[0033] Finally, use the swish function to Perform feature enhancement and linear activation.

[0034] Furthermore, in the multi-attention fusion branch module, the outputs of the spectral attention branch module and the bidirectional spatial attention branch module are first fused by element-by-element multiplication to obtain a feature weight matrix , the expression is:

[0035] ; (8)

[0036] The feature weight matrix With feature map After element-by-element multiplication, they are input into the BN layer and the Leaky-ReLU layer in turn. The output of the Leaky-ReLU layer is consistent with the feature weight matrix. After element-by-element multiplication, it is input into the first feature extraction module. The output of the first feature extraction module is combined with the feature weight matrix After element-by-element multiplication, it is input into the second feature extraction module. The output of the second feature extraction module is consistent with the feature weight matrix After element-by-element multiplication, it is input into the third feature extraction module. The output of the third feature extraction module is consistent with the feature map. After element-by-element multiplication, the data is input into the convolutional layer and the classifier in turn, and the remaining service life of the thruster is output;

[0037] The first feature extraction module includes a depth-separable convolution layer, an EM-IBi-LSTM network, a self-attention module, a BN layer and a Leaky-ReLU layer in sequence; the second feature extraction module includes a depth-expanded convolution layer, an EM-IBi-LSTM network, a self-attention module, a BN layer and a Leaky-ReLU layer in sequence; the third feature extraction module includes a convolution layer, an EM-IBi-LSTM network, a self-attention module, a BN layer and a Leaky-ReLU layer in sequence.

[0038] Furthermore, the EM-IBi-LSTM network is specifically adapted to the clustering task of sequence data by combining the expectation maximization EM algorithm and the improved bidirectional long short-term memory network IBi-LSTM, with the purpose of optimizing the distribution of IBi-LSTM model parameters;

[0039] IBi-LSTM adapts to the complexity of different input data by dynamically adjusting the number of hidden units in the LSTM layer. When the input data presents a more complex pattern, the number of hidden units is increased; when the input data is relatively simple, the number of hidden units is reduced.

[0040] First, the model parameters are initialized, including input size, hidden layer size, number of layers, output size, and standard deviation δ; then the ImprovedBi-LSTM class is defined, in which the bidirectional LSTM layer is initialized, and additional layers and activation functions are defined; its forward propagation function is written, which implements the forward propagation of the LSTM layer, combines the outputs of forward and backward propagation, applies additional layers and activation functions, and finally returns the final output; an optimized ImprovedBi-LSTM instance is created, the training data, labels, and hyperparameters are specified, and the model is trained using backpropagation and optimization algorithms; subsequently, clustering is performed through the E / M algorithm, different δ values ​​are tried, the number of clusters in the LSTM layer is configured, and the model is retrained using the specified cluster configuration, and finally the model performance is evaluated. This process is repeated under various δ values ​​to select the model with the best performance, so that the EM-IBi-LSTM algorithm can achieve better results in the clustering task of sequence data.

[0041] Furthermore, in the self-attention module, the query matrix is ​​calculated , key matrix Sum Matrix , and then through the dot product operation Generate an attention score matrix and normalize the score through the softmax function to get the attention score Finally, the hidden state at the current moment is concatenated with the attention-weighted context vector to form a new feature vector ;

[0042] The expression is as follows:

[0043] ; (11)

[0044] ; (12)

[0045] ; (13)

[0046] in, is the hidden state at the current moment, is the hidden state at the previous moment, , , is the weight matrix, is the factor used to scale the dot product result in the dot product attention mechanism; Indicates The attention scores for time steps, Indicates that the corresponding The value of the time step, is the total number of time steps.

[0047] A life prediction system for a vector thruster, using the life prediction method for a vector thruster as described above, is composed of a vector thruster, a sensor, a data receiving unit, a data processing unit, a data storage unit and a display;

[0048] The sensors include hydrophones, temperature sensors, pressure sensors, vibration signal sensors and current and voltage sensors;

[0049] The data receiving unit is used to receive real-time data from the sensor and perform preprocessing;

[0050] The data processing unit is used to perform feature extraction, model training and prediction analysis on the pre-processed data, and predict the remaining service life of the vector thruster;

[0051] The data storage unit is used to store preprocessed data and prediction results, so as to facilitate users to review and analyze historical data;

[0052] The display is used to visualize the prediction results and implementation monitoring data to the staff.

[0053] Beneficial technical effects brought by the present invention:

[0054] 1. Efficient data preprocessing: Through timestamp alignment, data cleaning and normalization, the consistency and quality of data are ensured, and the accuracy of subsequent modeling is improved.

[0055] 2. Flexible data partitioning: By customizing the time window size, number of sensors, time, and number of sampling points, the construction of the data cube can be flexibly adjusted according to actual conditions to meet the needs of different scenarios.

[0056] 3. Powerful feature extraction capability: The three-branch fusion convolutional neural network model combines the spectral attention branch module, the spatial attention branch module and the multi-attention fusion branch module, which can effectively extract and fuse features from multiple angles such as spectrum and space, greatly improving the recognition ability and generalization performance of the model.

[0057] 4. Introduction of attention mechanism: By introducing the attention mechanism, the model can automatically focus on the most important features when processing large amounts of data, reducing the impact of redundant information and improving the efficiency and accuracy of the model.

[0058] 5. High-precision life prediction: Through the comprehensive application of the above-mentioned technologies and methods, the present invention can provide a more accurate prediction of the life of the vector thruster, which helps to detect potential faults in advance, reduce maintenance costs, and extend the service life of the equipment.

[0059] Therefore, the present invention effectively solves the problems existing in traditional life prediction methods through innovative design and technical means, and provides a more accurate and efficient vector thruster life prediction solution, which has important practical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of a life prediction system for a vector thruster.

[0061] Figure 2 The figure is a flow chart of a life prediction method for a vector thruster. DETAILED DESCRIPTION

[0062] The specific implementation of the present invention is further described below in conjunction with specific embodiments:

[0063] A life prediction method for vector thrusters, such as Figure 2 As shown, the following steps are included:

[0064] Step 1: Collect historical sensor data of the vector thruster, including current and voltage, vibration signals, pressure, acceleration, temperature, and sound data, and add a remaining life label to the historical sensor data;

[0065] Step 2: Preprocess the historical sensor data collected in step 1;

[0066] First, the timestamps of historical sensor data are aligned to ensure data consistency; then data cleaning is performed, namely data deduplication, noise elimination, missing completion and anomaly detection; finally, the data is normalized to ensure that all sensor data are on the same scale, thereby improving the training efficiency and performance of the model.

[0067] Step 3: Construct a data cube and divide it into training set and test set;

[0068] In step 3, first define a time window size every hour, then determine the number of sensors, the number of time windows, and the number of sampling points respectively. The number of sampling points in each time window is determined by the sampling frequency of the sensor. Finally, construct a three-dimensional array for each time window. The first dimension represents the number of time windows, the second dimension represents the number of sensors, and the third dimension represents the number of sampling points in each time window. Divide the three-dimensional array into training sets and test sets according to a certain ratio. Usually, 80% of the data is used as the training set and 20% of the data is used as the test set.

[0069] Step 4: Construct a life prediction model based on a three-branch fusion convolutional neural network model. The model embeds an attention mechanism to more effectively capture and process key features in the data. The three-branch fusion convolutional neural network includes a spectral attention branch module, a bidirectional spatial attention branch module, and a multi-attention fusion branch module.

[0070] The spectral attention branch module is used to extract useful features at the spectral level and focus on processing features on the spectral dimension. By introducing the attention mechanism to reduce redundant information in spectral features and enhance useful signal components, it can more accurately extract useful features at the spectral level and improve the recognition ability of the model.

[0071] The bidirectional spatial attention branch module focuses on solving the feature correlation problem in the spatial dimension. Through the spatial attention mechanism, this branch can dynamically adjust the weight distribution at different positions to better capture the spatial layout characteristics of the image or signal.

[0072] The multi-attention fusion branch module is responsible for comprehensively processing the information extracted by the above two branches, fusing spectral and spatial features through the multi-attention mechanism to ensure that the network can fully understand the input data from multiple perspectives. This design not only improves the network's ability to learn complex patterns, but also enhances its robustness and generalization performance in different application scenarios.

[0073] In step 4, the original data is first processed using a 3D convolution operation to extract low-order spectral features, and the output feature map is defined as , The expression is:

[0074] ; (1)

[0075] in, Represents a three-dimensional input data block, where is the size of the input block, is the number of spectral features, i.e., the spectral dimension; and Represent weight and bias parameters, respectively, which will be optimized during training to improve model performance; Represents a three-dimensional convolution operation, which is to convolve the input data through the convolution kernel to generate a new feature map;

[0076] Vector thrusters usually have a large number of spectral features, but not all spectral features are useful for life prediction. Therefore, it is necessary to highlight the significant spectral features for feature extraction. Input into the spectral attention branch module, perform maximum pooling operation and average pooling operation respectively, the expression is:

[0077] ; (2)

[0078] in, is the pixel in the input block The elements at represents the output of the average pooling operation, Represents the output of the maximum pooling operation; these two pooling operations can help reduce the spatial size of the feature map while retaining important feature information.

[0079] Then and The shared selective multilayer perceptron is used to process spectrum data from different sensors, and each data stream represents a spectrum feature that may affect the life of the device. The traditional multilayer perceptron consists of an input layer, a simple hidden layer, and an output layer. The hidden layer usually reduces the number of parameters through compression operations, which may cause the loss of important features. The shared selective multilayer perceptron of the present invention considers local neighborhoods to refine the hidden layer, thereby modeling the long-term dependencies of all spectral features. The value is set to 9, and the output vector of the hidden layer ,Depend on constitute, , The output vector The mth element in ;

[0080] The expression is:

[0081] ; (3)

[0082] in, represents the ceiling function, which rounds a given number to the nearest integer; Represents the first Elements adjacent to A collection of spectra; Every The shared parameters of Represents the first Elements adjacent to The value of each spectral band;

[0083] For the feature vector After deconvolution, Leaky-ReLU activation function and batch normalization, it is finally output from the output layer;

[0084] The outputs of the two shared selective multilayer perceptrons are merged together by element-by-element addition to obtain a feature map, which is nonlinearly activated by a sigmoid function to generate an attention weight matrix , the expression is:

[0085] ; (4)

[0086] This attention weight matrix can more effectively highlight important spectral features, thereby enhancing the model's sensitivity to key signals.

[0087] Finally, use the swish function to Perform feature enhancement and linear activation. In this way, the weights between different features can be amplified while keeping the relative size constant, thus avoiding feature dissipation.

[0088] The location or distribution of sensors in the network is called spatial information, which is helpful for life prediction of vector thrusters because the data of adjacent sensors are likely to belong to the same category or show similar characteristics, and the spatial features from multiple adjacent sensors can suppress noise interference and redundant information. In the present invention, a bidirectional spatial attention branch module is designed to obtain an abstract spatial representation for life prediction of vector thrusters. It does not use 2D pooling operations in spatial feature extraction, which may cause the loss of position information. Instead, it separates spatial attention into two parallel 1D feature encoding processes. The two independent attention feature maps are embedded independently and have direction-specific information. They each capture the long-range dependencies of the input feature map along one of the spatial directions, while retaining the position information of the other direction in the generated attention map.

[0089] The bidirectional spatial attention branch module includes horizontal and vertical channels. First, two spatial dimensions of kernels are used on each channel. and Perform two independent global average pooling operations to encode the feature map ,nuclear Indicates that the size in the horizontal direction is , a pooling kernel of size 1 in the vertical direction, kernel Indicates that the size in the horizontal direction is 1 and the size in the vertical direction is The pooling kernel of . Global average pooling is a commonly used pooling technique that reduces the size of the feature map by calculating the average value of all elements in a region, thereby reducing the number of parameters and maintaining the statistical information of important features. H and V represent the pooling kernel sizes in the horizontal and vertical directions, respectively. These sizes are equivalent to the size of the original data block S×S. In order to clearly distinguish the operations in the horizontal and vertical directions during the description process, different symbols are used to represent the pooling kernel sizes in these two directions. The expression is:

[0090] ; (5)

[0091] in, Indicates at height and Input feature pixels at channels, Indicates the width and Input feature pixels at channels, Indicates the horizontal direction The global average pooling result is Indicates the vertical direction The global average pooling operation in two directions generates paired direction-aware feature maps. These feature maps not only capture the direction information in a wide range, but also preserve the location information.

[0092] Will and Spliced ​​into and ,Will And the reshaped feature map Perform feature fusion, which includes channel compression and feature nonlinear reconstruction, to generate feature maps , U has a height and width of (h+v), a channel number of 1, and a number of B spectral features; the feature map U is split into two tensor matrices along the horizontal and vertical directions of the space through a cascade operation and , and then input into the convolution layer with 1×1 convolution kernel, BN layer and Leaky-ReLU activation function layer, and processed by Sigmoid nonlinear activation function, and then feature fusion is performed by element-by-element multiplication to finally obtain the spatial attention weight matrix The same shape as the input data block X, is the size of the input block, expressed as:

[0093] ; (6)

[0094] Among them, Conv represents a 1×1 convolutional layer, which is used to further extract and transform the feature map.

[0095] Finally, use the swish function to Perform feature enhancement and linear activation.

[0096] Compared with the traditional ReLU activation function, the swish activation function is closer to the activation model of brain neurons and solves the Dead ReLU problem. The equation of the swish activation function is as follows:

[0097] (7).

[0098] In the multi-attention fusion branch module, the outputs of the spectral attention branch module and the bidirectional spatial attention branch module are first fused by element-by-element multiplication to obtain the feature weight matrix , the expression is:

[0099] ; (8)

[0100] The feature weight matrix With feature map After element-by-element multiplication, they are input into the BN layer and the Leaky-ReLU layer in turn. The output of the Leaky-ReLU layer is consistent with the feature weight matrix. After element-by-element multiplication, it is input into the first feature extraction module. The output of the first feature extraction module is combined with the feature weight matrix After element-by-element multiplication, it is input into the second feature extraction module. The output of the second feature extraction module is consistent with the feature weight matrix After element-by-element multiplication, it is input into the third feature extraction module. The output of the third feature extraction module is consistent with the feature map. After element-by-element multiplication, the data is input into the convolutional layer and the classifier in turn, and the remaining service life of the thruster is output;

[0101] The first feature extraction module includes a depthwise separable convolutional layer, an EM-IBi-LSTM network, a self-attention module, a BN layer, and a Leaky-ReLU layer in sequence; the second feature extraction module includes a depthwise dilated convolutional layer, an EM-IBi-LSTM network, a self-attention module, a BN layer, and a Leaky-ReLU layer in sequence; the third feature extraction module includes a convolutional layer, an EM-IBi-LSTM network, a self-attention module, a BN layer, and a Leaky-ReLU layer in sequence.

[0102] IBi-LSTM (Improved Bidirectional LSTM) adapts to the complexity of different input data by dynamically adjusting the number of hidden units. This improvement enables the model to better handle input data with different feature patterns, thereby improving the flexibility and adaptability of the model. Specifically, the improvements of IBi-LSTM include the following aspects:

[0103] (1) Clustering parameters:

[0104] Definition: The clustering parameter determines the number of hidden units in each LSTM layer. The main responsibility of these hidden units is to capture different features and patterns from the input data and transform this information into a higher-level representation.

[0105] Dynamic Adjustment: Different from the traditional LSTM that uses a fixed number of hidden units, IBi-LSTM dynamically adjusts the number of hidden units according to the characteristics of the input data and the complexity of the pattern by introducing an adjustable clustering parameter.

[0106] (2) Adapting to different input characteristics:

[0107] Complex patterns: When the input data presents a more complex pattern, the network's representation ability can be enhanced by increasing the clustering value. More hidden units can capture richer features and patterns.

[0108] Simple mode: When the input data is relatively simple, the complexity of the network can be reduced by reducing the clustering value, thereby improving the generalization ability of the model. Fewer hidden units can reduce the risk of overfitting.

[0109] (3) Improve model flexibility:

[0110] Flexibility: By dynamically adjusting the number of hidden units, IBi-LSTM is able to better adapt to the patterns of different input features, thereby improving the flexibility and adaptability of the model.

[0111] Performance optimization: This dynamic adjustment mechanism enables the model to be more efficient when processing data of different complexities, avoiding the limitations of fixed structures.

[0112] IBi-LSTM inherits the gating mechanism of LSTM, including input gate, forget gate and output gate. These gating mechanisms control the flow of information, allowing the model to selectively remember or ignore different parts of the input sequence. The specific formula is as follows:

[0113] Input Gate: ; Forget gate: ; Output gate: ;Candidate memory cell status: ; Memory unit status update: ;Hide status update: ;

[0114] in, is the Sigmoid function, tanh is the hyperbolic tangent function, W and b are the weight and bias respectively.

[0115] The expectation maximization (EM) algorithm is introduced into the training process. The EM algorithm dynamically updates the model parameters by dynamically adjusting the number of hidden units in the bidirectional LSTM to adapt to the complexity of different input data features. The flexibility and robustness of the model are further enhanced by adjusting the hyperparameters of the EM algorithm (such as the standard deviation δ) and using cross-validation and appropriate evaluation indicators.

[0116] EM clustering is an iterative algorithm used to classify data points into different clusters. The algorithm is based on the expectation maximization principle and maximizes the likelihood function of the data by iteratively updating the parameters of the cluster. The standard deviation δ is a hyperparameter of EM clustering, which is used to control the shape and compactness of the cluster, thereby affecting the quality of the clustering results. The E step (expectation step) is used to calculate the probability that each data point belongs to each cluster, expressed as:

[0117] ; (9)

[0118] in is a given data point Belong to The probability of a cluster, In the given The probability density under the clusters is It is In addition, the M-step (maximization step) uses the probabilities calculated in the E-step to update the parameters of the clusters.

[0119] ; (10)

[0120] in is the number of data points, is the number of clusters, are model parameters.

[0121] The EM-IBi-LSTM algorithm aims to optimize the parameter distribution of the BiLSTM model by combining the expectation maximization (EM) algorithm and the bidirectional long short-term memory network (Bi-LSTM) to improve model performance. The algorithm is mainly used for clustering tasks of sequence data, in which each data point is assigned to a different cluster, each cluster has its own specific parameters (mean μ and covariance matrix δ), and a standard deviation δ for adjusting the probability density of the Gaussian distribution. The initialization phase of the algorithm first randomly initializes the parameters of each cluster, including the mean and covariance matrix, and sets the initial value of the standard deviation δ. Next, the E step (expectation step) is used to calculate each data point The probability of belonging to each cluster is calculated, and the probability density of the Gaussian distribution is calculated based on the current cluster parameters and standard deviation δ. In the M step (maximization step), the results of the E step are used to update the probability of each data point belonging to each cluster, and to update the cluster parameters, including the mean and covariance matrix. Specifically, the mean The update is done by weighted averaging, and the covariance matrix Σ j The update of takes into account the deviation of the data points from the mean. In addition, the standard deviation δ is also updated to maximize the likelihood function. The whole process is repeated until convergence is reached or the maximum number of iterations is reached. Through this process, the EM-IBi-LSTM algorithm aims to effectively adjust the distribution of model parameters by jointly using the EM algorithm and the Bi-LSTM model to better fit the complex structure of the sequence data, thereby improving the clustering performance.

[0122] Specifically, the EM-IBi-LSTM algorithm is adapted to the clustering task of sequence data by combining the expectation maximization (EM) algorithm and the bidirectional long short-term memory network (Bi-LSTM), with the aim of optimizing the distribution of the Bi-LSTM model parameters. First, the model parameters are initialized, including the input size, hidden layer size, number of layers, output size, and standard deviation δ; then the ImprovedBi-LSTM class is defined, in which the bidirectional LSTM layer is initialized, and the additional layer and activation function are defined; its forward propagation function is written, which implements the forward propagation of the LSTM layer, combines the outputs of the forward and backward propagations, applies the additional layer and activation function, and finally returns the final output; an optimized ImprovedBi-LSTM instance is created, the training data, labels, and hyperparameters are specified, and the model is trained using backpropagation and optimization algorithms. Subsequently, clustering is performed by the E / M algorithm, different δ values ​​are tried, the number of clusters in the LSTM layer is configured, and the model is retrained with the specified cluster configuration, and finally the model performance is evaluated. This process is repeated under various δ values ​​to select the model with the best performance, so that the EM-IBi-LSTM algorithm can achieve better results in the clustering task of sequence data.

[0123] In the EM-IBi-LSTM model, an attention mechanism is used to dynamically adjust the importance of each time step to better capture the key patterns and features in time series data. The mechanism calculates the similarity between the query and the keyword, and then uses the similarity weighted value to weight the value matrix to finally obtain the context vector. This self-attention module obtains the attention-weighted context vector by calculating the correlation between the query, key, and value, and weights the value according to the correlation to obtain the attention-weighted context vector.

[0124] In the self-attention module, the query matrix is ​​calculated , key matrix Sum Matrix , the expression is:

[0125] ; (11)

[0126] in, is the hidden state at the current moment, is the hidden state at the previous moment, , , is the weight matrix;

[0127] To calculate the correlation between queries Q and K, we use the dot product operation to measure the similarity between the query and the key, and through the scaling operation To control the value range, is a factor used to scale the dot product result in the dot product attention mechanism to control the range of values. Its purpose is to avoid the problem of gradient disappearance or explosion caused by excessive dot product results. Finally, the score is normalized by the softmax function to obtain the attention score, which is expressed as:

[0128] ; (12)

[0129] The purpose of this is to allow the model to automatically learn the associations between different time steps, so that it can better selectively focus on important information at different moments. By combining the attention score with the value V, a weighted sum of the values ​​is achieved. Since the attention score indicates the importance of the value pair, the weighted sum can make the model pay more attention to important time steps, thereby generating more accurate context information. Subsequently, the hidden state at the current moment is concatenated with the attention-weighted context vector to form a new feature vector ; The expression is as follows:

[0130] ; (13)

[0131] in, Indicates The attention scores for time steps, Indicates that the corresponding The value of the time step, is the total number of time steps; this sum operation Σ represents the weighted sum of all time step values, and the weight is determined by the attention score. It can be seen that through the concatenation operation, the model considers both the hidden state of the current moment and the contextual information weighted by attention, thereby better capturing the key features and patterns in the time series.

[0132] Step 5: Use the training set to train the life prediction model. After the training is completed, freeze all trainable parameters and input the remaining 20% ​​of the test set into the trained model for performance testing. In actual application, the multi-source sensor data acquired in real time is input into the model to obtain the life prediction results, thereby realizing the life prediction of the vector thruster.

[0133] A life prediction system for vector thrusters, such as Figure 1 As shown, a life prediction method for a vector thruster is adopted, which is composed of a vector thruster, a sensor, a data receiving unit, a data processing unit, a data storage unit and a display;

[0134] The sensors include hydrophones, temperature sensors, pressure sensors, vibration signal sensors, and current and voltage sensors;

[0135] The vector thruster is equipped with a variety of sensors to monitor key data to ensure its long-term reliability and safety. The motor is the core component of the vector thruster and generates a lot of heat during operation. Therefore, a temperature sensor is installed on the motor to monitor the temperature changes during the operation of the motor, which can prevent damage caused by overheating and ensure that the motor operates within a safe temperature range. The propeller under the protective cover will generate noise during operation. The hydrophone is installed on the base near the propeller to monitor the noise level during operation. Abnormal sounds can be identified by analyzing sound data, potential fault points can be found in time, and the normal operation of the vector thruster can be ensured. In addition, when the propeller is running in the water, the surrounding water pressure changes will affect the performance of the thruster. Therefore, installing a pressure sensor on the base near it can monitor the water pressure changes, detect potential pressure anomalies in time, and ensure that the thruster operates within a safe pressure range. The flange base is an important supporting part of the vector thruster. Installing a vibration signal sensor can monitor its mechanical state, detect potential mechanical failures in time, ensure the smoothness and reliability of the thruster operation, and monitor its acceleration changes. By monitoring this data, the dynamic performance of the vector thruster can be evaluated to ensure its stability and safety under different working conditions. The current and voltage sensors are installed on a base equipped with a rotatable electric rod to monitor the voltage and current changes of the power supply in real time. They can detect the stability of the power supply system and ensure that the thruster obtains a stable power supply during operation, thereby avoiding failures caused by power supply problems.

[0136] Through the monitoring and data collection of these sensors, the operating status of the vector thruster can be evaluated in real time, and potential faults can be discovered and handled in a timely manner, thereby extending the service life of the thruster and improving its operational reliability and safety.

[0137] The data receiving unit is used to receive the real-time data from the sensor and perform pre-processing;

[0138] The sensor data is transmitted to the sensor data collection node, and then sent to the data receiving unit through wireless transmission technology. The data receiving unit is responsible for receiving real-time data from the sensor and performing preliminary data cleaning and formatting to ensure data consistency and integrity. The data verification function can automatically identify and eliminate abnormal data to improve data quality.

[0139] The data processing unit is used to extract features, train models and conduct predictive analysis on the preprocessed data, and predict the remaining service life of the vector thruster. The data processing unit can monitor the operating status of the thruster in real time, detect potential faults in time, and issue early warnings. Through adaptive learning, the model can be continuously optimized and adjusted according to new data to improve prediction performance.

[0140] The data storage unit is used to store pre-processed data and prediction results, which is convenient for users to review and analyze historical data; the unit has a large storage capacity and can store a large amount of historical data, supporting long-term data analysis. It uses encryption technology and backup mechanisms to ensure the security and integrity of data. In addition, the data storage unit provides flexible data query and analysis tools to facilitate users to review and analyze historical data.

[0141] The display is used to visualize the prediction results and implementation monitoring data to the staff, including the remaining service life, key parameter change trends, early warning information, etc. The display uses charts, curves, etc. to intuitively display the prediction results and real-time data, which is convenient for the staff to quickly understand and make decisions. The display updates the displayed data in real time to ensure the timeliness of the information. In addition, the display supports user interactive operations, such as data screening, parameter setting, etc., to improve the user experience.

[0142] Through this system, the staff can detect and predict the life of the vector thruster in real time, so as to take maintenance measures in advance and ensure the reliability and safety of the equipment. The advantages of this system include the following points: First, high-precision sensors and deep learning technology ensure the accuracy of data and the reliability of prediction; second, the real-time monitoring function can detect potential faults in time, give early warning, and improve the reliability and safety of the system; third, the large-capacity storage capacity and data security mechanism of the data storage unit ensure the long-term preservation and security of data; fourth, the intuitive display and real-time update function of the display improve the work efficiency and accuracy of the staff; finally, the adaptive learning ability of the system enables the model to be continuously optimized, improve the prediction performance, and ensure the long-term stable operation of the system.

[0143] 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 life prediction method for a vector thruster, characterized in that: The following steps are involved: Step 1: Collect historical sensor data of the vector thruster, including current and voltage, vibration signals, pressure, acceleration, temperature, and sound data, and add a remaining life label to the historical sensor data; Step 2: Preprocess the historical sensor data collected in step 1; Step 3: Construct a data cube and divide it into training set and test set; Step 4: Construct a life prediction model based on a three-branch fusion convolutional neural network, which includes a spectral attention branch module, a bidirectional spatial attention branch module, and a multi-attention fusion branch module; The spectral attention branch module is used to extract useful features at the spectral level, the bidirectional spatial attention branch module is used to dynamically adjust the weight distribution at different positions, and the multi-attention fusion branch module is used to fuse spectral and spatial features through a multi-attention mechanism; Step 5: Use the training set and test set to train and test the life prediction model. In actual application, the multi-source sensor data acquired in real time is input into the model to obtain the life prediction results, thereby realizing the life prediction of the vector thruster.

2. The life prediction method of a vector thruster according to claim 1, characterized in that: In step 2, the timestamps of the historical sensor data are first aligned to ensure data consistency; then data cleaning is performed, namely, data deduplication, noise elimination, missing data completion, and anomaly detection; finally, the data is normalized to ensure that all sensor data are on the same scale.

3. The life prediction method of a vector thruster according to claim 2 is characterized in that: In step 3, first, a time window size is defined every hour, and then the number of sensors, the number of time windows and the number of sampling points are determined respectively, and the number of sampling points in each time window is determined by the sampling frequency of the sensor; finally, a three-dimensional array is constructed for each time window; wherein the first dimension represents the number of time windows, the second dimension represents the number of sensors, and the third dimension represents the number of sampling points in each time window; and the three-dimensional array is divided into a training set and a test set according to a certain ratio.

4. The life prediction method of a vector thruster according to claim 3 is characterized in that: In step 4, the original data is first processed using a 3D convolution operation to extract low-order spectrum features, and the output feature map is defined as , The expression is: ；(1) in, Represents a three-dimensional input data block, where is the size of the input block, is the number of spectral features, i.e., the spectral dimension; and Represent weight and bias parameters respectively; Represents a three-dimensional convolution operation, which is to convolve the input data through the convolution kernel to generate a new feature map; The feature map Input into the spectral attention branch module, perform maximum pooling operation and average pooling operation respectively, the expression is: ；(2) in, is the pixel in the input block The elements at represents the output of the average pooling operation, represents the output of the maximum pooling operation; Then and They are input into the shared selective multilayer perceptron, which is constructed by considering local neighborhoods to refine the hidden layer, thereby modeling the long-term dependencies of all spectral features. The value is set to 9, and the output vector of the hidden layer ,Depend on constitute, , The output vector The mth element in ; The expression is: ；(3) in, represents the ceiling function, which rounds a given number to the nearest integer; Represents the first Elements adjacent to A collection of spectra; Every The shared parameters of Represents the first Elements adjacent to The value of each spectral band; right After deconvolution, Leaky-ReLU activation function and batch normalization, it is finally output from the output layer; The outputs of the two shared selective multilayer perceptrons are merged together by element-by-element addition to obtain a feature map, which is nonlinearly activated by a sigmoid function to generate an attention weight matrix , the expression is: ；(4) Finally, use the swish function to Perform feature enhancement and linear activation.

5. A life prediction method for a vector thruster according to claim 4, characterized in that: The bidirectional spatial attention branch module includes a horizontal channel and a vertical channel. First, a kernel of two spatial dimensions is used on each channel. and Perform two independent global average pooling operations to encode the feature map ,nuclear Indicates that the size in the horizontal direction is , a pooling kernel of size 1 in the vertical direction, kernel Indicates that the size in the horizontal direction is 1 and the size in the vertical direction is The pooling kernel is expressed as: ；(5) in, Indicates at height and Input feature pixels at channels, Indicates the width and Input feature pixels at channels, Indicates the horizontal direction The global average pooling result is Indicates the vertical direction Global average pooling results; Will and Spliced ​​into and ,right Reshape ,Will and Perform feature fusion, which includes channel compression and feature nonlinear reconstruction, to generate feature maps , through the cascade operation, the feature map U is divided into two tensor matrices along the horizontal and vertical directions of the space and , and then input into the convolution layer with 1×1 convolution kernel, BN layer and Leaky-ReLU activation function layer, and processed by Sigmoid nonlinear activation function, and then feature fusion is performed by element-by-element multiplication to finally obtain the spatial attention weight matrix The same shape as the input data block X, is the size of the input block, expressed as: ；(6) Finally, use the swish function to Perform feature enhancement and linear activation.

6. A life prediction method for a vector thruster according to claim 5, characterized in that: In the multi-attention fusion branch module, the outputs of the spectral attention branch module and the bidirectional spatial attention branch module are first fused by element-by-element multiplication to obtain a feature weight matrix , the expression is: ； (8) The feature weight matrix With feature map After element-by-element multiplication, they are input into the BN layer and the Leaky-ReLU layer in turn. The output of the Leaky-ReLU layer is consistent with the feature weight matrix. After element-by-element multiplication, it is input into the first feature extraction module. The output of the first feature extraction module is combined with the feature weight matrix After element-by-element multiplication, it is input into the second feature extraction module. The output of the second feature extraction module is consistent with the feature weight matrix After element-by-element multiplication, it is input into the third feature extraction module. The output of the third feature extraction module is consistent with the feature map. After element-by-element multiplication, the data is input into the convolutional layer and the classifier in turn, and the remaining service life of the thruster is output; The first feature extraction module includes a depth-separable convolution layer, an EM-IBi-LSTM network, a self-attention module, a BN layer and a Leaky-ReLU layer in sequence; the second feature extraction module includes a depth-expanded convolution layer, an EM-IBi-LSTM network, a self-attention module, a BN layer and a Leaky-ReLU layer in sequence; the third feature extraction module includes a convolution layer, an EM-IBi-LSTM network, a self-attention module, a BN layer and a Leaky-ReLU layer in sequence.

7. A life prediction method for a vector thruster according to claim 6, characterized in that: The EM-IBi-LSTM network is specifically adapted to the clustering task of sequence data by combining the expectation maximization EM algorithm and the improved bidirectional long short-term memory network IBi-LSTM, with the purpose of optimizing the distribution of IBi-LSTM model parameters; IBi-LSTM adapts to the complexity of different input data by dynamically adjusting the number of hidden units in the LSTM layer. When the input data presents a more complex pattern, the number of hidden units is increased; when the input data is relatively simple, the number of hidden units is reduced. First, the model parameters are initialized, including input size, hidden layer size, number of layers, output size and standard deviation δ; then the ImprovedBi-LSTM class is defined, in which the bidirectional LSTM layer is initialized, and the additional layer and activation function are defined; its forward propagation function is written, which implements the forward propagation of the LSTM layer, combines the outputs of forward and backward propagation, applies the additional layer and activation function, and finally returns the final output; an optimized ImprovedBi-LSTM instance is created, the training data, labels and hyperparameters are specified, and the model is trained using backpropagation and optimization algorithms; subsequently, clustering is performed through the E / M algorithm, different δ values ​​are tried, the number of clusters in the LSTM layer is configured, and the model is retrained using the specified cluster configuration, and finally the model performance is evaluated; this process is repeated under various δ values ​​to select the model with the best performance, so that the EM-IBi-LSTM algorithm can achieve better results in the clustering task of sequence data.

8. The life prediction method of a vector thruster according to claim 7, characterized in that: In the self-attention module, the query matrix is ​​calculated , key matrix Sum Matrix , and then through the dot product operation Generate an attention score matrix and normalize the score through the softmax function to get the attention score Finally, the hidden state at the current moment is concatenated with the attention-weighted context vector to form a new feature vector ; The expression is as follows: ；(11) ；(12) ；(13) in, is the hidden state at the current moment, is the hidden state at the previous moment, , , is the weight matrix, is the factor used to scale the dot product result in the dot product attention mechanism; Indicates The attention scores for time steps, Indicates that the corresponding The value of the time step, is the total number of time steps.

9. A life prediction system for a vector thruster, using a life prediction method for a vector thruster as claimed in any one of claims 1 to 8, characterized in that: It consists of a vector thruster, a sensor, a data receiving unit, a data processing unit, a data storage unit and a display; The sensors include hydrophones, temperature sensors, pressure sensors, vibration signal sensors and current and voltage sensors; The data receiving unit is used to receive real-time data from the sensor and perform preprocessing; The data processing unit is used to perform feature extraction, model training and prediction analysis on the pre-processed data, and predict the remaining service life of the vector thruster; The data storage unit is used to store preprocessed data and prediction results, so as to facilitate users to review and analyze historical data; The display is used to visualize the prediction results and implementation monitoring data to the staff.

Images