A method for constructing a health index system for underwater vector thrusters

Abstract

The present invention provides a method for constructing a health index system of an underwater vector thruster, which relates to the technical field of underwater vector thrusters, and specifically includes the following steps: obtaining a full-cycle vibration signal, performing data preprocessing on the full-cycle vibration signal, and forming a preprocessed data set; dividing the data set into a training set and a test set, and simultaneously inputting the test set and the training set into BiLSTM, CNN‑SE, DNN, and information coding neural network, and extracting feature information using different neural networks; fusing the extracted multiple feature information using MVSBA to obtain a weighted fusion vector; smoothing the weighted fusion vector using a smoothing algorithm; and using a weighted mean as an evaluation index of an evaluation function for measuring the health status of an underwater vector thruster. The technical solution of the present invention overcomes the problem of the lack of a stable and reliable health index system for underwater vector thrusters in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater vector thrusters, and in particular to a method for constructing a health indicator system for underwater vector thrusters. Background Art

[0002] Seabed resource development and underwater engineering construction rely on reliable underwater thrusters, and the construction of a health indicator system has become one of the key links of vector thrusters. The construction of a health indicator system is mainly used to monitor, evaluate and predict the working status of vector thrusters. These indicators can help operators discover and deal with potential problems in a timely manner to ensure the reliability, safety and efficient operation of vector thrusters.

[0003] The construction of the health indicator system is not limited to the mission execution period, but also covers the design stage and the maintenance stage. In the design stage, the performance and reliability of the vector thruster during use can be predicted by establishing a health indicator system, thereby optimizing the design scheme and reducing the risk of failure. In the maintenance stage, through the analysis of the health indicator system, a reasonable maintenance strategy can be specified, the maintenance time and frequency can be optimized, the maintenance cost can be reduced, and the reliability of the equipment can be improved.

[0004] Therefore, a reliable and stable method for constructing a health index system for underwater vector thrusters is needed. Summary of the invention

[0005] The main purpose of the present invention is to provide a method for constructing a health indicator system of an underwater vector thruster, so as to solve the problem of lack of a stable and reliable health indicator system of an underwater vector thruster in the prior art.

[0006] To achieve the above object, the present invention provides a method for constructing a health index system of an underwater vector thruster, which specifically comprises the following steps:

[0007] S1, using the sensor carried on the underwater vector thruster to obtain the full-cycle vibration signal, the full-cycle vibration signal is preprocessed to form a preprocessed data set.

[0008] S2, divide the data set into training set and test set, and input the test set and training set into the bidirectional long short-term memory network BiLSTM, the convolutional neural network CNN-SE based on the attention mechanism, the deep neural network DNN, and the information encoding neural network at the same time, and use different neural networks to extract feature information.

[0009] S3, using the feature fusion algorithm MVSBA based on the marked area to fuse the various feature information extracted in step S2 to obtain a weighted fusion vector.

[0010] S4, using the smoothing algorithm SG, smoothing the weighted fusion vector.

[0011] S5, , and The indicators are integrated into a weighted mean , as the evaluation index of the evaluation function HI to measure the health status of the underwater vector thruster.

[0012] Furthermore, the preprocessing in step S1 specifically includes the following steps:

[0013] S1.1, use continuous wavelet transform to denoise, detect abnormal points and abnormal segments on the denoised data, and clean the abnormal segments.

[0014] S1.2, considering the missing original data caused by sensor failure factors and the missing original data caused by abnormal segment cleaning, the regression interpolation method is used to fill the missing data according to the relationship between the data points before and after the missing data.

[0015] S1.3, use KNN algorithm to remove outliers.

[0016] S1.4, normalize the data, the formula is as follows:

[0017] ;

[0018] in, represents the normalized data, For input data, To minimize, To take the maximum.

[0019] S1.5, use principal component analysis to reduce the dimensionality of the data.

[0020] Furthermore, the marked area-based feature fusion algorithm MVSBA includes: a first module Block-1 and a second module Block-2 which are connected in sequence.

[0021] Furthermore, in step S3, the following steps are performed in the first module Block-1:

[0022] S3.1, map feature vectors of different dimensions to the same dimensional space, and standardize the mapped new vectors.

[0023] S3.2, Initialize the first load matrix , M is the number of load vectors;

[0024] , To generate a random number function, take 4 load vectors as an example, recorded as the load matrix ,in, , , , is the first load vector, and the initial value of the load vector follows Gaussian distribution.

[0025] S3.3, perform an inner product operation on the standardized vector and the first load vector to obtain the first labeling matrix :

[0026] ;

[0027] in, is the feature vector dimension.

[0028] S3.4, the labeling matrix Sum by row to get the first sum matrix .

[0029] S3.5, calculate the first fusion vector obtained by MVSBA of the feature vectors under different viewing angles :

[0030] The four components of are: , then the first fusion vector for:

[0031] ;

[0032] in, is the timing information mapping vector, is the global mapping vector, is the local saliency information mapping vector, Map vectors for sensors.

[0033] Furthermore, step S3.1 specifically includes the following steps:

[0034] S3.1.1, let the initial vector be , is the average, is the standard deviation, and the initial eigenvector , mapped to ,…, , denoted as .

[0035] S3.1.2, taking 4 eigenvectors as an example, the standardized eigenvectors are ,but , , , are all unit vectors, let ,but is the first matrix after concatenating the eigenvectors.

[0036] Furthermore, step S3.4 is specifically as follows:

[0037] Will Sum by row .

[0038] Furthermore, in step S3, the following steps are performed in the second module Block-2:

[0039] S3.6, Calculate the second load matrix :

[0040] ;

[0041] in, is the transformation matrix.

[0042] S3.7, calculate the second matrix after concatenating the eigenvectors , and calculate the second labeling matrix :

[0043] ;

[0044] , .

[0045] in, , , and The second matrix The normalized feature vector is , , and is the second load vector;

[0046] .

[0047] Will Sum the rows to get the second sum matrix :

[0048] ;

[0049] The four components of are: , then the second fusion vector for:

[0050] .

[0051] Furthermore, step S4 specifically includes the following steps:

[0052] S4.1, construct a Order polynomial:

[0053] ;

[0054] In the formula, are the coefficients of the polynomial; is the order of the polynomial; is the weight of the polynomial; For about Polynomial function of order; .

[0055] S4.2, using least squares to fit the residuals :

[0056] ;

[0057] In the formula, is a set of data points, for a given data point.

[0058] To make the residual Minimum, let The derivative of each coefficient is 0, that is:

[0059] ;

[0060] From this we can get, .

[0061] Furthermore, step S5 specifically includes the following steps:

[0062] S5.1, the health indicator sequence is ,in, Indicates the first Time point The corresponding eigenvalues ​​are decomposed into two parts: the steady trend and the random residual using the sliding average method:

[0063] ;

[0064] In the formula, For the stable part, is the random margin.

[0065] S5.2, the calculation formulas for the three indicators are as follows:

[0066] ; ;

[0067] ;

[0068] in, represents the number of data points, is the trend value, is the monotonicity value, is the robustness value.

[0069] In the formula, Defined as:

[0070] ;

[0071] in, is a step function.

[0072] S5.3, Calculation , and The fusion weighted mean , the calculation formula is as follows:

[0073] ;

[0074] in , , Respectively , , The weight of the three.

[0075] The present invention has the following beneficial effects:

[0076] The method proposed in the present invention realizes efficient and accurate monitoring and evaluation of the health status of the thruster by comprehensively utilizing the full-cycle vibration signal carried on the vector thruster. The core of the method lies in the first module Block-1 and the second module Block-2 structure based on MVSBA. Through multiple links such as data preprocessing and model training, the health index curve of the health status and operating status of the vector thruster is finally constructed more fully, comprehensively and accurately, thereby effectively evaluating the health status of the vector thruster and predicting the RUL of the vector thruster. In the data preprocessing stage, the present invention adopts dirty data cleaning and filling, KNN algorithm to eliminate outliers, PCA dimensionality reduction and other methods to enhance data quality. This step not only improves the utilization rate of data, but also provides more accurate and comprehensive input for subsequent feature fusion; in the feature fusion stage, the present invention adopts MVSBA for effective fusion to extract key features that can reflect the health indicators of the vector thruster. By retaining useful information in multidimensional features and removing redundant information, the model can learn from multiple perspectives and construct a comprehensive feature representation. At the same time, the evaluation function of correlation and monotonicity health indicators (HI-SpeaMon) is introduced to screen out sensitive indicators. The higher the index value, the better the feature can represent the performance degradation curve of the vector thruster. Through the evaluation and screening process, it can ensure that the constructed health indicator system has high representativeness and independence, and can truly reflect the health status of the vector thruster. In addition, the channel attention mechanism is adopted to give different weights to channel features according to their importance, thereby further enhancing the effect of feature extraction. In general, this method of constructing a health indicator system provides a strong guarantee for the safe and stable operation of the vector thruster. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] 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:

[0078] Figure 1 A flow chart of a method for constructing a health indicator system for an underwater vector thruster of the present invention is shown. DETAILED DESCRIPTION

[0079] 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.

[0080] like Figure 1 A method for constructing a health indicator system for an underwater vector thruster is shown, and specifically comprises the following steps:

[0081] S1, using the sensors on the underwater vector thruster to obtain the full-cycle vibration signal, preprocessing the full-cycle vibration signal to form a preprocessed data set. The sensors on the vector thruster, such as accelerometer, gyroscope, magnetometer, GPS module, and barometer, respectively measure the full-cycle vibration signals such as acceleration, Euler angle (pitch angle, roll angle, yaw angle), magnetic field strength and direction, position and displacement, height, and pressure.

[0082] S2, divide the data set into training set and test set, usually initially set to 80% training set and 20% test set, and input the test set and training set into the bidirectional long short-term memory network BiLSTM, the convolutional neural network CNN-SE based on the attention mechanism, the deep neural network DNN, and the information encoding neural network at the same time, and use different neural networks to extract feature information.

[0083] BiLSTM is an improved model of LSTM. BiLSTM is composed of a forward and a reverse LSTM. This combination of forward propagation and reverse propagation takes into account the impact of past and future information on the current situation. Compared with the unidirectional LSTM, the bidirectional processing method of BiLSTM is more outstanding.

[0084] Vector thruster data is a typical time series data with time-series dependency. Recurrent neural networks (RNNs) can usually effectively capture the time series characteristics of data. In the RNN model, each time step integrates past memory information and current real-time information, and the formula is:

[0085] ;

[0086] ;

[0087] in, is the hidden layer output, and is the activation function, For the time step Enter data when The weight matrix of the connection, For the time step The hidden layer output The weight matrix of the connection, For the time step The hidden layer output The weight matrix of the connection, and is the offset, For downstream tasks.

[0088] Although RNN can retain certain historical information when processing sequence data, its ability to retain early information is weak when facing longer sequences, and it is easy to encounter the problem of gradient disappearance or explosion. In order to effectively capture dependencies over a longer time span, it is a better choice to use the LSTM model to extract time series features. The LSTM neural network uses input gates, output gates, forget gates, and internal memory units to control the behavior of the hidden layer. The forget gate determines whether to retain previous memory information. If the gate is open, the historical information is saved; if the gate is closed, the historical information is forgotten. The input gate controls whether new input information is allowed to enter the current memory unit, and the output gate determines whether the current input signal is passed to the next layer.

[0089] In the present invention, the input time series data is processed by the forward and reverse LSTM models to obtain information of two cell states. These two states encode the forward and reverse time series information respectively. The two cell states are spliced ​​along the last dimension to obtain the time series information mapping vector .

[0090] The DNN model is the basic model in deep learning and is an extension of the perceptron. The perceptron allows multiple inputs but only one output. The perceptron learns how to map multiple inputs to outputs:

[0091] ;

[0092] in, The bias term allows the model to have a certain prediction value even without input features. It can change the center position of the activation function and improve the flexibility and expressiveness of the model. In a neural network, each node (neuron) usually has a bias term associated with it. Representatives and Input The associated weights, Indicates input features, is an index used to label specific inputs and their corresponding weights, Refers to the number of inputs, that is, how many Input to the neuron currently being considered.

[0093] Then, output the result Through the activation function :

[0094] .

[0095] The perceptron model has only two outputs: 1 or -1, so it is only suitable for binary classification tasks. Since the perceptron is a linear classification model, it cannot solve more complex tasks, such as the XOR problem. Therefore, DNN is improved on the basis of the perceptron, and the formula is:

[0096] .

[0097] In the present invention, for each time step, DNN is first used to map it into a high-dimensional dense vector. Then, an average pooling operation is performed along the time step to obtain the global mapping vector In the output layer, the activation function related to the task is usually used, such as the Sigmoid activation function in the binary classification task or the Softmax activation function in the multi-classification task. In the hidden layer, the commonly used activation functions are Relu, Tanh, etc.

[0098] The present invention utilizes two convolutional layers and two maximum pooling layers in the CNN module for feature extraction. The convolutional layer uses a small convolution kernel to reduce parameters and prevent overfitting. The two convolutional layers use different step sizes to capture more spatial local features. After the convolution operation, the maximum pooling process is performed to achieve dimensionality reduction of the convolutional layer output, while retaining the main features and reducing the amount of calculation and parameters. In the convolutional layer, the ReLU activation function is used to increase the nonlinear ability of the model, and the L2 regularization operation is used to balance the model performance and complexity to avoid overfitting. Parameter initialization can speed up the gradient convergence speed.

[0099] In order to enhance the spatial feature extraction capability, the features extracted by CNN are input into the SE attention module to strengthen important feature information and suppress irrelevant information. Subsequently, the re-weighted features are input into a two-layer BIGRU network to extract the spatiotemporal features of the entire vibration signal. A dropout layer is connected after each layer of BIGRU to reduce the number of network parameters and avoid overfitting problems. Since the BIGRU network is affected by past and future bidirectional information, a dropout layer is used to balance the number of parameters and performance of the network. This method combines the advantages of CNN and BIGRU, can effectively extract deep spatiotemporal features from vibration signals, and ultimately generate a feature vector , in order to achieve accurate prediction of underwater vector thruster life.

[0100] The sensor mapping vector model is based on information encoding technology, in which each type of sensor can reflect the operating status characteristics of the device from its own unique perspective. This step aims to enhance the generalization ability of the model so that it can still make accurate predictions or classifications when faced with data from different sensor combinations. Each time period is associated with specific port attributes, which are derived from different sensor data. By encoding the port information of different sensors, richer and more detailed information can be provided for the analysis time period.

[0101] At the beginning of the present invention, a trainable feature vector is randomly set for each sensor. These vectors are trainable parameters. As the training process progresses, the sensor's representation vector is adjusted through the back propagation algorithm so that they gradually approach a state that can better represent the sensor information. After multiple iterations of optimization, the feature vector corresponding to each sensor gradually converges to the optimal state, and finally forms the optimal vector that can accurately represent the characteristics of the sensor. This processing method allows the model to understand the thruster operating status more comprehensively from different perspectives, thereby improving the accuracy of the thruster condition analysis. Afterwards, these optimized sensor feature vectors are integrated into the neural network model. Dynamic weighted fusion is performed through different fusion strategies, allowing the model to receive sensor information related to the time period. The sensor representation vector obtained in the training phase can be used in scenarios containing only a single sensor in the testing phase.

[0102] S3, in view of the high coupling and real-time requirements of data, integrate the features from multiple neural networks and select features from different perspectives. The multiple feature information extracted in step S2 is fused using the feature fusion algorithm MVSBA based on the marked area to obtain a weighted fusion vector.

[0103] MVSBA first initializes multiple load vectors, which together form a load region. Then, the feature vectors in this load region are compared with each load vector to calculate the similarity score. The higher the sum of these similarity scores, the greater the weight value of the corresponding feature vector. In this way, feature vectors located at different positions in the load region will be assigned different weights. By adjusting and optimizing the load region, that is, adjusting and optimizing the load vector, the best way to fuse various different feature vectors can be found.

[0104] The different dimensions of the eigenvector reflect the significance of the feature in different directions. Simply adding the eigenvalues ​​of different dimensions cannot fully reflect its significance in the overall direction, because different eigenvectors have different influences on the final result. Therefore, it is necessary to consider the weight ratio of different eigenvectors.

[0105] To solve this problem, MVSBA is proposed. Through the training process, the algorithm enables the model to automatically find the most suitable feature weight allocation and realize the organic combination of different feature vectors according to their proportions. This method avoids excessive reliance on prior knowledge and improves the model's adaptability. Through training, the model can learn how to allocate weights of different feature vectors according to the actual situation of the problem, thereby more effectively extracting key features that reflect the health indicators of underwater vector thrusters.

[0106] S4, using the smoothing algorithm SG, the weighted fusion vector, i.e., the second fusion vector Perform smoothing.

[0107] S5, , and The indicators are integrated into a weighted mean , as the evaluation index of the evaluation function HI to measure the health status of the underwater vector thruster.

[0108] Specifically, according to the characteristics of abnormal data, incomplete data and other data, a dirty database is established, and dirty data purification is achieved with the help of statistics, clustering and other algorithms to separate the usable database from the source database. The preprocessing in step S1 specifically includes the following steps:

[0109] S1.1, using continuous wavelet transform to denoise, detect abnormal points and abnormal segments of the denoised data, and clean the abnormal segments. Among them, abnormal points usually refer to noise points and isolated points in the signal, and abnormal segments refer to continuous segments composed of continuous abnormal points.

[0110] Continuous wavelet transform (CWT) uses a wavelet basis function called "mother wave" and performs operations such as scaling and translation on it to match the input signal to obtain a continuous wavelet function.

[0111] The principle of small fundamental wave transform can be explained by its formula. Controlling the compression or expansion degree of the wavelet basis function determines the stretching of the wavelet basis function on the time axis. When it is large, the wavelet basis function will be stretched to correspond to the lower frequency components; when When it is smaller, the wavelet basis function will be compressed, corresponding to the higher frequency components. This determines the translation of the wavelet basis function on the time axis, that is, it determines the starting position of the wavelet basis function. By changing the translation factor , signals in different time periods can be analyzed.

[0112] In the present invention, the process of wavelet transform can be divided into two steps: decomposition and reconstruction. First, the original signal is decomposed by combining different scale and translation parameters to obtain a series of wavelet coefficients, which represent signal components of different frequencies and time ranges. Then, these wavelet coefficients are reconstructed into the original signal through inverse wavelet transform.

[0113] S1.2, considering the missing original data caused by sensor failure factors and the missing original data caused by abnormal segment cleaning, the regression interpolation method is used to fill the missing data according to the relationship between the data points before and after the missing data.

[0114] S1.3, use KNN algorithm to remove outliers: First, set a distance threshold. If the distance between a sample and its nearest neighbor exceeds the threshold, it is considered an outlier and removed. Then calculate the Euclidean distance between the test sample and the training sample, select the nearest K neighbors for classification, sort the samples according to the calculated distance, and select the nearest K neighbors for output.

[0115] S1.4, normalize the data, the formula is as follows:

[0116] ;

[0117] in, represents the normalized data, For input data, To minimize, To take the maximum.

[0118] The purpose of normalization is to map the value range of the original data to a common scale, so as to facilitate comparison and analysis between different indicators. Normalization can help eliminate the dimensional differences between different indicators, making the data more standardized and comparable.

[0119] S1.5, use principal component analysis to reduce the dimensionality of the data. Principal component analysis (PCA) is a commonly used data dimensionality reduction technique that maps high-dimensional data to a low-dimensional space through linear transformation while retaining as much key information as possible in the data. The key to PCA is to perform eigendecomposition on the covariance matrix of the data, find the eigenvector with the largest eigenvalue as the new coordinate system, and then project the data into these new coordinate systems to achieve dimensionality reduction.

[0120] The covariance matrix is ​​used to measure the linear correlation between features in the data. First, the covariance matrix C is calculated for the standardized data, and the eigendecomposition is performed to obtain the eigenvalues ​​and corresponding eigenvectors. Then, the eigenvectors are sorted according to the size of the eigenvalues, and the first K eigenvectors are selected as principal components. Finally, the original data is projected onto these principal components to obtain the data representation after dimensionality reduction.

[0121] Specifically, the feature fusion algorithm MVSBA based on the marked area includes: a first module Block-1 and a second module Block-2 which are connected in sequence.

[0122] Specifically, in step S3, the following steps are performed in the first module Block-1:

[0123] S3.1, map feature vectors of different dimensions to the same dimensional space, and standardize the mapped new vectors.

[0124] S3.2, Initialize the first load matrix , M is the number of load vectors.

[0125] , To generate a random number function, take 4 load vectors as an example, recorded as the load matrix ,in, , , , is the first load vector. The initial value of the load vector follows the Gaussian distribution, and the modulus of the load vector is small and remains stable.

[0126] S3.3, perform inner product operation on the standardized vector and the first load vector to obtain the first labeling matrix :

[0127] ;

[0128] in, is the feature vector dimension.

[0129] The result of the inner product operation of the standardized vector and a vector with a fixed modulus can reflect the similarity between the two, so That is to say, The eigenvector and The similarity of the loading vectors.

[0130] S3.4, the labeling matrix Sum by row to get the first sum matrix .

[0131] in, No. The component is The sum of similarities with each component vector in the load area represents The total score. Input to the Softmax function to get the vector , where the Softmax function formula is as follows:

[0132] ;

[0133] In the formula, For the The output value of the node, is the total number of output nodes. Node refers to the virtual node of the output layer, and the number of nodes is equal to When calculating, first calculate the exponential value of each output node value, and then sum them as the denominator to be divided by the final result. Node When outputting a value, first calculate the exponential value of the output value, and then divide it by the previous summation value to get the final result.

[0134] S3.5, calculate the first fusion vector obtained by MVSBA of the feature vectors under different viewing angles :

[0135] The four components of are: , then the first fusion vector for:

[0136] ;

[0137] in, is the temporal information mapping vector obtained by BiLSTM; is the global mapping vector obtained by the deep neural network; is the local saliency information mapping vector, obtained by a convolutional neural network based on the attention mechanism; is the sensor mapping vector obtained by the information encoding neural network.

[0138] Specifically, step S3.1 includes the following steps:

[0139] S3.1.1, since the feature vectors generated by various methods usually have different dimensions, it is necessary to first map these feature vectors of different dimensions to the same dimensional space and standardize the mapped new vectors. Suppose the initial vector is , is the average, is the standard deviation, and the initial eigenvector , mapped to ,…, , denoted as .

[0140] S3.1.2, taking 4 eigenvectors as an example, the standardized eigenvectors are ,but , , , are all unit vectors, let ,but is the first matrix after concatenating the eigenvectors.

[0141] Specifically, step S3.4 is as follows:

[0142] Will Sum by row .

[0143] Specifically, the load matrix of block-2 is obtained by linear mapping the fusion vector obtained from block-1, and the new feature vector matrix to be fused is obtained by multiplying the feature matrix obtained from block-1 by their respective weights. The load matrix B2 in the block-2 model structure is initialized to the output fusion vector Z of block-1, which is transposed and right-multiplied by the matrix In addition, the two blocks remain consistent. Through the hierarchical structure of two blocks, the model can often obtain more abstract and robust information than a single block structure.

[0144] In step S3, the following steps are performed in the second module Block-2:

[0145] S3.6, Calculate the second load matrix :

[0146] ;

[0147] in, is the transformation matrix.

[0148] S3.7, calculate the second matrix after concatenating the eigenvectors , and calculate the second labeling matrix :

[0149] ;

[0150] , ;

[0151] in, , , and The second matrix The normalized feature vector is , , and is the second load vector.

[0152] ;

[0153] Will Sum the rows to get the second sum matrix :

[0154] .

[0155] The four components of are: , then the second fusion vector for:

[0156] .

[0157] Specifically, in order to eliminate the local fluctuations of the health indicators generated by the MVSBA model, correctly maintain the long-term trend and effectively capture mutations, the SG smoothing algorithm is introduced. The Savitzky-Golay (SG) filter is a low-pass filter. The most characteristic feature is that it can ensure that the shape and width of the signal remain unchanged while filtering. Step S4 specifically includes the following steps:

[0158] S4.1, construct a Order polynomial:

[0159] ;

[0160] In the formula, are the coefficients of the polynomial; is the order of the polynomial; is the weight of the polynomial; About Polynomial function of order; .

[0161] S4.2, using least squares to fit residuals :

[0162] ;

[0163] In the formula, is a set of data points, for a given data point.

[0164] To make the residual Minimum, let The derivative of each coefficient is 0, that is:

[0165] ;

[0166] From this we can get, .

[0167] Given a data point and order The polynomial can be determined, and whenever the window is moved, the output value of SG is the value of the center point of the window.

[0168] Specifically, the eigenvalues ​​output by the SG filter contain a variety of features that can characterize the health status of the underwater vector thruster, and different features have different characterization strengths. The health index system should have three characteristics: correlation, monotonicity, and robustness. Correlation measures the degree of association between the health indicator sequence and the time series, monotonicity describes the trend of continuous growth or decline of the health indicator sequence, and robustness evaluates the resistance of the health indicator sequence to noise. The present invention uses these three indicators to evaluate the performance of health indicators.

[0169] Step S5 specifically includes the following steps:

[0170] S5.1, the health indicator sequence is ,in, Indicates the first Time point The corresponding eigenvalues ​​are decomposed into two parts: the steady trend and the random residual using the sliding average method:

[0171] ;

[0172] In the formula, For the stable part, is the random margin.

[0173] S5.2, the calculation formulas for the three indicators are as follows:

[0174] ; ;

[0175] ;

[0176] in, represents the number of data points, is a trend value, indicating the linear correlation between the health HI value and the sampling time t. The closer the absolute value is to 1, the stronger the linear correlation with the sampling time. It is the monotonicity value, indicating the monotonicity of the curve. The positive and negative numbers of the difference between two adjacent HI values ​​are used to evaluate the monotonicity of the HI curve. It is a robustness value that quantifies the degree to which the model's output changes relative to the expected output when it is affected by disturbances or uncertainties.

[0177] In the formula, Defined as:

[0178] ;

[0179] in, is a step function.

[0180] S5.3, Calculation , and The fusion weighted mean , the calculation formula is as follows:

[0181] ;

[0182] in , , Respectively , , The weight of the three, in the invention, is , , ,in The value range of is from 0 to 1, and it can be concluded that the value range of CMR is from 0 to 1. The higher the value, the better the correlation, monotonicity and robustness performance. The feature can better represent the performance degradation curve of the vector thruster.

[0183] Finally, all trainable parameters in the model are frozen, and the remaining 20% ​​of the test set is input into the trained model for performance testing. In actual application, the real-time acquired and fused full-cycle sensor data is directly input into the model to output health indicators, thereby realizing the construction of the underwater vector thruster health indicator system.

[0184] 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 method for constructing a health index system for an underwater vector thruster, characterized in that: The specific steps include: S1, using the sensor carried by the underwater vector thruster to obtain the full-cycle vibration signal, performing data preprocessing on the full-cycle vibration signal to form a preprocessed data set; S2, divide the data set into a training set and a test set, and input the test set and the training set into the bidirectional long short-term memory network BiLSTM, the convolutional neural network CNN-SE based on the attention mechanism, the deep neural network DNN, and the information encoding neural network at the same time, and use different neural networks to extract feature information; S3, fusing the various feature information extracted in step S2 using a feature fusion algorithm MVSBA based on a marked area to obtain a weighted fusion vector; S4, using smoothing algorithm SG, smoothing the weighted fusion vector; S5, , and The indicators are integrated into a weighted mean , as the evaluation index of the evaluation function HI to measure the health status of the underwater vector thruster.

2. A method for constructing a health index system of an underwater vector thruster according to claim 1, characterized in that: The preprocessing in step S1 specifically includes the following steps: S1.1, using continuous wavelet transform to denoise, detect abnormal points and abnormal segments of the denoised data, and clean the abnormal segments; S1.2, considering the missing original data caused by sensor failure and the missing original data caused by abnormal segment cleaning, the regression interpolation method is used to fill the missing data according to the relationship between the data points before and after the missing data; S1.3, use KNN algorithm to remove outliers; S1.4, normalize the data, the formula is as follows: ； in, represents the normalized data, For input data, To minimize, To obtain the maximum; S1.5, use principal component analysis to reduce the dimensionality of the data.

3. The method for constructing a health index system of an underwater vector thruster according to claim 1, characterized in that: The feature fusion algorithm MVSBA based on the marked area includes: a first module Block-1 and a second module Block-2 connected in sequence.

4. The method for constructing a health index system of an underwater vector thruster according to claim 3, characterized in that: In step S3, the following steps are performed in the first module Block-1: S3.1, map feature vectors of different dimensions to the same dimensional space, and standardize the mapped new vectors; S3.2, Initialize the first load matrix , M is the number of load vectors; , To generate a random number function, take 4 load vectors as an example, recorded as the load matrix ,in, , , , is the first load vector, and the initial value of the load vector follows Gaussian distribution; S3.3, perform an inner product operation on the standardized vector and the first load vector to obtain the first labeling matrix : ； in, is the feature vector dimension; S3.4, the labeling matrix Sum by row to get the first sum matrix ; S3.5, calculate the first fusion vector obtained by MVSBA of the feature vectors under different viewing angles : The four components of are: , then the first fusion vector for: ； in, Mapping vector for timing information, is the global mapping vector, is the local saliency information mapping vector, Map vectors for sensors.

5. The method for constructing a health index system of an underwater vector thruster according to claim 4, characterized in that: Step S3.1 specifically includes the following steps: S3.1.1, let the initial vector be , is the average, is the standard deviation, and the initial eigenvector , mapped to ,…, , denoted as ; S3.1.2, taking 4 eigenvectors as an example, the standardized eigenvectors are ,but , , , are all unit vectors, let ,but is the first matrix after concatenating the eigenvectors.

6. A method for constructing a health index system of an underwater vector thruster according to claim 5, characterized in that: Step S3.4 is specifically as follows: Will Sum by row .

7. A method for constructing a health index system for an underwater vector thruster according to claim 6, characterized in that: In step S3, the following steps are performed in the second module Block-2: S3.6, Calculate the second load matrix : ； in, is the transformation matrix; S3.7, calculate the second matrix after concatenating the eigenvectors , and calculate the second labeling matrix : ； ， ； in, , , and The second matrix The normalized feature vector is , , and is the second load vector; ； Will Sum the rows to get the second sum matrix : ； The four components of are: , then the second fusion vector for: 。 8. The method for constructing a health index system of an underwater vector thruster according to claim 1, characterized in that: Step S4 specifically includes the following steps: S4.1, construct a Order polynomial: ； In the formula, are the coefficients of the polynomial; is the order of the polynomial; is the weight of the polynomial; For about Polynomial function of order; ; S4.2, using least squares to fit residuals : ； In the formula, is a set of data points, For a given data point; To make the residual Minimum, let The derivative of each coefficient is 0, that is: ； From this we can get, .

9. The method for constructing a health index system of an underwater vector thruster according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1, the health indicator sequence is ,in, Indicates the first Time point The corresponding eigenvalues ​​are decomposed into two parts: the steady trend and the random residual using the sliding average method: ； In the formula, For the stable part, is the random margin; S5.2, the calculation formulas for the three indicators are as follows: ； ； ； in, represents the number of data points, is the trend value, is the monotonicity value, is the robustness value; In the formula, Defined as: ； in, is a step function; S5.3, Calculation , and The fusion weighted mean , the calculation formula is as follows: ； in , , Respectively , , The weight of the three.

Images