Marine composite material propeller working strain prediction method based on LSTM and PINN fusion
Through the fusion method of LSTM and PINN, combined with the embedded sensor and the loss function of physical constraints, the dynamic adaptability and physical constraint problems of marine composite propeller strain prediction are solved, and high-precision strain prediction is achieved.
Patent Information
- Application Number
- CN202510532484.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-25
- Publication Date
- 2025-08-01
AI Technical Summary
The existing LSTM-based propeller strain prediction method for marine composite materials cannot dynamically adapt to complex marine conditions, and lacks explicit constraints on physical laws, resulting in inaccurate prediction results and low engineering credibility.
Using the method of fusion of LSTM and PINN, we use the finite element model to establish a finite element model, embed an embedded flexible sensor to collect strain data, perform data preprocessing and multi-scale decomposition, and build a loss function with physical constraints, and combine the online incremental learning mechanism for model training and prediction.
It improves the accuracy and engineering credibility of strain prediction, can dynamically adapt to complex maritime working conditions, reduces data acquisition and processing costs, and ensures that the prediction results comply with physical laws.
Smart Images
Figure CN120409250A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of composite material structural health monitoring, and particularly to a method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN. Background Technique
[0002] With the development of composite material technology, composite propellers have been widely used in the ship field. Due to the advantages of high strength and light weight of composite materials, they show good performance in the actual working process. However, due to the relatively complex working environment of marine propellers (such as fluid impact, load fluctuation, sudden vibration, temperature change, etc.), the working strain of the propeller is easily affected by various factors, and traditional strain prediction methods relying on finite element analysis and empirical formulas are difficult to accurately and real-time predict the working state of the propeller, resulting in the inability to effectively monitor it during the working process. In recent years, deep learning technology has provided new solutions for time series prediction. Among them, the Long Short-Term Memory (LSTM) neural network shows potential in the field of industrial monitoring due to its excellent ability to capture strain time series characteristics. However, the existing LSTM-based strain prediction methods still have two major limitations: First, relying on offline data preprocessing and static model training, it cannot dynamically adapt to the real-time changes of complex marine working conditions; Second, the pure data-driven model lacks explicit constraints on physical laws (such as material mechanics equilibrium equations, fluid dynamics equations), resulting in the prediction results may deviate from the actual physical laws and reduce the engineering credibility. Summary of the Invention
[0003] The present invention provides a method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN to overcome the above technical problems.
[0004] To achieve the above object, the technical solution of the present invention is:
[0005] A method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN, specifically including the following steps:
[0006] S1: Based on finite element software, according to the attribute data of the marine composite propeller, establish a finite element model of the propeller, and perform mesh division on the finite element model of the propeller to obtain the model area mesh of the finite element model of the propeller;
[0007] S2: Define each area mesh in the model area mesh as a mesh node, and call the static analysis module of the finite element software to obtain the stress / strain distribution point cloud of each mesh node;
[0008] According to the concentration degree of stress / strain distribution point clouds in the model area grid, confirm the key monitoring areas of the marine composite propeller, and embed the preset inlaid flexible sensors on the maximum thickness line inside the propeller blade corresponding to the key monitoring areas of the real propeller. Through the set underwater propeller test system, collect and obtain the internal original strain time series data of the propeller blade under different working conditions;
[0009] S3: Perform data preprocessing on the internal original strain time series data of the propeller blade to obtain optimized strain time series data;
[0010] S4: Extract the time series data features of different frequency bands in the optimized strain time series data to obtain multi-scale decomposed time series data;
[0011] And the multi-scale decomposed time series data includes high-frequency time series data and low-frequency time series data;
[0012] The high-frequency time series data includes instantaneous vibration data and impact data;
[0013] The low-frequency time series data includes the slow-varying effect caused by temperature and the structural fatigue data of fatigue accumulation;
[0014] And according to the components of the multi-scale decomposed time series data, reconstruct the optimized strain time series data to obtain the reconstructed strain time series data;
[0015] S5: Build a propeller working strain prediction model based on the LSTM neural network;
[0016] And introduce a loss function with physical constraints improved based on the PINN algorithm, and obtain an optimized propeller working strain prediction model according to the reconstructed strain time series data;
[0017] S6: According to the optimized propeller working strain prediction model, realize the prediction of the working strain of the marine composite propeller.
[0018] Furthermore, the S3 specifically includes the following steps:
[0019] S31: Perform data filling on the internal original strain time series data of the propeller blade to obtain filled time series data, and the expression for data filling is
[0020]
[0021] In the formula: represents the filled time series data; x i represents the i-th data point of the internal original strain time series data; N represents the number of data samples of the internal original strain time series data;
[0022] S32: Remove the outliers from the filled time-series data, and interpolate the data outliers based on the linear interpolation method to obtain the optimized time-series data;
[0023] And the method for removing outliers from the filled time-series data is specifically
[0024] Calculate and obtain the Z-score value of each data point in the filled time-series data;
[0025] And the formula for obtaining the Z-score value of each data point is
[0026]
[0027] In the formula: x i ′ represents the data point in the filled time-series data; μ represents the mean of the filled time-series data; σ represents the standard deviation of the filled time-series data;
[0028] Preset the Z-score threshold range, and determine whether the Z-score value of each data point satisfies the Z-score threshold range;
[0029] If it is satisfied, the corresponding data point is regarded as a normal data point and retained;
[0030] Otherwise, the corresponding data point is regarded as an outlier data point and deleted;
[0031] S33: Standardize the optimized time-series data to obtain the optimized strain time-series data;
[0032] And the expression for standardizing the optimized time-series data is
[0033]
[0034] In the formula: represents the optimized time-series data after standardization; x i ″ represents the data point in the optimized strain time-series data.
[0035] Furthermore, the expression for reconstructing the optimized strain time-series data in S4 is
[0036]
[0037] In the formula: x(t) represents the reconstructed optimized strain time-series data; a3(t) represents the low-frequency time-series data; d j (t) represents the high-frequency time-series data and j = 1, 2; d1(t) represents the instantaneous vibration data of the propeller; d2(t) represents the impact data of the propeller.
[0038] Further, the method for obtaining the loss function with physical constraints improved based on the PINN algorithm in S5 specifically includes the following steps:
[0039] S001: Assuming that the structural deformation of the propeller during underwater operation is negligible, obtain the material mechanics equilibrium equation of static equilibrium as the physical constraint operator of the propeller, and its expression is
[0040]
[0041] In the formula: x represents the position vector, that is, the spatial position of the key monitoring area on the propeller blade; t represents the specific time point when the propeller operates at a certain rotational speed; σ represents the stress tensor of the propeller blade at the spatial position x and time t; f represents the body force density vector, that is, the additional hydrodynamic force per unit volume of the propeller blade;
[0042] S002: Denote the physical constraint operator as where u represents the predicted value of the regional network for the strain field, and construct the physical consistency loss function, and its expression is
[0043]
[0044] In the formula: Loss phys represents the physical consistency index of the physical consistency loss function; x i ,t i represents the data at the i-th spatial-temporal sampling position point in the reconstructed strain time series data;
[0045] S003: Construct the data fitting loss function of the LSTM neural network, and its expression is
[0046]
[0047] In the formula: Loss data represents the data fitting loss function; represents the predicted value of the strain time series data output by the propeller working strain prediction model; y(t) represents the true value of the strain time series data;
[0048] S004: According to the physical consistency loss function and the data fitting loss function, obtain the loss function with physical constraints, and its expression is
[0049]
[0050] In the formula: Loss total represents the loss function with physical constraints; λ represents the balance coefficient that can adjust the trade-off between data fitting and physical constraints.
[0051] Further, the method for obtaining an optimized propeller working strain prediction model in S5 specifically includes the following steps:
[0052] S51: Obtain the characteristic data of the propeller blade;
[0053] And the characteristic data includes any one of the high-frequency time series data or low-frequency time series data in the reconstructed strain time series data, or a combined data of high-frequency time series data and low-frequency time series data;
[0054] Take the stress inside the propeller blade as the label data, and randomly divide it into a test set and a training set according to the reconstructed strain time series data;
[0055] S52: Construct a propeller working strain prediction model based on the LSTM neural network;
[0056] And the propeller working strain prediction model includes an input layer, an LSTM layer, and a fully connected output layer connected in sequence;
[0057] The input layer is used to transmit the reconstructed strain time series data to the LSTM layer;
[0058] The LSTM layer is used to extract the strain time series features of the reconstructed strain time series data, and the time series features are used to describe the correlation between the strain fluctuations inside the propeller blade and the sudden load;
[0059] The fully connected output layer is used to predict and output the strain prediction value inside the propeller blade according to the strain time series features;
[0060] S53: Based on the online incremental learning mechanism, train the propeller working strain prediction model according to the training set to obtain the trained propeller working strain prediction model;
[0061] S54: Based on the loss function with physical constraints, confirm whether the output of the trained propeller working strain prediction model converges according to the test set to evaluate the trained propeller working strain prediction model;
[0062] If it is confirmed that the output of the trained propeller working strain prediction model converges, then confirm that the trained propeller working strain prediction model at this time is the optimized propeller working strain prediction model;
[0063] Otherwise, adaptively adjust the parameter weights of the trained propeller working strain prediction model based on the backpropagation method, and repeat step S53.
[0064] Beneficial effects: The present invention provides a method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN. By identifying the key monitoring areas of the marine composite propeller and using an embedded flexible sensor, the original internal strain time series data of the propeller blade under different working conditions can be obtained, so as to obtain representative strain time series data and reduce the cost of collecting and processing a large amount of data. By preprocessing the original internal strain time series data of the propeller blade, it is possible to detect and eliminate the error data caused by communication interference, sensor drift or environmental factors that may occur during the acquisition process, reduce high-frequency noise, and retain the main strain signal to improve the accuracy of the strain time series data. By extracting and optimizing the high-frequency and low-frequency time series data in the strain time series data, multi-scale decomposed time series data can be obtained, and the optimized strain time series data can be reconstructed to obtain the reconstructed strain time series data. It can fully improve the accuracy of the strain time series data, construct a propeller working strain prediction model based on the LSTM neural network, and introduce a loss function with physical constraints improved by the PINN algorithm. According to the reconstructed strain time series data, an optimized propeller working strain prediction model can be obtained. Since the traditional data-driven LSTM prediction lacks direct constraints on the engineering physical mechanism, it may lead to prediction results that do not conform to the basic physical laws. Therefore, by fusing the LSTM neural network and the PINN algorithm, and adding a physical constraint operator in the form of a partial differential equation (PDE) to the loss function, the output of the model not only has a fitting effect on the data, but also satisfies the basic laws physically, greatly improving the engineering credibility. Description of the Drawings
[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0066] Figure 1 It is a flowchart of the method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN of the present invention;
[0067] Figure 2 It is a schematic installation diagram of the flexible sensor of the composite propeller in this embodiment;
[0068] Figure 3 It is a flowchart of obtaining an optimized propeller working strain prediction model in this embodiment;
[0069] Figure 4This is the core block diagram of the working strain prediction method for marine composite propellers based on the fusion of LSTM and PINN in this embodiment;
[0070] Figure 5 This is a schematic diagram of an underwater experiment for obtaining the internal original strain time series data of the propeller blade under different working conditions in this embodiment. Detailed implementation manners
[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0072] This embodiment provides a working strain prediction method for marine composite propellers based on the fusion of LSTM and PINN, as Figure 1 shown in Figure 4 the following, and specifically includes the following steps:
[0073] S1: Based on finite element software, according to the property data of the marine composite propeller, establish a propeller finite element model; the property data includes material properties, geometric shapes, and expected load conditions;
[0074] And perform mesh division on the propeller finite element model to obtain the model area mesh of the propeller finite element model; specifically, the method of establishing a propeller finite element model through finite element software and performing mesh division on the propeller finite element model is a well-known prior art means and will not be elaborated here.
[0075] S2: Define each area mesh in the model area mesh as a mesh node, and call the static analysis module of the finite element software to obtain the stress / strain distribution point cloud of each mesh node;
[0076] According to the concentration degree of the stress / strain distribution point cloud in the model area mesh, confirm the key monitoring area of the marine composite propeller. Among them, according to expert experience, the area mesh with a high concentration degree of the stress / strain distribution point cloud is used as the key monitoring area of the propeller;
[0077] And embed the preset embedded flexible sensor on the maximum thickness line inside the propeller blade corresponding to the key monitoring area of the real propeller, and through the set propeller underwater test system, collect and obtain the internal original strain time series data of the propeller blade under different working conditions; among them, the method of embedding the preset embedded flexible sensor inside the propeller blade is a well-known prior art means and will not be elaborated here. AsFigure 2 As shown in the figure, 1 is the maximum blade thickness line, and 2 is the embedded flexible sensor;
[0078] In this embodiment, it also includes a propeller underwater test system for assembling a composite propeller, a connecting shaft, a conductive slip ring, a marine thruster, and a strain tester, connecting the strain tester to a computer terminal, and conducting an underwater test on the propeller; A total of three groups of experiments were carried out. According to the power of the marine thruster, three different propeller rotation speeds were changed. The propeller rotation speed n was set as the independent variable, and the strain ε change of the strain gauge inside the propeller was measured within the set time T to obtain the original data; The methods for assembling the composite propeller, the connecting shaft, the conductive slip ring, the marine thruster, and the strain tester and obtaining the original data are well-known technical means in the art and will not be elaborated here; As Figure 5 shown in the figure, 3 is the propeller, 4 is the conductive slip ring, 5 is the marine thruster, 6 is the computer terminal, and 7 is the strain tester;
[0079] S3: Perform data preprocessing on the internal original strain time series data of the propeller blade to obtain optimized strain time series data;
[0080] In this embodiment, since the internal original strain time series data of the propeller blade often contains noise, missing values, and outliers, it is necessary to perform data preprocessing to ensure the high quality of the model input. Screen the time interval (t) during which the propeller operates stably. The stable operating time interval is the time interval that removes the instantaneous states of the start and stop of the marine thruster and has a small strain amplitude fluctuation of the propeller within the set time T. Perform preprocessing operations such as data cleaning and denoising, standardization, and multi-scale analysis on this section of data to ensure the quality of the input data;
[0081] Specifically, it includes the following steps:
[0082] S31: The model used is a statistical method, namely the Mean Imputation method, to fill the internal original strain time series data of the propeller blade to obtain the filled time series data, and the expression for data filling is
[0083]
[0084] In the formula: represents the filled time series data; x i represents the i-th data point of the internal original strain time series data; N represents the number of data samples of the internal original strain time series data;
[0085] S32: Use the Z-score method to detect and remove outliers in the data, and perform interpolation processing on the data outliers based on the linear interpolation method to obtain the optimized time series data;
[0086] And a method for removing outliers from the filled time-series data, specifically as follows
[0087] Calculate and obtain the Z-score value of each data point in the filled time-series data;[[ID=))]]
[0088] And the formula for obtaining the Z-score value of each data point is
[0089]
[0090] In the formula: x i ′ represents the data point in the filled time-series data; μ represents the mean of the filled time-series data; σ represents the standard deviation of the filled time-series data;
[0091] Preset the Z-score threshold range, and determine whether the Z-score value of each data point meets the Z-score threshold range;
[0092] If it meets, regard the corresponding data point as a normal data point and retain it;
[0093] Otherwise, regard the corresponding data point as an abnormal data point and delete it;
[0094] S33: Perform normalization processing on the optimized time-series data to obtain the optimized strain time-series data;
[0095] Since the LSTM neural network is sensitive to the scale of the input data, it is necessary to perform normalization operations on the experimental data. Normalization can map the data to the same scale range, thereby improving the convergence speed and prediction accuracy of the model; the data normalization method in this embodiment is the Z-score normalization method, and the expression for performing normalization processing on the optimized time-series data is
[0096]
[0097] In the formula: represents the optimized time-series data after normalization processing; x i ″ represents the data point in the optimized strain time-series data;
[0098] S4: In this embodiment, through multi-scale analysis of the optimized strain time-series data, the goal is to decompose the collected composite propeller strain time-series data into components of different frequency bands, so as to capture the high-frequency dynamic changes (such as instantaneous vibration, impact) in the short term and long-term trends (such as slow change effects caused by temperature, fatigue accumulation) respectively, and apply them to the strain prediction of the composite propeller. Specifically, it includes:
[0099] Extract the high-frequency time-series data and low-frequency time-series data from the optimized strain time-series data to obtain the multi-scale decomposition time-series data;
[0100] Moreover, the high-frequency time series data includes instantaneous vibration data and impact data, which are used to capture sudden situations during the operation of the propeller;
[0101] The low-frequency time series data includes the slow-varying effect caused by temperature and the structural fatigue data of fatigue accumulation, which are used to judge the overall trend and the situation of fatigue accumulation;
[0102] And according to the components of the multi-scale decomposed time series data, the reconstructed and optimized strain time series data is obtained by reconstructing the optimized strain time series data;
[0103] In a specific embodiment, the expression of the reconstructed and optimized strain time series data is
[0104]
[0105] In the formula: x(t) represents the reconstructed and optimized strain time series data; a3(t) represents the low-frequency time series data; d j (t) represents the high-frequency time series data and j = 1, 2; d1(t) represents the instantaneous vibration data of the propeller operation; d2(t) represents the impact data of the propeller operation;
[0106] This embodiment can input the separated high-frequency time series data and low-frequency time series data into the constructed propeller working strain prediction model separately or in combination, so that the model can capture short-term fluctuations and long-term trends at the same time. At the same time, for different scales, based on the preset adaptive update mechanism to confirm the parameter weights of the model, the high-frequency part can be used to detect instantaneous anomalies, while the low-frequency part is used for the correction of the overall trend, and both can trigger targeted early warnings and parameter adjustments;
[0107] S5: This embodiment relies on the Python platform to complete the construction of the LSTM neural network model to obtain a propeller working strain prediction model based on the LSTM neural network;
[0108] And introduce a loss function with physical constraints improved based on the PINN algorithm, and obtain an optimized propeller working strain prediction model according to the reconstructed strain time series data;
[0109] In a specific embodiment, the method for obtaining an optimized propeller working strain prediction model is as Figure 3 shown, and specifically includes the following steps:
[0110] S51: Obtain the characteristic data of the propeller blade;
[0111] And the characteristic data includes any one of the high-frequency time series data or the low-frequency time series data in the reconstructed strain time series data, or a combined data of the high-frequency time series data and the low-frequency time series data;
[0112] Take the stress inside the propeller blade as labeled data, and randomly divide it into a test set and a training set according to the reconstructed strain time series data;
[0113] S52: Construct a propeller working strain prediction model based on the LSTM neural network;
[0114] And the propeller working strain prediction model includes an input layer, an LSTM layer, and a fully connected output layer connected in sequence;
[0115] The input layer is used to transmit the reconstructed strain time series data to the LSTM layer;
[0116] The LSTM layer is used to extract the strain time series features of the reconstructed strain time series data, and the time series features are used to describe the correlation between the strain fluctuations inside the propeller blade and the sudden load;
[0117] The fully connected output layer is used to predict and output the strain prediction value inside the propeller blade according to the strain time series features;
[0118] In this embodiment, the input layer is responsible for inputting the short-term high-frequency time series data and long-term low-frequency time series data after data cleaning, denoising, and multi-scale decomposition into the lower LSTM layer, and setting the shape of the input data of the LSTM layer. Its shape includes the number of time steps and the number of features. The default number of time steps is set to 1, and the number of features is 1; the LSTM layer captures the time series features in the data through memory units, and at the same time outputs high-dimensional time series features, and expresses the long-term trend and short-term mutations in the input signal by continuously iteratively updating the cell state. Since the propeller rotates in water, its strain change is affected by various sudden situations. Therefore, the LSTM layer can capture the correlation between the strain fluctuations of the propeller and the sudden load, ensuring the tracking consistency of the prediction; at the same time, the number of neurons in the LSTM layer is set according to the complexity of the input data and the length of the time series to enhance the accuracy of the prediction. The fully connected output layer converts the abstract features learned by the neural network into specific strain prediction values, and calculates the difference between the prediction value and the true value through the loss function, so as to backpropagate and optimize the network parameters;
[0119] S53: Based on the online incremental learning mechanism, train the propeller working strain prediction model according to the training set to obtain the trained propeller working strain prediction model;
[0120] In this embodiment, based on historical data, i.e., the training set, a comprehensive model integrating LSTM and PINN is constructed and trained. The network parameters are continuously optimized through backpropagation to ensure the initial accuracy and physical consistency of the propeller strain prediction. Meanwhile, an online incremental learning mechanism is utilized to quickly update the model parameters according to the real-time collected data, ensuring that the model can adapt to the frequently changing working conditions in the marine environment. The online incremental learning mechanism adopts a sliding window mechanism to fine-tune the model parameters based on the latest collected data every fixed time step (10 seconds), ensuring that the model dynamically adapts to the changes in working conditions;
[0121] S54: Based on the loss function with physical constraints, confirm whether the output of the trained propeller working strain prediction model converges according to the test set to evaluate the trained propeller working strain prediction model;
[0122] In this embodiment, since the traditional data-driven LSTM prediction lacks direct constraints on engineering physical mechanisms, it may lead to prediction results that do not conform to basic physical laws. Therefore, the LSTM neural network is fused with the PINN algorithm, and a physical constraint operator in the form of a partial differential equation (PDE) is added to the loss function to construct a loss function with physical constraints, so that the output of the model not only has a fitting effect on the data but also satisfies the basic laws physically;
[0123] Specifically, the method for obtaining the loss function with physical constraints includes the following steps:
[0124] S001: Assume that when the structural deformation of the propeller during underwater operation is negligible, obtain the material mechanics equilibrium equation of static equilibrium as the physical constraint operator of the propeller, and its expression is
[0125]
[0126] In the formula: x represents the position vector, that is, the spatial position of the key monitoring area on the propeller blade; t represents the specific time point when the propeller operates at a certain rotational speed; σ represents the propeller blade stress tensor at the spatial position x and time t; f represents the body force density vector, that is, the additional hydrodynamic force per unit volume of the propeller blade;
[0127] S002: Denote the physical constraint operator as where u represents the predicted value of the strain field by the regional network, and construct a physical consistency loss function, and its expression is
[0128]
[0129] In the formula: Loss phys represents the physical consistency index of the physical consistency loss function; x i ,t iRepresents the data at the \(i\)-th spatio-temporal sampling position point in the reconstructed strain time series data;
[0130] In this embodiment, the training set is gradually passed through the LSTM network layer. After internal state update and feature extraction, it is finally mapped to the final output value through the fully connected layer. The output value is usually the propeller strain distribution at consecutive moments. At the same time, the physical consistency index Loss is calculated according to the defined physical constraint operator. phys , which is used to measure the degree of violation of the physical law by the prediction result; if Loss phys approaches 0, it indicates that the prediction result has good physical rationality; if the value is large, it indicates that there is a physical violation behavior, which can be used to trigger the system warning or the subsequent correction mechanism of the model parameter weights. The final predicted value is the strain field distribution of the propeller monitoring points at consecutive times, which reflects the actual stress state of the blade in the dynamic environment;
[0131] S003: Construct the data fitting loss function of the LSTM neural network, and its expression is
[0132]
[0133] In the formula: Loss data represents the data fitting loss function; represents the predicted value of the output strain time series data of the propeller working strain prediction model; \(y(t)\) represents the true value of the strain time series data;
[0134] In this embodiment, during the training process, the model prediction result will be compared with the actually measured strain data, and the fitting accuracy of the model will be evaluated through statistical indicators such as the mean square error (MSE). When it is found that the physical consistency index significantly deviates from the normal level, the system will automatically enter the online adaptive update process to finely adjust the model parameter weights to adapt to the changes in the current sea conditions or working conditions; at the same time, the abnormal time periods and index information recorded during the abnormal detection process will also be stored for subsequent abnormal analysis, model performance evaluation and feedback optimization, so as to realize the continuous improvement and evolution of the prediction system;
[0135] S004: According to the physical consistency loss function and the data fitting loss function, obtain the loss function with physical constraints, and its expression is
[0136]
[0137] In the formula: Loss total represents the loss function with physical constraints; \(\lambda\) represents the balance coefficient that can adjust the trade-off between data fitting and physical constraints;
[0138] If it is confirmed that the output of the trained propeller working strain prediction model converges, then confirm that the trained propeller working strain prediction model at this time is the optimized propeller working strain prediction model;
[0139] Otherwise, adaptively adjust the parameter weights of the trained propeller working strain prediction model based on the backpropagation method, and repeat step S53;
[0140] S6: Based on the optimized propeller working strain prediction model, realize the prediction of the working strain of the marine composite propeller.
[0141] This embodiment is to make up for the deficiencies of the prior art, and proposes a method for predicting the working strain of a marine composite propeller that combines the LSTM neural network, the PINN algorithm, and the real-time online adaptive anomaly detection technology, that is, a method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN, to solve the problems of low strain prediction accuracy, poor real-time performance, and complex calculation in the prior art. This method not only uses historical data to train the LSTM network to extract strain time series features, but also introduces physical constraint information, and ensures that the model prediction conforms to physical laws such as material mechanics and fluid mechanics through the PINN algorithm. At the same time, a loss function with physical constraints improved based on the PINN algorithm is also introduced to update the model parameters in real time in practical applications, monitor the abnormal signals under sudden working conditions of the marine composite propeller, and ensure the robustness and real-time performance of the prediction system.
[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN, characterized in that, Specifically, it includes the following steps: S1: Based on finite element software, according to the property data of the marine composite propeller, establish a finite element model of the propeller, and perform mesh division on the finite element model of the propeller to obtain the model area mesh of the finite element model of the propeller; S2: Define each area mesh in the model area mesh as a mesh node, and call the static analysis module of the finite element software to obtain the stress / strain distribution point cloud of each mesh node; According to the concentration degree of the stress / strain distribution point cloud in the model area mesh, confirm the key monitoring area of the marine composite propeller, and embed the preset embedded flexible sensor on the maximum thickness line inside the propeller blade corresponding to the key monitoring area of the real propeller. Through the set underwater test system of the propeller, collect and obtain the internal original strain time series data of the propeller blade under different working conditions; S3: Perform data preprocessing on the internal original strain time series data of the propeller blade to obtain optimized strain time series data; S4: Extract the time series data features of different frequency bands in the optimized strain time series data to obtain multi-scale decomposition time series data; And the multi-scale decomposition time series data includes high-frequency time series data and low-frequency time series data; The high-frequency time series data includes instantaneous vibration data and impact data; The low-frequency time series data includes the slow-varying effect caused by temperature and the structural fatigue data of fatigue accumulation; And according to the multi-scale decomposition time series data, reconstruct the optimized strain time series data to obtain the reconstructed strain time series data; S5: Build a propeller working strain prediction model based on the LSTM neural network; And introduce a loss function with physical constraints improved based on the PINN algorithm, and obtain an optimized propeller working strain prediction model according to the reconstructed strain time series data; S6: According to the optimized propeller working strain prediction model, realize the prediction of the working strain of the marine composite propeller.
2. The working strain prediction method for a marine composite propeller based on the fusion of LSTM and PINN according to claim 1, wherein, The specific steps of S3 are as follows: S31: Perform data filling on the internal original strain time series data of the propeller blade to obtain filled time series data, and the expression for data filling is In the formula: represents the filled time series data; x i represents the i-th data point of the internal original strain time series data; N represents the number of data samples of the internal original strain time series data; S32: Remove outliers from the filled time series data, and perform interpolation processing on the data outliers based on the linear interpolation method to obtain optimized time series data; And the method for removing outliers from the filled time series data is specifically Calculate and obtain the Z-score value of each data point in the filled time series data; And the formula for obtaining the Z-score value of each data point is where: x i ' represents a data point in the filled time series data; μ represents the mean of the filled time series data; σ represents the standard deviation of the filled time series data; Preset the Z-score threshold range, and judge whether the Z-score value of each data point meets the Z-score threshold range; If it meets, retain the corresponding data point as a normal data point; Otherwise, delete the corresponding data point as an abnormal data point; S33: Perform standardization processing on the optimized time series data to obtain optimized strain time series data; And the expression for performing standardization processing on the optimized time series data is In the formula: represents the optimized timing data after normalization; x i ″ represents the data points in the optimized strain timing data.
3. A method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN according to claim 2, characterized in that, The expression for reconstructing the optimized strain time series data in S4 is where: x(t) represents the reconstructed optimized strain time series data; a3(t) represents the low-frequency time series data; d j (t) represents the high-frequency time series data and j = 1, 2; d1(t) represents the instantaneous vibration data of the propeller during operation; d2(t) represents the impact data of the propeller during operation.
4. A method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN according to claim 3, characterized in that, The method for obtaining the loss function with physical constraints improved based on the PINN algorithm in S5 specifically includes the following steps: S001: When the structural deformation of the propeller during underwater operation is negligible, obtain the material mechanics equilibrium equation for static equilibrium as the physical constraint operator of the propeller. Its expression is In the formula: x represents the position vector, that is, the spatial position of the key monitoring area on the propeller blade; t represents the specific time point when the propeller operates at a certain rotational speed; σ represents the propeller blade stress tensor at the spatial position x and time t; f represents the body force density vector, that is, the additional hydrodynamic force per unit volume of the propeller blade; S002: Denote the physical constraint operator as where u represents the predicted value of the regional network for the strain field, and construct a physical consistency loss function, and its expression is where: Loss phys denotes the physical consistency index of the physical consistency loss function; x i , t i denotes the data at the i-th spatio-temporal sampling position point in the reconstructed strain time series data; S003: Construct the data fitting loss function of the LSTM neural network. Its expression is Where: Loss data represents the data fitting loss function; represents the predicted value of the output strain time series data of the propeller working strain prediction model; y(t) represents the true value of the strain time series data; S004: According to the physical consistency loss function and the data fitting loss function, obtain the loss function with physical constraints. Its expression is where: Loss total represents a loss function with physical constraints; λ represents a balance coefficient that can adjust the trade-off between data fitting and physical constraints.
5. A method for predicting the working strain of a marine composite propeller based on the fusion of LSTM and PINN according to claim 4, characterized in that, The method for obtaining the optimized propeller working strain prediction model in S5 specifically includes the following steps: S51: Obtain the characteristic data of the propeller blade; And the characteristic data includes any one of the high-frequency time series data or the low-frequency time series data in the reconstructed strain time series data, or the combined data of the high-frequency time series data and the low-frequency time series data; Take the stress inside the propeller blade as the label data, and randomly divide it into a test set and a training set according to the reconstructed strain time series data; S52: Construct a propeller working strain prediction model based on the LSTM neural network; And the propeller working strain prediction model includes an input layer, an LSTM layer, and a fully connected output layer connected in sequence; The input layer is used to transmit the reconstructed strain time series data to the LSTM layer; The LSTM layer is used to extract the strain time series characteristics of the reconstructed strain time series data, and the time series characteristics are used to describe the correlation between the strain fluctuation inside the propeller blade and the sudden load; The fully connected output layer is used to predict and output the strain prediction value inside the propeller blade according to the strain time series characteristics; S53: Based on the online incremental learning mechanism, train the propeller working strain prediction model according to the training set to obtain the trained propeller working strain prediction model; S54: Based on the loss function with physical constraints, confirm whether the output of the trained propeller working strain prediction model converges according to the test set to realize the evaluation of the trained propeller working strain prediction model; If it is confirmed that the output of the trained propeller working strain prediction model converges, then confirm that the trained propeller working strain prediction model at this time is the optimized propeller working strain prediction model; Otherwise, adaptively adjust the parameter weights of the trained propeller working strain prediction model based on the backpropagation method, and repeat step S53.