Method for stress prediction of manned submersible observation window based on machine learning

Abstract

The present application belongs to the technical field of electric digital data processing, and particularly relates to a manned submersible observation window stress prediction method based on machine learning, steps of which comprise: collecting strain and pressure data, and establishing a finite element model of the observation window; constructing a training set based on the strain and pressure data, and then constructing a machine learning hybrid model in correspondence; training the machine learning hybrid model through the training set, and obtaining a stress prediction model; and predicting stress data of the manned submersible observation window under different environmental loads through the stress prediction model. The data-driven stress prediction method of the present application does not depend on a complex physical modeling process, can more quickly and accurately predict stress distribution under complex working conditions, not only can greatly reduce calculation time and resource consumption, but also can provide robust stress prediction results in a relatively complex load environment; reduce calculation time, improve prediction accuracy, and provide efficient and reliable prediction results for structural safety evaluation.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing technology, specifically relating to a method for predicting the stress of a manned submersible observation window based on machine learning. Background Technology

[0002] Deep-sea manned submersibles, as crucial tools for marine scientific research and resource development, require safe operation under extreme high-pressure environments. The observation window, a vital component of the submersible, is typically made of transparent polymethyl methacrylate (PMMA). While providing operators with visibility, it also withstands immense deep-sea pressure. During descent, the observation window experiences significant external loads and stress accumulation due to increased water depth and pressure, particularly the maximum principal stress (MPS). This stress is a critical factor affecting the structural safety of the observation window. When the MPS exceeds the tensile or compressive strength of the PMMA material, the observation window may fail under tensile or compressive stress, threatening the overall safety of the submersible. Therefore, accurately predicting the MPS of the observation window to ensure its ability to withstand extreme external pressure is crucial for guaranteeing the structural integrity of the manned submersible and the safety of its crew.

[0003] In stress prediction for observation windows, traditional methods mainly rely on numerical simulation techniques such as finite element analysis (FEA). Finite element analysis is an effective means of obtaining the overall stress distribution by decomposing the structure into many tiny elements and performing stress, strain, and displacement analysis on each element. Finite element analysis can provide a certain degree of accurate stress distribution of the observation window under high-pressure loads and is widely used in structural design and safety assessment. However, due to the complexity of the working environment of deep-sea manned submersibles and the high nonlinearity and uncertainty of external loads, finite element analysis faces many limitations in practical applications. First, the computational cost of finite element analysis is high, typically requiring significant computational resources and time, making real-time prediction difficult. Second, under complex loading conditions, the adaptability and accuracy of the finite element model may be limited, especially under multiple dynamic loads or irregular loads, where traditional finite element methods struggle to provide stable and accurate prediction results. Therefore, finding more efficient and adaptable stress prediction methods under complex deep-sea conditions has become a research hotspot.

[0004] In summary, traditional analysis methods need to be improved to more quickly and accurately predict the stress distribution of the submersible observation window under complex working conditions, especially the maximum principal stress; reduce computation time and resource consumption; and provide robust stress prediction results in more complex load environments, greatly improving the efficiency and applicability of stress prediction. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by providing a machine learning-based method for predicting the stress of a manned submersible observation window, thereby solving the aforementioned problems.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a machine learning-based method for predicting the stress of a manned submersible observation window, comprising the following steps:

[0008] Strain and pressure data of the manned submersible's observation window were collected, and a finite element model of the manned submersible's observation window was established.

[0009] A training set is constructed based on the strain and pressure data, and a machine learning hybrid model is then constructed based on the training set and the finite element model.

[0010] The machine learning hybrid model is trained using the training set to obtain a stress prediction model;

[0011] The stress prediction model is used to predict the stress data of the manned submersible's observation window under different environments.

[0012] Furthermore, the steps also include:

[0013] A validation set and a test set were also constructed based on the strain and pressure data.

[0014] During the training of the machine learning hybrid model, the model hyperparameters are adjusted using the validation set;

[0015] After obtaining the stress prediction model, the accuracy of the stress prediction model is evaluated using the test set.

[0016] Furthermore, the method for evaluating the accuracy includes:

[0017] Calculate at least one of the following: mean square error, average mean square error, and root mean square residual, based on the predicted stress data and the stress data of the test set.

[0018] Furthermore, when the accuracy falls below a threshold, the machine learning hybrid model is reconstructed, and the training set is augmented with additional data.

[0019] Furthermore, the data augmentation method involves increasing the amount of data collected and / or adding simulation quantities from finite element analysis.

[0020] Furthermore, the data augmentation method involves incorporating simulated quantities from finite element analysis, and the steps include:

[0021] The training set is sorted based on the pressure sequence, and a mapping relationship between the pressure and strain data sets is established to obtain the initial dataset;

[0022] The pressure sequence is divided into high-density and low-density intervals based on a data density threshold.

[0023] The first pressure point is inserted in the high-density region using linear interpolation.

[0024] A second pressure point is inserted in the low-density zone based on the midpoint interpolation method;

[0025] Input the pressure values ​​corresponding to the first and second pressure points into the finite element analysis model to obtain the corresponding strain simulation data;

[0026] The pressure values ​​corresponding to the first and second pressure points are mapped to the strain simulation data and inserted into the initial dataset to obtain the expanded training set.

[0027] Furthermore, the strain data from the manned submersible's observation window is multi-directional strain data;

[0028] The outer surface of the manned submersible's observation window is provided with at least one key point, and biaxial strain data is collected at the key point, with directions of 0° and 90°.

[0029] The manned submersible's observation window has at least one key point on its internal facet, and triaxial strain data is collected at the key point in the directions of 0°, 45° and 90°.

[0030] Furthermore, the architecture of the machine learning hybrid model consists of a position encoding layer, a Transformer layer, a convolutional neural network layer, and a long short-term memory network layer.

[0031] Furthermore, the architecture of the machine learning hybrid model consists of a position encoding layer, a Transformer layer, a CNN layer (convolutional neural network layer), and an LSTM layer (long short-term memory network layer).

[0032] Furthermore, the convolutional neural network layer includes an input layer, a convolutional layer, a pooling layer, and a flattening layer;

[0033] The output of the convolutional neural network layer is the input of the long short-term memory network layer.

[0034] The long short-term memory network layer includes a fully connected layer and an output layer; the output result of the long short-term memory network layer is the maximum principal stress.

[0035] Furthermore, the hyperparameters of the machine learning hybrid model include the learning rate, the number of layers, and the number of neurons;

[0036] The long short-term memory network layer also includes a dropout layer.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This invention provides a new direction for the structural safety assessment of observation windows based on a data-driven stress prediction method. This data-driven method utilizes historical and experimental data, without relying on complex physical modeling processes, to predict stress distribution under complex working conditions more quickly and accurately, especially when dealing with nonlinear and high-dimensional characteristic data. Compared with traditional numerical simulation methods, its hybrid machine learning model not only significantly reduces computation time and resource consumption but also provides robust stress prediction results in complex load environments, greatly improving the efficiency and applicability of stress prediction, reducing computation time, and enhancing model prediction accuracy. This provides efficient and reliable prediction results for the structural safety assessment of manned submersible observation windows under extreme conditions and also provides important technical support for the design, maintenance, and real-time monitoring of deep-sea manned submersibles. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the method of the present invention.

[0041] Figure 2 This is a schematic diagram of the observation window structure in a specific embodiment;

[0042] Figure 3 This is a simulation diagram of the finite element model of the observation window in a specific embodiment;

[0043] Figure 4 This is a schematic diagram showing the pressure change in the observation window in a specific embodiment;

[0044] Figure 5 This is a schematic diagram of strain changes observed in the observation window in a specific embodiment;

[0045] Figure 6This is a schematic diagram of the observation window prediction CNN-LSTM model in a specific embodiment;

[0046] Figure 7 This is a schematic diagram of the observation window prediction Transformer-CNN-LSTM model in a specific embodiment;

[0047] Figure 8 This is a graph showing the variation of the loss value of the CNN-LSTM model in a specific embodiment;

[0048] Figure 9 This is a graph showing the variation of the loss value of the Transformer-CNN-LSTM model in a specific embodiment;

[0049] Figure 10 In a specific embodiment, the CNN-LSTM model predicts the maximum principal stress value and the actual principal stress value;

[0050] Figure 11 In a specific embodiment, the Transformer-CNN-LSTM model predicts the maximum principal stress value and the actual principal stress value;

[0051] Figure 12 This is a comparison chart of the metrics of Transformer-CNN-LSTM and CNN-LSTM in a specific embodiment;

[0052] Figure 13 This is a graph showing the changes in the metrics of Transformer-CNN-LSTM in a specific embodiment;

[0053] Figure 14 This is a graph showing the changes in CNN-LSTM metrics in a specific embodiment. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0056] It should also be noted that, unless otherwise specified, the methods used in this invention are conventional methods; and the raw materials and apparatus used are, unless otherwise specified, conventional commercially available products.

[0057] This embodiment proposes a machine learning-based method for predicting the stress of a manned submersible observation window, such as... Figure 1 As shown, the specific steps include:

[0058] 1. Collect strain and pressure data of the manned submersible's observation window and establish a finite element model of the manned submersible's observation window.

[0059] In this embodiment, Figure 2 The observation window (longitudinal section) of the submersible shown is used as an example for explanation. It includes an outer large surface (circular), an inner small surface (circular), and a lateral conical surface.

[0060] A finite element model of the manned submersible's observation window structure under high-pressure loads was constructed based on its structural dimensions. Furthermore, a finite element model simulating the effect of high-pressure loads on the observation window in a deep-sea environment was established, which can be used not only for subsequent analysis but also for simulation analysis, such as... Figure 3 As shown. Furthermore, the finite element model is constructed based on the material properties, geometry, and boundary conditions of the observation window. The observation window is typically made of polymethyl methacrylate (PMMA), and its material parameters (density, elastic modulus, Poisson's ratio, etc.) directly affect the stress distribution results. Therefore, these material parameters need to be accurately set when building the model. Simultaneously, the geometry (e.g., thickness, curvature) and boundary conditions (e.g., fixation, force direction) of the observation window should also be consistent with the real environment. The finite element model simulates the external high-pressure load under deep-sea conditions to reflect the actual stress situation under deep-sea conditions. Generally, the pressure in the deep-sea environment increases with depth; therefore, an external load conforming to deep-sea pressure is applied to the model to simulate the stress accumulation of the observation window during descent. The applied high-pressure load in the model should conform to the pressure gradient of the deep-sea environment to generate a stress distribution that matches the actual environment. The model is run in finite element analysis software (e.g., Abaqus) for dynamic stress analysis. Through simulation analysis, the preliminary stress distribution of the observation window under high-pressure conditions can be obtained, including the stress concentration at key points and the location of the maximum principal stress. This process provides a preliminary theoretical basis for the structural safety of the observation window and also serves as a reference for subsequent experiments.

[0061] Strain and Pressure Data: Stress-deformation experiments were conducted on the observation window structure under realistic high-pressure conditions. The observation window was placed in an experimental setup simulating the high-pressure environment of the deep sea. This setup gradually applied pressure equivalent to that of the deep sea to simulate the external loads experienced by the observation window at different depths. To obtain high-precision experimental data, the pressure was gradually increased during the experiment, ensuring that the rate of pressure increase matched the actual pressure change rate during descent. Strain and pressure sensors were used to collect strain and pressure change data at key points on the observation window surface, such as… Figure 4 and Figure 5 As shown, the pressure data serves as the initial data source for the model. During the experiment, pressure sensors are used to record pressure changes within the experimental setup in real time, allowing for synchronized analysis with strain data. The acquisition frequency of the pressure sensors is consistent with that of the strain sensors to ensure accurate correspondence between pressure and strain data at each time point. This pressure data reflects the intensity of the external load applied to the observation window, providing crucial load condition information for the model.

[0062] Preferably, the strain data of the manned submersible observation window in this embodiment is multi-directional strain data. Specifically, two key points are provided on the large outer surface of the manned submersible observation window structure, and two-dimensional strain data are collected at each key point, with directions of 0° and 90°, that is, two sets of strain sensors are vertically attached along the surface; thus, this embodiment performs 13,000 acquisitions, forming 4×13,000 data points. Three key points are also provided on the small inner surface of the manned submersible observation window structure, and three-dimensional strain data are collected at each key point, with directions of 0°, 45° and 90°; thus, 9×13,000 data points are formed. Combined with pressure data, the total length of the constructed training set input vector is 13×13,000 and the corresponding 13,000 pressure data points.

[0063] The collected multidimensional data undergoes preprocessing before being fed into the machine learning hybrid model, including data cleaning, normalization, and time-series alignment. Data cleaning removes outliers, normalization ensures consistent numerical ranges when inputting data into the model, and time-series alignment ensures complete correspondence between strain and pressure data at each time point, improving the model's learning performance. This processed data forms a standardized training set, providing clear and effective feature input to the machine learning model. A complete training dataset is created by using multidimensional data from each time point as a training sample. Each sample records the stress distribution of the observation window under specific high-pressure conditions. These training samples reflect how pressure changes affect the stress on the observation window surface, providing comprehensive and accurate data support for stress prediction by the machine learning model.

[0064] 2. Construct a training set based on the initial data (strain and pressure data), and then construct a machine learning hybrid model based on the training set and the finite element model.

[0065] In this embodiment, a hybrid model integrating Transformer, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) network is designed and constructed. The convolutional neural network is as follows: Figure 6 As shown, Long Short-Term Memory (LSTM) networks are as follows: Figure 7As shown, the machine learning hybrid model also incorporates a positional encoding layer. Specifically, the hybrid model includes a positional encoding layer, a Transformer layer, a convolutional neural network layer, and a long short-term memory network layer. The positional encoding layer enhances the model's sensitivity to the temporal features of stress data, while the self-attention mechanism learns patterns in the sequence by calculating the correlation between each position and other positions.

[0066] Optionally, the Transformer layer can be configured as follows:

[0067] 2.1 The Transformer layer's self-attention mechanism allows each input vector to "pay attention" to other vectors in the sequence. The formula for self-attention is as follows:

[0068] First, the input sequence strain data and pressure data are mapped to a query ( Q ),key( K ) and value ( V )matrix:

[0069] ;

[0070] In the formula, and It is a trainable weight matrix.

[0071] Calculate the dot product of the query and the key, and scale the result as follows:

[0072] ;

[0073] In the formula, It is a scaling factor to prevent gradient vanishing or exploding due to excessively large values. Softmax is used to normalize the result of the dot product.

[0074] Multi-head attention extends the self-attention mechanism to multiple heads:

[0075] ;

[0076] Each of them ,and It is the output projection matrix.

[0077] 2.2 Transformer Layer Position Encoding: Since the Transformer does not have a sequence structure, positional encoding is needed to incorporate sequence order information. The formula for the positional encoding vector is as follows:

[0078] Given position pos and dimensions i:

[0079] ;

[0080] In the formula, pos Indicates the position of an element in the sequence. i This represents the first dimension in the encoding. i Indexes. d This refers to the dimension of the encoding. Positional encoding uses sine and cosine functions to ensure the uniqueness of the encoding between different positions and to maintain information about the relative distance between them. This positional encoding, applied on a fixed dimension, helps the Transformer capture positional information within a sequence.

[0081] This enables the model to automatically capture the important relationships between different positions in the sequence. Since the Transformer model generally performs poorly in predicting stress, we considered combining the advantages of Transformer position encoding, bidirectional LSTM, and one-dimensional convolutional neural networks to enable it to provide high-precision stress prediction in real time under high-pressure environments.

[0082] This embodiment uses a CNN to extract features from the dataset for processing the spatial features of multi-dimensional data. The CNN can perform convolution operations on triaxial strain and pressure data to extract local spatial relationships and patterns, helping the model understand the local features of stress distribution at different observation points. These features are then input into an LSTM for prediction. The CNN in this embodiment mainly includes an input layer, convolutional layers, pooling layers, and a flattening layer. The output of the CNN primarily serves as the input to the LSTM. The 1D convolutional and pooling layers are used to extract the spatial features of the input data. The flattening layer is placed between the CNN and the LSTM to convert the CNN features into a one-dimensional array as the input to the LSTM.

[0083] In this embodiment, an LSTM layer is used to extract long-term dependency features of the time series. Because stress data has a temporal sequence, the LSTM captures the stress evolution trend and pattern during the pressure change process by memorizing and processing historical information.

[0084] The spatial features extracted by CNN, the temporal dependency features extracted by LSTM, and the global dependency features extracted by Transformer are fused to form a complete feature vector. These features can comprehensively represent the multidimensional information in the stress data, ensuring that the model can capture both local and global stress features.

[0085] Preferably, a validation set is also constructed based on strain and pressure data. During the training process of the machine learning hybrid model, the model's hyperparameters are adjusted using the validation set. Hyperparameter adjustment includes the learning rate, number of layers, and number of neurons to ensure optimal performance of the model in maximum principal stress prediction. The performance of different network structures in stress prediction is compared, and the network structure with higher prediction accuracy is selected. This embodiment also sets three sets of one-dimensional convolution and pooling operations, followed by a flattening operation for output prediction on two LSTM layers and the output layer. Each LSTM layer has a dropout layer to prevent overfitting, with the dropout value set to 0.5. The model's hyperparameter settings are as follows:

[0086] The learning rate (which controls the step size of model weight updates, and uses a learning rate scheduler and the adaptive optimization algorithm Adam to dynamically adjust the learning rate to speed up convergence) is set to 0.001;

[0087] The number of neurons between the convolutional layer and the pooling layer is set to 64;

[0088] The kernel size and pooling size are set to 1;

[0089] Both LSTM neurons have 100 neurons;

[0090] ReLU is used as the activation function.

[0091] In addition, to prevent gradient explosion during training, gradient clipping is added before training. The maximum range of the gradient to be clipped is 1, and the gradient is clipped before the optimization step is performed.

[0092] Thus, the maximum principal stress is predicted for the fully connected layer and the output layer of the LSTM.

[0093] Finally, the machine learning hybrid model is trained using the aforementioned training set, with the goal of obtaining the maximum principal stress, and a stress prediction model is obtained after training.

[0094] Optionally, during model training, different combinations of model architectures can be tried and compared. For example, CNN-LSTM and Transformer-CNN-LSTM can be compared, and cross-validation can be used to evaluate the prediction performance of each architecture to find the optimal model architecture. Specifically, the comparison results in this embodiment are as follows:

[0095] pass Figure 8 and Figure 9The graphs showing the changes in loss values ​​for the CNN-LSTM and Transformer-CNN-LSTM models reveal that the Transformer-CNN-LSTM model exhibits a more stable loss value change compared to the CNN-LSTM model. This indicates that the Transformer-CNN-LSTM model has a better learning effect and greater stability in predicting the maximum principal stress value.

[0096] Figure 10 In this embodiment, the CNN-LSTM model predicts the maximum principal stress and the actual principal stress. Figure 11 The figures show the predicted maximum principal stress and the actual principal stress values ​​for the Transformer-CNN-LSTM model in this embodiment. As can be seen from these two figures, both models perform well in predicting the maximum principal stress. However, the Transformer-CNN-LSTM model is significantly more accurate in predicting the high points of stress concentration within the observation window. Therefore, the Transformer-CNN-LSTM model has higher prediction accuracy than the CNN-LSTM model.

[0097] Figure 12 This is a comparison chart of the metrics between Transformer-CNN-LSTM and CNN-LSTM in this embodiment. The Transformer-CNN-LSTM model performs the best, specifically in that it has the smallest MSE, MAE, and RSR values ​​among the CNN-LSTM models. The smaller the MSE and MAE values, the better, and the closer the RSR value is to 0, the better.

[0098] Figure 13 This is a graph showing the changes in the metrics of Transformer-CNN-LSTM in this embodiment. Figure 14 The graph shows the changes in CNN-LSTM metrics in this embodiment. As can be seen from the graph, the evaluation metrics of the improved Transformer-CNN-LSTM model have been improved, and all evaluation metrics have been reduced.

[0099] As shown above, the deep learning algorithm verified the feasibility of machine learning methods in predicting the stress structure of the observation window. The improved and optimized Transformer-CNN-LSTM algorithm enhanced the learning effect and showed high consistency with the experimental data.

[0100] 3. Model performance evaluation.

[0101] Optionally, in order to effectively predict stress values, this embodiment also adds a model evaluation method, the specific steps of which are as follows:

[0102] A test set was also constructed based on strain and pressure data;

[0103] After obtaining the stress prediction model, the accuracy of the stress prediction model is evaluated using a test set.

[0104] Specifically, in this embodiment, when evaluating model performance based on predicted stress data and stress data from the test set, the deviation and distribution similarity between predicted and true values ​​should be considered. Therefore, this embodiment uses mean squared error (MSE), average mean squared error (MAE), and root mean square residual (RSR) to compare the deviation between true and predicted values ​​of different models and evaluate model performance. The formulas involved are as follows:

[0105] ;

[0106] ;

[0107] ;

[0108] In the formula, N It refers to the number of values. It is the actual value. It is a predicted value. It is the average of the true values. The mean squared error (MSE) and the average mean squared error (MAE) are generally better the smaller they are, and the root mean square residual (RSR) is even better if it is close to 0.

[0109] The above evaluation metrics are collectively defined as accuracy. A corresponding threshold is set for each metric according to the engineering needs. When any metric falls below the threshold, the above steps are repeated, including reconstructing the machine learning hybrid model and expanding the training set.

[0110] Alternatively, data augmentation methods include increasing the amount of data collected and adding simulation parameters for finite element analysis.

[0111] Considering the difficulty and long preparation period of obtaining actual measured data, this embodiment selects the method of adding simulated quantities from finite element analysis for data augmentation. The specific steps include:

[0112] First, since each set of data in the dataset contains the corresponding pressure, the pressure is arranged in ascending order. The training set is then sorted based on the pressure sequence, and a mapping relationship between pressure and strain data sets is established to obtain the initial dataset.

[0113] Then, the pressure sequence (a one-dimensional array of pressures) is divided into several intervals of equal length (intervals with the same pressure interval, such as 5 MPa as one interval length, then 10 MPa-15 MPa as one interval). Based on the data density threshold, the intervals are further divided into high-density intervals and low-density intervals. If there are a pressure points in interval n, and a is greater than the data density threshold k, then this interval is a high-density interval. If there are b pressure points in interval m, and b is less than the data density threshold k, then this interval is a low-density interval.

[0114] To further increase the simulation data, it is necessary to determine the corresponding pressure points. Therefore, different pressure point insertion rules are set according to different intervals, as follows:

[0115] (1) Insert the first pressure point in the high-density interval based on linear interpolation. Since the data sampling in the measured data is not completely linear, there are such high-density intervals with a large number of pressure points and dense distribution. The training set generated in this way can achieve better training results. Even if the simulation results of the expanded data are somewhat distorted, they will generally be corrected by training under the influence of the measured data. Therefore, a simple linear interpolation method can be used to insert several first pressure points in the interval.

[0116] (2) Inserting a second pressure point in a low-density interval using midpoint interpolation. Since there are few and sparsely distributed pressure points in such low-density intervals, using a general linear interpolation method will result in the second pressure point being close to the measured pressure point. If the strain variable in the simulation analysis is distorted, the closer erroneous data in the neural network will cause the weight update to be biased. Since the model will adjust the weights according to the loss function of the samples during training, distorted data will cause the loss function to generate an incorrect gradient signal. Especially during small-batch training, if the sample containing erroneous data is grouped with other correct samples, the direction of gradient descent may be "pulled off" by the erroneous sample. For example, when calculating the mean square error loss between the strain prediction value and the true value, the model will try to reduce this loss, so that the weight update will be made in the direction of adapting to the erroneous sample, resulting in a decrease in the prediction accuracy on the correct samples. Therefore, to perform interpolation in this interval, the inserted second pressure point should be as far away from the measured pressure point as possible. The ideal way is the special midpoint interpolation method in linear interpolation, that is, each second pressure point is equidistant from the two adjacent measured pressure points.

[0117] The pressures corresponding to the first and second pressure points are sequentially input into the constructed finite element analysis model to obtain the corresponding strain simulation data.

[0118] The pressure and strain simulation data corresponding to the first and second pressure points are then mapped to each other, and the data is inserted back into the initial dataset after the time is assigned. The time assignment can be calculated by referring to the interpolation points, and will not be elaborated further here. After data insertion, an expanded dataset containing both real-world collected data and finite element simulation data is obtained, forming a new training set to improve the accuracy of the stress prediction model.

[0119] 4. Output the results.

[0120] Based on actual needs, environmental parameters such as pressure are substituted into the stress prediction model to predict the stress data of the observation window of the manned submersible under different environments, and this can be used for further structural strength research or to verify the observation window design scheme.

[0121] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A method for predicting stress in the observation window of a manned submersible based on machine learning, characterized in that the steps include... include: Strain and pressure data of the observation window of the manned submersible were collected, and stress-deformation experiments of the observation window structure were conducted. The observation window was placed in an experimental device simulating the high-pressure environment of the deep sea. During the experiment, the pressure was gradually increased, and the rate of pressure increase was ensured to match the rate of pressure change during the actual descent. Strain and pressure data of key points on the surface of the observation window were collected using strain and pressure sensors. Establish a finite element model of the observation window of the manned submersible; A training set is constructed based on the strain and pressure data, and then the training set is expanded based on the simulated insertion data of the finite element model, and a machine learning hybrid model is constructed. The architecture of the machine learning hybrid model consists of a position encoding layer, a Transformer layer, a convolutional neural network layer, and a long short-term memory network layer in sequence. The machine learning hybrid model is trained using the training set to obtain a stress prediction model; Data needs to be preprocessed before it is fed into a machine learning hybrid model, including data cleaning, normalization and time alignment; The stress prediction model is used to predict the stress data of the manned submersible's observation window under different environments.

2. The method for predicting stress in the observation window of a manned submersible based on machine learning according to claim 1, characterized in that, The steps also include: A validation set and a test set were also constructed based on the strain and pressure data. During the training of the machine learning hybrid model, the model hyperparameters are adjusted using the validation set; After obtaining the stress prediction model, the accuracy of the stress prediction model is evaluated using the test set.

3. The method for predicting stress in the observation window of a manned submersible based on machine learning according to claim 2, characterized in that, The methods for evaluating the accuracy include: Calculate at least one of the following: mean square error, average mean square error, and root mean square residual, based on the predicted stress data and the stress data of the test set.

4. The method for predicting stress in a manned submersible observation window based on machine learning according to claim 2 or 3, characterized in that, When the accuracy rate is below the threshold, the machine learning hybrid model is reconstructed, and the training set is augmented with additional data.

5. The method for predicting stress in the observation window of a manned submersible based on machine learning according to claim 4, characterized in that, The data augmentation method involves increasing the amount of data collected and / or adding simulation data from finite element analysis.

6. The method for predicting stress in the observation window of a manned submersible based on machine learning according to claim 5, characterized in that, The data augmentation method involves incorporating simulated quantities from finite element analysis, and the steps include: The training set is sorted based on the pressure sequence, and a mapping relationship between the pressure and strain data sets is established to obtain the initial dataset; The pressure sequence is divided into high-density and low-density intervals based on a data density threshold. The first pressure point is inserted in the high-density region using linear interpolation. A second pressure point is inserted in the low-density zone based on the midpoint interpolation method; Input the pressure values ​​corresponding to the first and second pressure points into the finite element analysis model to obtain the corresponding strain simulation data; The pressure values ​​corresponding to the first and second pressure points are mapped to the strain simulation data and inserted into the initial dataset to obtain the expanded training set.

7. The method for predicting stress in the observation window of a manned submersible based on machine learning according to claim 1, characterized in that, The strain data from the observation window of the manned submersible are multi-directional strain data; The outer surface of the manned submersible's observation window is provided with at least one key point, and biaxial strain data is collected at the key point, with directions of 0° and 90°. The manned submersible's observation window has at least one key point on its internal facet, and triaxial strain data is collected at the key point in the directions of 0°, 45° and 90°.

8. The method for predicting stress in the observation window of a manned submersible based on machine learning according to claim 1, characterized in that, The convolutional neural network layer includes an input layer, a convolutional layer, a pooling layer, and a flattening layer; The output of the convolutional neural network layer is the input of the long short-term memory network layer. The long short-term memory network layer includes a fully connected layer and an output layer; the output result of the long short-term memory network layer is the maximum principal stress.

9. The method for predicting stress in a manned submersible observation window based on machine learning according to claim 8, characterized in that, The hyperparameters of the machine learning hybrid model include the learning rate, the number of layers, and the number of neurons; The long short-term memory network layer also includes a dropout layer.

Images