Satellite structure state intelligent evaluation method based on optical fiber sensor

Abstract

This invention relates to the field of on-orbit satellite structural condition monitoring technology, specifically disclosing an intelligent assessment method for satellite structural condition based on fiber optic sensors. The method includes: Step S1, conducting ablation simulations of the rubber material on the surface of a flexible spacecraft to obtain different damage states and corresponding strain data; Step S2, training a time recurrent neural network (LSTM) using the different damage states and corresponding strain data to obtain a classification prediction model; Step S3, inputting the data to be predicted into the classification prediction model to obtain the predicted damage state output by the model. This invention achieves intelligent assessment of laser ablation damage modes of flexible satellite surface materials through fiber Bragg grating sensing networks and long short-term memory (LSTM) network technology.

Description

Technical Field

[0001] This invention relates to the field of on-orbit satellite structural status monitoring technology, and specifically to an intelligent assessment method for satellite structural status based on fiber optic sensors. Background Technology

[0002] To address the current limitations in damage monitoring and sensing capabilities for flexible spacecraft surface materials, the following technologies have been developed both domestically and internationally: 1) Monitoring technology based on fiber Bragg grating sensor networks. The harsh environment of space and the limitations of space launch vehicles place stringent requirements on spacecraft structural condition monitoring systems. FBG technology, through optical principles, overcomes the technical challenges of traditional electronic sensors in terms of size, interference resistance, distributed measurement, and environmental adaptability, becoming the most promising technical solution for spacecraft structural condition monitoring.

[0003] 2) Spacecraft damage detection technology. Spacecraft damage detection technology is crucial for ensuring the safe operation of spacecraft. Significant progress has been made in research both domestically and internationally, showing a trend towards intelligence and multimodal fusion. However, challenges remain, such as insufficient sensitivity in detecting defects in complex structures, low efficiency in fusing multi-source data, and poor adaptability to extreme environments.

[0004] 3) Damage detection technology for flexible spacecraft. Currently, there are few reports on research on the structural state perception and assessment of flexible spacecraft both domestically and internationally, and relevant literature and data are scarce.

[0005] Currently, flexible materials are widely used in satellite structural systems due to their lightweight, low cost, and variability. With the development of aerospace technology, there is an urgent need for more precise and reliable structural health monitoring technologies for flexible structures to ensure the stable operation of satellites in orbit. Summary of the Invention

[0006] To address the aforementioned problems, the purpose of this invention is to provide an intelligent assessment method for satellite structural status based on fiber optic sensors. This method utilizes fiber Bragg grating sensing networks and long short-term memory network technology to achieve intelligent assessment of laser ablation damage modes on the surface materials of flexible satellite structures.

[0007] This invention provides a method for intelligent assessment of satellite structural status based on fiber optic sensors, comprising: Step S1: Taking the rubber material on the surface of the flexible spacecraft as the object, conduct ablation simulation of the surface material of the flexible spacecraft to obtain different damage states and corresponding strain data. Step S2: Train the time recurrent neural network LSTM with the different damage states and corresponding strain data to obtain a classification prediction model; Step S3: Input the data to be predicted into the classification prediction model to obtain the predicted damage state output by the classification prediction model.

[0008] Step S1 includes: The temperature field induced by a laser heat source on the surface of a rubber material is simulated, and a temperature field cloud map is generated. Based on the temperature field cloud map, data acquisition points around the ablation point are selected; By changing the duration of the laser heat source during the simulation, the damage state of the rubber material under different temperature fields and the strain data at the data acquisition points are obtained.

[0009] In one possible implementation, step S2 includes: The preprocessed damage states and corresponding strain data are input into a time recurrent neural network LSTM. The input gate, forget gate, and output gate of the time recurrent neural network LSTM work together to output the predicted damage state.

[0010] In one possible implementation, the error between the predicted damage state and the actual label is calculated using the loss function of the classification prediction model; When the error reaches a preset threshold or the training rounds reach an upper limit, the final predicted damage state is output.

[0011] In one possible implementation, when the error does not reach a preset threshold and the number of training rounds does not reach the upper limit, the error is propagated backward along the network using the backpropagation algorithm. The weights and biases of the network are updated according to the gradient descent method, the network parameters are adjusted, and the training process is repeated until the error reaches the preset threshold or the number of training rounds reaches the upper limit.

[0012] In one possible implementation, the network parameters include at least: the number of hidden layer units, the learning rate, the number of training epochs, the weights and biases of the initialized network.

[0013] In one possible implementation, step S2 further includes: The different damage states and corresponding strain data are converted into cell arrays, and the cell arrays are input into the temporal recurrent neural network LSTM for training.

[0014] In one possible implementation, the classification prediction model includes an input layer, an LSTM layer, a fully connected layer, and an output layer.

[0015] In one possible implementation, the fully connected layer includes a leakyReluLayer function, a dropoutLayer function, and a ReLU activation function.

[0016] In one possible implementation, the output layer includes a softmaxLayer function and a classificationLayer function.

[0017] In one possible implementation, the damage state includes: deformation, pitting, and perforation.

[0018] The present invention provides a satellite structural state intelligent assessment method based on fiber optic sensors. It constructs a fiber optic grating sensor network to sense damage to the surface materials of flexible spacecraft, and uses an intelligent assessment method to convert the signals sensed by the sensors into structural damage modes, allowing the spacecraft to judge its own structural state based on the damage modes. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the intelligent assessment method for satellite structural status provided in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following detailed description of the embodiments and the accompanying drawings are used to illustrate the principles of the present invention by way of example, but should not be used to limit the scope of the present invention. That is, the present invention is not limited to the described preferred embodiments, and the scope of the present invention is defined by the claims.

[0021] In the description of this invention, it should be noted that, unless otherwise stated, "a plurality of" means two or more; the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance; those skilled in the art can understand the specific meaning of the above terms in this invention as appropriate.

[0022] Figure 1 A flowchart illustrating the intelligent satellite structural status assessment method provided for embodiments of the present invention is shown below. Figure 1 As shown, this invention provides a method for intelligent assessment of satellite structural status based on fiber optic sensors, comprising: Step S1: Taking the rubber material on the surface of the flexible spacecraft as the object, conduct ablation simulation of the surface material of the flexible spacecraft to obtain different damage states and corresponding strain data. In one possible implementation, the temperature field induced by the laser heat source on the surface of the rubber material is simulated to generate a temperature field cloud map; based on the temperature field cloud map, data acquisition points around the ablation point are selected; by changing the duration of the laser heat source during the simulation, the damage state of the rubber material under different temperature fields and the strain data of the data acquisition points are obtained.

[0023] Damage states include: deformation, pitting, and perforation.

[0024] Step S2: Train the time recurrent neural network LSTM using different damage states and corresponding strain data to obtain a classification prediction model; In one possible implementation, the preprocessed different damage states and corresponding strain data are input into a temporal recurrent neural network (LSTM); the LSTM's input gate, forget gate, and output gate work together to output the predicted damage state.

[0025] In one possible implementation, different damage states and corresponding strain data are converted into cell arrays, and the cell arrays are input into a time recurrent neural network LSTM for training.

[0026] Step S3: Input the data to be predicted into the classification prediction model to obtain the predicted damage state output by the classification prediction model.

[0027] In one possible implementation, the error between the predicted damage state and the actual label is calculated using the loss function of the classification prediction model. When the error reaches a preset threshold or the training epochs reach their limit, the final predicted damage state is output. When the error does not reach the preset threshold and the training epochs do not reach their limit, the error is propagated backward along the network using the backpropagation algorithm. The network weights and biases are updated using gradient descent, the network parameters are adjusted, and the training process is repeated until the error reaches the preset threshold or the training epochs reach their limit.

[0028] In one possible implementation, the network parameters include at least: the number of hidden layer units, the learning rate, the number of training epochs, the weights and biases used to initialize the network.

[0029] In one possible implementation, the classification prediction model includes an input layer, an LSTM layer, a fully connected layer, and an output layer. The fully connected layer includes the leakyReluLayer function, the dropoutLayer function, and the ReLU activation function. The output layer includes the softmaxLayer function and the classificationLayer function.

[0030] To facilitate understanding, the present invention will be further explained and illustrated below with reference to specific embodiments.

[0031] This invention focuses on flexible spacecraft surface rubber materials and uses the finite element analysis software ABAQUS to simulate the laser ablation process. The simulation consists of two steps: the first step simulates the temperature field induced by the laser heat source on the surface of the rubber material, and the second step realizes the ablation effect of the rubber material under the action of the temperature field.

[0032] The first step involves constructing a square solid component with sides of 100mm and a thickness of 5mm. This component is assigned parameters identical to the rubber material, including thermal conductivity, density, elastic modulus, and coefficient of thermal expansion. These material parameters are then assigned to the cross-section and the square solid component, completing the assembly process. Next, a transient heat transfer analysis step is defined, with a step duration of 40 seconds. This step duration determines the monitoring time window for the rubber material. In the load setting stage, a user subroutine, DFLUX, is edited using Visual Studio. Parameters such as Gaussian heat source power, effective radius, and application time are set within the subroutine. This subroutine serves as the input for the user-defined load, applying a laser heat source to the rubber material surface. Additionally, a predefined environmental field is set to define the initial ambient temperature of the material. The component is then meshed, optimizing computational efficiency while ensuring accuracy by selecting a linear heat transfer mesh type. Finally, the user subroutine DFLUX is appended to the job directory and the job is submitted. The job outputs the temperature field generated by the laser acting on the rubber material, completing the first step of the laser heat source simulation and generating a temperature field contour map. The temperature field follows a Gaussian heat source distribution. The laser heat source input causes a strong material response in the region near the ablation point, with the unit temperature value being significantly higher than the surrounding area. At this point, the origin temperature of the laser ablation region of the rubber material is the highest, at 431.4℃.

[0033] The second step involves replicating the solid component model established in the first step and changing the analysis step to "Dynamic, Explicit". To ensure the material remains fixed during ablation, six-degree-of-freedom constraints are applied to the four sides of the square component for complete fixation. Then, the laser temperature field generated in the first step is applied frame-by-frame to the material surface as a predefined field to realize the effect of the laser heat source on the component, and the component mesh properties are adjusted to a hexahedral three-dimensional stress type. Finally, the second step is submitted, importing a VUSDFLD user subroutine to update the state variables during ablation and obtain the real-time temperature values ​​of all material points. Based on the physicochemical properties of the rubber material, its melting point is set to 200℃, and the subroutine is configured to: when the component temperature exceeds 200℃, the state variables are reset to zero to mark material failure, thereby simulating the material ablation and melting process. The second step yields a damage pattern diagram of the laser acting on the rubber material.

[0034] From the perspective of stress distribution on the material surface, annular pits appear on the surface at this point, with high-strain areas concentrated at the edges of the pits, consistent with the boundary where laser ablation causes material melting and failure. From the perspective of temperature distribution within the material surface units, the material exhibits a concentric gradient cooling pattern, with the color changing from red to orange to yellow to green to blue from the ablation point center outwards. The highest temperature at the center reaches 242.9℃, while the temperature at the material edge is the same as the ambient temperature at 25℃. The temperature gradually decreases from the center outwards, reflecting the diffusion process of heat conduction.

[0035] After obtaining the cloud map of the laser heat source acting on the rubber material, this invention analyzes the damage state of the rubber material under different temperature fields by changing the action time of the laser heat source during the simulation process.

[0036] Adjusting the analysis step time and the DFLUX subroutine, the laser power was set to 20W and the action time to 5 seconds. At this point, annular pits appeared on the material surface, indicating that some elements had reached the material's melting point, leading to failure and removal. Looking at the element temperature distribution, the highest temperature at the laser ablation point of the rubber material reached 98℃, gradually decreasing from the center outwards until it matched the ambient temperature. From the strain distribution, high-strain areas (such as the red and yellow areas) were concentrated at the pit edges and the center of the ablation point, indicating that these areas underwent significant plastic deformation, consistent with the boundary where laser ablation caused material melting failure. This pit shape characteristic reflects that during laser ablation of the rubber material, the concentrated heat caused the material in the central region to reach its melting point first, while simultaneously causing plastic flow and deformation of the surrounding material under thermal stress and mechanical action, leading to increased stress and strain in adjacent elements, forming the high-strain pit morphology of the central and edge elements.

[0037] The analysis step time was adjusted to 15 seconds, while the laser power remained constant at 20W. At this point, the annular pits on the material surface deepened with increasing laser exposure time, and the radius of the deformed pits also increased. The highest temperature at the laser ablation point of the rubber material rose to 213℃, gradually decreasing from the center outwards until it reached the ambient temperature, with the area of ​​temperature increase expanding. The strain values ​​of the elements in the center and edge regions of the pits were higher, and the maximum strain value was also greater than when the laser exposure time was 5 seconds.

[0038] The analysis step time was adjusted to 25 seconds, while the laser power remained constant at 20W. At this point, the annular pits on the material surface deepened with increasing laser exposure time, reaching a depth exceeding the material thickness, thus forming perforation damage on the material surface. The highest temperature at the laser ablation point of the rubber material rose to 246℃, gradually decreasing from the center outwards until it reached the ambient temperature. Due to the effects of material removal and the temperature field, the strain values ​​of the elements around the perforation area and at the edge of the pit were relatively high.

[0039] From the perspective of unit temperature distribution, the longer the laser irradiation time, the higher the temperature within the ablation point region. Regarding shape changes, as the laser irradiation time increases, the material progresses from deformation to the formation of pits; the depth and radius of these pits continuously increase until perforation occurs, and the shape of the ablation point region becomes increasingly irregular. In terms of strain changes, the unit strain values ​​are higher in the central region of the ablation point and at the edge of the pits, while the area experiencing strain changes expands, and the strain values ​​also increase.

[0040] Based on the wavelength shift measurement principle of fiber Bragg grating sensors, this invention extracts strain data from simulation results. Training an LSTM classification prediction model requires obtaining the damage signal and corresponding damage mode at the ablation point. The simulation should be consistent with the experiment; therefore, strain data near the ablation point is used as the ablation damage signal, and strain data and corresponding damage modes are obtained from the finite element numerical simulation results. When extracting simulation results, the number of simulation elements is too large to obtain strain data for all elements. By observing the simulation contour plot, eight points with relatively drastic strain changes near the ablation point are selected as data acquisition points, numbered 1 to 8.

[0041] After determining the data acquisition points, the simulation was started to simulate the damage state and strain changes of the component under different laser irradiation times within a 40-second monitoring time window.

[0042] The strain changes at eight sampling points under different damage conditions. The horizontal axis represents the monitoring time window from 1 to 40 seconds, and the vertical axis represents the logarithmic strain at the sampling point. From the changes, all curves show a fluctuating upward trend. The strain changes are more drastic during the laser treatment period. The strain changes are also related to the distribution of the sampling points. The closer to the center of the ablation point, the greater the influence of material reduction and flaking, and the more drastic the strain changes.

[0043] LSTM (Long Short-Term Memory) is a type of temporal recurrent neural network suitable for processing and predicting important events with relatively long intervals and delays in time series. The strain data acquired at the sampling points increases non-linearly over time. For such long-term time series signals, the LSTM model can effectively memorize and process the data, enabling the classification of damage patterns.

[0044] During the simulation, it was found that the strain and temperature data acquired at the sampling points increased nonlinearly with time within a 40-second time window. For such long-term time series signals, the LSTM model can be used to fit and classify the time series signals.

[0045] The basic process of LSTM classification prediction involves inputting preprocessed data into an LSTM neural network, dividing it into training and test sets, setting network parameters including the number of hidden layer units, learning rate, training epochs, and initializing the network weights and biases. As data flows through the LSTM units, the input gate, forget gate, and output gate work together to selectively retain and update information. The hidden state at each time step is passed between units and participates in the computation, ultimately yielding the predicted output. The predicted output is compared with the actual labels, and a loss function (such as cross-entropy loss) is calculated to obtain the error value. If the error reaches a preset threshold or the training epochs reach their maximum limit, the prediction result is output; otherwise, the error is propagated back along the network using backpropagation, and the network weights and biases are updated using gradient descent. The network parameters are adjusted, and the training process is repeated until the error meets the conditions, completing the training and prediction task of the LSTM classification prediction model.

[0046] This invention constructs an LSTM-based classification prediction model in MATLAB software.

[0047] The first step is to prepare the input data. A total of 34 samples were collected for the laser ablation damage process. The samples were labeled 1, 2, and 3, corresponding to three damage modes: deformation, pitting, and perforation. Each sample contained strain and temperature time-series data for 8 monitoring points within a 40-second time window. To classify such multiple time-series data, a cell array was used to store the input data.

[0048] The second step is to design the LSTM classification model network structure. In the input layer, the `sequenceInputLayer` function is used to define the input feature dimensions, with 8 data collection points. `inputSize=8` indicates that the model receives sequence features from 8 dimensions as input. In the LSTM layer, the `lstmLayer` function is used to extract temporal features. `numHiddenUnits` is used to set the number of hidden units to capture temporal dependencies in the sequence. An appropriate number of neurons makes the prediction results more accurate; too many hidden units can lead to overfitting, while too few result in poor model data fitting. This invention explores and determines the most suitable number of neurons during model training. The output mode of the `lstmLayer` function is set to `last`, indicating that only the hidden state at the last moment of the sequence is output for time series classification. Then, the fully connected layer is set. The `fullyConnectedLayer` maps the LSTM output to the number of classification categories. This invention classifies damage patterns into three categories, therefore, the fully connected layer is set to 3. The fully connected layers use the leakyReluLayer and dropoutLayer functions, introducing a ReLU activation function with a leakage parameter of 0.2 to alleviate the gradient vanishing problem, and randomly dropping neurons with a 30% probability to prevent overfitting. Finally, the output layer uses softmaxLayer and classificationLayer to achieve probabilistic classification. When dealing with complex feature relationships, it is necessary to enhance the model's non-linear expressive power. ReLU activation layers can be added between the fully connected layers. The Sigmoid activation function, through a "soft-gated" signal with an output range of (0, 1), can control the memory gate, determining the proportion of information to be retained or forgotten from the previous time step's memory cell state; it controls the input gate, determining how much new information from the current input needs to be incorporated into the memory cell state for updating long-term memory; and it controls the output gate, determining how much information from the current cell state needs to be output to the hidden state for use in subsequent time steps. The Sigmoid function expression is as follows: The third step is configuring the training parameters. The optimizer chosen is the 'adam' algorithm, which has an adaptive learning rate and fast convergence. MaxEpochs is set to adjust the number of training epochs and control the training depth. MiniBatchSize is used to adjust the batch size to balance training speed and stability. The SequenceLength function is used to automatically handle the longest sequence and pad shorter sequences with zeros. Shuffle is enabled, randomly shuffling the data every epoch to prevent overfitting.

[0049] The fourth step is model training. The `trainNetwork` function is used to train the model in conjunction with the defined network layers and training options.

[0050] The fifth step is model classification prediction and performance evaluation. The `classify` function is used to predict the class of the test data, and the accuracy of the model's classification is verified by comparing the predicted results with the label values.

[0051] By using a constructed neural network model LSTM and inputting simulation data, the model is trained to classify and identify strain time-series data under different damage modes. The specific steps mainly include: inputting and processing simulation data, and adjusting network parameter configurations.

[0052] Data input: Taking strain data as an example, some strain time series data collected from 8 monitoring points when the laser is applied for 1 second are stored in an Excel spreadsheet for easy model training.

[0053] This sample dataset contains eight time series of equal length. When the laser exposure time increases from 0 seconds to 1.8 seconds, the material deforms. By adjusting the laser exposure time from 0 to 1.8 seconds, and taking strain data at 0.2-second intervals, nine sample data points under the deformation damage state can be obtained, and these data files are named z1 to z9. When the laser exposure time increases from 2 seconds to 20 seconds, the material develops pits. By adjusting the laser exposure time from 2 to 20 seconds, and taking strain data at 1-second intervals, 19 sample data points under the pitting damage state can be obtained, and these data files are named x1 to x19. When the laser exposure time increases from 21 seconds to 40 seconds, the material perforates. By adjusting the laser exposure time from 21 to 26 seconds, and taking strain data at 1-second intervals, six sample data points under the perforation damage state can be obtained, and these data files are named j1 to j6. All data files are saved in .csv format. This invention defines the label for deformation damage state as 1, the label for pitting damage state as 2, and the label for perforation damage state as 3. The corresponding labels are assigned to the damage states corresponding to these 38 sample data. The dataset is defined as X, and the label set as Y. Then, training sets XTrain and YTrain, and test sets XTest and YTest are created from the sample data of each type. 26 samples (z1 to z6, x1 to x15, j1 to j4) are used as the training set, and 8 samples (z7 to z9, x16 to x19, j5 to j6) are used as the prediction set. The training set XTrain corresponds to the classification label set YTrain, and the test set XTest corresponds to the predicted classification label set YPred. XTrain is a 26x1 cell array containing the time-series data of the 26 training set samples, and each cell array unit is an 8×201 matrix. 201 represents that 201 strain data points were collected within a 40-second time window, and 8 represents the dimension of the time series set, i.e., the number of data collection points for each sample. YTrain is a 26x1 numerical array, corresponding to the category of each cell in XTrain.

[0054] Network parameter configuration: During model training, the number of hidden layers and neurons can be adjusted. Setting the number of hidden layers too high will significantly increase the computational load of the network and lead to overfitting; setting the number of hidden layers too high or too low will fail to achieve the desired fitting effect for data with complex features, affecting the performance of the neural network. This invention explores and determines the most suitable number of hidden layers during model training. Through multiple training iterations, it was found that the prediction accuracy is highest when the number of hidden layers is 10.

[0055] In addition, the classify function in MATLAB can be used to predict the category of the test data, and the accuracy of the model classification can be verified by checking whether the prediction results are consistent with the label values.

[0056] This invention numerically simulates the damage process of flexible spacecraft surface structures subjected to laser ablation. The time-series strain and temperature data obtained during the simulation are converted into cell arrays, which are then input into a constructed LSTM neural network model for training. During training, the network parameters are adjusted based on the training results to improve the accuracy of the model's classification and prediction. Finally, simulation experimental data can be used to verify the classification and prediction accuracy of the evaluation model to ensure good damage prediction results.

[0057] The present invention provides a satellite structural state intelligent assessment method based on fiber optic sensors. It constructs a fiber optic grating sensor network to sense damage to the surface materials of flexible spacecraft, and uses an intelligent assessment method to convert the signals sensed by the sensors into structural damage modes, allowing the spacecraft to judge its own structural state based on the damage modes.

[0058] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent assessment of satellite structural status based on fiber optic sensors, characterized in that, include: Step S1: Taking the rubber material on the surface of the flexible spacecraft as the object, conduct ablation simulation of the surface material of the flexible spacecraft to obtain different damage states and corresponding strain data. Step S2: Train the time recurrent neural network LSTM using the different damage states and corresponding strain data to obtain a classification prediction model; Step S3: Input the data to be predicted into the classification prediction model to obtain the predicted damage state output by the classification prediction model; Step S1 includes: Simulate the temperature field induced by a laser heat source on the surface of a rubber material to generate a temperature field cloud map; Based on the temperature field cloud map, data acquisition points around the ablation point are selected; By changing the duration of the laser heat source during the simulation, the damage state of the rubber material under different temperature fields and the strain data at the data acquisition points are obtained.

2. The intelligent assessment method for satellite structural status according to claim 1, characterized in that, Step S2 includes: The preprocessed damage states and corresponding strain data are input into a time recurrent neural network LSTM. The input gate, forget gate, and output gate of the time recurrent neural network LSTM work together to output the predicted damage state.

3. The intelligent assessment method for satellite structural status according to claim 2, characterized in that, Also includes: The error between the predicted damage state and the actual label is calculated using the loss function of the classification prediction model. When the error reaches a preset threshold or the training rounds reach an upper limit, the final predicted damage state is output.

4. The intelligent assessment method for satellite structural status according to claim 3, characterized in that, Also includes: When the error does not reach the preset threshold and the number of training rounds does not reach the upper limit, the error is propagated backward along the network using the backpropagation algorithm. The weights and biases of the network are updated according to the gradient descent method, the network parameters are adjusted, and the training process is repeated until the error reaches the preset threshold or the number of training rounds reaches the upper limit.

5. The intelligent assessment method for satellite structural status according to claim 4, characterized in that, The network parameters include at least the number of hidden layer units, learning rate, number of training epochs, and the weights and biases of the initialized network.

6. The intelligent assessment method for satellite structural status according to claim 1, characterized in that, Step S2 further includes: The different damage states and corresponding strain data are converted into cell arrays, and the cell arrays are input into the temporal recurrent neural network LSTM for training.

7. The intelligent assessment method for satellite structural status according to claim 1, characterized in that, The classification prediction model includes an input layer, an LSTM layer, a fully connected layer, and an output layer.

8. The intelligent assessment method for satellite structural status according to claim 7, characterized in that, The fully connected layer includes the leakyReluLayer function, the dropoutLayer function, and the ReLU activation function.

9. The intelligent assessment method for satellite structural status according to claim 7, characterized in that, The output layer includes the softmaxLayer function and the classificationLayer function.

10. The intelligent assessment method for satellite structural status according to claim 1, characterized in that, The damage states include: deformation, pitting, and perforation.

Images