Fiber composite material damage diagnosis method based on physical neural network

Abstract

The invention discloses a fiber composite material damage diagnosis method based on a physical neural network, and the method comprises the steps: firstly constructing a multi-dimensional physical parameter fused physical neural network model, comprehensively considering material, mechanical, thermal and electrical parameters, and fusing a physical mechanism to construct a network structure; input features are determined and preprocessed, acoustic emission, strain, temperature and other signals are collected, and input layer features are constructed after filtering and standardization processing. In the aspect of model architecture design, a neural network comprising an input layer, a hidden layer and an output layer is constructed, and a physical mechanism is integrated in the neural network. A loss function under physical constraints is constructed to ensure that the model accords with a physical rule, a final model is obtained through model initialization and training optimization, finally, real-time dynamic monitoring of damage is achieved, a diagnosis result is output, and the reliability of material use is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fiber composite material damage detection, and in particular to a fiber composite material damage diagnosis method based on physical neural network. Background Art

[0002] Fiber composites hold a crucial position in modern industry due to their high strength-to-weight ratio, excellent corrosion resistance, and design flexibility. In aerospace, they are used in key components such as aircraft fuselages and wings, effectively reducing aircraft weight and improving fuel efficiency. In the automotive industry, they are used in the manufacture of vehicle bodies and parts, contributing to lightweighting and reducing energy consumption. In wind power generation, they are the core material for wind turbine blades, directly impacting power generation efficiency and blade lifespan.

[0003] However, in actual service, fiber composite materials are subjected to complex loads, environmental erosion and other effects for a long time, and are very prone to damage such as delamination, cracks, and debonding. These damages are often difficult to detect in the early stages, but once they develop and spread, they may cause catastrophic consequences and seriously threaten structural safety and stable system operation. Traditional damage detection methods, such as visual inspection, ultrasonic inspection, and radiographic inspection, can detect damage to a certain extent, but they have many limitations. Visual inspection relies on manual experience and is difficult to detect internal damage; ultrasonic inspection and radiographic inspection have high requirements for equipment, complex inspection processes, and cannot achieve real-time and dynamic monitoring, making it difficult to meet the needs of modern industry for efficient and accurate inspection of fiber composite materials.

[0004] With the cross-integration of artificial intelligence and materials science, data-driven machine learning methods are gradually being applied to the field of fiber composite material damage diagnosis. However, these methods typically rely solely on large amounts of data for model training and lack a deep understanding of the physical mechanisms of the materials. This results in insufficient generalization of the models under complex working conditions, making it difficult to guarantee the reliability and accuracy of the diagnostic results. Furthermore, in practical applications, the damage process of fiber composite materials involves the coupling of multidisciplinary physical parameters such as materials, mechanics, thermals, and electricity. A single data-driven model cannot effectively process this multidimensional information and cannot fully reveal the complex relationship between damage and various physical parameters. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a fiber composite material damage diagnosis method based on physical neural network, the specific steps are as follows:

[0006] Step 1: Construct a physical neural network model that integrates multi-dimensional physical parameters. This model comprehensively considers the material parameters, mechanical parameters, thermal parameters, and electrical parameters of fiber composite materials. The model architecture incorporates the physical mechanisms of material damage mechanics, thermal conduction, and electrical conduction. A network structure consisting of input, hidden, and output layers is constructed to provide a basic model framework for fiber composite material damage diagnosis.

[0007] Step 2: Feature data preprocessing: Filtering and noise reduction signal processing techniques are used to preprocess the acoustic emission signals and strain signals to remove interference noise. Acoustic emission sensors, vibrating wire surface strain sensors, and temperature sensors placed on the fiber composite material are used to collect multi-dimensional feature data from the damage process in real time, ensuring that the collected data accurately reflects the material's damage state and provides high-quality data support for model input.

[0008] Step 3: Conduct model training and optimization. Train the physical neural network model using multi-source data, optimize model parameters, and verify its generalization capability.

[0009] Step 4: Implement real-time dynamic monitoring of damage by embedding the optimized model into the monitoring system. By interacting with real-time data from sensors, real-time analysis and location of damage status can be achieved.

[0010] Step 5: Output standardized diagnostic results to provide strong technical support for the reliable application of fiber composite materials.

[0011] As a further improvement of the present invention, the physical neural network model for multi-dimensional physical parameter fusion constructed in step 1 can be expressed as:

[0012] Step 1.1 Determine input features and preprocess

[0013] Establish the input data of the model so that it can reflect the physical changes of fiber composite materials during the damage process;

[0014] Step 1.1.1 Data acquisition:

[0015] Acoustic emission signals of materials are collected through acoustic emission sensors; material strain information is obtained through vibrating wire surface strain gauges; and the temperature change rate of the material surface is recorded through temperature sensors.

[0016] Step 1.1.2. Data preprocessing:

[0017] A low-pass filter was used to remove environmental noise from the acoustic emission signal and strain data. Z-score standardization was used to adjust the data to a similar range.

[0018] Step 1.1.3. Construct input layer features:

[0019] Construct features for the pre-processed data, including: the amplitude and spectrum characteristics of the acoustic emission signal; the peak frequency and energy distribution of the strain signal; the temperature change rate;

[0020] Step 1.2 Model architecture design

[0021] Construct a neural network consisting of an input layer, a hidden layer, and an output layer, and incorporate physical mechanism constraints into it;

[0022] Step 1.2.1 Input layer design:

[0023] The pre-processed acoustic emission signal, strain signal, and temperature change rate are used as multi-dimensional features to form the input vector X:

[0024] X=[x1,x2,...,x n ]

[0025] Among them, each x i A numerical value representing a single-dimensional feature;

[0026] Step 1.2.2. Hidden layer construction:

[0027] Design a multi-layer fully connected network, perform multi-layer abstraction on the input data through nonlinear activation functions, and capture the nonlinear relationship between each feature; the network has L layers, where the input layer is the 0th layer, the hidden layers are the 1st to L-1th layers, and the output layer is the Lth layer. For the l=1, 2, ..., L-1th layer, the calculation formula is

[0028] h (l) =f(W (l) h (l-1) +b (l) )

[0029] Among them, h (0) =X is the input vector, W (l) and b (l) are the weight matrix and bias vector of the lth layer, f is the ReLU activation function, h (l) is the output of the lth hidden layer;

[0030] Step 1.2.3. Output layer design:

[0031] The L-th layer output layer uses the Softmax function to output the probability distribution of damage categories, and outputs the physical quantities of temperature, strain, and electric field strength through the linear layer for subsequent physical constraint calculations. The specific design is as follows:

[0032]

[0033] in, and is the weight matrix and bias vector of the output layer, y is the predicted probability distribution vector, which represents the confidence of different damage categories; T pred , ε pred They are temperature prediction, strain prediction, y class is the probability distribution vector of the damage category;

[0034] Step 1.3 Construct the loss function under physical constraints

[0035] Embedding physical mechanisms into neural networks ensures that model predictions not only conform to data trends but also satisfy the physical laws of the material.

[0036] Step 1.3.1 Heat conduction constraint:

[0037] According to Fourier's law of heat conduction, the change in temperature distribution in the material satisfies the following heat conduction equation:

[0038]

[0039] Where T is temperature, α is thermal diffusion coefficient, t is time variable, ▽ 2 T is the Laplace operator of the temperature field; for the temperature prediction T output by the network pred , add the loss term L to the loss function thermal :

[0040]

[0041] Automatically calculate partial derivatives and Laplace operators to ensure that model predictions meet the laws of heat conduction;

[0042] Step 1.3.2. Mechanical constraints:

[0043] Considering the stress-strain relationship of the material, Hooke's law of linear elasticity theory is quoted:

[0044] σ=E·ε

[0045] Where σ is stress, ε is strain, and E is the elastic modulus of the material; for the strain prediction ε output by the network pred , add the loss term L to the loss function mechanical :

[0046]

[0047] Make the predicted stress state conform to the basic laws of mechanics;

[0048] Step 1.3.4. Comprehensive physical loss function:

[0049] Integrate the above physical constraints into the total loss function L totalIn the example, the mean square error loss L is compared with the traditional data fitting data Co-optimization model:

[0050] L total =L data +λ thermal L thermal +λ mechanical L mechanical

[0051] Among them, λ is the corresponding weight factor, which is adjusted according to the importance of each part to the final task;

[0052] Step 1.4 Model integration and implementation details

[0053] Goal: Integrate the above design into a complete and operational neural network system;

[0054] Steps:

[0055] Step 1.4.1 Model initialization:

[0056] Follow step 1.2 to define the input layer, the hidden layer with a fully connected layer and ReLU activation, and the output layer with Softmax and linear output. Use the He initialization method to set the hidden layer weights to avoid vanishing gradients. Initialize the output layer weights to small random numbers and the bias term to zero.

[0057] Step 1.4.3. Training optimization:

[0058] Using the collected and preprocessed data, the total loss function L is optimized through back propagation total In each forward propagation, the preprocessed multimodal data is input and passed through the hidden layer and output layer in turn. Using the Adam optimizer, the learning rate and hyperparameters are adjusted according to the validation set so that the loss function gradually converges. Cross-validation and tuning are performed on different physical constraint weights λ to ensure a balance between physical constraints and data fitting. Finally, a physical neural network model that integrates multi-dimensional physical parameters is obtained.

[0059] As a further improvement of the present invention, the feature data preprocessing in step 2 can be expressed as follows:

[0060] Step 2.1 Data collection and sensor placement

[0061] Collect various physical signals reflecting the damage status of the material at key parts of the fiber composite material;

[0062] Step 2.1.1 Sensor selection and placement:

[0063] Acoustic emission sensors are used to detect instantaneous high-frequency signals generated when microcracks occur inside the material; vibrating wire surface strain sensors are used to collect structural strain data; temperature sensors are used to record temperature changes in real time to ensure the collection of temperature change rate data;

[0064] Step 2.1.2. Data recording and storage:

[0065] After data is collected, it is stored in real time with a timestamp for subsequent time series analysis and synchronization processing;

[0066] Step 2.2 Signal preprocessing

[0067] Use digital signal processing technology to reduce noise, filter, and normalize the collected multi-channel raw data to extract effective information and remove external interference noise;

[0068] Steps:

[0069] Step 2.2.1 Filtering and noise reduction:

[0070] Acoustic emission signal and strain signal filtering: A bandpass filter is used to eliminate low-frequency background noise and high-frequency interference that exceeds the main components of the signal; the filtering implementation formula is shown below:

[0071]

[0072] Among them, x[n] is the nth sampling value of the original signal, y[n] is the filtered signal, and b k 、a k is the filter coefficient, k is the index variable in the summation process, M is the maximum delay step number of the forward path, and N is the maximum delay step number of the feedback path;

[0073] Temperature signal processing: Since the temperature signal changes slowly, a moving average filter is used to smooth the data; the specific formula is as follows:

[0074]

[0075] Where L is the window length, T smooth [n] is the smoothed temperature value at the nth time point, and T[nk] is the original temperature sampling value at the nkth time point;

[0076] Step 2.2.2. Normalization:

[0077] In order to eliminate the impact of different sensor dimensions and amplitude differences on subsequent model training, the output of each sensor is normalized:

[0078]

[0079] Among them, x is the original data value, xmin is the minimum value that the sensor can take in the entire data set, x max is the maximum value that the sensor can obtain in the entire data set, x norm is the normalized result;

[0080] Step 2.3 Feature extraction and data construction

[0081] Extract key features that can reflect the material damage state from the preprocessed signal and construct the model input vector;

[0082] Step 2.3.1 Acoustic emission signal feature extraction:

[0083] Time domain features: extract peak amplitude, pulse duration, and envelope information; frequency domain features: perform discrete Fourier transform on the preprocessed signal:

[0084]

[0085] Where x[n] is the discrete time domain sampling value of the nth sampling point of the acoustic emission signal, N is the total number of sampling points; j is the imaginary unit, e is the base of the natural logarithm; X[k] is the discrete Fourier transform result of the acoustic emission signal, the kth frequency component; analyze the main frequency component and energy distribution in the spectrum and construct the eigenvector;

[0086] Step 2.3.2. Strain signal feature extraction:

[0087] Calculate the signal's root mean square value, crest factor, and spectrum energy concentration index. The RMS value calculation formula is as follows:

[0088]

[0089] Where x[n] is the discrete time domain sampling value of the nth sampling point of the acoustic emission signal, N is the total number of sampling points; RMS is the root mean square value of the strain signal, which measures the overall energy level of the signal;

[0090] Step 2.3.3. Calculation of temperature change rate:

[0091] Using the smoothed temperature data, the temperature change rate is calculated by time differentiation. The formula is:

[0092]

[0093] Where T[n] is the sample value of the temperature signal after smoothing at time n, Δt is the sampling period, is the rate of change of temperature with time;

[0094] Step 2.3.5. Construct a multidimensional input vector:

[0095] The features extracted by each sensor are combined into a set of multidimensional feature vectors, denoted as X:

[0096]

[0097] Among them, AE 幅值 is the maximum amplitude characteristic of the acoustic emission signal, AE 频谱特征 It is the frequency domain feature extracted by Fourier transform of acoustic emission signal, RMS 应变 is the RMS value of the strain signal, and the peak frequency is the main frequency component of the strain signal and the acoustic emission signal. is the temperature change rate; this vector serves as the input data of the subsequent physical neural network model to ensure that the data can accurately reflect the material damage state.

[0098] The fiber composite material damage diagnosis method based on physical neural network of the present invention has beneficial effects. The technical effects of the present invention are:

[0099] 1. The present invention comprehensively considers the material parameters, mechanical parameters, thermal parameters and electrical parameters of fiber composite materials to construct a physical neural network model that integrates multi-dimensional physical parameters. This model comprehensively reflects the various physical changes of the material during the damage process. Compared with traditional methods that only consider single or partial parameters, it can more accurately capture the material damage characteristics, thereby significantly improving the accuracy of damage diagnosis.

[0100] 2. The present invention collects multi-dimensional characteristic data during the damage process in real time, ensuring that the collected data can accurately reflect the damage state of the material and is timely input into the model for analysis.

[0101] 3. This method uses cross-validation to adjust the weights of heat conduction and mechanical constraints, balancing data fitting with physical constraints and avoiding overfitting or underfitting. This optimization approach enables the model to maintain good performance across different datasets, further enhancing its generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 is a flow chart of the present invention;

[0103] Figure 2 This is a diagram of the physical neural network structure of the present invention. DETAILED DESCRIPTION

[0104] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0105] The present invention relates to fiber composite material damage diagnosis technology, which constructs a physical neural network model that integrates multi-dimensional physical parameters and integrates multiple parameters and physical mechanisms. After data acquisition, preprocessing, and feature extraction, the model is trained and optimized to achieve real-time dynamic monitoring of damage, output standardized diagnostic results, and improve diagnostic accuracy and reliability. The invention flow chart is as follows: Figure 1 As shown, the steps of the present invention are described in detail below.

[0106] Step 1: Construct a physical neural network model that integrates multi-dimensional physical parameters, comprehensively considers the material parameters, mechanical parameters, thermal parameters and electrical parameters of fiber composite materials, integrates the physical mechanisms of material damage mechanics, heat conduction and electrical conduction into the model architecture, and constructs a network structure consisting of input layer, hidden layer and output layer to provide a basic model framework for fiber composite material damage diagnosis.

[0107] Step 1.1 Determine input features and preprocess

[0108] The input data of the model is established so that it can reflect the physical changes of fiber composite materials during the damage process.

[0109] Step 1.1.1 Data acquisition:

[0110] The acoustic emission signal of the material is collected by the acoustic emission sensor; the material strain information is obtained by the vibrating wire surface strain gauge; and the temperature change rate of the material surface is recorded by the temperature sensor.

[0111] Step 1.1.2. Data preprocessing:

[0112] Low-pass filters were used to remove environmental noise from the acoustic emission signals and strain data. Z-score normalization was used to adjust the data to a similar range.

[0113] Step 1.1.3. Construct input layer features:

[0114] Features are constructed for the preprocessed data, including: the amplitude and spectrum characteristics of the acoustic emission signal; the peak frequency and energy distribution of the strain signal; and the temperature change rate.

[0115] Step 1.2 Model architecture design

[0116] Construct a neural network consisting of an input layer, a hidden layer, and an output layer, and incorporate physical mechanism constraints into it.

[0117] Step 1.2.1 Input layer design:

[0118] The pre-processed acoustic emission signal, strain signal, and temperature change rate are used as multi-dimensional features to form the input vector X:

[0119] X=[x1,x2,...,x n ]

[0120] Among them, each x i A numerical value representing a single-dimensional feature.

[0121] Step 1.2.2. Hidden layer construction:

[0122] Design a multi-layer fully connected network, and use nonlinear activation functions to perform multi-layer abstraction on the input data to capture the nonlinear relationship between each feature. The network has L layers, where the input layer is layer 0, the hidden layers are layers 1 to L-1, and the output layer is layer L. For layers l = 1, 2, ..., L-1, the calculation formula is

[0123] h (l) =f(W (l) h (l-1) +b (l) )

[0124] Among them, h (0) =X is the input vector, W (l) and b (l) are the weight matrix and bias vector of the lth layer, f is the ReLU activation function, h (l) is the output of the lth hidden layer.

[0125] Step 1.2.3. Output layer design:

[0126] The L-th layer output layer uses the Softmax function to output the probability distribution of damage categories, and outputs the physical quantities of temperature, strain, and electric field strength through the linear layer for subsequent physical constraint calculations. The specific design is as follows:

[0127]

[0128] in, and is the weight matrix and bias vector of the output layer, and y is the predicted probability distribution vector, which represents the confidence of different damage categories. pred , ε pred They are temperature prediction, strain prediction, y class is the probability distribution vector of the damage category.

[0129] Step 1.3 Construct the loss function under physical constraints

[0130] Embedding physical mechanisms into neural networks ensures that model predictions not only conform to data trends but also satisfy the physical laws of the material.

[0131] Step 1.3.1 Heat conduction constraint:

[0132] According to Fourier's law of heat conduction, the change in temperature distribution in the material satisfies the following heat conduction equation:

[0133]

[0134] Where T is temperature, α is thermal diffusion coefficient, t is time variable, ▽ 2 T is the Laplace operator of the temperature field. For the temperature prediction T output by the network pred , add the loss term L to the loss function thermal :

[0135]

[0136] The partial derivatives and Laplace operators are calculated through automatic differentiation to ensure that the model predictions satisfy the laws of heat conduction.

[0137] Step 1.3.2. Mechanical constraints:

[0138] Considering the stress-strain relationship of the material, Hooke's law of linear elasticity theory is quoted:

[0139] σ=E·ε

[0140] Where σ is stress, ε is strain, and E is the elastic modulus of the material. pred , add the loss term L to the loss function mechanical :

[0141]

[0142] This makes the predicted stress state consistent with the basic laws of mechanics.

[0143] Step 1.3.4. Comprehensive physical loss function:

[0144] Integrate the above physical constraints into the total loss function L total In the example, the mean square error loss L is compared with the traditional data fitting data Co-optimization model:

[0145] L total =L data +λ thermal L thermal +λ mechanical L mechanical

[0146] Among them, λ is the corresponding weight factor, which is adjusted according to the importance of each part to the final task. The physical neural network structure is shown in the figure below: Figure 2 shown.

[0147] Step 1.4 Model integration and implementation details

[0148] Goal: Integrate the above designs into a complete and operational neural network system.

[0149] Steps:

[0150] Step 1.4.1 Model initialization:

[0151] Follow step 1.2 to define the input layer, a hidden layer with a fully connected layer and ReLU activation, and an output layer with Softmax and linear output. Use the He initialization method to set the hidden layer weights to prevent vanishing gradients. Initialize the output layer weights to small random numbers and the bias to zero.

[0152] Step 1.4.3. Training optimization:

[0153] Using the collected and preprocessed data, the total loss function L is optimized through back propagation total In each forward propagation, preprocessed multimodal data is input and passed through the hidden layer and output layer in sequence. Using the Adam optimizer, the learning rate and hyperparameters are adjusted based on the validation set to achieve gradual convergence of the loss function. Cross-validation and optimization are performed for different physical constraint weights λ to ensure a balance between physical constraints and data fit. Ultimately, a physical neural network model is obtained that integrates multi-dimensional physical parameters.

[0154] Step 2: Feature data preprocessing: Use filtering and noise reduction signal processing techniques to preprocess the acoustic emission signals and strain signals to remove interference noise; through the acoustic emission sensors, vibrating wire surface strain sensors, and temperature sensors arranged on the fiber composite material, multi-dimensional feature data of the damage process is collected in real time to ensure that the collected data can accurately reflect the damage state of the material and provide high-quality data support for model input.

[0155] Step 2.1 Data collection and sensor placement

[0156] Various physical signals reflecting the damage state of the material are collected at key parts of the fiber composite material.

[0157] Step 2.1.1 Sensor selection and placement:

[0158] Acoustic emission sensors are used to detect the instantaneous high-frequency signals generated when microcracks occur within the material. Vibrating wire surface strain sensors are used to collect structural strain data. Temperature sensors are used to record temperature changes in real time, ensuring the collection of temperature change rate data.

[0159] Step 2.1.2. Data recording and storage:

[0160] After data is collected, it is stored in real time with a timestamp for subsequent time series analysis and synchronization processing.

[0161] Step 2.2 Signal preprocessing

[0162] Digital signal processing technology is used to reduce noise, filter and normalize the collected multi-channel raw data to extract effective information and remove external interference noise.

[0163] Steps:

[0164] Step 2.2.1 Filtering and noise reduction:

[0165] Acoustic emission signal and strain signal filtering: A bandpass filter is used to eliminate low-frequency background noise and high-frequency interference that exceeds the main components of the signal. The filtering implementation formula is shown below:

[0166]

[0167] Among them, x[n] is the nth sampling value of the original signal, y[n] is the filtered signal, and b k 、a k is the filter coefficient, k is the index variable in the summation process, M is the maximum delay step number of the forward path, and N is the maximum delay step number of the feedback path.

[0168] Temperature signal processing: Since the temperature signal changes slowly, a moving average filter is used to smooth the data. The specific formula is as follows:

[0169]

[0170] Where L is the window length, T smooth [n] is the smoothed temperature value at the nth time point, and T[nk] is the original temperature sampling value at the nkth time point.

[0171] Step 2.2.2. Normalization:

[0172] In order to eliminate the impact of different sensor dimensions and amplitude differences on subsequent model training, the output of each sensor is normalized:

[0173]

[0174] Among them, x is the original data value, x min is the minimum value that the sensor can take in the entire data set, x max is the maximum value that the sensor can obtain in the entire data set, x norm is the normalized result.

[0175] Step 2.3 Feature extraction and data construction

[0176] The key features that can reflect the material damage state are extracted from the preprocessed signal and the model input vector is constructed.

[0177] Step 2.3.1 Acoustic emission signal feature extraction:

[0178] Time domain features: extract peak amplitude, pulse duration, and envelope information; frequency domain features: perform discrete Fourier transform on the preprocessed signal:

[0179]

[0180] Where x[n] is the discrete time-domain sample value of the acoustic emission signal at the nth sampling point, and N is the total number of sampling points. j is the imaginary unit, and e is the base of the natural logarithm. X[k] is the discrete Fourier transform result of the acoustic emission signal, representing the kth frequency component. The dominant frequency component and energy distribution in the spectrum are analyzed to construct the eigenvector.

[0181] Step 2.3.2. Strain signal feature extraction:

[0182] Calculate the signal's root mean square value, crest factor, and spectrum energy concentration index. The RMS value calculation formula is as follows:

[0183]

[0184] Where x[n] is the discrete time domain sample value of the acoustic emission signal at the nth sampling point, and N is the total number of sampling points. RMS is the root mean square value of the strain signal, which measures the overall energy level of the signal.

[0185] Step 2.3.3. Calculation of temperature change rate:

[0186] Using the smoothed temperature data, the temperature change rate is calculated by time differentiation. The formula is:

[0187]

[0188] Where T[n] is the sample value of the temperature signal after smoothing at time n, Δt is the sampling period, is the rate of change of temperature with time.

[0189] Step 2.3.5. Construct a multidimensional input vector:

[0190] The features extracted by each sensor are combined into a set of multidimensional feature vectors, denoted as X:

[0191]

[0192] Among them, AE 幅值 is the maximum amplitude characteristic of the acoustic emission signal, AE 频谱特征 It is the frequency domain feature extracted by Fourier transform of acoustic emission signal, RMS 应变is the RMS value of the strain signal, and the peak frequency is the main frequency component of the strain signal and the acoustic emission signal. is the temperature change rate. This vector serves as the input data for the subsequent physical neural network model to ensure that the data can accurately reflect the material damage state.

[0193] Step 3: Conduct model training and optimization, train the physical neural network model through multi-source data, optimize the model parameters and verify its generalization ability.

[0194] Step 3.1 Dataset division:

[0195] The preprocessed data is divided into training set, validation set and test set, and the samples cover different damage types, damage degrees and working conditions.

[0196] Step 3.2 Physical constraint weight tuning:

[0197] Tuning the heat conduction constraint λ via cross-validation thermal and the weight λ of the mechanical constraint mechanical , balancing data fitting and physical law constraints.

[0198] Step 3.3 Model verification and testing:

[0199] The model performance is evaluated on the test set by the damage classification accuracy and physical quantity prediction error to ensure the robustness of the model on unknown data.

[0200] Step 4: Implement real-time dynamic monitoring of damage, embed the optimized model into the monitoring system, and implement real-time analysis and positioning of damage status through real-time data interaction with sensors.

[0201] Step 5: Output standardized diagnostic results to provide strong technical support for the reliable application of fiber composite materials.

[0202] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A fiber composite material damage diagnosis method based on a physical neural network, comprising the following specific steps, characterized in that: Step 1: Construct a physical neural network model that integrates multi-dimensional physical parameters. This model comprehensively considers the material parameters, mechanical parameters, thermal parameters, and electrical parameters of fiber composite materials. The model architecture incorporates the physical mechanisms of material damage mechanics, thermal conduction, and electrical conduction. A network structure consisting of input, hidden, and output layers is constructed to provide a basic model framework for fiber composite material damage diagnosis. Step 2: Feature data preprocessing: Filtering and noise reduction signal processing techniques are used to preprocess the acoustic emission signals and strain signals to remove interference noise. Acoustic emission sensors, vibrating wire surface strain sensors, and temperature sensors placed on the fiber composite material are used to collect multi-dimensional feature data from the damage process in real time, ensuring that the collected data accurately reflects the material's damage state and provides high-quality data support for model input. Step 3: Conduct model training and optimization. Train the physical neural network model using multi-source data, optimize model parameters, and verify its generalization capability. Step 4: Implement real-time dynamic monitoring of damage by embedding the optimized model into the monitoring system. By interacting with real-time data from sensors, real-time analysis and location of damage status can be achieved. Step 5: Output standardized diagnostic results to provide strong technical support for the reliable application of fiber composite materials.

2. The fiber composite material damage diagnosis method based on physical neural network according to claim 1, characterized in that: The physical neural network model for multi-dimensional physical parameter fusion constructed in step 1 can be expressed as: Step 1.1 Determine input features and preprocess Establish the input data of the model so that it can reflect the physical changes of fiber composite materials during the damage process; Step 1.1.1 Data acquisition: Acoustic emission signals of materials are collected through acoustic emission sensors; material strain information is obtained through vibrating wire surface strain gauges; and the temperature change rate of the material surface is recorded through temperature sensors. Step 1.1.

2. Data preprocessing: A low-pass filter was used to remove environmental noise from the acoustic emission signal and strain data. Z-score standardization was used to adjust the data to a similar range. Step 1.1.

3. Construct input layer features: Construct features for the pre-processed data, including: the amplitude and spectrum characteristics of the acoustic emission signal; the peak frequency and energy distribution of the strain signal; the temperature change rate; Step 1.2 Model architecture design Construct a neural network consisting of an input layer, a hidden layer, and an output layer, and incorporate physical mechanism constraints into it; Step 1.2.1 Input layer design: The pre-processed acoustic emission signal, strain signal, and temperature change rate are used as multi-dimensional features to form the input vector X: X=[x1,x2,...,x n ] Among them, each x i A numerical value representing a single-dimensional feature; Step 1.2.

2. Hidden layer construction: Design a multi-layer fully connected network, perform multi-layer abstraction on the input data through nonlinear activation functions, and capture the nonlinear relationship between each feature; the network has L layers, where the input layer is the 0th layer, the hidden layers are the 1st to L-1th layers, and the output layer is the Lth layer. For the l=1, 2, ..., L-1th layer, the calculation formula is h (l) =f(W (l) h (l-1) +b (l) ) Among them, h (0) =X is the input vector, W (l) and b (l) are the weight matrix and bias vector of the lth layer, f is the ReLU activation function, h (l) is the output of the lth hidden layer; Step 1.2.

3. Output layer design: The L-th layer output layer uses the Softmax function to output the probability distribution of damage categories, and outputs the physical quantities of temperature, strain, and electric field strength through the linear layer for subsequent physical constraint calculations. The specific design is as follows: in, and is the weight matrix and bias vector of the output layer, y is the predicted probability distribution vector, which represents the confidence of different damage categories; T pred , ε pred They are temperature prediction, strain prediction, y class is the probability distribution vector of the damage category; Step 1.3 Construct the loss function under physical constraints Embedding physical mechanisms into neural networks ensures that model predictions not only conform to data trends but also satisfy the physical laws of the material. Step 1.3.1 Heat conduction constraint: According to Fourier's law of heat conduction, the change in temperature distribution in the material satisfies the following heat conduction equation: Where T is temperature, α is thermal diffusion coefficient, and t is time variable. is the Laplace operator of the temperature field; for the temperature prediction T output by the network pred , add the loss term L to the loss function thermal : Automatically calculate partial derivatives and Laplace operators to ensure that model predictions meet the laws of heat conduction; Step 1.3.

2. Mechanical constraints: Considering the stress-strain relationship of the material, Hooke's law of linear elasticity theory is quoted: σ=E·ε Where σ is stress, ε is strain, and E is the elastic modulus of the material; for the strain prediction ε output by the network pred , add the loss term L to the loss function mechanical : Make the predicted stress state conform to the basic laws of mechanics; Step 1.3.

4. Comprehensive physical loss function: Integrate the above physical constraints into the total loss function L total In the example, the mean square error loss L is compared with the traditional data fitting data Co-optimization model: L total =L data +λ thermal L thermal +λ mechanical L mechanical Among them, λ is the corresponding weight factor, which is adjusted according to the importance of each part to the final task; Step 1.4 Model integration and implementation details Goal: Integrate the above design into a complete and operational neural network system; Steps: Step 1.4.1 Model initialization: Follow step 1.2 to define the input layer, the hidden layer with a fully connected layer and ReLU activation, and the output layer with Softmax and linear output. Use the He initialization method to set the hidden layer weights to avoid vanishing gradients. Initialize the output layer weights to small random numbers and the bias term to zero. Step 1.4.

3. Training optimization: Using the collected and preprocessed data, the total loss function L is optimized through back propagation total In each forward propagation, the preprocessed multimodal data is input and passed through the hidden layer and output layer in turn. Using the Adam optimizer, the learning rate and hyperparameters are adjusted according to the validation set so that the loss function gradually converges. Cross-validation and tuning are performed on different physical constraint weights λ to ensure a balance between physical constraints and data fitting. Finally, a physical neural network model that integrates multi-dimensional physical parameters is obtained.

3. The fiber composite material damage diagnosis method based on physical neural network according to claim 1, characterized in that: The feature data preprocessing in step 2 can be expressed as follows: Step 2.1 Data collection and sensor placement Collect various physical signals reflecting the damage status of the material at key parts of the fiber composite material; Step 2.1.1 Sensor selection and placement: Acoustic emission sensors are used to detect instantaneous high-frequency signals generated when microcracks occur inside the material; vibrating wire surface strain sensors are used to collect structural strain data; temperature sensors are used to record temperature changes in real time to ensure the collection of temperature change rate data; Step 2.1.

2. Data recording and storage: After data is collected, it is stored in real time with a timestamp for subsequent time series analysis and synchronization processing; Step 2.2 Signal preprocessing Use digital signal processing technology to reduce noise, filter, and normalize the collected multi-channel raw data to extract effective information and remove external interference noise; Steps: Step 2.2.1 Filtering and noise reduction: Acoustic emission signal and strain signal filtering: A bandpass filter is used to eliminate low-frequency background noise and high-frequency interference that exceeds the main components of the signal; the filtering implementation formula is shown below: Among them, x[n] is the nth sampling value of the original signal, y[n] is the filtered signal, and b k 、a k is the filter coefficient, k is the index variable in the summation process, M is the maximum delay step number of the forward path, and N is the maximum delay step number of the feedback path; Temperature signal processing: Since the temperature signal changes slowly, a moving average filter is used to smooth the data; the specific formula is as follows: Where L is the window length, T smooth [n] is the smoothed temperature value at the nth time point, and T[nk] is the original temperature sampling value at the nkth time point; Step 2.2.

2. Normalization: In order to eliminate the impact of different sensor dimensions and amplitude differences on subsequent model training, the output of each sensor is normalized: Among them, x is the original data value, x min is the minimum value that the sensor can take in the entire data set, x max is the maximum value that the sensor can obtain in the entire data set, x norm is the normalized result; Step 2.3 Feature extraction and data construction Extract key features that can reflect the material damage state from the preprocessed signal and construct the model input vector; Step 2.3.1 Acoustic emission signal feature extraction: Time domain features: extract peak amplitude, pulse duration, and envelope information; frequency domain features: perform discrete Fourier transform on the preprocessed signal: Where x[n] is the discrete time domain sampling value of the nth sampling point of the acoustic emission signal, N is the total number of sampling points; j is the imaginary unit, e is the base of the natural logarithm; X[k] is the discrete Fourier transform result of the acoustic emission signal, the kth frequency component; analyze the main frequency component and energy distribution in the spectrum and construct the eigenvector; Step 2.3.

2. Strain signal feature extraction: Calculate the signal's root mean square value, crest factor, and spectrum energy concentration index. The RMS value calculation formula is as follows: Where x[n] is the discrete time domain sampling value of the nth sampling point of the acoustic emission signal, N is the total number of sampling points; RMS is the root mean square value of the strain signal, which measures the overall energy level of the signal; Step 2.3.

3. Calculation of temperature change rate: Using the smoothed temperature data, the temperature change rate is calculated by time differentiation. The formula is: Where T[n] is the sample value of the temperature signal after smoothing at time n, Δt is the sampling period, is the rate of change of temperature with time; Step 2.3.

5. Construct a multidimensional input vector: The features extracted by each sensor are combined into a set of multidimensional feature vectors, denoted as X: Among them, AE 幅值 is the maximum amplitude characteristic of the acoustic emission signal, AE 频谱特征 It is the frequency domain feature extracted by Fourier transform of acoustic emission signal, RMS 应变 is the RMS value of the strain signal, and the peak frequency is the main frequency component of the strain signal and the acoustic emission signal. is the temperature change rate; this vector serves as the input data of the subsequent physical neural network model to ensure that the data can accurately reflect the material damage state.

Images