Method for predicting residual bearing capacity of composite material sample based on dynamic data flow

Abstract

The present application relates to the field of composite material nondestructive testing and health monitoring, and aims to provide a method for predicting the residual bearing capacity of composite laminated material sample based on dynamic data flow, which comprises: collecting acoustic emission signal data and arranging; evaluating the importance of the traditional time-frequency characteristics and cumulative characteristics of the signal; establishing and training a convolutional neural network model, and using the model to predict the residual bearing capacity of the composite laminated material sample. The present application establishes a residual bearing capacity prediction model for end notched FRP laminated plates based on feature evaluation and convolutional neural network, and realizes the prediction of the residual bearing capacity of the sample according to the acoustic emission data flow generated by the sample. Compared with conventional non-cumulative characteristics, the present application can more accurately qualitatively describe the damage process of the ENF sample by using cumulative characteristics. The present application can continuously correct the model based on the NCA algorithm to ensure the accuracy of the prediction, so that it gradually tends to the actual value.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing and health monitoring of composite materials, and in particular to a method for predicting the remaining load-bearing capacity of composite layup material specimens based on dynamic data streams. Background Technology

[0002] With the continuous progress of society and the constant development of science and technology, single materials are increasingly unable to meet the stringent requirements of advanced technologies in terms of material strength, stiffness, density, durability, and other aspects. Therefore, composite materials composed of two or more materials with different properties are playing an increasingly important role, among which fiber-reinforced polymer composites (FPR) are the most widely used.

[0003] Harsh service environments and complex load conditions pose significant challenges to the long-term stable operation of composite material structures. Therefore, it is necessary to develop non-destructive damage detection and structural health monitoring technologies, among which acoustic emission (AE) technology shows considerable potential. Damage diagnosis and prediction of composite materials based on AE technology largely rely on complex signal processing. Compared to damage diagnosis based on AE signals, damage assessment based on AE signals is an emerging field with relatively limited reported research results. Damage assessment refers to roughly judging the real-time health status of composite material structures and predicting their future state based on existing AE data streams. Specifically, the main objectives of damage assessment include damage tolerance estimation, remaining service life prediction, remaining strength prediction, remaining load-bearing capacity prediction, and structural health status assessment. Achieving these objectives requires certain mathematical methods, commonly including regression analysis and artificial neural networks.

[0004] While artificial neural networks and support vector regression models provide relatively effective methods for damage assessment of composite structures, they are not adept at capturing sequential information in acoustic emission data streams. Compared to damage pattern recognition and singular signal detection, quantitatively estimating the current damage level of composite materials based on real-time acquired dynamic acoustic emission data streams remains a significant challenge. The key lies in revealing the hidden correlation between the acoustic emission signals generated in the specimen and the specimen's mechanical properties.

[0005] Currently, research on damage prediction based on acoustic emission data streams and deep learning methods is still limited, and the applicability of the models and influencing factors remain unclear. Furthermore, research on feature extraction that is sensitive to the damage process and possesses certain physical significance is relatively scarce. Therefore, it is necessary to develop techniques for predicting the residual load-bearing capacity of composite materials based on acoustic emission data streams. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for predicting the remaining load-bearing capacity of composite material layup specimens based on dynamic data flow.

[0007] To solve the technical problem, the solution of the present invention is:

[0008] A method for predicting the remaining load-bearing capacity of composite layup material specimens based on dynamic data stream is provided, including: collecting and processing acoustic emission signal data; evaluating the importance of the traditional time-frequency characteristics and cumulative characteristics of the signals; establishing and training a convolutional neural network model, and using the model to predict the remaining load-bearing capacity of the composite layup material specimens.

[0009] As a preferred embodiment of the present invention, the collection and processing of acoustic emission signal data specifically includes the following steps:

[0010] (1) Conduct acoustic emission tests on composite material layup specimens with end notches under three-point bending loads and collect acoustic emission signal data;

[0011] (2) Use multi-view learning techniques to increase the amount of dataset and use acoustic emission data from different sensors to characterize the damage persistence process of the specimen.

[0012] As a preferred embodiment of the present invention, the evaluation of the importance of the traditional time-frequency characteristics and cumulative characteristics of the signal specifically includes the following steps:

[0013] (1) Extract multiple traditional time-frequency features from the collected acoustic emission signal waveforms, and add multiple cumulative features with interpretable physical meaning as objects of importance evaluation;

[0014] (2) The traditional time-frequency features and cumulative features are evaluated by weighting using the RReliefF algorithm, and then cross-validation is performed using the NCA algorithm to minimize the mean square error. The features with relatively higher weights are used to predict the degree of damage to the sample.

[0015] As a preferred embodiment of the present invention, the conventional time-frequency characteristics include at least the following 15 characteristics: rise time, count, energy, duration, peak amplitude, average frequency, root mean square, average signal level, initial frequency, signal strength, absolute energy, centroid frequency, peak frequency, rise angle, and decay angle; the cumulative characteristics include at least the following 9 characteristics: cumulative impact count, cumulative rise time, cumulative duration, cumulative energy, cumulative signal strength, cumulative absolute energy, cumulative count, cumulative average signal level, and cumulative root mean square.

[0016] As a preferred embodiment of the present invention, the establishment and training of the convolutional neural network model specifically includes the following steps:

[0017] (1) Construct a one-dimensional deep convolutional neural network model, which includes four convolutional layers and two fully connected layers arranged in sequence; the kernel size of the first three convolutional layers is 5×1, and the kernel size of the last convolutional layer is 3×1; the four convolutional layers use 20, 20, 10 and 5 filters respectively; the first fully connected layer has 100 hidden units, and the second fully connected layer has only one unit; except for the last fully connected layer, each layer is followed by a ReLU activation function; after the first fully connected layer, the second convolutional layer and the fourth convolutional layer, there is a dropout layer with a dropout rate of 0.5.

[0018] (2) The high-dimensional sequence input of the model consists of 50 impact signals, i.e., the time window length is 50. The model can be trained for a maximum of 50 periods. Each training period refers to the process of all training data completing one forward and backward propagation in the model. For each training period, the input sequence is randomly shuffled into several batches, with the smallest batch containing 250 sequences. During the training process, the model uses the adaptive moment estimation algorithm to update and optimize the weights in each layer. The initial learning rate for the first ten training periods is set to 0.001 to achieve rapid optimization. After that, the learning rate is reduced to three-tenths of the original rate every ten periods to achieve convergence.

[0019] As a preferred embodiment of the present invention, the method of predicting the remaining load-bearing capacity of the composite layup material sample using a model specifically includes the following steps:

[0020] Acoustic emission tests were conducted on the composite material layup specimens to be tested, and acoustic emission signal data was collected. Acoustic emission characteristics and time series data were extracted from the acoustic emission signal waveforms, input into the model, and after calculation, the predicted results of the remaining load-bearing capacity of the specimens were output.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. This invention establishes a prediction model for the remaining load-bearing capacity of end-notched FRP laminates based on feature evaluation and convolutional neural networks, and realizes the prediction of the remaining load-bearing capacity of the specimen based on the acoustic emission data stream generated by the specimen.

[0023] 2. This invention utilizes multi-view learning technology to increase the amount of data, introduces two supervised algorithms to evaluate the importance of each acoustic emission feature in the regression problem, and establishes a convolutional network prediction model based on feature evaluation. Compared with conventional non-cumulative features, this invention uses cumulative features to more accurately qualitatively describe the damage process of ENF specimens.

[0024] 3. Existing damage assessment methods, including regression analysis and artificial neural networks, are not adept at capturing sequential information in acoustic emission data streams, making it difficult to quantitatively estimate the current damage level of composite materials in real time using dynamic data streams. The convolutional neural network used in this invention can continuously refine the model based on the NCA algorithm as the sample failure progresses, ensuring the accuracy of the prediction and gradually bringing it closer to the actual value. In later stages of actual testing, the prediction error was within 10N. Attached Figure Description

[0025] Figure 1 These are geometric dimensions and morphological images of an end-notch bending (ENF) specimen.

[0026] Figure 2 It is the weight evaluation of 24 features in each dataset under the RReliefF algorithm;

[0027] Figure 3 It is the weight evaluation of 24 features in each dataset under the NCA algorithm;

[0028] Figure 4 This is a framework diagram of the deep CNN model established in this invention;

[0029] Figure 5 The following is a comparison of the predicted and actual remaining carrying capacity values ​​for each dataset: (a) ENF1, (b) ENF2, (c) ENF3, (d) ENF4, (e) ENF5. Detailed Implementation

[0030] The present invention will be described in detail below with reference to the accompanying drawings and embodiments, and the objectives and effects of the present invention will become more apparent.

[0031] This invention uses acoustic emission technology to monitor the failure process of end-notched FRP laminates under three-point bending loads. Multi-view learning technology is used to increase the amount of data. RReliefF and NCA are used to determine the sensitivity of conventional features to the degree of damage. A CNN model for predicting the remaining load-bearing capacity of the laminate is constructed.

[0032] The method for predicting the remaining load-bearing capacity of composite material layup specimens based on dynamic data streams as described in this invention includes the following steps:

[0033] 1. Collect and organize acoustic emission signal data.

[0034] The test specimens were composite end-notched flexure (ENF) specimens prepared according to the international standard ASTM D7905. These specimens were made from medium-temperature curing prepreg ACTECH 1203 / EW301F / 38. Initial delamination cracks were pre-initiated by inserting a 50 mm long polytetrafluoroethylene (PTFE) film into the mid-plane of the laminate. Three-point bending mechanical testing was controlled using an electronic universal testing machine at a loading rate of 1 mm / min. The specimen width was 20 mm, and five repeated tests were performed. During each compression test, the load-displacement curve was recorded until the initial crack in the ENF specimen had fully propagated.

[0035] Acoustic emission (AE) testing was conducted simultaneously with mechanical testing to monitor the failure process of the specimen under bending loads. The AE testing system consisted of two piezoelectric sensors, two 2 / 4 / 6 AST preamplifiers, and a PAC-SAMOS 8 acoustic emission analyzer responsible for data acquisition and recording. The broadband sensors operated in the 0-1000kHz frequency range, with a particularly sensitive region of 120-1000kHz, characterized by a high frequency response amplitude and small fluctuation within this range. A maximum response amplitude of -60dB was observed near 520kHz. Figure 1 As shown, two sensors were symmetrically fixed on the ENF sample, and their surfaces were coated with silicone grease. The acquisition threshold was set to 40 dB, the preamplifier gain to 40 dB, the sampling frequency to 1000 kHz, and the low-frequency of the analog filter to 5 kHz. The acquired waveforms were recorded in the form of acoustic emission impact signals. The four important parameters for defining the impact, including peak definition time, impact definition time, impact latch-up time, and maximum duration, were set to 50 μs, 200 μs, 300 μs, and 500 ms, respectively. Multiple lead-breaking tests were performed to calibrate the acquisition parameter settings.

[0036] Acoustic emission signals are continuously generated as internal damage accumulates in the specimen, meaning these signals may indicate the degree of degradation in the specimen's mechanical properties. Therefore, the model aims to predict the remaining load-bearing capacity of the specimen based on real-time acquired local data. Let the i-th impact signal of specimen j be a high-dimensional vector composed of acoustic emission features. The r-th feature is f r (r = 1, 2, ..., p), where p is the total number of characteristics. Impact signals recorded before reaching the maximum load are retained to reflect the degree of damage inside the specimen and to prevent sudden failure of the specimen. The remaining load-bearing capacity is defined as shown in equation (1).

[0037]

[0038] In the formula: F i j N is the actual load applied to the specimen when the i-th impact signal occurs.j This represents the total number of impact signals generated in specimen j. It can be seen that when the actual load reaches its maximum value, the remaining load-bearing capacity... Reduced to zero.

[0039] After data preparation, a multi-view learning technique is used to increase the dataset size, utilizing acoustic emission data from different sensors to characterize the damage progression process of the specimen. During testing, the failure process of the specimen is monitored by multiple sensors (i.e., multiple channels), thus the remaining load-bearing capacity of the specimen can be interpreted from multiple different recording perspectives. Sensors closer to the initial crack may detect more impact signals, while sensors farther away collect fewer signals, because particularly weak signals can only excite the nearest sensor, while other signals can excite more sensors.

[0040] 2. Evaluate the importance of the traditional time-frequency characteristics and cumulative characteristics of the signal.

[0041] (1) Extract multiple traditional time-frequency features from the collected acoustic emission signal waveforms, and add multiple cumulative features with interpretable physical meaning as objects of importance evaluation;

[0042] In acoustic emission testing, 15 traditional time-frequency features are extracted from the waveform: rise time (RT), count (C), energy (E), duration (TD), peak amplitude (PA), average frequency (AF), root mean square (RMS), average signal level (ASL), initial frequency (IF), signal strength (SS), absolute energy (AbE), centroid frequency (FG), peak frequency (PF), rise angle (RA), and decay angle (DA). Additionally, nine physically interpretable cumulative features are considered: cumulative AE hits (CHit), cumulative rise time (CRT), cumulative time of duration (CTD), cumulative energy (CE), cumulative signal strength (CSS), cumulative absolute energy (CAbE), cumulative counts (CC), cumulative average signal level (CASL), and cumulative root mean square (RMS). The Squared Representational Study (CRMS) evaluates the importance of these 24 features and uses high-scoring features for prediction.

[0043] During testing, the failure process of the ENF specimen was monitored by two sensors (i.e., two channels), thus the remaining load-bearing capacity of the ENF specimen can be interpreted from two different recording perspectives. The impact signal from each test was divided into two parts by channel, resulting in a total of 10 datasets. The sensor closer to the initial crack may detect more impact signals, while the sensor further away collects fewer signals, because particularly weak signals can only excite the nearest sensor, while other signals can excite both sensors. Therefore, the same signal received by different sensors behaves differently, but both can serve as training datasets, thereby significantly increasing the amount of data—that is, multi-view learning.

[0044] (2) The traditional time-frequency features and cumulative features are evaluated by weighting using the RReliefF algorithm, and then cross-validation is performed using the NCA algorithm to minimize the mean square error. The features with relatively higher weights are used to predict the degree of damage to the sample.

[0045] This invention introduces the Regressional ReliefF (RReliefF) algorithm and the Neighborhood Component Analysis (NCA) algorithm, and uses these two supervised algorithms to evaluate the importance of each acoustic emission feature in the regression problem.

[0046] The RReliefF algorithm provides a global view of feature quality evaluation in regression problems, where dependencies may exist between different features. The key idea of ​​this algorithm is to score a feature using contextual information, specifically based on the difference in the predicted values ​​of data points with the same response value and data points with different response values ​​for that feature. In other words, the quality of the feature is approximated by the probability that two data points have different predicted values. The basic steps of the algorithm are described below.

[0047] Let w be the weight for different residual carrying capacity response values. dy The weights for different features are w. df The weights for different response values ​​and different features are w. dy&df The final weight of the r-th feature is w. r Initialize all four weights to zero. Randomly select an impact signal x. i Find its k nearest neighbor signals The weights are updated for each neighbor signal, each feature, and each collision signal. Therefore, the intermediate weights can be calculated according to Equation (2).

[0048]

[0049] In the formula: N represents the total number of impact signals, and They are xi and The differences between the corresponding remaining carrying capacity response values, the normalized distance, and the differences in values ​​on the characteristic dimension can be calculated by equation (3).

[0050]

[0051] In the formula: rank(i,i n )express In x i The position of each feature is sorted by distance among all neighboring signals, and β is a user-defined parameter that controls the influence of distance, set to 50. The final weight estimate of each feature is calculated according to equation (4).

[0052] w r =(w dy&df ) r / w dy -((w df ) r -(w dy&df ) r ) / (Nw dy (4)

[0053] NCA is a simple and effective nonlinear decision rule developed based on the nearest neighbor concept. The algorithm uses gradient ascent to maximize the expected leave-one-out regression accuracy with a regularization term, thereby learning the feature weights. Let T = {(x1, y1), (x2, y2), ..., (x...} i y i ), ..., (x N y N Let be a training dataset containing N impact signals and their corresponding residual load capacity response values. The goal is to obtain a weight vector w, which helps in selecting a feature subset for optimizing the nearest neighbor regression problem. i and x j The weighted distance between them can be denoted as Equation (5).

[0054]

[0055] In the formula: w r is the weight of the r-th acoustic emission feature, and p is the total number of features. An effective approximation method for selecting the nearest neighbor as the reference point is to determine the reference point through a probability distribution. For example, x j Selected as x i The probability of the reference point can be calculated according to equation (6).

[0056]

[0057] In the formula: κ represents a kernel function or similarity function, which can be defined as κ(i)=exp(-i / σ) κ ), where the kernel width σ κ This is an input parameter that influences the probability that the impact signal is selected as the reference point. Let... This represents the predicted value from the random regression model; 1 represents the measure of the predicted value. Compared with the actual value y i The loss function is the deviation between the two values. The loss function can be the mean absolute deviation, the mean squared error, or other user-defined functions. The average loss can be calculated according to equation (7).

[0058]

[0059] From equation (7), we can obtain the approximate accuracy of the leave-one-out regression method, as shown in equation (8).

[0060]

[0061] To avoid overfitting, a regularization term is introduced in equation (8), resulting in a regularization objective function, as shown in equation (9).

[0062]

[0063] In the formula: λ is a regularization parameter, which can be optimized through cross-validation. Considering that the regularization objective function is differentiable, its value with respect to the feature weights w... r The derivative of the gradient update direction can be calculated according to equation (10).

[0064] From equation (10), the corresponding gradient update equation can be obtained, as shown in equation (11).

[0065]

[0066] In the formula: the superscript t represents the update time, i.e. the iteration order, and α is the update step size, which can be calculated according to formula (12).

[0067]

[0068] From this, the weight vector w of the acoustic emission characteristics can be obtained.

[0069] The weight evaluation of the 24 features in each dataset under the RReliefF algorithm is as follows: Figure 2As shown, the center line and the hollow squares within the quadrilateral box represent the median and mean, respectively. The box boundary consists of two percentiles: the first quartile and the third quartile, with the difference between them called the interquartile range. The variability of the data outside the box is represented by two lines extending to the upper and lower limits, respectively. These limits are numerically equal to the third quartile plus 1.5 times the interquartile range and the first quartile minus 1.5 times the interquartile range, respectively. Unconnected short horizontal lines represent the maximum and minimum values. Except for CRMS, the weights of cumulative features are generally greater than zero; the weights of non-cumulative features are generally less than zero. Among cumulative features, CHit, CRT, CTD, CC, and CASL have relatively higher weights and are more suitable for describing the damage process of the specimen.

[0070] The results of the NCA algorithm are highly correlated with the regularization parameter λ, therefore fine-tuning the regularization parameter is necessary to obtain the optimal value. Specifically, cross-validation is used to minimize the mean squared error. A five-fold cross-validation method is employed, dividing the dataset into five equal parts, each of which is used as the validation set, while the remaining four parts are used as the training set. Under the optimal regularization parameter, the weights of the 24 features in each dataset are calculated, and the evaluation results are as follows: Figure 3 As shown, the cumulative features, except for CAbE, are significantly more sensitive to residual load-bearing capacity than non-cumulative features. The weights of CE, CSS, CAbE, and CRMS are closer to zero than those of the other cumulative features, indicating that they have a smaller impact on residual load-bearing capacity. Therefore, CHit, CRT, CTD, CC, and CASL are more suitable for quantifying the damage level of specimens, consistent with the assessment results of RReliefF.

[0071] 3. Establish and train a convolutional neural network model, and use the model to predict the remaining load-bearing capacity of composite layup material samples.

[0072] After identifying the features sensitive to remaining load-bearing capacity, a damage prediction model based on a convolutional neural network (CNN) was established. The difference between the CNN model and the traditional artificial neural network (ANN) model lies in the fact that neurons in the convolutional layers are connected to sub-regions of the previous layer, rather than connecting to all neurons in the previous layer.

[0073] A CNN model consists of multiple repeating layers with different functions, mainly including convolutional layers, pooling layers, and fully-connected layers. The convolutional layer performs convolution operations between multiple filters and the input data, passing the resulting feature map to the next layer. A non-linear layer is added after the convolutional layer to introduce non-linear factors not considered in the convolution operation. This layer typically uses the Rectified Linear Unit (ReLU) activation function, which retains non-negative values ​​and removes negative values. Subsequently, the pooling layer performs downsampling by dividing the input map into rectangular pooling regions and calculating the maximum or average value of each region, thus reducing the number of connections to the next layer and improving the model's fault tolerance. Dropout layers prevent overfitting. Finally, the fully-connected layer combines all the features learned from the previous layers in the input data, reducing information loss, and connects to the final regression target value to achieve prediction.

[0074] Acoustic emission sequences defined by impact generally have two dimensions: acoustic emission features and time series. However, considering the different physical meanings of each acoustic emission feature, convolution operations on the feature dimension are not considered. Therefore, actual convolution operations are performed only on the time series dimension. The basic principles of the one-dimensional CNN framework are introduced below.

[0075] The i-th impact signal in the training dataset is x i The corresponding actual remaining bearing capacity is y. i The goal is to obtain data from a time window of length T. w The predicted value of the remaining bearing capacity is obtained from the impact signal sequence. This sequence is linked to T w A series of impact signals in time, and expressed as x i With the endpoint as the endpoint, a two-dimensional input matrix containing time-varying information is formed, as shown in Equation (13).

[0076]

[0077] This matrix serves as the input to the first convolutional layer, unlike the input to an ANN at a single time stamp. Through convolution operations and activation functions, the feature maps learned by the kernel filters on the input sequence are obtained, also known as intermediate outputs. While pooling layers' downsampling of these intermediate outputs saves computational cost, they also filter out some useful information. Therefore, one-dimensional CNN models do not incorporate pooling layers; their basic steps are briefly described below.

[0078] First, initialize the weights and biases from the visible units to the hidden units of the CNN model, and calculate the output of each layer through forward propagation, as shown in Equation (14).

[0079]

[0080] In the formula: and These represent the input and bias of the k-th neuron in the l-th layer, respectively. It is the output of the m-th neuron in the (l-1)-th layer. M is the filter kernel from the m-th neuron in layer (l-1) to the k-th neuron in layer l. l-1 and This represents the number of neurons and the convolution operator in layer (l-1). Since pooling layers are not considered, the intermediate output... It is also the output of the kth neuron in the lth layer, which can be obtained by passing the input to the activation function, as shown in Equation (15).

[0081]

[0082] Using a one-dimensional backpropagation technique, the error terms of each layer are calculated layer by layer from the last fully connected layer to the first convolutional layer. The loss function or loss layer describes how the training process penalizes the deviation between the predicted residual capacity and the corresponding actual residual capacity, and is defined as the mean square error (MSE) as shown in Equation (16).

[0083]

[0084] Where: MSE n u n and y n These represent the mean squared error, output, and true label corresponding to the nth input sequence, respectively. The total number of input sequences is N. x Here, the backpropagation of error is illustrated using only one input sequence as an example, since the error on the training set is simply the sum of the errors of all sequences. By using the chain rule, the weights and biases of each neuron in each layer can be calculated and updated using equation (17).

[0085]

[0086] In the formula: Indicates the error increment Therefore, the backward propagation of error from the first fully connected layer to the last convolutional layer can be expressed as Equation (18).

[0087]

[0088] The backpropagation of the input error increment from layer (l+1) to layer l can be expressed as equation (19).

[0089]

[0090] In the formula: f′ is the derivative of the activation function. The backpropagation of the input error from the (l+1)th layer to the lth layer is shown in Equation (20).

[0091]

[0092] In the formula: Let represent the array after rotating array c by 180°. Calculate the partial derivatives of the error with respect to the weights and biases according to equation (21).

[0093]

[0094] Equation (20) can be used to update the weights and biases at time t to the weights and biases at time (t+1) with a certain learning rate ε, as shown in Equation (21).

[0095]

[0096] Constructing a one-dimensional deep CNN model, such as Figure 4 The diagram primarily consists of four convolutional layers and two fully connected layers. The first three convolutional layers have a kernel size of 5×1 because convolution operations are performed only in the time series dimension, while the last convolutional layer has a kernel size of 3×1. These four convolutional layers use 20, 20, 10, and 5 filters, respectively. The first fully connected layer has 100 hidden units, while the second fully connected layer has only one unit because the final residual capacity is one-dimensional. Furthermore, each layer except the last one is followed by a ReLU activation function, and dropout layers with a dropout rate of 0.5 are placed after the first fully connected layer, the second convolutional layer, and the fourth convolutional layer.

[0097] After completing model training, the steps for predicting the remaining load-bearing capacity of composite plywood samples using the model are as follows:

[0098] Following the aforementioned method, acoustic emission tests were conducted on the composite material layup specimens to be tested, and acoustic emission signal data was collected. Acoustic emission characteristics and time series data were extracted from the acoustic emission signal waveforms, input into the model, and after calculation, the predicted results of the specimen's remaining load-bearing capacity were output.

[0099] The root mean squared error (RMSE) can be used as an evaluation index for prediction performance, and its definition is shown in equation (23).

[0100]

[0101] In the formula: and y i These represent the predicted and actual values ​​of the remaining carrying capacity, respectively.

[0102] This test used 5 ENF samples, yielding 10 datasets from different perspectives of the two sensors. Two datasets from the same test group were used as test datasets to verify the predictive performance of the model trained on the other 8 datasets. The high-dimensional sequence input of the model consisted of 50 concatenated impact signals, i.e., a time window length of 50. The model was trained for a maximum of 50 phases, with each phase referring to one forward and backward propagation of all training data within the model. For each phase, the input sequence was randomly shuffled into several batches, with the smallest batch containing 250 sequences. During training, the model used an adaptive moment estimation algorithm to update and optimize the weights in each layer. The initial learning rate for the first ten phases was set to 0.001 for rapid optimization, and thereafter decreased to three-tenths of the original learning rate every ten phases to achieve convergence.

[0103] The predicted results of the remaining bearing capacity when the two datasets for each specimen are used as the test set are as follows: Figure 5 As shown, the model predicts the continuous degradation process of the specimen's load-bearing capacity, not just the final value. It is evident that the predicted value of the remaining load-bearing capacity during loading is relatively close to the corresponding actual value, especially when the ENF specimen is nearing failure, i.e., the initial crack is about to propagate. This is because the load is greater in the later stages of testing, and interlaminar delamination and fiber fracture within the specimen begin to occur more frequently, generating more impact signals and significantly enhancing the characteristic accumulation, thus implying more information about the ENF specimen's load-bearing capacity. However, in the early stages of testing, the predicted remaining load-bearing capacity differs significantly from the actual value. At this time, the internal damage of the specimen mainly manifests as matrix cracking and fiber / matrix interface debonding. This is because acoustic emission signals are not continuously emitted from the specimen in the early stages of loading, resulting in less information in the input sequence. In the middle stages of testing, the trend of the predicted value differs from the actual trend. Specifically, the actual remaining load-bearing capacity of the specimen decreases monotonically with loading time, while the predicted value shows a fluctuating downward trend with loading time. During this stage, the damage evolves from matrix cracking and fiber / matrix interface debonding into a complex mixed damage mode. The generation rate of impact signals tends to stabilize, but is lower than in the later stages. Therefore, this is a period of continuous adjustment and correction of prediction accuracy. As the test continues, acoustic emission data streams are continuously generated, and the predicted values ​​gradually converge to the actual values. In summary, the prediction accuracy of the remaining load-bearing capacity of the ENF specimen varies depending on the stage of the test, but is higher in the later stages of the test, which is of most interest, with the final prediction error within 10N.

[0104] In summary, the prediction performance of the residual load-bearing capacity prediction model based on convolutional neural networks exhibits characteristics that vary with the damage stage of the sample. In the early stage, dominated by matrix cracking and fiber / matrix interface debonding, there is a large deviation between the predicted and actual values. In the middle stage, where the damage mode is more complex, the prediction deviation fluctuates significantly and gradually decreases. In the later stage, when the damage reaches a certain level, the predicted and actual values ​​tend to be consistent.

Claims

1. A method for predicting the remaining load-bearing capacity of composite layup material specimens based on dynamic data stream, characterized in that, include: Collect and process acoustic emission signal data; evaluate the importance of the signal's traditional time-frequency characteristics and cumulative characteristics; A convolutional neural network model was established and trained, and the model was used to predict the remaining load-bearing capacity of composite layup material specimens. The importance evaluation of the traditional time-frequency characteristics and cumulative characteristics of the signal specifically includes the following steps: (1) Extract multiple traditional time-frequency features from the collected acoustic emission signal waveforms, and add multiple cumulative features with interpretable physical meaning as objects of importance evaluation; (2) The traditional time-frequency features and cumulative features are evaluated by weight using the RReliefF algorithm, and then cross-validation is performed using the NCA algorithm to minimize the mean square error. The features with relatively higher weights are used to predict the degree of damage to the sample. The conventional time-frequency characteristics include at least the following 15 characteristics: rise time, count, energy, duration, peak amplitude, average frequency, root mean square, average signal level, initial frequency, signal strength, absolute energy, centroid frequency, peak frequency, rise angle, and decay angle; the cumulative characteristics include at least the following 9 characteristics: cumulative impact count, cumulative rise time, cumulative duration, cumulative energy, cumulative signal strength, cumulative absolute energy, cumulative count, cumulative average signal level, and cumulative root mean square. The establishment and training of the convolutional neural network model specifically includes the following steps: (a) Construct a one-dimensional deep convolutional neural network model, which consists of four convolutional layers and two fully connected layers arranged sequentially; the kernel size of the first three convolutional layers is 5×1, and the kernel size of the last convolutional layer is 3×1; the four convolutional layers use 20, 20, 10 and 5 filters respectively; the first fully connected layer has 100 hidden units, and the second fully connected layer has only one hidden unit; except for the last fully connected layer, each layer is followed by a ReLU activation function; after the first fully connected layer, the second convolutional layer and the fourth convolutional layer, there is a dropout layer with a dropout rate of 0.

5. (b) The high-dimensional sequence input of the model consists of 50 impact signals, i.e., the time window length is 50. The model can be trained for a maximum of 50 periods. Each training period refers to the process of all training data completing one forward and backward propagation in the model. For each training period, the input sequence is randomly shuffled into several batches, with the smallest batch containing 250 sequences. During the training process, the model uses an adaptive moment estimation algorithm to update and optimize the weights in each layer. The initial learning rate for the first ten training periods is set to 0.001 to achieve rapid optimization. After that, the learning rate is reduced to three-tenths of the original rate every ten periods to achieve convergence.

2. The method according to claim 1, characterized in that, The collection and processing of acoustic emission signal data specifically includes the following steps: (1) Conduct acoustic emission tests on composite material layup specimens with end notches under three-point bending loads and collect acoustic emission signal data; (2) Use multi-view learning techniques to increase the number of datasets and use acoustic emission data from different sensors to characterize the damage persistence process of the specimen.

3. The method according to claim 1, characterized in that, The method of predicting the remaining load-bearing capacity of composite layup material samples using this model specifically includes the following steps: Acoustic emission tests were conducted on the composite material layup specimens to be tested, and acoustic emission signal data was collected. Acoustic emission characteristics and time series data were extracted from the acoustic emission signal waveforms, input into the model, and after calculation, the predicted results of the remaining load-bearing capacity of the specimens were output.

Images