Online Intelligent Prediction Method for Remaining Life of Aircraft Structures Based on Acoustic Emission

Abstract

An online intelligent prediction method for the remaining life of an aircraft structure based on acoustic emission is proposed. First, fatigue experiments are conducted on the aircraft structure to obtain raw data of multiple full-life acoustic emission waveform streams. Then, the raw data undergoes sliding window sampling feature extraction, and the parts with good monotonicity are selected to construct health indicators. A portion is randomly selected as the test set, and the remainder is used as the training set. The dataset is then segmented according to the 3σ principle to find the starting prediction point. Finally, in the testing phase, the test set data is input into the trained life prediction model to achieve online intelligent prediction of the remaining life of the aircraft structure. This invention utilizes a combination of DenseNet and contrastive learning to predict the remaining life of the aircraft structure online using intelligent algorithms, achieving good prediction results.

Description

Technical Field

[0001] This invention belongs to the field of aircraft structural health monitoring technology, specifically relating to an online intelligent prediction method for the remaining life of aircraft structures based on acoustic emission. Background Technology

[0002] Traditional aircraft structural maintenance relies on regular manual inspections and component replacements, which is not only time-consuming and labor-intensive but also sometimes involves unnecessary costs. Among various non-destructive testing techniques, acoustic emission signals can continuously monitor the actual health status of aircraft structures. By analyzing the characteristics of acoustic emission signals, such as signal frequency, energy distribution, and waveform characteristics, researchers can establish health status monitoring models. These models not only help identify existing damage but can also be used to predict potential future fatigue development, thus providing an important basis for assessing the remaining life of the structure. The introduction of this dynamic health monitoring technology can help aviation professionals analyze structural behavior in depth and enhance their control over structural safety. However, during the fatigue failure process of aircraft structural components, due to the complexity of the aircraft structure and harsh working conditions, it is easily affected by many factors such as internal and external dynamic loads and atmospheric pressure forces. As a result, the weak acoustic emission signals generated by fatigue crack initiation and propagation are submerged by massive background noise signals, leading to low-quality acoustic emission signal data.

[0003] Currently, acoustic emission signal processing methods can be mainly divided into two categories: parametric analysis and waveform analysis. Parametric analysis involves multi-parameter quantitative description of the acoustic emission signal. This method extracts characteristic parameters of the signal, such as energy, burst count, and frequency, for rapid, simple, and intuitive analysis. The advantages of parametric analysis lie in its speed and ease of understanding, allowing engineers to quickly obtain information about the structural health status. However, this method usually relies on certain prior knowledge, and the presence of environmental noise or other interference may affect the accuracy of the analysis results. Waveform analysis involves detailed analysis of the time-domain waveform and frequency-domain characteristics of the acoustic emission signal to identify and assess potential damage and defects in the material. This method mainly focuses on various characteristics such as waveform morphology, spectral distribution, duration, and amplitude. The advantages of waveform analysis are its strong anti-interference capability and applicability to various signal types. However, due to its complexity and the need for extensive computation and data processing, waveform analysis often requires a high level of technical expertise and a long analysis time, thus potentially posing challenges in large-scale engineering applications.

[0004] In recent years, data-driven structural health monitoring has shown superior performance in feature extraction and structural life prediction, providing a new approach for life prediction based on acoustic emission signals of aircraft structures. This method does not rely on a specific physical model, but automatically extracts features from the data through machine learning algorithms to establish a mapping relationship between the data and the structural health status, which has wider applicability and higher real-time performance. However, existing supervised learning methods often require a large amount of labeled data to establish a monitoring model ([1] Huang Yong, Bao Yuequan, Li Hui. Research progress on machine learning methods for structural state identification and evaluation [J]. Progress in Mechanics, 2023, 53(04): 774-792.), which greatly increases the cost of experiments.

[0005] Contrastive learning, as an unsupervised learning technique, does not rely on explicit labels, enabling it to learn even in the absence of labeled data. This characteristic makes it particularly suitable for handling large-scale unlabeled datasets, thereby reducing the cost and time of data labeling. Currently, there is no published literature on the application of contrastive learning for data-driven structural health monitoring and lifespan prediction. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide an online intelligent prediction method for the remaining life of aircraft structures based on acoustic emission. By combining DenseNet with contrastive learning, an intelligent algorithm is used to predict the remaining life of aircraft structures online, achieving good prediction results.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An online intelligent prediction method for the remaining life of an aircraft structure based on acoustic emission is proposed. First, fatigue experiments are conducted on the aircraft structure to obtain raw data of multiple acoustic emission waveforms throughout the entire lifespan. Then, the raw data is subjected to sliding window sampling feature extraction, and the parts with good monotonicity are selected to construct health indicators. A portion is randomly selected as the test set, and the remainder is used as the training set. The dataset is divided according to the 3σ principle to find the starting prediction point. Finally, in the testing phase, the test set data is input into the trained life prediction model to achieve online intelligent prediction of the remaining life of the aircraft structure.

[0009] An online intelligent prediction method for the remaining life of an aircraft structure based on acoustic emission includes the following steps:

[0010] Step 1: Collect raw data of the acoustic emission waveform flow of the aerospace structural component throughout its entire lifespan, from healthy state to failure process, denoted as O = {O1, O2, ..., O}. k}, which contains k distinct data records;

[0011] Step 2: Perform sliding window sampling on the raw data. During the sampling process, calculate 21 feature values ​​for each window, including waveform index, variance, standard deviation, mean, median, peak-to-peak value, maximum value, minimum value, root mean square frequency, root mean square value, mean square frequency, mean square value, skewness, skewness index, spectral variance, spectral skewness, spectral kurtosis, kurtosis, kurtosis index, impulse index, and margin index. Let the raw data O contain L O Using a sliding window sampling method with a window size of w, the calculated feature values ​​are combined into a feature vector and arranged in chronological order to obtain the full lifecycle dataset O. e ={V k}, 1≤k≤l, where each eigenvector V consists of n eigenvalues. This represents the length of the original data after feature extraction. The formula for calculating the features of the original data is as follows:

[0012]

[0013] In the formula, y i Let M represent the i-th point in each window, and S be the mean. 2 Here, ρ is the variance, PI is the impulse index, P is the peak value index, G is the margin index, w is the sampling window size, and n is the number of eigenvalues.

[0014] Step 3: Based on the 21 feature values ​​obtained in Step 2, plot the feature trend graph, and find features with good monotonicity to construct health indicators. The final dataset consists of eight features: impulse index, peak index, margin index, average amplitude, root square amplitude, maximum value, standard deviation, and spectral energy. V k ={f1,f2,…f8}, from the dataset One data point is randomly selected as the test set, and the remaining data is used as the training set.

[0015] Step 4: Construct the Health Indicator (HI) model. The basic network structure framework adopts DenseNet (Densely connected convolutional networks). The modules in DenseNet consist of "DenseBlock + Transition". The HI model includes 2 Dense Blocks, 2 Transition Layers, and 2 fully connected layers. After inputting the training set obtained in Step 3 into a single-layer convolution, the data undergoes dimensionality reduction in the Dense Blocks to decrease the number of parameters. Then, the Transition Layer reduces the size of the feature maps. Finally, a 3×3 max pooling operation with a stride of 2 is performed.

[0016]

[0017] Obtain the model output result N represents the total number of sample points, and x n,k is the health index corresponding to the sample point, and y n,k is the label of the sample;

[0018] Step 5: Calculate the loss using contrastive learning. Randomly select two different sample points from the model output result Y k obtained in Step 4 and denote them as a and b, where a < b. The specific calculation formula is as follows:

[0019] L d = -θ × (x a,k - x b,k )

[0020] θ is the amplification factor used to amplify the distance loss L d value;

[0021] Step 6: Repeat Step 4 - Step 5, gradually reduce the distance loss L d in Step 5, iterate and update the model parameters until the maximum number of iterations is reached, the training is completed, and the final prediction model is obtained;

[0022] Step 7: Input the test set into the prediction model trained in Step 6 in chronological order to construct the health index HI, obtain the image of the health index HI, and take the mean M p of the first p% of the data as the health state value of the aircraft structure and calculate the standard deviation σ p of the first p% of the data, and calculate the health index critical value P h of the aircraft structure health state. The specific calculation formula is as follows:

[0023] P h = M p - 3σ p

[0024] Step 8: Take the first sample point below the health index critical value P h as the starting prediction point, and input the data set O' e formed by the sample points after the starting prediction point into the prediction model trained in Step 6, so that it enters the fully connected layer after passing through 2 Dense Blocks and 2 Transition Layers, and finally obtain the remaining life prediction of the damaged section of the test set.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] This invention proposes an online intelligent prediction method for the remaining life of aircraft structures based on acoustic emission. First, features are extracted from the raw data of multiple full-lifetime acoustic emission waveform streams of the aircraft structure. Several features with good monotonicity are selected to construct health indicators, which significantly reduces the complexity of the raw data and shortens model training and testing time. The 3σ principle is used to find the initial prediction point, and a distance loss is constructed based on contrastive learning, improving the model's generalization ability and reducing the need for specialized knowledge. This method demonstrates good performance in predicting the remaining life of aircraft structural components under different environments, accurately determining the time of damage to the aircraft structure, thereby reducing the accident rate. Attached Figure Description

[0027] Figure 1 This is a flowchart of an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the lifetime prediction model according to an embodiment of the present invention.

[0029] Figure 3 This is a diagram illustrating the construction and initial prediction point location of the health indicator HI in an embodiment of the present invention.

[0030] Figure 4 This is a diagram illustrating the remaining lifetime prediction effect of an embodiment of the present invention. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings.

[0032] like Figure 1 As shown, an online intelligent prediction method for the remaining life of an aircraft structure based on acoustic emission includes the following steps:

[0033] Step 1: Collect the raw dataset of acoustic emission waveforms of aerospace structural components throughout their entire lifespan, from healthy state to failure process, denoted as O = {O1, O2, ..., O}. k}, which contains k distinct data records;

[0034] Step 2: Perform sliding window sampling on the original data. During the sampling process, calculate 21 characteristic values ​​for each window, including waveform index, variance, standard deviation, mean, median, peak-to-peak value, maximum value, minimum value, root mean square frequency, root mean square value, mean square frequency, mean square value, skewness, skewness index, spectral variance, spectral skewness, spectral kurtosis, kurtosis, kurtosis index, impulse index, and margin index. Specifically, let the original data O contain L O Using a sliding window sampling method with a window size of w, the calculated feature values ​​are combined into a feature vector and arranged in chronological order to obtain the full lifecycle dataset O. e ={V k}(1 ≤ k ≤ l), where each eigenvector V consists of n eigenvalues, represents the length of the original data after feature extraction. The calculation formula for the original data features is as follows:

[0035]

[0036] In the formula, y i represents the i-th point in each window, M is the mean, S 2 is the variance, PI is the pulse index, P is the peak index, G is the margin index, w is the sampling window size, and n is the number of eigenvalues;

[0037] Step 3: Draw a feature trend graph based on the 21 eigenvalues obtained in Step 2, find features with good monotonicity to construct a health index, and finally obtain a data set constructed with eight features: pulse index, peak index, margin index, average amplitude, root mean square amplitude, maximum value, standard deviation, and spectral energy V k = {f1, f2,... f8}. Randomly select one piece of data from the data set as the test set, and the remaining data as the training set;

[0038] Step 4: Construct a health index HI model. The basic network structure framework uses a densely connected network DenseNet (Densely connected convolutional networks). The module in the DenseNet network consists of "DenseBlock + Transition". The health index HI model contains 2 Dense Blocks, 2 Transition Layers, and 2 fully connected layers. After inputting the training set obtained in Step 3 into a single-layer convolution, the data reduces the number of parameters through dimensionality reduction in the Dense Block, and then reduces the size of the feature map by Trasition. Finally, a maximum pooling operation with a size of 3×3 and a stride of 2 is performed,

[0039]

[0040] to obtain the model output result N represents the total number of sample points, x n,k is the health index corresponding to the sample point, and y n,k is the label of the sample;

[0041] Step 5: Calculate the loss using contrastive learning. Randomly select two different sample points from the model output result Y k obtained in Step 4 and denote them as a and b (a < b). The specific calculation formula is as follows:

[0042] L d=-θ×(x a,k -x b,k )

[0043] θ is the magnification factor used to amplify the distance loss L. d The value;

[0044] Step 6: Repeat steps 4-5, gradually reducing the distance loss L in step 5. d The model parameters are iterated and updated until the maximum number of iterations is reached, training is complete, and the final prediction model is obtained.

[0045] Step 7: Input the test set into the prediction model trained in Step 6 in chronological order to construct the health index HI, obtain the image of the health index HI, and take the mean M of the top p% of data. p As a health status value of the aircraft structure, and the standard deviation σ of the first p% of the data is calculated. p Calculate the critical value P of the health index for the structural health status of the aircraft. h The specific calculation formula is as follows:

[0046] P h =M p -3σ p

[0047] Step 8: The first value below the health indicator threshold P h The sample points are used as the starting prediction points, and the dataset D′ is composed of the sample points after the starting prediction points. e Input the trained prediction model from step 6 into the fully connected layer after passing through two Dense Blocks and two Transition Layers, and finally obtain the remaining lifetime prediction of the damaged segments in the test set.

[0048] The embodiments use an experiment in a laboratory environment to test an aluminum alloy plate from a healthy state to fracture under cyclic loads to verify the feasibility of the present invention.

[0049] The experimental design involved collecting data on an aluminum alloy plate under cyclic loading from its healthy state to fracture. Acoustic emission sensors were placed at both ends of the plate to collect the data. The experiment was conducted four times, yielding eight raw acoustic emission waveform streams of the aluminum alloy plate throughout its lifespan. Feature extraction was performed on each of the eight raw data streams, and feature trend graphs were created. Eight features with good monotonicity were selected to form a feature vector, creating a new feature dataset. This dataset underwent time-overlap sampling, resulting in 8400 training samples and 1200 test samples. Each sample contained 300 data points. One sample was randomly selected as the test set, and the remaining seven were used as the training set. The training set was then input into a lifespan prediction model for training. The lifespan prediction model consisted of two dense blocks, two transition layers, and two fully connected layers, as shown below. Figure 2 As shown in Table 1, the model training parameters are shown in Table 2, and the model structure parameters are shown in Table 3.

[0050] Table 1 Model training parameters

[0051]

[0052] Table 2 Model Structure Parameters

[0053]

[0054] After the model is trained, the normalized dataset is used as the test set and input into the model to obtain the acoustic emission health indicators of the alloy plate, such as... Figure 3 As shown, by Figure 3 It is known that the aluminum alloy plate is in a transitional period from a healthy state to a damaged state. Since more attention is paid to the data of the aluminum alloy plate during the damage stage, the mean M of the first 40% of the data is taken. Then, the training set data is segmented based on the 3σ principle. The segmented dataset is then input into the trained DenseNet network to predict the remaining life of the aircraft structure. The prediction results are as follows: Figure 4 As shown, the prediction results are consistent with the overall trend of the actual data, and the lifespan prediction results have high accuracy, which verifies the superiority of this model.

Claims

1. A method for online intelligent prediction of the remaining life of an aircraft structure based on acoustic emission, characterized in that: First, fatigue tests were conducted on the aircraft structure to obtain raw data of multiple full-lifetime acoustic emission waveform streams. Then, the raw data were subjected to sliding window sampling feature extraction, and the parts with good monotonicity were selected to construct health indicators. A portion was randomly selected as the test set, and the remainder was used as the training set. The dataset was then divided according to the 3σ principle to find the starting prediction point. Finally, during the testing phase, the test set data is input into the trained life prediction model to achieve online intelligent prediction of the remaining life of the aircraft structure. The method includes the following steps: Step 1: Collect raw data of the acoustic emission waveform flow of the aerospace structural component throughout its entire lifespan, from healthy state to failure process, denoted as O = {O1, O2, ..., O}. k }, which contains k distinct data records; Step 2: Perform sliding window sampling on the raw data. During the sampling process, calculate 21 feature values ​​for each window, including waveform index, variance, standard deviation, mean, median, peak-to-peak value, maximum value, minimum value, root mean square frequency, root mean square value, mean square frequency, mean square value, skewness, skewness index, spectral variance, spectral skewness, spectral kurtosis, kurtosis, kurtosis index, impulse index, and margin index. Let the raw data O contain L O Using a sliding window sampling method with a window size of w, the calculated feature values ​​are combined into a feature vector and arranged in chronological order to obtain the full lifecycle dataset O. e ={V k }, 1≤k≤l, where each eigenvector V consists of n eigenvalues. This represents the length of the original data after feature extraction. The formula for calculating the features of the original data is as follows: In the formula, y i Let M represent the i-th point in each window, and S be the mean. 2 Here, ρ is the variance, PI is the impulse index, P is the peak value index, G is the margin index, w is the sampling window size, and n is the number of eigenvalues. Step 3: Based on the 21 feature values ​​obtained in Step 2, plot the feature trend graph, and find features with good monotonicity to construct health indicators. The final dataset consists of eight features: impulse index, peak index, margin index, average amplitude, root square amplitude, maximum value, standard deviation, and spectral energy. 1≤k≤l, V k ={f1,f2,…f8}; from the dataset One data point is randomly selected as the test set, and the remaining data is used as the training set. Step 4: Construct the Health Indicator (HI) model. The basic network structure framework adopts DenseNet (Densely connected convolutional networks). The modules in the DenseNet network consist of "DenseBlock + Transition". The HI model includes 2 Dense Blocks, 2 Transition Layers, and 2 fully connected layers. After inputting the training set obtained in Step 3 into a single-layer convolution, the data undergoes dimensionality reduction in the DenseBlock to decrease the number of parameters. Then, the Transition layer reduces the size of the feature maps. Finally, a 3×3 max pooling operation with a stride of 2 is performed. Obtain the model output results N represents the total number of sample points, x n,k y represents the health indicator corresponding to the sample point. n,k Labels for the samples; Step 5: Calculate the loss using contrastive learning, and obtain the model output result Y in Step 4 k Randomly select two different sample points from it and denote them as a and b, where a < b. The specific calculation formula is as follows: L d =-θ×(x a,k -x b,k ) θ is the magnification factor used to amplify the distance loss L. d The value; Step 6: Repeat steps 4-5, gradually reducing the distance loss L in step 5. d The model parameters are iterated and updated until the maximum number of iterations is reached, training is complete, and the final prediction model is obtained. Step 7: Input the test set into the prediction model trained in Step 6 in chronological order to construct the health index HI, obtain the image of the health index HI, and take the mean M of the top p% of data. p As a health status value of the aircraft structure, and the standard deviation σ of the first p% of the data is calculated. p Calculate the critical value P of the health index for the structural health status of the aircraft. h The specific calculation formula is as follows: P h =M p -3σ p Step 8: The first value below the health indicator threshold P h The sample points are used as the starting prediction points, and the dataset O' is composed of the sample points after the starting prediction points. e Input the trained prediction model from step 6 into a fully connected layer after passing through two DenseBlocks and two TransitionLayers, and finally obtain the remaining lifetime prediction of the damaged segments in the test set.

Images