Method for intelligent extraction of multi-source data damage features and construction of health index of aircraft structure monitoring

Abstract

A method for intelligent extraction of damage features and construction of health indicators from multi-source data in aircraft structural monitoring is proposed. First, raw monitoring signal data from multiple sources, including acoustic emission waveform stream data and guided wave data, are acquired. A convolutional neural network (CNN) is then constructed to automatically learn deep features from the raw data. Expert knowledge features are extracted from the raw data and fused with the deep features of the CNN. The fused feature vector is then input into a classifier for classification. Next, a feature vector sequence is extracted using sliding window sampling to construct a full-lifecycle feature dataset. A health indicator construction model based on a deep CNN is established, employing a contrastive learning loss function as the optimization objective and iteratively training the model using gradient descent. Finally, the health indicators from the acoustic emission waveform stream data and guided wave data are fused from multiple sources. By utilizing the consistent changes in the data at the time of damage, the damage evolution process is comprehensively judged, and a fused health indicator is output. This invention improves the accuracy and reliability of the health indicators.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring and intelligent fault diagnosis technology, specifically relating to a method for intelligent extraction of multi-source data damage features and construction of health indicators for aircraft structural monitoring. Background Technology

[0002] Aircraft operate in harsh environments for extended periods, subjected to external and internal dynamic loads, atmospheric forces, and other factors. Damage can accumulate gradually without being detected, eventually leading to structural failure and even catastrophic accidents. Traditional structural health monitoring methods primarily rely on periodic manual inspections, which are costly, time-consuming, and require highly skilled operators. However, with the rapid development of sensing and signal analysis technologies, structural health assessment based on real-time monitoring data has become possible.

[0003] Deep learning methods can automatically extract deep features directly from raw data, but their diagnostic effectiveness is highly dependent on a large amount of training data. Traditional signal processing methods based on expert prior knowledge have strong feature targeting (Sun Zhijun, Xue Lei, Xu Yangming, et al. A review of deep learning research [J]. Computer Applications Research, 2012, 29(08):2806-2810.), but manual extraction is time-consuming, labor-intensive, and overly reliant on prior knowledge. Relying solely on either method has limitations, and there is an urgent need to organically integrate the two.

[0004] In terms of health indicator construction, existing supervised learning methods require a large amount of damage-labeled data, which is costly to acquire. Contrastive learning, as an unsupervised learning method, can learn degradation information from a small amount of unlabeled raw data and has strong generalization ability, making it effective for structural health monitoring. (Zhang Chongsheng, Chen Jie, Li Qilong, et al. A review of deep contrastive learning [J]. Acta Automatica Sinica, 2023, 49(01):15-39. DOI:10.16383 / j.aas.c220421.) In addition, traditional health indicator construction methods mostly rely on single-type data, while the acquisition of various types of sensor data involves randomness and contingency, resulting in insufficient data reliability. Research on health indicator construction based on multi-source data fusion is urgently needed. (She Daoming. Research on rolling bearing health assessment and remaining life prediction based on deep learning [D]. Southeast University, 2020. DOI:10.27014 / d.cnki.gdnau.2020.002903.) Summary of the Invention To overcome the shortcomings of the existing technologies, this invention provides a method for intelligent extraction of damage features and construction of health indicators from multi-source data in aircraft structure monitoring. By integrating deep features of convolutional neural networks with expert prior knowledge features, and an unsupervised health indicator construction framework based on comparative learning, this method achieves the extraction of damage evolution features and intelligent construction of trend-based health indicators from multi-source monitoring signals such as acoustic emission and guided waves, thereby improving the accuracy and reliability of health indicators.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent extraction of damage features and construction of health indicators from multi-source data in aircraft structural monitoring includes the following steps: Step 1: Acquire raw data of multi-source monitoring signals, including acoustic emission waveform stream data and guided wave data; Step 2: Construct a convolutional neural network to automatically learn deep features from the original data; extract expert knowledge features in the time domain, frequency domain, and time-frequency domain from the original data respectively; Step 3: Fuse the deep features extracted by the convolutional neural network with the expert knowledge features to form a fused feature vector; Step 4: Input the fused feature vector into the classifier for classification and recognition, and complete the extraction of damage evolution features; Step 5: Based on the original data of multi-source monitoring signals, feature vector sequences are extracted using sliding window sampling to construct a full life cycle feature dataset; a health indicator construction model based on deep convolutional neural network is established, using contrastive learning loss function as optimization objective, and the model is iteratively trained using gradient descent method; Step 6: Perform multi-source fusion of acoustic emission waveform stream data health indicators and guided wave data health indicators, utilize the consistent changes of various data types at the time of damage occurrence, comprehensively judge the damage evolution process, and output fused health indicators.

[0006] In step 2, the convolutional neural network consists of three network units, each of which comprises a convolutional layer and a pooling layer; the output feature map of the first convolutional layer is as follows: In the formula, These are the parameters of convolutional layer 1; For convolution kernel; Step size; For activation functions; after the first... The output after convolution and pooling is: In the formula, For pooling functions, The total number of layers; the input samples are processed through three layers of convolution and pooling to obtain deep features. The vector is flattened into a one-dimensional vector for feature fusion; lowercase l represents the convolutional layer number, i represents the feature map number, and s represents the pooling window size, where L=3.

[0007] In step 2, the time-frequency domain features are extracted using wavelet packet decomposition; the low-pass filter of wavelet packet decomposition... and high-pass filter Defined as: In the formula, It is a wavelet transform function; It is a scaling function; Indicates inner product calculation; , For system variables; The formulas for calculating the wavelet coefficients of each layer are as follows: In the formula, For length is The original data vibration signal; For the first Detail factor of the high-frequency part of the layer; For the first Approximation coefficients for the low-frequency portion of the layer.

[0008] Step 5, sliding window feature extraction, includes the mean. Pulse Indicators Peak indicators Margin indicators The calculation formulas are as follows: In the formula, In each window One amplitude point; The mean; For pulse indicators; Peak indicator; It is a margin indicator.

[0009] The contrastive learning loss function in step 5 is: In the formula, Amplification factor; For distance loss; and These are health indicator values ​​randomly selected from the same data at two different times. Update model parameters using gradient descent: In the formula, For model parameters, This is the learning rate.

[0010] In step 6, the signal characteristic changes are quantified using the Damage Index (DI); three damage factors are selected to characterize the damage features of the guided wave signal: The three damage factors are: cross-correlated damage factor DI1: In the formula, As a reference signal; For monitoring signals; This is the start time of the signal band; This is the end time of the signal band; Energy damage factor DI2 based on differential curve: In the formula, The sequence number of the signal sampling point; To select the total number of sampling points for the band; This is the difference curve between the reference signal and the monitoring signal; Difference signal energy damage factor DI3: In the formula, and The normalized forms of the reference signal and the monitoring signal are defined as follows: In the formula, As a reference signal; For monitoring signals; and These represent the start and end times of the signal band, respectively.

[0011] In step 6, the multi-source data fusion is based on the consistent changes of various types of data when damage occurs to make a comprehensive judgment: when a crack occurs, the acoustic emission sensor collects acoustic emission waveform data, and at the same time, the health index curve constructed by the guided wave data shows an inflection point. The two are compared in a comprehensive manner to accurately determine the time of crack initiation.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a feature extraction method that integrates expert knowledge and deep learning. It leverages prior expert knowledge to enhance the diagnostic accuracy of deep learning models on a small number of samples, reducing the dependence of deep learning on the amount of data. Furthermore, this invention utilizes an unsupervised framework based on contrastive learning to construct health indicators without requiring large amounts of labeled data. The constructed health indicators exhibit good trend characteristics and can detect structural damage earlier. Finally, this invention proposes a multi-source data fusion method for constructing health indicators, comprehensively utilizing acoustic emission waveform stream data and guided wave data. This overcomes the shortcomings of insufficient reliability from a single data source, improving the accuracy and reliability of health indicators. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method flow of an embodiment of the present invention.

[0014] Figure 2 The diagram shows the results of constructing health indicators for acoustic emission data as an example.

[0015] Figure 3 The diagram shows the results of constructing health indicators for guided wave data in an example.

[0016] Figure 4 The figure shows the result of constructing multi-source data index fusion for the test specimen in the example. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0018] Reference Figure 1 A method for intelligent extraction of damage features and construction of health indicators from multi-source data in aircraft structural monitoring includes the following steps: Step 1: Acquire raw data of multi-source monitoring signals, including acoustic emission waveform stream data and guided wave data; Step 2: Construct a convolutional neural network to automatically learn deep features from the original data; extract expert knowledge features in the time domain, frequency domain, and time-frequency domain from the original data respectively; The convolutional neural network constructed in this embodiment consists of three network units with similar basic structures, each network unit consisting of one convolutional layer and one pooling layer; for input samples The convolutional neural network first convolves it, and the feature map output by the first convolutional layer... for: In the formula, These are the parameters of convolutional layer 1; The kernel size; This is the step size value; The activation function is Leaky ReLU; in this embodiment, the kernel size is set to 3 and the stride value is set to 1. To reduce data dimensionality, the feature map obtained from the convolution layer is subjected to max pooling, with the following formula: In the formula, For the first The first convolutional feature map Each unit value; After the first pooling layer The first sampled feature map Each unit function value; This is the pooling window size; Therefore, after the [number]th [year], we can deduce that... The output after convolution and pooling is: In the formula, This is the pooling function; This is the total number of convolutional and pooling layers; in this embodiment, it is taken as... Input Sample After the feature extraction process of convolution C1 → pooling S1 → convolution C2 → pooling S2 → convolution C3 → pooling S3, the deep features learned by the convolutional neural network are obtained. This is flattened into a one-dimensional vector for subsequent feature fusion operations; lowercase l represents the convolutional layer number, i represents the feature map number, and s represents the pooling window size, where L=3.

[0019] This embodiment extracts expert knowledge features, including time-domain features, frequency-domain features, and time-frequency-domain features; Temporal feature extraction: In this embodiment, 14 types of temporal feature parameters are selected for feature fusion, including dimensional and dimensionless temporal features, mainly including maximum value, standard deviation, root square amplitude, kurtosis, impulse index, margin index, skewness index and kurtosis index. Frequency domain feature extraction: In this embodiment, eight types of frequency domain feature parameters are selected for feature fusion. Taking spectral energy as an example, its calculation formula is as follows: The center of gravity of the spectrum is In addition, it also includes three special indicators: spectral second moment, spectral kurtosis, and spectral skewness. Among them, spectral kurtosis is very sensitive to transient impulse components in the signal and can effectively identify impulses and their frequency band distribution from signals containing background noise. Time-frequency domain feature extraction: In this embodiment, wavelet packet decomposition is selected to extract time-frequency domain features from the vibration signal in the original data. Wavelet packet decomposition decomposes the vibration signal to obtain high- and low-frequency components, then obtains approximate coefficients from the low-frequency components and detail coefficients from the high-frequency components; the low-pass filter of wavelet packet decomposition... and high-pass filter The expression is: In the formula, It is a wavelet transform function; It is a scaling function; Indicates inner product calculation; , For system variables; The formula for calculating the wavelet coefficients of the vibration signal at each layer of wavelet packet decomposition is as follows: In the formula, For length is The original data vibration signal; For the first Detail factor of the high-frequency part of the layer; For the first Approximation coefficients for the low-frequency portion of the layer; Step 3: Fuse the deep features extracted by the convolutional neural network with the time domain, frequency domain, and time-frequency domain expert knowledge features respectively to form a fused feature vector; Step 4: Input the fused feature vector into the classifier for classification and recognition, complete the extraction of damage evolution features, and use expert knowledge features to enhance the diagnostic accuracy of the health indicator model on a small number of samples. Step 5: Based on the original data of multi-source monitoring signals, feature vector sequences are extracted using sliding window sampling to construct a full life cycle feature dataset; a health indicator construction model based on deep convolutional neural network is established, using contrastive learning loss function as optimization objective, and the model is iteratively trained using gradient descent method; This embodiment uses acoustic emission signals as an example to construct health indicators based on a contrastive learning framework. The specific process is as follows: Step 5.1: Obtain raw acoustic emission waveform stream data of aerospace structural components throughout their entire lifespan, from healthy to damaged. , indicating shared ownership From a set of different original data, select one data point as the test set and the rest as the training set. Step 5.2: Process the raw data of the full-lifetime acoustic emission waveform stream. (Include Sliding window sampling is performed on 1 data point, with a window size of 1. Calculate the feature values ​​(mean, variance, impulse index, peak index, margin index) within each window; arrange the feature vectors in chronological order to obtain the full lifecycle data. ,in Given a single feature vector, the length of the data sequence after feature extraction is... ; The formulas for calculating each feature are as follows, including the mean. : Pulse index : Peak Indicator : margin index : In the formula, In each window One amplitude point; The mean; For pulse indicators; Peak indicator; As a margin indicator; Step 5.3: Perform overlap sampling on the feature vector sequence obtained in Step 5.2 and label it in chronological order; the sampling window size is denoted as... The sampling interval is denoted as , obtain dataset Number of samples after oversampling ;sample ,in Indicates the first The length of each sample is The number of channels is , For sample labels; Step 5.4: Establish a health indicator model, using a deep convolutional neural network (DCNN) as the basic framework; in each training session, the sample set... Feature maps are calculated in the convolutional layers of the input model. The output result is then obtained through calculation by a fully connected layer. ,in , Indicates sample Corresponding health indicator values; Step 5.5: Output results from Step 5.4 Two different samples were randomly selected from the corresponding index values. , The contrastive learning loss is calculated using the following formula: In the formula, Amplification factor; For distance loss; and These are health indicator values ​​randomly selected from the same data at two different times. ; Step 5.6: During the training phase, using the contrastive learning loss from Step 5.5 as the optimization objective, update the model parameters using gradient descent. In the formula, These are model parameters; The learning rate; Step 5.7: Repeat steps 5.4 to 5.6 to iteratively optimize the health indicator model until the maximum number of iterations or the minimum loss value is reached; Step 5.8: Input the test set samples into the trained health indicators in chronological order to build the model and obtain the corresponding health indicators; Step 6: Perform multi-source fusion of acoustic emission waveform stream data and guided wave data health indicators. Utilize the consistent changes in various data types at the time of damage occurrence to comprehensively judge the damage evolution process and output reliable fused health indicators.

[0020] This embodiment uses guided wave data as an example, and utilizes the Damage Index (DI) to quantify signal characteristic changes; the data in sensor_17 file is selected as the reference signal, and the rest are monitoring signals. Three damage factors are selected to characterize the damage characteristics of the guided wave signal: Cross-correlated damage factors, denoted as DI1: In the formula, As a reference signal; For monitoring signals; This is the start time of the signal band; This is the end time of the signal band; The energy damage factor of the difference curve, denoted as DI2: In the formula, The sequence number of the signal sampling point; To select the total number of sampling points for the band; This is the difference curve between the reference signal and the monitoring signal; The difference signal energy damage factor, denoted as DI3: In the formula, and The normalized forms of the reference signal and the monitoring signal are defined as follows: In the formula, As a reference signal; For monitoring signals; and These represent the start and end times of the signal band, respectively.

[0021] This embodiment performs multi-source fusion of health indicators from acoustic emission waveform stream data and guided wave data, and comprehensively analyzes the consistency changes of various types of data when damage occurs, thereby improving the accuracy of crack initiation time prediction. By comparing the changes in acoustic emission waveform data and guided wave data when a crack occurs, it can be seen that when a crack occurs, the acoustic emission sensor collects acoustic emission waveform data, and at the same time, the health index curve constructed based on the guided wave data shows an inflection point, that is, different types of data change consistently when a crack occurs; by combining the two analyses, the time of crack occurrence can be accurately determined.

[0022] This embodiment uses a fracture test of an aluminum alloy plate in a laboratory environment as an example to verify the feasibility of the method of the present invention.

[0023] Acoustic emission experimental scheme: An aluminum alloy plate was subjected to cyclic loading from healthy to fracture. Acoustic emission waveform data was collected throughout the entire lifespan using two acoustic emission sensors. The experiment was conducted in four groups, with two channels collected in each group, resulting in a total of eight groups of full-lifetime acoustic emission waveform data. The acoustic emission waveform data was processed, retaining only the data from 30 minutes before crack initiation to the fracture of the test piece. One data point was randomly selected as the test set, and the other seven were used as the training set. Feature extraction (mean, variance, impulse index, peak index, margin index, etc.) was performed on the eight data points. Overlap sampling was performed on the feature sequences to obtain 4800 training samples and 1600 test samples, with each sample containing 300 sample points. The training set was input into the health index construction model for training. The training parameters of the health index construction model are shown in Table 1, and the model network structure parameters are shown in Table 2.

[0024] Table 1 Training parameters for the health indicator model Basic parameters Values Basic parameters Values Number of iterations 500 Batch size 64 Initial learning rate 0.0005 Learning rate decay rate 0.6 Learning rate update step 100 Table 2 Network structure parameters of the health indicator construction model Number Network layer Convolution kernel size Step Number of convolution kernels 1 Conv1d_1 1×7 1 256 2 Conv1d_2 1×5 1 256 3 Conv1d_3 1×5 1 128 4 Linear_1 5 Linear_2 After the model training is complete, the test set data is input into the health metric construction model to obtain the output results of the fully connected layer, such as... Figure 2 As shown, the health indicators constructed by the health indicator construction model in this embodiment show a clear upward trend over time, exhibiting good trend and monotonicity. Compared with the result of feature extraction only from the original data, the method of this invention can detect damage to the aluminum alloy plate earlier, verifying the superiority and applicability of the method of this invention.

[0025] Guided wave experiment scheme: Aluminum alloy plates were subjected to cyclic loading from healthy to fracture. Guided wave signals were acquired throughout the entire process using guided wave sensors. Four sets of experiments were conducted, yielding four sets of raw guided wave data for the entire lifespan. The raw data were analyzed, with the sensor response file at the maximum load value selected as the analysis object. Using the data in file #17 as the baseline signal, the damage characteristics of the guided wave signal were quantified using three damage factors (DI1, DI2, DI3). The changes in damage factors of each channel with fatigue crack propagation were analyzed, and the channel with significant monotonicity (channel S12) was selected as the main analysis object. Furthermore, a comparative learning method was used to construct guided wave signal health indicators, such as... Figure 3 As shown in the figure, the health index established by the model proposed in this study for guided wave data also shows a relatively obvious upward trend over time, indicating that the method is still effective for guided wave data.

[0026] Multi-source fusion experimental scheme: The experiment was conducted twice. Acoustic emission signals of the aluminum alloy plate were collected using an acoustic emission sensor, while a camera recorded the surface crack conditions. The time-domain variation of the acoustic emission signals was plotted using the raw acoustic emission data, and a health index was constructed using guided wave data. The two were combined and compared to predict the crack initiation time. Finally, the accuracy of the model was verified using the camera recording results. The experimental results are as follows: Figure 4 As shown, concentrated noise appears in the acoustic emission signal during crack propagation, and the health index constructed from guided wave data also shows a significant inflection point. Both of these correspond to the time of crack initiation and propagation. The comprehensive analysis results show that the multi-source data fusion method has high accuracy in predicting crack initiation time, overcomes the drawback of the single data source in the traditional health index construction method, and has good application prospects.

Claims

1. A method for intelligent extraction of damage features and construction of health indicators from multi-source data in aircraft structural monitoring, characterized in that, Includes the following steps: Step 1: Acquire raw data of multi-source monitoring signals, including acoustic emission waveform stream data and guided wave data; Step 2: Construct a convolutional neural network to automatically learn deep features from the raw data; Extract time-domain, frequency-domain, and time-frequency-domain expert knowledge features from the original data; Step 3: Fuse the deep features extracted by the convolutional neural network with the expert knowledge features to form a fused feature vector; Step 4: Input the fused feature vector into the classifier for classification and recognition, and complete the extraction of damage evolution features; Step 5: Based on the original data of multi-source monitoring signals, feature vector sequences are extracted using sliding window sampling to construct a full life cycle feature dataset; a health indicator construction model based on deep convolutional neural network is established, using contrastive learning loss function as optimization objective, and the model is iteratively trained using gradient descent method; Step 6: Perform multi-source fusion of acoustic emission waveform stream data health indicators and guided wave data health indicators, utilize the consistent changes of various data types at the time of damage occurrence, comprehensively judge the damage evolution process, and output fused health indicators.

2. The method according to claim 1, characterized in that, In step 2, the convolutional neural network consists of three network units, each of which comprises a convolutional layer and a pooling layer; the output feature map of the first convolutional layer is as follows: In the formula, These are the parameters of convolutional layer 1; For convolution kernel; Step size; For activation functions; after the first... The output after convolution and pooling is: In the formula, For pooling functions, The total number of layers; the input samples are processed through three layers of convolution and pooling to obtain deep features. The vector is flattened into a one-dimensional vector for feature fusion. Lowercase l represents the convolutional layer number, i represents the feature map number, and s represents the pooling window size. Here, L=3.

3. The method according to claim 1, characterized in that, In step 2, the time-frequency domain features are extracted using wavelet packet decomposition; the low-pass filter of wavelet packet decomposition... and high-pass filter Defined as: In the formula, It is a wavelet transform function; It is a scaling function; Indicates inner product calculation; , For system variables; The formulas for calculating the wavelet coefficients of each layer are as follows: In the formula, For length is The original data vibration signal; For the first Detail factor of the high-frequency part of the layer; For the first Approximation coefficients for the low-frequency portion of the layer.

4. The method according to claim 1, characterized in that, Step 5, sliding window feature extraction, includes the mean. Pulse Indicators Peak indicators Margin indicators The calculation formulas are as follows: In the formula, In each window One amplitude point; The mean; For pulse indicators; Peak indicator; It is a margin indicator.

5. The method according to claim 1, characterized in that, The contrastive learning loss function in step 5 is: In the formula, Amplification factor; For distance loss; and These are health indicator values ​​randomly selected from the same data at two different times. ; Update model parameters using gradient descent: In the formula, For model parameters, This is the learning rate.

6. The method according to claim 1, characterized in that, In step 6, the signal characteristic changes are quantified using the Damage Index (DI); three damage factors are selected to characterize the damage features of the guided wave signal: The three damage factors are: cross-correlated damage factor DI1: In the formula, As a reference signal; For monitoring signals; This is the start time of the signal band; This is the end time of the signal band; Energy damage factor DI2 based on differential curve: In the formula, The sequence number of the signal sampling point; To select the total number of sampling points for the band; This is the difference curve between the reference signal and the monitoring signal; Difference signal energy damage factor DI3: In the formula, and The normalized forms of the reference signal and the monitoring signal are defined as follows: In the formula, As a reference signal; For monitoring signals; and These represent the start and end times of the signal band, respectively.

7. The method according to claim 1, characterized in that, In step 6, the multi-source data fusion is based on the consistent changes of various types of data when damage occurs to make a comprehensive judgment: when a crack occurs, the acoustic emission sensor collects acoustic emission waveform data, and at the same time, the health index curve constructed by the guided wave data shows an inflection point. The two are compared in a comprehensive manner to accurately determine the time of crack initiation.

Images