Intelligent prediction method for residual life of aircraft structure based on fusion of acoustic emission and strain data

Abstract

The application relates to an aircraft structure residual life intelligent prediction method based on acoustic emission and strain data fusion, and belongs to the technical field of aerospace structure health monitoring. First, a fatigue machine is used to perform a fatigue damage test on an aircraft structure, original acoustic emission and strain data are collected, and feature extraction and data dimensionality increase are performed through sliding window sampling. Then, acoustic emission and strain data features are extracted through an LSTM network model and a DCNN network model respectively, and the features are spliced and fused through an attention mechanism. Finally, residual service life prediction of the aircraft structure is realized. The application realizes higher-precision real-time residual life prediction under complex service conditions.

Description

Technical Field

[0001] This invention belongs to the field of aircraft structural health monitoring technology, specifically relating to an intelligent prediction method for the remaining life of aircraft structures by fusing acoustic emission and strain data. Background Technology

[0002] With the rapid development of modern aircraft towards lightweight design, high thrust-to-weight ratio, and long service life, titanium alloys, aluminum alloys, and advanced composite materials are widely used in main load-bearing structures. However, in complex and variable service environments, aircraft structures are subjected to the coupled effects of aerodynamic loads, thermal stress, and mechanical vibrations over long periods, making them highly susceptible to inducing microscopic defects that gradually evolve into fatigue cracks. These cracks propagate continuously under alternating loads, leading to a significant decrease in structural stiffness and strength. This not only threatens flight safety but also drastically shortens the structural service life and may even cause catastrophic failure. To achieve early warning and accurate assessment of structural damage, structural health monitoring technology has become a key means of ensuring aircraft reliability.

[0003] Currently, acoustic emission (AE) monitoring and strain monitoring are two mainstream technologies in the field of structural health monitoring. AE technology captures elastic waves released from the propagation of microcracks within materials, enabling highly sensitive identification of early damage dynamics. Strain monitoring, on the other hand, uses strain gauges or fiber optic sensors to acquire real-time strain distribution on the structural surface, reflecting the overall mechanical response. However, single monitoring technologies have inherent limitations: AE signals are susceptible to environmental noise interference (such as engine vibration or airflow noise), leading to missed or false alarms of microcrack events; while strain monitoring provides macroscopic deformation information, it is insensitive to early microcracks and struggles to distinguish damage type and location. Furthermore, sensor deployment is limited by the complex curved surfaces of aircraft, and single data sources often only describe local damage characteristics, lacking multi-dimensional correlation analysis and failing to support high-precision remaining life prediction. Therefore, there is an urgent need to develop multi-source data fusion methods to improve the robustness and comprehensiveness of monitoring.

[0004] Although multi-source data fusion has attracted much attention in structural health monitoring, existing research mainly focuses on the combination of sensing technologies such as vibration, temperature, and pressure (Niu Kaixuan. Research on Real-time Health Monitoring System of Aero-engine Based on Multi-Source Sensor Data Fusion [J]. Modern Manufacturing Technology and Equipment, 2025, 61(04):72-74.DOI:10.16107 / j.cnki.mmte.2025.0239.), while research on the deep fusion of acoustic emission and strain monitoring data is still insufficient. Acoustic emission data contains information on the dynamic evolution of cracks, while strain data characterizes the static load-bearing state of the structure. The two differ significantly in time scale, signal characteristics, and noise mechanisms, and innovative fusion strategies are urgently needed to explore complementary features. Currently, few methods in the literature can effectively integrate the characteristics of acoustic emission events (such as energy and amplitude) with strain evolution trends (such as peak strain and strain variance) and directly apply them to the quantitative prediction of the remaining life of aircraft structures. This technological gap seriously restricts the accuracy of health assessment and makes it difficult to meet the high reliability maintenance requirements of modern aircraft. Therefore, developing a method for predicting the remaining life of aircraft structures based on the fusion of acoustic emission and strain monitoring data is of urgent engineering value for overcoming the limitations of single technologies and achieving global perception and accurate prediction of damage status. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention aims to provide an intelligent prediction method for the remaining life of aircraft structures by fusing acoustic emission and strain data. It employs Long Short-Term Memory (LSTM) networks and Deep Convolutional Neural Networks (DCNNs) to extract multi-scale features from acoustic emission signals and strain data, respectively. Then, it utilizes an attention mechanism for collaborative fusion to construct a remaining life prediction model capable of accurately quantifying the evolution trend of structural damage. This effectively solves the inherent defects of single monitoring technologies in terms of prediction accuracy, robustness, and time consistency, thereby significantly improving the accuracy and reliability of remaining life prediction. The present invention not only inherits the high sensitivity response advantage of acoustic emission technology to crack initiation and propagation but also integrates the stable characterization capability of strain monitoring for the overall mechanical state of the structure. By complementing multi-source information, it overcomes the shortcomings of single-source data in characterizing monotonic degradation trends and environmental adaptability. Compared with traditional single-technology prediction schemes, the present invention achieves higher-precision real-time prediction of remaining life under complex service conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An intelligent prediction method for the remaining service life of aircraft structures by fusing acoustic emission and strain data is proposed. First, fatigue failure tests are conducted on aircraft structural components using a fatigue machine to collect raw acoustic emission and strain data. Feature extraction and data dimensionality upscaling are performed through sliding window sampling. Then, acoustic emission and strain data features are extracted using LSTM and DCNN network models respectively, and attention mechanisms are used to stitch and fuse the features. Finally, the remaining service life of the aircraft structure is predicted.

[0007] A method for intelligent prediction of the remaining life of an aircraft structure by fusing acoustic emission and strain data includes the following steps: Step 1: Conduct a full life cycle test on the perforated component of the aircraft structural sample on a fatigue tester. The perforated component is subjected to cyclic loads from a healthy state until fracture. Acoustic emission and strain data are collected and recorded as follows: and This contains k distinct data records; Step 2: For the acquired acoustic emission and strain raw data, outliers are removed using the interquartile range (IQR) method; for scalar data sequences of any monitoring channel... The empirical cumulative distribution function determines the first quartile. and the third and fourth quartiles satisfy: in, Let be the empirical distribution function. For indicator functions; Define the interquartile range (IQR) as: The interquartile range (ICM) method is defined to set the outlier detection threshold range as follows: Among them, the threshold coefficient Only retain those that meet the requirements. Data points are used to generate purified data sequences. ; Step 3: Perform sliding window sampling on all purified datasets. During the sampling process, calculate 25 characteristic values ​​for each window of the acoustic emission data, including peak value, variance, standard deviation, mean, peak-to-peak value, maximum value, minimum value, root mean square frequency, root mean square value, mean square frequency, mean square value, skewness, skewness factor, spectral variance, spectral skewness, spectral kurtosis, kurtosis index, mean absolute value, root mean absolute value, kurtosis, waveform factor, impulse factor, peak factor, and spectral energy. Assume one original acoustic emission dataset. Include There are 10 data points, and the window size for sliding window sampling is 1000. The calculated eigenvalues ​​are combined into an eigenvector and arranged in chronological order to obtain a dataset representing the entire lifecycle of acoustic emission data. Each feature vector V consists of 25 feature values. The feature calculation formula for acoustic emission data is as follows: Here, PTP is the signal mean, PTP is the peak-to-peak value, and RootMAV is the root mean square absolute value. For skewness, For kurtosis, Standard deviation; For each strain signal, calculate its window. The five features—maximum, minimum, average, peak, and standard deviation—form a full lifecycle dataset of strain data. This completes the preprocessing of all raw acoustic emission and strain data; Step 4: For the preprocessed acoustic emission dataset It contains k datasets, and long-term dependencies are modeled using a multi-layer long short-term memory network. Its core recursive formula is: in This is the weight matrix. For bias terms, σ For sigmoid function, ⊙ represents the Hadamard product. , The hidden states are the process parameters during network transmission; a one-way LSTM model is used to preserve the causal temporal nature of crack propagation, and finally the hidden states are... h Enter a fully connected network: in As a Dropout mask (70% retention), output a 32-dimensional robust feature vector. ; For preprocessed strain datasets This dataset contains k data points. A one-dimensional convolutional neural network (1D-CNN) model is used to mine spatial-temporal coupled features. The output C of the convolutional layer is: in For convolution operations, The kernel size is [size]. For batch normalization, This represents the number of sensor nodes. Represents the first convolutional layer One output channel, For bias; eventually hide the state. g Input a fully connected network; in As a Dropout mask (70% retention), output a 32-dimensional robust feature vector. ; Step 5: Use an attention mechanism model to fuse the two types of data features; dynamically quantify the contribution weights of acoustic emission and strain data at each stage of structural damage evolution, and combine the spliced ​​features. Input a two-layer perceptron to generate normalized attention weights: in satisfy Here This represents the kurtosis characteristics of the AE signal. θ =4.5 is the standard critical threshold for crack propagation. k =0.8 is the attenuation coefficient; when It was considered to be the crack initiation stage at that time. →0 makes the model focus on strain data to capture overall stiffness degradation; when β 2> θ At this time, the crack propagation stage begins. α AE →1 Automatically enhances the weight of AE data to track dynamic crack evolution; Step 6: Repeat steps 4-5, gradually reducing the training loss, iterating and updating the optimized model parameters until the maximum number of iterations is reached, training is complete, and the final prediction model is obtained. Step 7: Acquire acoustic emission and strain data Data mean As a health status value of the aircraft structure and before calculation Standard deviation of data Calculate the critical values ​​of health indicators for the structural health status of the aircraft. The specific calculation formula is as follows: Step 8: Start with the first value below the health indicator threshold. The sample points are used as the starting prediction points. The dataset composed of the sample points after the starting prediction points is input into the prediction model trained in step 6. The acoustic emission and strain data are processed by the LSTM model and the 1D-CNN model respectively and then enter the fully connected layer to finally obtain the remaining lifetime prediction of the damaged segment of the test set.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention innovatively integrates acoustic emission dynamic damage signals and structural strain field static response data, achieving precise quantification of the structural degradation process through a deep feature synergy mechanism. First, time-frequency domain damage features are extracted based on the transient event characteristics of the acoustic emission signals, while spatial gradient analysis is performed on the strain data to obtain a characterization of the overall mechanical state of the structure. Then, a dual-channel neural network architecture is used to adaptively align and weightedly fuse the two types of heterogeneous data, constructing a physically meaningful joint damage characterization vector. This method successfully addresses the inherent limitations of acoustic emission signals under environmental noise and strain monitoring's limited sensitivity to early micro-damage—the acoustic emission branch accurately captures the dynamic process of crack initiation and propagation, while the strain branch stably reflects the structural stiffness degradation trend. The two complement each other through a physically driven attention mechanism. The final life prediction model not only possesses excellent monotonic degradation characteristics but also significantly improves robustness and damage identification sensitivity under complex operating conditions, providing an innovative solution for aircraft structural health management that combines physical interpretability and engineering practicality. Attached Figure Description

[0009] Figure 1 This is a flowchart of an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of the neural network model used in an embodiment of the present invention.

[0011] Figure 3 This is a comparison chart of the lifetime prediction effect of acoustic emission and strain data fusion and the effect of single-source lifetime prediction model in an embodiment of the present invention. Detailed Implementation

[0012] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings.

[0013] Reference Figure 1 A method for intelligent prediction of the remaining life of an aircraft structure by fusing acoustic emission and strain data includes the following steps: Step 1: Conduct a full life cycle test on the perforated component of the aircraft structural sample on a fatigue tester. The perforated component is subjected to cyclic loads from a healthy state until fracture. Acoustic emission and strain data are collected and recorded as follows: and This contains k distinct data records; Step 2: For the acquired acoustic emission and strain raw data, outliers are removed using the interquartile range (IQR) method; for scalar data sequences of any monitoring channel... The empirical cumulative distribution function determines the first quartile. and the third and fourth quartiles satisfy: in, Let be the empirical distribution function. For indicator functions; Define the interquartile range (IQR) as: Define the interquartile range (ICM) to set the outlier detection threshold range as follows: Among them, the threshold coefficient Only retain those that meet the requirements. Data points are used to generate purified data sequences. ; Step 3: Perform sliding window sampling on all purified datasets. During the sampling process, calculate 25 characteristic values ​​for each window of the acoustic emission data, including peak value, variance, standard deviation, mean, peak-to-peak value, maximum value, minimum value, root mean square frequency, root mean square value, mean square frequency, mean square value, skewness, skewness factor, spectral variance, spectral skewness, spectral kurtosis, kurtosis index, mean absolute value, root mean absolute value, kurtosis, waveform factor, impulse factor, peak factor, and spectral energy. Assume one original acoustic emission dataset. Include There are 10 data points, and the window size for sliding window sampling is 1000. The calculated eigenvalues ​​are combined into an eigenvector and arranged in chronological order to obtain a dataset representing the entire lifecycle of acoustic emission data. Each feature vector V consists of 25 feature values. The feature calculation formula for acoustic emission data is as follows: Here, PTP is the signal mean, PTP is the peak-to-peak value, and RootMAV is the root mean square absolute value. For skewness, For kurtosis, Standard deviation; For each strain signal, calculate its window. The five features—maximum, minimum, average, peak, and standard deviation—form a full lifecycle dataset of strain data. This completes the preprocessing of all raw acoustic emission and strain data; Step 4: For the preprocessed acoustic emission dataset It contains k datasets, and long-term dependencies are modeled using a multi-layer long short-term memory network. Its core recursive formula is: in This is the weight matrix. For bias terms, σ For sigmoid function, ⊙ represents the Hadamard product. , The hidden states are the process parameters during network transmission; a one-way LSTM model is used to preserve the causal temporal nature of crack propagation, and finally the hidden states are... h Enter a fully connected network: in As a Dropout mask (70% retention), output a 32-dimensional robust feature vector. The parameters of the LSTM model are shown in the table below: Table 1 LSTM Model Parameters serial number Network layer Input dimensions Output size 1 LSTM (batch, seq_len, 25) (batch, seq_len, 32) 2 Fully connected layer + activation function (batch, 32) (batch, 32) For the processed strain dataset This dataset contains k data points. A one-dimensional convolutional neural network (1D-CNN) model is used to mine spatial-temporal coupled features. The output C of the convolutional layer is: in For convolution operations, The kernel size is [size]. For batch normalization, This represents the number of sensor nodes. Represents the first convolutional layer One output channel, For bias; eventually hide the state. g Input a fully connected network; in As a Dropout mask (70% retention), output a 32-dimensional robust feature vector. The parameters of the 1D-CNN model are shown in the table below: Table 2 1D-CNN Model Parameters Serial Number Network layer Input dimensions Output size Parameter configuration 1 One-dimensional convolution + batch normalization (batch, 4, L) (batch, 16, L) Conv1d(4→16,kernel=3,padding=) 2 One-dimensional convolution + batch normalization (batch, 16, L) (batch, 32, L) Conv1d(16→32,kernel=3,padding=1) 3 Adaptive average pooling (batch, 32, L) (batch,32,1) AdaptiveAvgPool1d(1) 4 Fully connected layer + activation function (batch, 32) (batch, 32) Linear(32→32)+ReLU+Dropout(0.3) Step 5: Use an attention mechanism model to fuse the two types of data features; dynamically quantify the contribution weights of acoustic emission and strain data at each stage of structural damage evolution, and combine the spliced ​​features. Input a two-layer perceptron to generate normalized attention weights: in satisfy Here This refers to the kurtosis characteristics of the AE signal calculated in step 3. θ =4.5 is the standard critical threshold for crack propagation. k =0.8 is the attenuation coefficient; when It was considered to be the crack initiation stage at that time. →0 makes the model focus on strain data to capture overall stiffness degradation; when β 2> θ At this time, the crack propagation stage begins. α AE →1 Automatically enhances AE data weights to track dynamic crack evolution; attention mechanism model parameters are shown in the table below: Table 3 Parameters of the Attention Mechanism Model serial number Network layer Input dimensions Output size Parameter configuration 1 Fully connected layer + Tanh (batch, 64) (batch, 16) Linear(64→16)+Tanh 2 Fully connected layer + Softmax (batch, 16) (batch, 2) Linear(16→2)+Softmax Step 6: Repeat steps 4-5, gradually reducing the training loss, iterating and updating the optimized model parameters until the maximum number of iterations is reached, training is complete, and the final prediction model is obtained. Step 7: Acquire acoustic emission and strain data Data mean As a health status value of the aircraft structure and before calculation Standard deviation of data Calculate the critical values ​​of health indicators for the structural health status of the aircraft. The specific calculation formula is as follows: Step 8: Start with the first value below the health indicator threshold. The sample points are used as the starting prediction points. The dataset composed of the sample points after the starting prediction points is input into the prediction model trained in step 6. After the acoustic emission and strain data are processed by the LSTM model and the 1D-CNN model respectively, they enter the fully connected layer. Then, the attention mechanism dynamically allocates the weights of the two, and finally obtains the remaining lifetime prediction of the damaged segment of the test set. The overall neural network structure is as follows: Figure 2 As shown.

[0014] The performance of the lifetime prediction model using fused acoustic emission and strain data was compared with the results of lifetime prediction using either acoustic emission or strain data alone. The mean absolute error of each model was calculated to verify its superiority. Figure 3 As shown, the fusion model (red curve) demonstrates significantly better lifetime prediction performance than the acoustic emission model (green curve) and the strain model (blue curve). Furthermore, by calculating the mean absolute error between the predicted results and the actual lifetimes, the prediction performance of the three models can be further quantified and compared. It can be intuitively observed that the fusion model has the smallest mean absolute error, indicating the best prediction performance.

Claims

1. A method for intelligent prediction of the remaining life of an aircraft structure by fusing acoustic emission and strain data, characterized in that: First, fatigue failure tests are conducted on aircraft structural components using a fatigue machine to collect raw acoustic emission and strain data. Feature extraction and data dimensionality upscaling are then performed using sliding window sampling. Next, acoustic emission and strain data features are extracted using LSTM and DCNN network models respectively, and the features are spliced ​​and fused using an attention mechanism. Finally, the remaining service life of the aircraft structure is predicted.

2. The intelligent prediction method for the remaining life of an aircraft structure according to claim 1, characterized in that, Includes the following steps: Step 1: Conduct a full life cycle test on the perforated component of the aircraft structural sample on a fatigue tester. The perforated component is subjected to cyclic loads from a healthy state until fracture. Acoustic emission and strain data are collected and recorded as follows: and This contains k distinct data records; Step 2: For the acquired acoustic emission and strain raw data, outliers are removed using the interquartile range (IQR) method; for scalar data sequences of any monitoring channel... The empirical cumulative distribution function determines the first quartile. and the third and fourth quartiles satisfy: in, Let be the empirical distribution function. For indicator functions; Define the interquartile range (IQR) as: The interquartile range (ICM) method is defined to set the outlier detection threshold range as follows: Among them, the threshold coefficient Only retain those that meet the requirements. Data points are used to generate purified data sequences. ; Step 3: Perform sliding window sampling on all purified datasets. During the sampling process, calculate 25 characteristic values ​​for each window of the acoustic emission data, including peak value, variance, standard deviation, mean, peak-to-peak value, maximum value, minimum value, root mean square frequency, root mean square value, mean square frequency, mean square value, skewness, skewness factor, spectral variance, spectral skewness, spectral kurtosis, kurtosis index, mean absolute value, root mean absolute value, kurtosis, waveform factor, impulse factor, peak factor, and spectral energy. Assume one original acoustic emission dataset. ,Include There are 10 data points, and the window size for sliding window sampling is 1000. The calculated eigenvalues ​​are combined into an eigenvector and arranged in chronological order to obtain a dataset representing the entire lifecycle of acoustic emission data. Each feature vector V consists of 25 feature values. The feature calculation formula for acoustic emission data is as follows: Here, PTP is the signal mean, PTP is the peak-to-peak value, and RootMAV is the root mean square absolute value. For skewness, For kurtosis, Standard deviation; For each strain signal, calculate its window. The five features—maximum, minimum, average, peak, and standard deviation—form a full lifecycle dataset of strain data. This completes the preprocessing of all raw acoustic emission and strain data; Step 4: For the preprocessed acoustic emission dataset It contains k datasets, and long-term dependencies are modeled using a multi-layer long short-term memory network. Its core recursive formula is: in This is the weight matrix. For bias terms, σ For sigmoid function, ⊙ represents the Hadamard product. , The hidden states are the process parameters during network transmission; a one-way LSTM model is used to preserve the causal temporal nature of crack propagation, and finally the hidden states are... h Enter a fully connected network: in As a Dropout mask (70% retention), output a 32-dimensional robust feature vector. ; For preprocessed strain datasets This dataset contains k data points. A one-dimensional convolutional neural network (1D-CNN) model is used to mine spatial-temporal coupled features. The output C of the convolutional layer is: in For convolution operations, The kernel size is [size]. For batch normalization, This represents the number of sensor nodes. Represents the first convolutional layer One output channel, For bias; eventually hide the state. g Input a fully connected network; in As a Dropout mask (70% retention), output a 32-dimensional robust feature vector. ; Step 5: Use an attention mechanism model to fuse the two types of data features; dynamically quantify the contribution weights of acoustic emission and strain data at each stage of structural damage evolution, and combine the spliced ​​features. Input a two-layer perceptron to generate normalized attention weights: in satisfy Here This represents the kurtosis characteristics of the AE signal. θ =4.5 is the standard critical threshold for crack propagation. k =0.8 is the attenuation coefficient; when It was considered to be the crack initiation stage at that time. →0 makes the model focus on strain data to capture overall stiffness degradation; when β 2> θ At this time, the crack propagation stage begins. α AE →1 Automatically enhances the weight of AE data to track dynamic crack evolution; Step 6: Repeat steps 4-5, gradually reducing the training loss, iterating and updating the optimized model parameters until the maximum number of iterations is reached, training is complete, and the final prediction model is obtained. Step 7: Acquire acoustic emission and strain data Data mean As a health status value of the aircraft structure and before calculation Standard deviation of data Calculate the critical values ​​of health indicators for the structural health status of the aircraft. The specific calculation formula is as follows: Step 8: Start with the first value below the health indicator threshold. The sample points are used as the starting prediction points. The dataset composed of the sample points after the starting prediction points is input into the prediction model trained in step 6. The acoustic emission and strain data are processed by the LSTM model and the 1D-CNN model respectively and then enter the fully connected layer to finally obtain the remaining lifetime prediction of the damaged segment of the test set.

Images