Aviation Composite Workpiece Damage Diagnosis Method Based on Transfer Learning

Abstract

A damage diagnosis method for aircraft composite workpieces based on transfer learning, belonging to the fields of structural health monitoring and machine learning technology. The purpose of the present invention is to use the metric learning theory in transfer learning to transfer from the source domain of the aluminum plate damage signal obtained through simulation software to the target domain of the cantilever beam damage signal obtained through simulation software and experiments, a damage diagnosis method for aircraft composite workpieces based on transfer learning. The present invention takes the damage signal of the aluminum plate obtained in the simulation software as the source domain, uses the redistribution method to process the received signal, constructs a convolutional neural network model, transfers the convolutional neural network model trained with the source domain knowledge to the target domain, and applies the nearest neighbor method to classify the sample data in the target domain to obtain the damage diagnosis result. The present invention addresses the actual situation of few damage samples and unbalanced sample distribution, and realizes high-precision and high-accuracy identification of damage to aircraft composites.

Description

Technical Field

[0001] The present invention belongs to the technical fields of structural health monitoring and machine learning. Background Art

[0002] Advanced aircraft are developing towards the directions of high overload, complexity, precision, and strong robustness. There are extremely high requirements for the structural safety and reliability throughout the entire life cycle from production and manufacturing to service and retirement. However, during the long-term service process of advanced aircraft, they are inevitably affected by unsafe factors such as fatigue, perturbation, and impact, which damage the airframe structure and seriously affect the safety and reliability of the aircraft. Severe airframe structure damage may lead to accidents and cause damage to our own equipment and personnel. With the in-depth concept of structural health monitoring, in order to meet the requirement of ensuring the safe and stable operation of the structure, it is necessary to conduct research on intelligent damage diagnosis.

[0003] Currently, many traditional machine learning methods are applied to the damage diagnosis of aviation workpieces. These methods often use the damage signals received by sensors after Fourier transform as signal classification features and rely on some simple machine learning algorithms and neural network models. However, advanced aircraft face a harsh environment of high temperature, high speed, high load, strong perturbation, and strong corrosion during service. Under the combined influence of the harsh service conditions of the outside world, the key aluminum alloy components of the aircraft show complex and diverse failure modes. This makes it very difficult to directly obtain a dataset with sufficient data volume and balanced distribution of various fault samples through experiments on aircraft components. Traditional machine learning methods will face problems such as low diagnostic accuracy and poor generalization ability due to the lack of typical damage samples when conducting aircraft structural damage diagnosis. Summary of the Invention

[0004] The purpose of the present invention is to provide a damage diagnosis method for aviation composite workpieces based on transfer learning, which uses the metric learning theory in transfer learning to transfer from the source domain of aluminum plate damage signals obtained through simulation software to the target domain of cantilever beam damage signals obtained through simulation software and experiments.

[0005] The steps of the present invention are as follows:

[0006] S1. Use the redistribution method to process the received signals, and adopt data fusion means to obtain the time-frequency representation of the damage signals received by multiple sensors in the RGB image;

[0007] To approximate the actual service conditions of the aircraft, Gaussian noise is added to the damage signals, which is defined by its probability density function:

[0008] (1)

[0009] Where is the expectation, is the standard deviation, is the variance;

[0010] Relocate and clarify the fuzzy time-frequency points, which are defined by the following formula:

[0011] (2);

[0012] S2. Construct a convolutional neural network model, which includes a feature extractor, a long short-term memory mechanism, and an attention mechanism;

[0013] (1) The feature extractor includes 2 convolutional kernels of 2*2 and 3 convolutional kernels of 3*3, 5 ReLU layers, and a max pooling layer with a pooling window size of 3*3;

[0014] (2) The bidirectional long short-term memory network embeds new samples. After passing through the embedding function and processing, the output passes through a recurrent neural network again to strengthen the relationship between the source domain and the selected individuals. The formula is:

[0015] (3);

[0016] (3) The attention mechanism is based on the cosine similarity between the embedded representations of the new sample data and the sample data in the support set, as well as the softmax function. The formula is:

[0017] (4);

[0018] S3. Divide the source domain and the target domain into a training set, a validation set, and a test set according to the ratio of 7:2:1. Use the data in the source domain to train the convolutional neural network model to obtain a pre-trained convolutional neural network model;

[0019] S4. Transfer the convolutional neural network model trained with source domain knowledge to the target domain, and use the nearest neighbor method to classify the sample data in the target domain to obtain the damage diagnosis result;

[0020] The idea of classification is to map the support set in the source domain to a classifier . For a given target domain sample , Define a probability distribution about i.e., , where is parameterized by the network. Just use the network defined by to predict the label distribution of each target domain sample . .

[0021] The present invention processes damage signals using the redistribution method and trains a model using transfer learning. The damage signals of aerospace aluminum alloy workpieces are used as the source domain for pre-training. The distance relationship between the source domain data and the target domain data is analyzed using metric learning theory, and a damage diagnosis model for aerospace composite materials is quickly constructed to address the actual situation of a small number of damage samples and unbalanced sample distribution, achieving high-precision and high-accuracy identification of damage to aerospace composite materials. Description of the Drawings

[0022] Figure 1 is the flowchart of the specific implementation of the present invention;

[0023] Figure 2 is the workpiece drawing of the composite material cantilever beam;

[0024] Figure 3 is the flowchart of data preprocessing of damage signals by the redistribution method;

[0025] Figure 4 is the diagram of the convolutional neural network signal feature extractor. Detailed Implementation Manner

[0026] The technical problem to be solved by the present invention is to provide a damage diagnosis method for aerospace composite material workpieces based on transfer learning, which uses the metric learning theory in transfer learning to transfer from the source domain of aluminum plate damage signals obtained through simulation software to the target domain of cantilever beam damage signals obtained through simulation software and experiments. When facing the problems of a small number of samples and unbalanced sample distribution, this method can alleviate the dependence of the neural network on the target working condition fault training data.

[0027] The present invention adopts the following technical solutions to solve the above technical problems:

[0028] Step 1: Take the damage signals of the aluminum plate obtained in the simulation software as the source domain, and take the damage signals obtained from the composite material cantilever beam in the simulation software and the composite material cantilever beam in the experiment as the target domain.

[0029] Step 2: Use the redistribution method to process the received signals, and adopt data fusion means to obtain the time-frequency representation of the damage signals received by multiple sensors in the RGB image.

[0030] Step 3: Construct a convolutional neural network model, which includes a feature extractor, a long short-term memory mechanism, and an attention mechanism.

[0031] Step 4: Divide the source domain and the target domain into a training set, a validation set, and a test set according to the ratio of 7:2:1. Use the data in the source domain to train the convolutional neural network model to obtain a pre-trained convolutional neural network model.

[0032] Step 5: Transfer the convolutional neural network model trained with source domain knowledge to the target domain, and use the nearest neighbor method to classify the sample data in the target domain to obtain the damage diagnosis result.

[0033] In step 2, to approximate the actual service conditions of the aircraft, Gaussian noise is added to the damage signal, which is defined by its probability density function:

[0034] (1)

[0035] where is the expectation, is the standard deviation, is the variance.

[0036] In step 2, the redistribution method used can reveal the time and frequency characteristics of non-stationary signals in the time-frequency plane, estimate the local instantaneous frequency and group delay, and reposition and clarify the blurred time-frequency points, which is defined by the following formula:

[0037] (2).

[0038] In step 2, the data fusion method assigns red, blue, and green colors to multiple damage signals that have completed the short-time Fourier transform and combines them.

[0039] In step 3, the feature extractor designed in the model includes 2 convolutional kernels of 2*2 and 3 convolutional kernels of 3*3, with a padding depth of 0 and a stride value of 2; it includes 5 ReLU layers with a negative slope of 0.001; it includes a max pooling layer with a pooling window size of 3*3, a padding depth of 1, and a stride value of 2.

[0040] In step 3, the bidirectional long short-term memory network designed in the model embeds new samples. After passing through the embedding function and processing, the output is passed through a recurrent neural network again to strengthen the relationship between the source domain and the selected individuals. Its mathematical formula is:

[0041] (3).

[0042] In step 3, the attention mechanism designed in the model is based on the cosine similarity between the embedded representations of the new sample data and the sample data in the support set and the softmax function. Its mathematical formula is:

[0043] (4).

[0044] In step 5, the classification idea is to map the support set in the source domain to a classifier , for a given target domain sample , Define a probability distribution over i.e., , where is parameterized by the network, and only use the network defined by to predict the label distribution of each target domain sample . .

[0045] The embodiments of the present invention will be described in detail below, and examples of the embodiments are shown in the drawings:

[0046] According to the metric learning theory in transfer learning, the present invention extracts the damage signals of aluminum alloy workpieces obtained in the simulation software, and after the redispersion method - short Fourier transform and multi-sensor data fusion processing, forms a pre-trained model by loading, so as to realize the rapid diagnosis of the damage degree of aircraft composite material workpieces. Aiming at the problem that it is difficult to obtain all typical damage data samples under the actual service conditions of the aircraft due to the coupling effect of the external environment and internal stress, and it is impossible to train a traditional neural network. The present invention can complete the high-speed and high-accuracy diagnosis of the damage degree of aircraft composite material workpieces on the basis of a short training time and fewer training samples by introducing transfer learning.

[0047] The process of a method for diagnosing damage to aircraft composite material workpieces based on transfer learning according to the present invention is as Figure 1 shown, and the specific steps in this embodiment are as follows:

[0048] S1. Use the simulation software to construct an aircraft aluminum alloy workpiece and a composite material cantilever beam workpiece. The workpiece size is the same as the internal components of the actual aircraft airfoil NACA0012. The aluminum alloy is Al-2024, and the composite material grade is T300 / QY9512, and the corresponding ply sequence is [0° / 45° / 0° / 45° / 0°]. Design and manufacture a fully composite cantilever beam according to the above parameters, and the cantilever beam size and material settings are the same as those in the simulation software.

[0049] S2. Set different degrees of circular perforation damage on the two workpieces built in the simulation software. The damage size increases from a 1 mm diameter to 5 mm, forming six workpiece preset states including the healthy state. Perform damage pretreatment on the composite material cantilever beam built in the experiment, and set a 5 mm diameter circular perforation damage at the same damage position as the cantilever beam in the simulation software.

[0050] S3. Experimentally apply an excitation voltage signal at sensor 1 on the cantilever beam workpiece. The excitation signal is a 5-peak narrowband sine function with a center frequency of 100 kHz, and the time step is set to 10 -8s. After testing, guided waves in mode A0 can be obtained at this frequency. Sensors 2, 3, and 4 on the workpiece are responsible for signal reception. The signal excitation steps in the simulation software are the same as those in the experiment.

[0051] S4. Using the redistribution method, convert the received damage signals in the time domain to the time-frequency domain, and assign the signals received by the three sensors the colors red, yellow, and blue respectively. Then perform data fusion processing to obtain the RGB image of the damage signal, with the image size being 4101*247*3.

[0052] S5. Allocate the preprocessed damage signal images to the training set, validation set, and test set according to the ratio of 7:2:1.

[0053] S6. According to the Matching Networks transfer learning network model, change the feature extractor network to establish an aviation composite workpiece damage diagnosis network model. The main body of the updated feature extractor network consists of 2 3*3 convolutional kernels and 3 3*3 convolutional kernels. Then, add a pooling layer at the end of each convolutional kernel, and add a fully connected layer and a global average pooling layer at the end of the network structure.

[0054] S7. Send all types of damage pictures (assuming there are K pictures) in the source domain and 5 pictures of each of the 5 types of damage in the target domain into the designed feature extractor network to obtain their shallow variable representations. Input all the obtained K + 25 shallow variables into the long short-term memory network to obtain K + 25 outputs. Then use the cosine distance to judge the similarity between each of the first K outputs and each variable in the last 25 outputs. According to the calculated similarity and the label information in the source domain, solve the class labels in the target domain.

[0055] The present invention performs pre-training on a large amount of damage signal data obtained in the simulation software, avoiding the problem of insufficient prior knowledge caused by too small damage samples, and being able to fully learn the feature data of various damage degrees.

[0056] The present invention uses the transfer learning method and uses the metric learning method for transfer learning, reducing the demand for target domain parameters and effectively improving the accuracy and generalization ability of the model.

[0057] In order to enable the model to effectively distinguish damage categories, use a deeper convolutional neural network for feature extraction operations, increasing the receptive field of the model and deepening the network structure level. This enables the network to have a stronger ability to distinguish noise features and damage features in the signal, improving the accuracy of damage diagnosis.

[0058] The present invention is designed for composite workpieces with less damage data. Through repeated training of a large amount of data and continuous optimization of the algorithm, it provides a reliable auxiliary and reference tool for the structural health monitoring of aircraft.

Claims

1. A damage diagnosis method for aviation composite workpiece based on transfer learning, characterized in that: The steps are as follows: S1. Use the redistribution method to process the received signal, and adopt the data fusion means to obtain the time-frequency representation of the damage signals received by multiple sensors in the RGB image: （1） wherein is the expectation, is the standard deviation, is the variance; （2）； S2. Construct a convolutional neural network model, which includes a feature extractor, a long short-term memory mechanism, and an attention mechanism; (1) The feature extractor includes 2 2*2 and 3 3*3 convolutional kernels, 5 ReLU layers, and a max pooling layer with a pooling window size of 3*3; (2)The bidirectional long short-term memory network embeds the new samples. After passing through the embedding functions and processing, the output is passed through a recurrent neural network again to strengthen the relationship between the source domain and the selected individuals. The formula is: （3） (3) The attention mechanism is based on the cosine similarity between the embedded representations of the new sample data and the sample data in the support set and the softmax function, and its formula is: （4）； S3. Divide the source domain and the target domain into a training set, a validation set, and a test set according to the ratio of 7:2:1, and use the data in the source domain to train the convolutional neural network model to obtain a pre-trained convolutional neural network model; S4. Transfer the convolutional neural network model trained with the source domain knowledge to the target domain, and use the nearest neighbor method to classify the sample data in the target domain to obtain the damage diagnosis result; The idea of classification is to map the support set in the source domain to a classifier . For a given target domain sample , define a probability distribution over which is . Among them is parameterized by the network. Just use the network defined by to predict the label distribution of each target domain sample .​

Images