Acoustic emission waveform screening method based on auto-encoder and BP neural network
Through the acoustic emission waveform screening method based on the autoencoder and BP neural network, the problem of difficulty in dealing with complex nonlinear signals and noise interference in the prior art is solved, and a more accurate and efficient waveform screening effect is achieved.
Patent Information
- Application Number
- CN202411617160.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-13
- Publication Date
- 2025-03-11
AI Technical Summary
Existing waveform screening algorithms are difficult to process complex nonlinear signals and are susceptible to noise interference, resulting in inaccurate screening results.
The acoustic emission waveform screening method based on the autoencoder and BP neural network is adopted. By constructing the autoencoder and BP neural network, the training data set is created using a combination of loading experiments under various stressed states, the data is preprocessed and annotated, and the Adam optimization algorithm is used for training until the loss function converges.
Improve the generalization of the data, obtain more characteristic data as training data, and improve the accuracy and processing capabilities of waveform screening.
Smart Images
Figure CN119598282A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of waveform screening technology, and in particular to an acoustic emission waveform screening method based on an autoencoder and a BP neural network. Background Art
[0002] Acoustic emission (AE) technology is a non-destructive detection method that monitors the health of structures by capturing acoustic signals generated by crack propagation, cavity generation or other stress release processes inside the material. In tunnel lining monitoring, acoustic emission technology can detect the damage development of linings in real time and dynamically. However, due to the complex on-site environment of tunnels, a large amount of noise and irrelevant waveforms are often mixed in the acoustic emission signals, which increases the difficulty of signal processing and analysis. Therefore, how to screen out effective waveforms related to structural damage from massive acoustic emission data has become a research focus. Current waveform screening algorithms, such as threshold method, frequency domain analysis and principal component analysis, have difficulty in processing complex nonlinear signals and are easily interfered by noise, resulting in inaccurate screening results. Traditional machine learning methods rely on manual feature extraction, have limited processing capabilities, and cannot fully cope with multimodal and low-amplitude signals. Summary of the invention
[0003] The technical problem to be solved by the present invention is how to provide a method for screening effective waveforms with stronger processing capability.
[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: an acoustic emission waveform screening method based on an autoencoder and a BP neural network, comprising the following steps:
[0005] Step 1) The concrete test block is loaded with various stress forms, and the loading tests under various stress states are combined to create a training data set;
[0006] Step 2) Build an autoencoder and set autoencoder parameters;
[0007] Step 3) Determine the training strategy and use the training data set to train the autoencoder;
[0008] Step 4) Construct a BP neural network. The BP neural network structure includes 1 input layer, 2 hidden layers,
[0009] Step 5) labeling and classifying the training data set, recombining, and preprocessing the data as the input of the autoencoder, and the data of the hidden layer in the autoencoder is used as the input of the BP neural network to train the BP neural network;
[0010] Step 6) When the loss function converges, a model capable of screening effective waveforms of acoustic emission is obtained.
[0011] The beneficial effects of adopting the above technical solution are: loading concrete in various stress states, combining various loading tests, and forming a training database, so that the data sources in the database cover all stress states of concrete, improving the generalization of the data; constructing a suitable autoencoder to preprocess the data for training the BP neural network, and obtaining data with more characteristics as training data; refining the hyperparameters of the Adam optimization algorithm in the training of the autoencoder and the BP neural network, so that the training results are more likely to approach the optimal solution. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0013] Figure 1 It is a flow chart of the acoustic emission waveform screening method using autoencoder and BP neural network;
[0014] Figure 2 It is the structure diagram of the autoencoder model;
[0015] Figure 3 is the loss function graph of the autoencoder training;
[0016] Figure 4 This is the structure diagram of the BP neural network. DETAILED DESCRIPTION
[0017] In order to make the technical solution and advantages of the present invention clearer, the present invention is further described below. The present invention is specifically implemented according to the following steps:
[0018] like Figure 1 As shown, the embodiment of the present invention discloses an acoustic emission waveform screening method based on an autoencoder and a BP neural network, comprising the following steps:
[0019] First, create a training dataset.
[0020] The concrete test blocks were loaded with various stress forms including tension, compression, shear and bending. The size of the concrete test block is 460mm in length, 100mm in width and height, and the strength grade of the concrete test block is C35. The loading tests under various stress states were combined to form a training database, including 30 concrete axial compression tests, 20 axial tensile tests, 20 shear tests, and 20 concrete bending tests, as shown in Table 1.
[0021] Table 1 Statistics of various types of test parameters
[0022]
[0023] Then, the autoencoder structure is constructed. The structure of the autoencoder model is determined by experimental methods. First, the basic autoencoder model framework is built. The number of neurons in the input layer and output layer is 1024, and the number of neurons in the hidden layer starts from 512, and then gradually decreases by multiples. After each change, try to modify the model.
[0024] Initial training, and calculate its loss function value, the loss function calculation formula is shown in 1:
[0025]
[0026] Where n is the number of signal sample points, n is 1024; t is the correction factor, the value range is 0-1; x i is the predicted value, y i is the acoustic emission signal waveform input.
[0027] The loss function values corresponding to the number of neurons in different intermediate layers can be calculated, as shown in Table 2. It can be seen from the table that when the number of intermediate layers is 256, the loss function value is the smallest. Thus, the structure of the autoencoder model is determined as follows Figure 2 shown.
[0028] Table 2 Number of hidden layer neurons and loss function values
[0029]
[0030]
[0031] Then train the autoencoder. First, we need to determine the number of samples to be taken for one training. Here, we set the batch size to 1000 based on the total number of samples, randomly divide all samples into 1681 batches, and then perform 1681 trainings.
[0032] Secondly, the Adam improved optimization algorithm is used to perform first-order gradient optimization on the sample. The algorithm is shown in Formula 2:
[0033] m t =β1β2m t-1 +(1-β1)g t (2)
[0034] Where: m t is the momentum of the gradient, i.e., the first-order moment estimate at the current time t, g t is the gradient at the current time t, β1 is the decay rate of the first-order moment, which is taken as 0.90, and β2 is the correction coefficient, which is taken as 0.50.
[0035] Finally, the condition for stopping training is that the loss function converges. The training loss function graph is as follows: Figure 3 As shown in the figure, the training process stops after 1681 epochs.
[0036] When constructing a BP neural network, the number of hidden layers is generally one or two, and three or more layers are rarely used. The BP neural network structure includes 1 input layer, 2 hidden layers, and 1 output layer. The model structure is as follows Figure 4 shown.
[0037] The input layer contains 256 neurons, which are used to receive input data after standard data preprocessing and autoencoder processing; the middle part contains two hidden layers for feature extraction. The number of nodes in the hidden layer has a direct impact on the output of the BP neural network (i.e., it affects the recognition effect and recognition efficiency). If the number of nodes in the hidden layer is too small, the network will be underfitted and the error will be large; if the number of nodes in the hidden layer is too large, the network training time will be too long, and even "overfitting" will occur, resulting in increased errors. The number of neurons decreases in a 1 / 4 manner, to 64 and 16 respectively; the output layer contains 2 neurons, which are used for the output of classification judgment.
[0038] The weight matrix size between the input layer and the first hidden layer is [64,256], and the bias number of the first hidden layer neurons is 64; the weight matrix size between the first hidden layer and the second hidden layer is [16,64], and the bias number of the second hidden layer neurons is 16; the weight matrix size between the second hidden layer and the output layer is [2,16], and the bias number of the output layer neurons is 2. The model has 17522 weight parameters.
[0039] The specific steps of training the BP neural network are as follows:
[0040] (1) After classification, the training data samples are divided into two categories and marked with 0 and 1. Secondly, the marked samples are put back into the same directory to be mixed again into a data set with classification marks. Among them, the training data set is randomly shuffled separately, and the test data set is simply merged.
[0041] (2) Set the batch size of the training process to 1000 and the sampling method to random sampling. In the subsequent training process, the small batch data used is 12 batches, each with 1000 samples, and is obtained by random sampling. Construct a DataLoader class object to conveniently complete this task, and the small batch data used in each training can be retrieved through it.
[0042] (3) Preprocess the data. In this process, all waveform samples are put into the standard acoustic emission data preprocessing process. After conversion in this process, the output is the standard waveform data with 1024-dimensional features that are de-averaged, absolute, and normalized. This data is then input into the trained autoencoder for denoising and feature extraction, so that an output feature vector with a dimension of 256 can be obtained. The data in this form can be used as the input data of the BP neural network.
[0043] (4) Adjust the hyperparameters. The Adam improved optimization algorithm is still used here, and the training can be stopped when the loss function converges.
Claims
1. A method for screening acoustic emission waveforms based on autoencoders and BP neural networks, characterized in that: The following steps are involved: Step 1) The concrete test block is loaded with various stress forms, and the loading tests under various stress states are combined to create a training data set; Step 2) Build an autoencoder and set autoencoder parameters; Step 3) Determine the training strategy and use the training data set to train the autoencoder; Step 4) construct a BP neural network and select appropriate parameters; Step 5) labeling and classifying the training data set, recombining, and preprocessing the data as the input of the autoencoder, and the data of the hidden layer in the autoencoder is used as the input of the BP neural network to train the BP neural network; Step 6) After the training is completed, a model capable of screening effective waveforms is obtained.
2. The acoustic emission waveform screening method based on autoencoder and BP neural network according to claim 1 is characterized in that: The specific steps of step 1) are to load a concrete specimen with a size of 460 mm in length, 100 mm in width and height, and a strength grade of C35 using a variety of stress forms including tension, compression, shear, and bending, and combine the loading experiments under different stress states into a database.
3. The acoustic emission waveform screening method based on autoencoder and BP neural network according to claim 1 is characterized in that: The specific steps of step 2) are to select the number of nodes in the input layer and output layer of the autoencoder as 1024, the number of hidden layers as 1, and select the loss function: Where n is the number of signal sample points, n is 1024; t is the correction factor, the value range is 0-1; x i is the predicted value, y i is the acoustic emission signal waveform input; the loss function value corresponding to the number of neurons in different hidden layers is calculated by experiment, and it is determined that the number of hidden layer neurons when the loss function is the smallest is 256.
4. The acoustic emission waveform screening method based on autoencoder and BP neural network according to claim 1 is characterized in that: The specific steps of step 3) are to determine the number of samples extracted at a time; use an optimization algorithm to perform first-order gradient optimization, and the algorithm is as follows: m t =β1β2m t-1 +(1-β1)g t (2) Where: m t is the momentum of the gradient, i.e., the first-order moment estimate at the current time t; g t is the gradient at the current moment t; β1 is the decay rate of the first-order moment, which is 0.90; β2 is the correction coefficient, which is 0.50; training stops when the loss function converges.
5. The acoustic emission waveform screening method based on autoencoder and BP neural network according to claim 1 is characterized in that: In step 4), select appropriate parameters. The specific steps are as follows: Step 4.1) The number of hidden layers is two; Step 4.2) The number of neurons in the input layer is 256, and the number of neurons in the hidden layer is reduced in a 1 / 4 manner, to 64 and 16 respectively, and the number of neurons in the output layer is 2, which is used to output the classification result; Step 4.3) The weight matrix size between the input layer and the first hidden layer is [64, 256], and the bias number of the first hidden layer neurons is 64; the weight matrix size between the first hidden layer and the second hidden layer is [16, 64], and the bias number of the second hidden layer neurons is 16; the weight matrix size between the second hidden layer and the output layer is [2, 16], and the bias number of the output layer neurons is 2. The model has 17522 weight parameters.
6. The acoustic emission waveform screening method based on autoencoder and BP neural network according to claim 1 is characterized in that: The specific steps of step 5) are: Step 5.1) The classified training data samples are divided into two categories and marked as 0 and 1, and then reshuffled and combined into a training data set; Step 5.2) Set the batch size of the training process to 1000 and the sampling method to random sampling; Step 5.3) All waveform samples are put into the standard acoustic emission data preprocessing process and output as standard waveform data; this data is then input into the trained autoencoder for denoising and feature extraction to obtain an output feature vector with a dimension of 256; the data in this form is used as the input data of the BP neural network; Step 5.4) Adjust the hyperparameters and use the optimization algorithm for training. When the loss function converges, stop training.