Alpha / beta signal discrimination method based on improved bp neural network model

Abstract

The application discloses an alpha / beta signal discrimination method based on an improved BP neural network model, extracts corresponding signal features after reading and preprocessing original signal data which needs to be discriminated, trains a BP neural network model based on the extraction result, and realizes discrimination of alpha / beta signals through the trained BP neural network model; wherein, in the feature extraction process, the PCA (principal component analysis) is used to reduce the dimension of the preprocessed signal, so as to extract signal data with most signal features and realize compression of the data set. The application provides an alpha / beta signal discrimination method based on an improved BP neural network model, uses the PCA to extract features of the preprocessed signal data, and then inputs the features into the BP neural network model for continuous iteration training, so as to reduce the training time of the model and improve the recognition accuracy of the model to the alpha / beta signal.

Description

Technical Field

[0001] This invention relates to the field of deep learning. More specifically, this invention relates to an α / β signal discrimination method based on an improved BP neural network algorithm. Background Technology

[0002] Currently, there are two main methods for identifying α / β signals both domestically and internationally: pulse shape discrimination and pulse amplitude discrimination. Pulse amplitude discrimination is simple in principle and easy to implement, but it also has drawbacks such as excessively high crosstalk ratios, which can introduce significant errors into the discrimination results.

[0003] Pulse shape discrimination methods can be mainly divided into traditional analog electronic circuit discrimination methods and digital discrimination methods. Traditional methods for signal discrimination mainly include rise time methods, charge comparison methods, and zero-crosing methods, which are based on analog technology. However, these require specialized electronic equipment, lack flexibility, and are easily affected by external noise. Digital discrimination methods, on the other hand, primarily use hardware circuits to convert analog signals into digital signals and perform pulse reconstruction in an FPGA. They extract relevant pulse information for analysis and discrimination, including pulse gradient methods and wavelet transform methods. However, these methods require pre-setting feature points for input into the program, which is a significant drawback for datasets with unclear patterns and results in low signal discrimination accuracy. Summary of the Invention

[0004] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0005] To achieve these objectives and other advantages of the present invention, an α / β signal discrimination method based on an improved BP neural network model is provided. The method involves reading and preprocessing the original signal data to be discriminated, extracting the corresponding signal features, training a BP neural network model based on the extraction results, and then using the trained BP neural network model to discriminate α / β signals.

[0006] In the feature extraction process, the preprocessed signal is reduced in dimensionality using principal component analysis (PCA) to extract signal data with most of the signal features, thereby compressing the dataset.

[0007] Preferably, the signal features are configured to include:

[0008] The signal's sampling point data, the amplitude information of the signal's rising and falling edges, and the data information corresponding to the fast and slow components in the signal.

[0009] Preferably, the preprocessing is configured to include:

[0010] The input signal is smoothed and denoised using a smoothing filter.

[0011] The amplitude data of the input signal is normalized using the minimax method.

[0012] Preferably, the dimensionality reduction process is configured to include:

[0013] S10. Based on the preprocessed signal dataset, construct an n-dimensional covariance matrix R for the samples, used to store the covariance values ​​σ between different features in pairs. jk ,in:

[0014]

[0015] Where, μ j With μ k These are the means of features j and k, respectively. These are the k-th feature value in the i-th row and the j-th feature value in the i-th row, respectively;

[0016] S11. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix by solving |λI-R|=0 to obtain the eigenvalues ​​λ. j Its corresponding eigenvector L j In |λI-R|=0, I is a matrix composed of k identity orthogonal eigenvectors arranged column-wise;

[0017] S12. Select the eigenvectors corresponding to the first k largest eigenvalues, and construct a mapping matrix W using the first k eigenvectors, where k is the dimension of the new feature space;

[0018] S13. Transform the signal dataset into a new k-dimensional feature space using the mapping matrix W, where k≦n;

[0019] When selecting feature vectors containing more information to form a subset, the variance contribution rate of each feature vector is calculated using the following formula, the cumulative variance is calculated, and the principal components with a cumulative variance of more than 95% are selected as the feature vectors of the mapping matrix.

[0020]

[0021] Among them, e j This represents the variance contribution rate of each eigenvector.

[0022] Preferably, the BP neural network structure includes an input layer, a hidden layer, and an output layer, and the training method of the BP neural network model is configured to include:

[0023] S20. Forward propagation of training data: Before training the BP neural network, the connection weights between the layers of the network structure and the thresholds of the hidden and output layers are initialized.

[0024] S21. Use the extracted signal data as training data and propagate it forward layer by layer from the input layer to the output layer. During the propagation process, continuously optimize the connection weights and threshold data. When the output value of the output layer meets the expected value or the number of iterations reaches the predetermined value, proceed to S23. When the difference between the output value of the output layer and the expected value exceeds the predetermined requirement, proceed to S22 to propagate the error backward.

[0025] S22. The initial weight values ​​obtained from the training data reaching the output layer are propagated backward from the output layer to the next layer. The difference between the output value and the expected value is gradually corrected through backward propagation layer by layer. At the same time, the error is fed forward to correct the weights between neurons in each layer. When the output value of the output layer meets the expected value or the number of iterations reaches the predetermined value, the process enters S23.

[0026] S23, save the currently trained network model as the judgment model, and use the other part of the signal data after feature extraction as test data to test the judgment model.

[0027] The present invention has at least the following beneficial effects: The identification method of the present invention mainly utilizes the BP neural network algorithm, which does not rely on any external electronic circuits or preset any feature points. It can quickly and accurately identify the type of signal by directly inputting the effective signal segment into the model. It can effectively identify datasets with unclear data patterns. Specifically, the method of the present invention uses principal component analysis (PCA) to extract features from the preprocessed signal data, and then inputs them into the BP neural network model for continuous iterative training, thereby reducing the training time of the model and improving the model's recognition accuracy of α / β signals.

[0028] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the α / β signal discrimination method based on the improved BP neural network model of the present invention;

[0030] Figure 2 This is a structural diagram of the BP neural network model of the present invention;

[0031] Figure 3 This is a flowchart of the BP neural network algorithm of the present invention. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0033] like Figure 1 As shown, an α / β signal discrimination method based on an improved BP neural network mainly consists of the following steps:

[0034] Step 1: Read signal data: The BP neural network model includes an input layer, which takes in the raw signal data to be identified. The signal information contained in the raw signal data includes: the sampling point data of the signal, the amplitude information of the rising and falling edges of the signal, and the data information of the fast and slow components in the signal.

[0035] Step 2: Data Preprocessing: The collected raw data inevitably contains electronic white noise, and the magnitude of the signal amplitude has a certain impact on the signal discrimination result. It is necessary to denoise and normalize the raw data for further optimization. In this step, the input signal is denoised by smoothing filtering to reduce the impact of electronic white noise on signal discrimination. Since the amplitude difference between α and β signals is large, in order to avoid the impact of amplitude difference on the discrimination effect, the amplitude data of the input raw signal is normalized by the minimax method to facilitate the extraction of the effective signal part in the next step.

[0036] Step 3: Feature Extraction: Principal Component Analysis (PCA) is used to reduce the dimensionality of the preprocessed signal, extracting signal data with most of the signal features and further compressing the dataset. This step uses PCA to extract features from the preprocessed data, reducing the dimensionality of the data input to the BP neural network model.

[0037] Step 4: Training the BP Neural Network Model: The BP neural network structure consists of three parts: an input layer, hidden layers, and an output layer. Its training process includes two steps: forward propagation and backward propagation of the error. First, before training the BP neural network, the connection weights between each layer and the thresholds of the hidden and output layers are initialized. Then, the connection weights and thresholds are continuously optimized and adjusted using training data to achieve the purpose of training the network structure. When the training data reaches the output layer, an initial weight value is obtained. Then, the data is propagated backward from the output layer to the previous layer, gradually correcting the difference between the output value and the expected value from one layer to the next, and feeding the error forward to continuously adjust the weights between neurons in each layer, making the output value of the neural network closer to the expected value. The network model at which training is complete (training completion means the error accuracy reaches the requirement, or the iteration ends) is saved as the judgment model. This step trains the BP neural network model by continuously reducing the error between the predicted value and the true value.

[0038] Step 5: Signal identification: Use the judgment model in the BP neural network model to identify the type of the input signal data.

[0039] like Figure 2 The illustrated BP neural network topology is a multilayer feedforward network with n input neurons, m output neurons, and l hidden neurons. The learning rate of the neural network model is η, and the activation of the hidden layer and output layer neurons uses the Sigmoid function and the Tanh function, respectively. The threshold of the j-th hidden layer neuron is denoted by a. j This indicates that the threshold of the k-th neuron in the output layer is represented by b. k This indicates that the connection weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer is ω. ij The connection weight between the j-th neuron in the hidden layer and the k-th neuron in the output layer is ω. jk Let H be the input received by the j-th neuron in the hidden layer. j for:

[0040]

[0041] The input O received by the k-th neuron in the output layer k for:

[0042]

[0043] For training example (x) k y k Assume the output of the neural network is y. k = (y1, y2, ..., y m ),Right now:

[0044]

[0045] Then the network is in (x k y k The error on ) is:

[0046] e k =y k -O k (3-4)

[0047] exist Figure 2 In the network structure, (m+n+1)*l+n parameters need to be determined: n*l weights from the input layer to the hidden layer, l*m weights from the hidden layer to the output layer, l hidden layer thresholds, and m output layer thresholds. The BP neural network algorithm is an iterative learning algorithm; in each iteration, it uses a generalized perceptron learning rule to update and estimate the parameters. Based on the error e... k For network connection weights ω ij With ω jk To perform updates and iterations, that is:

[0048]

[0049] ω jk =ω jk +ηH j e k (3-6)

[0050] Based on the network prediction error e k Update the threshold a of the network node j b k ,Right now:

[0051]

[0052] b k =b k +e k (3-8)

[0053] The implementation steps of a standard BP neural network are as follows: Figure 3 as follows:

[0054] Step 1: Create a BP neural network structure and initialize the network parameters. Set a random number (-1, 1) for the threshold and weights of the network, set the learning rate η to a constant (0, 1), and give a maximum number of iterations D and a computational precision ε.

[0055] Step 2: Calculate the output. Input the training sample data into the neural network, and calculate the outputs of the hidden layer and the output layer using formulas (3-1) and (3-2) respectively;

[0056] Step 3: Calculate the network error E using formula (3-4);

[0057] Step 4: Adjust the weights and thresholds of the network accordingly based on the feedback of the error E. Adjust the weights of the hidden layer and the output layer according to formulas (3-5) and (3-6), and then adjust the thresholds of the two layers according to formulas (3-7) and (3-8).

[0058] Step 5: End Judgment. Check if the network error meets the accuracy requirements. If E < ε or the maximum number of training iterations D is reached, the algorithm training ends; otherwise, retrain all samples and proceed to Step 2.

[0059] In summary, this invention is based on the BP neural network method and uses the identification of 10,000 α / β raw signals (5,000 α signals and 5,000 β signals, with 500 signal feature points for each signal) collected by the liquid scintillator as an example to illustrate the usage of this invention.

[0060] The acquired raw dataset is used as input data into the input layer. Since the acquired raw liquid scintillator signal data inevitably contains some electronic noise, a smoothing filter is needed to further denoise the signal and reduce the impact of electronic white noise on signal discrimination. Furthermore, the amplitude difference between the α and β signals is large. To avoid the impact of amplitude difference on the discrimination effect, the minimax method is used to further normalize the denoised signal.

[0061] Principal component analysis (PCA) was used to extract features from the preprocessed signal, extracting feature points that contain 95% of the signal information data, resulting in the final dataset A, which has a size of 10,000, with 11 feature points for each signal.

[0062] The PCA algorithm utilizes information compression and dimensionality reduction techniques to replace a large dataset with fewer principal components, transforming high-dimensional data into low-dimensional data. This ensures that these independent principal components represent as much information as possible from the original data. The implementation steps are as follows:

[0063] Step 1: Standardize the original dataset (using the minimax method);

[0064] Step 2: Construct the covariance matrix R of the samples. Construct a 10000*500 dimensional covariance matrix R that is symmetric along the main diagonal, where 500 is the dimension of the matrix. This matrix stores the covariance values ​​σ between different features in pairs. jk ,in:

[0065]

[0066] Where, μ j With μ k These are the means of features j and k, respectively. These are the k-th feature value in the i-th row and the j-th feature value in the i-th row, respectively, where the values ​​of j and k are both less than the number of feature points in the input signal;

[0067] Step 3: Calculate the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​λ are obtained by solving |λI-R|=0. j Its corresponding eigenvector L j In |λI-R|=0, I is a matrix composed of k identity orthogonal eigenvectors arranged column-wise;

[0068] In the covariance matrix, the eigenvectors represent the principal components (i.e., the directions of maximum variance), and the magnitude of their corresponding eigenvalues ​​determines the importance of the eigenvectors.

[0069] Step 4: Select the eigenvectors corresponding to the first k largest eigenvalues, and construct a mapping matrix W using the first k eigenvectors, where k is the dimension of the new feature space, k≦500;

[0070] Step 5: Transform the 500-dimensional original dataset to a new k-dimensional feature space using the mapping matrix W. When selecting the feature vectors containing the most information to form a subset, calculate the variance contribution rate of each feature vector (principal component) using the following formula, calculate the cumulative variance, and select the principal components with a cumulative variance of 95% or higher as the feature vectors of the mapping matrix. This ensures that the extracted low-dimensional data contains most of the feature information of the original signal data. The transformed k value is 11.

[0071]

[0072] Among them, e j This represents the variance contribution rate of each eigenvector.

[0073] Build a BP neural network model with an input layer, one hidden layer, and an output layer. Based on the optimized dataset A, determine the number of input nodes to be 11 and the number of hidden nodes to be 7. Set a random number (-1, 1) for the threshold and weights of the network, set the learning rate η to be 0.01, and give the maximum number of iterations D = 10000 and the computational precision ε = 0.01. Determine the number of output points to be 2 as needed to identify the number of signal categories. Output 0 represents the α signal and output 1 represents the β signal.

[0074] The final dataset A is divided into a test set and a training set in a 2:8 ratio. The divided training set is then input into the established BP neural network model. By continuously adjusting the parameters and iteratively training, the final test set achieves a discrimination accuracy of over 98%.

[0075] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0076] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0077] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for α / β signal discrimination based on an improved BP neural network model, characterized in that, After reading and preprocessing the original signal data that needs to be identified, the corresponding signal features are extracted. Based on the extraction results, a BP neural network model is trained, and the α / β signal is identified through the trained BP neural network model. In the feature extraction process, the preprocessed signal is reduced in dimensionality using principal component analysis (PCA) to extract signal data with most of the signal features, thereby compressing the dataset. The dimensionality reduction process is configured to include: S10. Based on the preprocessed signal dataset, construct an n-dimensional covariance matrix R for the samples, used to store the covariance values ​​between different features in pairs. ,in: in, and These are features j and k The mean, 、 They are the first i Line number k The eigenvalue and the eigenvalue i Line number j One eigenvalue; S11. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix by solving... Obtain eigenvalues Its corresponding feature vector ,in, middle I It is a matrix composed of k identity orthogonal eigenvectors arranged column by column; S12, Select and Previous k The eigenvectors corresponding to the largest eigenvalues ​​are used to construct a mapping matrix from the first k eigenvectors. W ,in k The dimension of the new feature space; S13, via mapping matrix W Transform the signal dataset into a new k-dimensional feature space, where k ≦ n; When selecting feature vectors containing more information to form a subset, the variance contribution rate of each feature vector is calculated using the following formula, the cumulative variance is calculated, and the principal components with a cumulative variance of more than 95% are selected as the feature vectors of the mapping matrix. in, This represents the variance contribution rate of each eigenvector; Given a multilayer feedforward network structure with n input neurons, m output neurons, and l hidden neurons in a BP neural network topology, for training examples... Assuming the output of the neural network is Then the BP neural network in Error on for: In the above formula, This is the input received by the k-th neuron in the output layer; Based on network prediction error The following formula is used to assign network connection weights. and Update and iterate: Based on network prediction error The threshold of network nodes is updated using the following formula. , : In the above formula, For the hidden layer j Threshold of a neuron For the output layer k The threshold of a neuron; For the hidden layer j The nth neuron and the output layer k The connection weights between neurons For the hidden layer j The input received by each neuron η This represents the learning rate of the neural network model. The BP neural network model determines the number of input nodes to be 11 and the number of hidden layer nodes to be 7 based on the optimized dataset A. A random number (-1, 1) is set for the threshold and weights of the network, the learning rate η is set to 0.01, and the maximum number of iterations is given as D=10000 and the calculation precision ε=0.

01. The number of output points of the network is determined to be 2 according to the number of signal categories to be identified. Output 0 represents the α signal and output 1 represents the β signal.

2. The α / β signal discrimination method based on the improved BP neural network model as described in claim 1, characterized in that, The signal features are configured to include: The signal's sampling point data, the amplitude information of the signal's rising and falling edges, and the data information corresponding to the fast and slow components in the signal.

3. The α / β signal discrimination method based on the improved BP neural network model as described in claim 1, characterized in that, The preprocessing is configured to include: The input signal is smoothed and denoised using a smoothing filter. The amplitude data of the input signal is normalized using the minimax method.

4. The α / β signal discrimination method based on the improved BP neural network model as described in claim 1, characterized in that, The BP neural network structure includes an input layer, a hidden layer, and an output layer, and the training method of the BP neural network model is configured to include: S20. Forward propagation of training data: Before training the BP neural network, the connection weights between the layers of the network structure and the thresholds of the hidden and output layers are initialized. S21. Use the extracted signal data as training data and propagate it forward layer by layer from the input layer to the output layer. During the propagation process, continuously optimize the connection weights and threshold data. When the output value of the output layer meets the expected value or the number of iterations reaches the predetermined value, proceed to S23. When the difference between the output value of the output layer and the expected value exceeds the predetermined requirement, proceed to S22 to propagate the error backward. S22. The initial weight values ​​obtained from the training data reaching the output layer are propagated backward from the output layer to the next layer. The difference between the output value and the expected value is gradually corrected through backward propagation layer by layer. At the same time, the error is fed forward to correct the weights between neurons in each layer. When the output value of the output layer meets the expected value or the number of iterations reaches the predetermined value, the process enters S23. S23, save the currently trained network model as the judgment model, and use the other part of the signal data after feature extraction as test data to test the judgment model.

Images