Nuclear pulse accumulation signal peak value identification method based on machine learning
Through the peak recognition method of nuclear pulse accumulation signal based on machine learning, using ladder formation algorithm and CNN convolutional neural network model, the problems of low stacking pulse processing efficiency and difficulty in peak recognition in the existing technology are solved, and more efficient pulse peak value is achieved Identification and energy resolution improves.
Patent Information
- Application Number
- CN202310719703.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-17
- Publication Date
- 2024-12-27
AI Technical Summary
When dealing with the phenomenon of nuclear pulse accumulation, the accumulation pulse deletion method causes the energy spectrum counting rate to be reduced, and the pulse forming separation method is difficult to identify the exponential nuclear pulse peak, and the formed pulses are prone to accumulate, affecting the energy resolution.
The peak recognition method of nuclear pulse accumulation signal based on machine learning is used to generate the initial data set, the ladder forming algorithm forming process, and the CNN convolutional neural network model is constructed, and the pulse peak data is extracted by combining the ReLU activation function and the Adam optimizer.
The recognition rate of pulse peaks is improved, the impact of noise is reduced, the recognition ability of pulse peaks is enhanced, and the energy resolution is improved.
Smart Images
Figure BDA0004290818510000021 
Figure BDA0004290818510000041 
Figure BDA0004290818510000051
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of government data management, and in particular to a method for identifying a nuclear pulse pileup signal peak value based on machine learning. Background Art
[0002] Nuclides decay randomly, and the signal output by the detector is also a randomly generated nuclear pulse signal, so when two or more random pulses are generated simultaneously within the shortest resolution time of the measurement system, pulse pileup will occur.
[0003] At present, the following two methods are mainly used to deal with the accumulation phenomenon:
[0004] The first method is the pile-up pulse deletion method first proposed by Wang Jingjin et al. This method directly discards the pile-up pulses to achieve the purpose of improving energy resolution.
[0005] The second method is to separate the piled-up pulses by pulse shaping. The core idea of this method is to select appropriate shaping methods and shaping parameters to improve the energy resolution by identifying and separating the piled-up pulses.
[0006] Although the above two methods can reduce the interference of pile-up pulses on energy resolution to a certain extent. However, in practical applications, the pile-up pulse deletion method discards some pulses, which reduces the final count rate of the energy spectrum, resulting in poor energy resolution improvement effect, and is rarely used. The pulse shaping and separation pile-up pulse method uses pulse shaping to identify and separate pile-up pulses. This method needs to select an appropriate pulse shaping method to identify the peak of the pulse signal according to the actual situation. The main difficulty is that after the exponential nuclear pulse is digitally pulse-shaped, the signal will be broadened to facilitate the identification of the amplitude, so the pulses after shaping will also be piled up, making the pulse peak identification very difficult. Summary of the invention
[0007] In view of the above problems, the present invention provides a nuclear pulse pileup signal peak recognition method based on machine learning to solve the above problems.
[0008] The technical solution of the present invention is:
[0009] A method for identifying a nuclear pulse pileup signal peak based on machine learning specifically comprises the following steps:
[0010] S1, generate an initial data set and preprocess the initial data set;
[0011] S2, using a trapezoidal shaping algorithm to shape the preprocessed data set;
[0012] S3, generating a label set of the data set according to the formed data, integrating the data set and the label set and dividing them into a training set and a test set;
[0013] S4. Build a machine learning model;
[0014] S5. Use the training set to train the machine learning model constructed in S4;
[0015] S6. Input the test set into the trained machine learning model and extract the pulse peak data.
[0016] As a preferred solution of the present invention: the initial data set in S1 is based on an ideal exponential signal pulse sequence V (t) Generate the ideal exponential signal pulse sequence V (t) It generates random numbers that conform to Gaussian distribution and exponential distribution by specifying the exponential distribution mean and random values that conform to Gaussian distribution, and adds appropriate offsets to generate discrete stacked pulse signals.
[0017] As a preferred solution of the present invention: the preprocessing operation in S1 refers to dividing the discrete piled-up pulse signals into groups according to their boundaries until discrete data of each group being a complete piled-up pulse signal is formed.
[0018] As a preferred solution of the present invention: the trapezoidal shaping algorithm expression in S2 is:
[0019]
[0020] where t a is the rising edge width, t b is the flat top width, t c is the falling edge width, V o is the time domain expression of the shaped pulse.
[0021] As a preferred solution of the present invention: the trapezoidal shaping algorithm in S2 needs to assume that the nuclear pulse signal is an ideal exponential signal, then the time domain expression is: V i (t) = V max *e -t / τ μ(t), where V max is the pulse amplitude, τ is the time constant of the front-end amplifier, μ(t) is the standard unit step function, and T s The pulse train expression can be obtained as follows:
[0022] As a preferred solution of the present invention: the data set and label set in S3 need to be randomly divided in a ratio of 9:1.
[0023] As a preferred solution of the present invention: the machine learning model in S4 is based on the CNN convolutional neural network model, which defines two convolutional layers, two pooling layers and a fully connected layer in the form of a list.
[0024] As a preferred solution of the present invention: the convolution layer performs convolution calculation on the input data through the convolution kernel, and generates a corresponding "feature data" through the convolution operation of the convolution layer. These feature data can be connected into a higher-level feature map. The pooling layer performs dimensionality reduction processing on the output feature data of the convolution layer by sampling or aggregation. The fully connected layer expands the feature map output by the previous convolution layer, pooling layer and other feature extraction layers into a vector form, and performs linear transformation through matrix multiplication and bias terms, and finally outputs the classification prediction result.
[0025] As a preferred solution of the present invention: the activation functions of the convolutional layer and the fully connected layer are ReLU.
[0026] As a preferred solution of the present invention: the machine learning model training process in S5 uses Adam as the optimizer.
[0027] The beneficial effects of the present invention are:
[0028] 1. Use the trapezoidal shaping algorithm to perform trapezoidal shaping on the accumulated nuclear pulse signals to obtain trapezoidal pulse signal data that is easy to identify;
[0029] 2. Use machine learning to extract the peak data of the pile-up core pulse signal to obtain pulse peak data that can be directly used;
[0030] 3. An activation function with noise reduction function is added to improve the recognition rate of pulse peak value. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0032] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.
[0033] A method for identifying a nuclear pulse pileup signal peak based on machine learning specifically comprises the following steps:
[0034] S1, generate an initial data set and preprocess the initial data set;
[0035] S2, using a trapezoidal shaping algorithm to shape the preprocessed data set;
[0036] S3, generating a label set of the data set according to the formed data, integrating the data set and the label set and dividing them into a training set and a test set;
[0037] S4. Build a machine learning model;
[0038] S5. Use the training set to train the machine learning model constructed in S4;
[0039] S6. Input the test set into the trained machine learning model and extract the pulse peak data.
[0040] Further: The initial data set in S1 is based on an ideal exponential signal pulse sequence V (t) Generate the ideal exponential signal pulse sequence V (t) It generates random numbers that conform to Gaussian distribution and exponential distribution by specifying the exponential distribution mean and random values that conform to Gaussian distribution, and adds appropriate offsets to generate discrete stacked pulse signals.
[0041] Further: the preprocessing operation in S1 refers to dividing the discrete pile-up pulse signals into groups through boundaries until each group is formed as discrete data of a complete pile-up pulse signal.
[0042] Further: The trapezoidal shaping algorithm expression in S2 is:
[0043]
[0044] where t a is the rising edge width, t b is the flat top width, t c is the falling edge width, V o is the time domain expression of the shaped pulse.
[0045] Further: The trapezoidal shaping algorithm in S2 needs to assume that the nuclear pulse signal is an ideal exponential signal, so the time domain expression is: V i (t) = V max *e -t / τμ(t), where V max is the pulse amplitude, τ is the time constant of the front-end amplifier, μ(t) is the standard unit step function, and T s The pulse train expression can be obtained as follows:
[0046] Furthermore: the data set and label set in S3 need to be randomly divided in a ratio of 9:1.
[0047] Further: The machine learning model in S4 is based on the CNN convolutional neural network model, which defines two convolutional layers, two pooling layers and a fully connected layer in the form of a list.
[0048] The role of the above machine learning model is to be able to identify the pulse peak from the formed accumulated pulse signal.
[0049] Furthermore: the convolution layer performs convolution calculation on the input data through the convolution kernel, and generates a corresponding "feature data" through the convolution operation of the convolution layer. These feature data can be connected into a higher-level feature map, which can further improve the abstraction ability of the input data;
[0050] The pooling layer performs dimensionality reduction processing on the output feature data of the convolutional layer by sampling or pooling. This has the following functions: reducing the number of model parameters and data storage volume, avoiding overfitting and improving classification efficiency. Overfitting refers to making the assumptions too strict in order to obtain consistent assumptions, which will lead to poor model performance.
[0051] The fully connected layer expands the feature graph output by the previous convolutional layer, pooling layer and other feature extraction layers into a vector form, performs linear transformation through matrix multiplication and bias terms, and finally outputs the classification prediction result.
[0052] Furthermore: the activation function of the convolutional layer and the fully connected layer is ReLU.
[0053] The above activation function ReLU has the characteristic of suppressing input signals less than 0, which makes it suppress random noise in the input data. Because in the output result of the ReLU function, if the input is less than 0, the output value will be set to 0, and the original signal will become weaker or disappear, thereby reducing the impact on the subsequent network layers. Secondly, the ReLU function can also increase the response difference between neurons. Since the ReLU function is a piecewise linear function, when the input value is greater than 0, the output is a constant value, which brings a more obvious difference in the response between neurons. When the input changes only slightly, the output change will be more significant than other activation functions. This response difference can also play a role in noise resistance.
[0054] Further: The machine learning model training process in S5 uses Adam as the optimizer.
[0055] The above Adam optimizer can realize adaptive adjustment of learning rate. It adaptively calculates the learning rate of each parameter according to the historical gradient size and variance of each parameter. In each iteration, it can give a learning rate suitable for the parameter.
[0056] like Figure 1 As shown, the working principle of the above technical solution is as follows:
[0057] A method for implementing a nuclear pulse pileup signal peak recognition method based on machine learning comprises the following steps:
[0058] S1, generate pulse sequence V based on ideal exponential signal (t) The initial data set is divided into groups by boundaries until each group is formed into a complete discrete data of the accumulated pulse signal;
[0059] S2. Assuming that the nuclear pulse signal is an ideal exponential signal, the time domain expression is: V i (t) = V max *e -t / τμ(t), where V max is the pulse amplitude, τ is the time constant of the front-end amplifier, μ(t) is the standard unit step function, and T s The pulse train expression can be obtained as follows: Then substitute into the trapezoidal shaping algorithm expression:
[0060]
[0061] Perform shaping processing on the preprocessed data set;
[0062] S3, generate a label set of the data set according to the formed data, integrate the data set and the label set and divide them into a training set and a test set, randomly dividing them in a ratio of 9:1;
[0063] S4. Build a machine learning model based on CNN convolutional neural network model;
[0064] S5. Use the training set to train the machine learning model built in S4, and use Adam as the optimizer;
[0065] S6. Input the test set into the trained machine learning model and extract the pulse peak data.
[0066] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.
Claims
1. A method for identifying the peak value of a nuclear pulse pileup signal based on machine learning, characterized in that: The specific steps include: S1, generate an initial data set and preprocess the initial data set; S2, using a trapezoidal shaping algorithm to shape the preprocessed data set; S3, generating a label set of the data set according to the formed data, integrating the data set and the label set and dividing them into a training set and a test set; S4. Build a machine learning model; S5. Use the training set to train the machine learning model constructed in S4; S6. Input the test set into the trained machine learning model and extract the pulse peak data.
2. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 1, characterized in that: The initial data set in S1 is based on an ideal exponential signal pulse sequence V (t) Generate the ideal exponential signal pulse sequence V (t) It generates random numbers that conform to Gaussian distribution and exponential distribution by specifying the exponential distribution mean and random values that conform to Gaussian distribution, and adds appropriate offsets to generate discrete stacked pulse signals.
3. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 2, characterized in that: The preprocessing operation in S1 refers to dividing the discrete pile-up pulse signal into groups according to the boundary until each group is formed as discrete data of a complete pile-up pulse signal.
4. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 1, characterized in that: The trapezoidal shaping algorithm expression in S2 is: where t a is the rising edge width, t b is the flat top width, t c is the falling edge width, V o is the time domain expression of the shaped pulse.
5. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 4, characterized in that: The trapezoidal shaping algorithm in S2 needs to assume that the nuclear pulse signal is an ideal exponential signal, so the time domain expression is: V i (t) = V max *e -t / τ μ(t), where V max is the pulse amplitude, τ is the time constant of the front-end amplifier, μ(t) is the standard unit step function, and T s The pulse train expression can be obtained as follows:
6. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 1, characterized in that: The data set and label set in S3 need to be randomly divided in a ratio of 9:
1.
7. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 1, characterized in that: The machine learning model in S4 is based on the CNN convolutional neural network model, which defines two convolutional layers, two pooling layers and a fully connected layer in the form of a list.
8. The method for identifying the peak value of a nuclear pulse pile-up signal based on machine learning according to claim 7, characterized in that: The convolution layer performs convolution calculation on the input data through the convolution kernel, and generates a corresponding "feature data" through the convolution operation of the convolution layer. These feature data can be connected into a higher-level feature map. The pooling layer performs dimensionality reduction processing on the output feature data of the convolution layer by sampling or aggregation. The fully connected layer expands the feature map output by the previous convolution layer, pooling layer and other feature extraction layers into a vector form, and performs linear transformation through matrix multiplication and bias terms, and finally outputs the classification prediction result.
9. The method for identifying a peak value of a nuclear pulse pile-up signal based on machine learning according to claim 8, characterized in that: The activation function of the convolutional layer and the fully connected layer is ReLU.
10. The method for identifying the peak value of a nuclear pulse pile-up signal based on machine learning according to claim 1, characterized in that: The machine learning model training process in S5 uses Adam as the optimizer.