High-speed real-time nuclear signal pulse detection method based on machine learning

Abstract

The invention provides a high-speed real-time nuclear signal pulse detection method based on machine learning, and the method specifically comprises the following steps: 1, carrying out the filtering and shaping of an input signal, and shaping the input signal into a Gaussian-like shape; 2, analog-to-digital conversion is carried out on the signals filtered and shaped in the step 1; 3, sequentially carrying out preprocessing, pulse forming and accumulation identification and positioning on the signals subjected to analog-to-digital conversion in the step 2, and counting the accumulation of the signals; 4, identifying the accumulation signals in the step 2 by adopting deep learning; and step 5, evaluating the accuracy of accumulation identification in the step 3 and the step 4, and optimizing parameters of deep learning according to an evaluation result. According to the invention, the accuracy is improved while the calculation cost is saved.

Description

Technical Field

[0001] The present invention belongs to the field of radiation nuclear detectors, and in particular relates to a high-speed real-time nuclear signal pulse detection method based on machine learning. Background Art

[0002] With the widespread application of nuclear technology, nuclear detectors are increasingly used in scientific research, production and life. In the field of nuclear signal detection, it is crucial to detect a large amount of data in real time and process the detection signals. However, the current processing methods of most nuclear detectors usually store the data first and then perform batch processing and optimization. The traditional solution is to collect the input signal and then perform analog or digital shaping, and then store it in a digital storage module. After the entire data is processed, the waveform is output and the number of pulses is checked. The traditional processing method cannot display the results in real time, and if there is a problem or inaccuracy after a large amount of data is stored, it often needs to be calculated repeatedly, resulting in a waste of time and resources.

[0003] With the development of technology, the application of nuclear detection has an increasing demand for real-time data processing and rapid response, especially in the fields of nuclear safety monitoring and radiation environment monitoring. Traditional data processing methods have been unable to meet the increasing requirements of real-time and accuracy. In order to improve the detection efficiency and accuracy of nuclear signal pulses, more and more researchers have begun to try to learn and process signals in real time through algorithms, especially machine learning methods.

[0004] In the prior art, although there have been methods that attempt to perform signal detection through threshold judgment or pattern recognition-based algorithms, these methods are easily affected by adverse factors such as noise and signal stacking, and it is difficult to provide real-time and accurate results in complex environments. Summary of the invention

[0005] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention provides a high-speed real-time nuclear signal pulse detection method based on machine learning.

[0006] Technical solution: The present invention discloses a high-speed real-time nuclear signal pulse detection method based on machine learning, which specifically includes the following steps:

[0007] Step 1: Filter and shape the input signal into a Gaussian-like shape;

[0008] Step 2: Perform analog-to-digital conversion on the signal filtered and shaped in step 1;

[0009] Step 3: Preprocessing, pulse shaping and accumulation identification are performed on the signal after analog-to-digital conversion in step 2, and the accumulation of the signal is counted;

[0010] Step 4: Use deep learning to identify the accumulation signal in step 2;

[0011] Step 5: Evaluate the accuracy of the accumulation recognition in steps 3 and 4, and optimize the parameters of deep learning based on the evaluation results.

[0012] Furthermore, the step 3 is specifically as follows: firstly, an asynchronous FIFO is used to buffer the signal, then the signal is reshaped, accumulation identification and positioning are performed on the shaped signal, and finally the accumulation of the signal is counted.

[0013] Furthermore, the deep learning in step 4 adopts a random forest regression model.

[0014] Furthermore, the maximum value, variance, number of peaks and half-height width in the input signal are used as features for learning, and the training set sample data obtained by learning is used to train the random forest regression model.

[0015] Furthermore, when training the random forest regression model, if a time period is selected, training and testing are performed according to the time period; otherwise, the random forest regression model is established using the default time period data.

[0016] Furthermore, the method for constructing a random forest regression model is as follows: a new data set is formed by randomly extracting data from the input data set, and a regression decision tree is generated for each new data set; the construction of the decision tree is specifically as follows: m features are randomly selected for the current node, all possible split values ​​of these features are traversed, and the effect of each split is evaluated based on the mean square error standard, and the split feature with the lowest mean square error and its corresponding split point are selected.

[0017] Furthermore, the method also includes adjusting the method in step 3 according to the evaluation result.

[0018] A detection system for a high-speed real-time nuclear signal pulse detection method based on machine learning, the system comprises a front-end analog signal processing module, an analog-to-digital conversion module, a dual-path processing module and a result evaluation module; the dual-path processing module comprises a first path module and a second path module;

[0019] The front-end analog signal processing module is used to filter and shape the input signal;

[0020] The analog-to-digital conversion module is used to perform analog-to-digital conversion on the signal after line filtering and shaping;

[0021] The first path module is used to perform accumulation identification and positioning of the signal using the method of step 3;

[0022] The second path module uses deep learning to identify and locate signal accumulation;

[0023] The result evaluation module is used to evaluate the accuracy of the recognition results of the first path module and the recognition results of the second path module, and to optimize the system in real time according to the evaluation results.

[0024] Beneficial effects: The present invention adopts a deep learning method to sample and learn the input pulse signal with replacement, and at the same time, combined with the input of the original signal, detects the pulse signal in real time through two-channel signal processing, identifies the accumulation signal and gives the correction result. This method of combining physics and machine learning integrates their respective advantages and characteristics, and improves accuracy while saving computing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the method of the present invention;

[0026] Figure 2 This is the flow chart of the random forest regression algorithm;

[0027] Figure 3 This is a test result diagram of the present invention. DETAILED DESCRIPTION

[0028] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0029] The system of the present invention comprises a front-end analog signal processing module, an analog-to-digital conversion module, a dual-path processing module and a result evaluation module; the dual-path processing module comprises a first path module and a second path module; the modules of the system are connected via a high-speed data bus to ensure smooth data transmission and real-time processing.

[0030] Figure 1 The flowchart of the pulse detection system based on machine learning of the present invention is as follows:

[0031] When the system starts, the entire detection system is initialized to ensure that all modules are working properly. During the initialization process, the system automatically detects the quality of the input signal and the adaptability of the front-end filter parameters. When the environmental conditions and input signals are the same, the parameters of the deep learning model do not need to be further adjusted.

[0032] The front-end analog signal processing module is used to filter and shape the input signal. This module provides adjustable filtering types and processes the nuclear pulse signal by optimizing the filtering algorithm to ensure that the signal is converted into a Gaussian shape at the front end of the system, while effectively removing high-frequency noise and low-frequency interference. Common front-end waveform shaping methods include SK filters and CR-RC filters. Both solutions have mature implementation technologies and can perform targeted shaping of the original signal according to actual needs.

[0033] The first path uses signal preprocessing, pulse shaping and accumulation identification, which is mainly used to identify signal accumulation. Since the clock frequencies of analog signals and digital signals may be inconsistent, asynchronous FIFO is used to buffer data to ensure data integrity. The digitized pulse data is reshaped, often using a trapezoidal shaping method. The processed signal can more accurately represent the waveform of the nuclear pulse signal. Accumulation identification and positioning are performed on the shaped signal, and the accumulation of the signal is counted using a suitable judgment mechanism.

[0034] The second path uses a deep learning solution channel. First, the signal features are extracted and preprocessed. The extracted features include the maximum value, variance, number of peaks, and half-height width of the signal, which are used to represent the key features of the nuclear signal. Then, the extracted features are trained and verified using a random forest regression model to identify abnormal states in the signal. This process uses an integrated decision tree to improve prediction accuracy and ensure accurate detection of abnormal conditions.

[0035] The specific steps of building a random forest regression model are as follows:

[0036] Establish a random forest regression model and directly import the input signal data into the deep learning model for calculation;

[0037] Identify anomalies in signals through feature extraction and model training. Feature extraction includes calculation of maximum value, variance, number of peaks, and half-height width to extract key features of accumulation and anomalies;

[0038] After selecting features, customize the training set and start training.

[0039] Continue to input and substitute the test set into the prediction model to compare the accuracy of the prediction model.

[0040] Figure 2It is the training flow chart inside the random algorithm. During the processing, preprocessing is first performed, and then the system is asked whether training is required. If so, it proceeds to the stage of selecting time periods. If not, it directly performs stacking detection and training based on the model. Next, the system is asked whether a time period has been selected. If a time period has been selected, training and testing are performed according to the time period. If not, a random forest regression model is established using the default time period data. After establishing the training set and the test set, stacking detection is performed on the data and the labels are exported.

[0041] The generation of the decision tree in this embodiment is specifically as follows: First, after data preprocessing, new data sets D1 to DN are formed by randomly extracting data from the input data set R. For each new data set Di (1 ≤ i ≤ N), the i-th regression decision tree is generated. N regression decision trees are constructed according to the method of the present invention. The construction process of each decision tree requires selecting node splitting, and the core step of determining the optimal node splitting is to determine the best splitting feature and its corresponding splitting value. In this embodiment, it is assumed that m features (where m < M, and M is the total number of features) are randomly selected for the current node. By traversing all possible splitting values of these features and evaluating the effect of each splitting based on the mean squared error (MSE) criterion, the splitting feature with the lowest mean squared error and its corresponding splitting point are finally selected. Specifically: For any divided feature A in the input space R, the two data sets R1 and R2 are divided by any variable value corresponding to it, that is, the splitting point s. Then, when the sum of the output mean squared errors of R1 and R2 is the smallest, the corresponding feature and the feature value splitting point are the best splitting nodes. Once A and s are determined, the current input space R is divided into two sub-spaces.紧接着, recursive splitting is performed again in these two sub-spaces until the conditions for stopping splitting are met, such as the maximum depth reaching the default value or the number of available samples being less than the set minimum number of samples required, etc.

[0042] The number of decision trees in this embodiment is 2178, the minimum number of samples for node splitting is 5, the minimum number of samples for leaf nodes is 3. Each time node splitting occurs, the square root of the number of features (i.e., the square root of the total number of features) is randomly selected to participate in the decision of node splitting, rather than using all features. The maximum depth is 10, and bootstrap is True.

[0043] Since the random forest cannot make predictions beyond the range of the training set data, when establishing the training model, it is necessary to ensure the determined prediction range. During the training process, feature parameters need to be set so as to detect the influence relationship between features.

[0044] The results in the dual-path processing module compare and integrate the first path and the second path. The output results of the first path and the second path are compared, combining the advantages of both to improve the accuracy and robustness of nuclear signal detection. The final result is used for counting, positioning and abnormal judgment of nuclear pulses.

[0045] By comparing and analyzing the output results of the dual-channel detection model, the accuracy of pile-up recognition and the precision of signal positioning are evaluated (using regression evaluation indicators such as MSE (mean square error), MAE (mean absolute error), RMSE (root mean square error), R^2 (coefficient of determination), etc.). The system will optimize the model based on the evaluation results, update the training parameters of the deep learning model or adjust the processing methods of the traditional solution to improve the detection performance of the overall system.

[0046] During the operation of the system, the model is continuously trained and optimized based on new data. The system automatically updates the deep learning model and adjusts the parameters of the traditional algorithm based on the new input signal to ensure that the system can adapt to signal changes in different environments.

[0047] Figure 3 This is a partial result diagram of this embodiment. The input signals in the figure are three random input signals with the same amplitude and noise. The stacking is divided into three cases, namely 1, 2 and 3. Stacking 1 means no stacking, and stacking 2 and stacking 3 have two signals and three signals at this location respectively. From the stacking samples, it can be seen that this method can simply and directly learn the stacking situation and predict the results. The actual value is directly derived in the calculation, which makes it convenient to compare with the results of the first path.

[0048] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A high-speed real-time nuclear signal pulse detection method based on machine learning, characterized in that: The specific steps include: Step 1: Filter and shape the input signal into a Gaussian-like shape; Step 2: Perform analog-to-digital conversion on the signal filtered and shaped in step 1; Step 3: Preprocessing, pulse shaping and accumulation identification are performed on the signal after analog-to-digital conversion in step 2, and the accumulation of the signal is counted; Step 4: Use deep learning to identify the accumulation signal in step 2; Step 5: Evaluate the accuracy of the accumulation recognition in steps 3 and 4, and optimize the parameters of deep learning based on the evaluation results.

2. A high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 1, characterized in that: The step 3 is specifically as follows: firstly, using an asynchronous FIFO to buffer the signal, then performing secondary shaping on the signal, performing accumulation identification and positioning on the shaped signal, and finally counting the accumulation of the signal.

3. The high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 1, characterized in that: The deep learning in step 4 adopts a random forest regression model.

4. The high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 3, characterized in that: The maximum value, variance, number of peaks and half-height width in the input signal are learned as features, and the random forest regression model is trained using the training set sample data obtained through learning.

5. The high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 1, characterized in that: When training the random forest regression model, if a time period is selected, training and testing are performed according to the time period; otherwise, the default time period data is used to establish the random forest regression model.

6. The high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 1, characterized in that: The method for constructing a random forest regression model is as follows: a new data set is formed by randomly extracting data from the input data set, and a regression decision tree is generated for each new data set; the construction of the decision tree is specifically as follows: m features are randomly selected for the current node, all possible split values ​​of these features are traversed, and the effect of each split is evaluated based on the mean square error standard, and the split feature with the lowest mean square error and its corresponding split point are selected.

7. The high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 1, characterized in that: The method also includes and adjusts the method in step 3 based on the evaluation results.

8. A detection system based on the high-speed real-time nuclear signal pulse detection method based on machine learning according to claim 1, characterized in that: The system includes a front-end analog signal processing module, an analog-to-digital conversion module, a dual-path processing module and a result evaluation module; the dual-path processing module includes a first path module and a second path module; The front-end analog signal processing module is used to filter and shape the input signal; The analog-to-digital conversion module is used to perform analog-to-digital conversion on the signal after line filtering and shaping; The first path module is used to perform accumulation identification and positioning of the signal using the method of step 3; The second path module uses deep learning to identify and locate signal accumulation; The result evaluation module is used to evaluate the accuracy of the recognition results of the first path module and the recognition results of the second path module, and to optimize the system in real time according to the evaluation results.

Images