Training method, application method and related system of waveform data processing model

Abstract

The invention discloses a training method, an application method and a related system of a waveform data processing model, and relates to the technical field of environmental radioactivity detection and particle physics, and the training method comprises the steps: obtaining original waveform data and corresponding particle real label data through a TPC detector; performing down-sampling processing and data compression processing on the original waveform data in sequence; inputting the preprocessed original waveform data into each channel of a waveform data processing model according to the spatial dimension of the channel, and determining the network loss according to the output prediction probability value of each category of particles and the real label of the particles; and according to the network loss, network parameters of the waveform data processing model are updated through back propagation, and a trained waveform data processing model is obtained after multiple iterations. And the trained waveform data processing model is used for particle identification, so that data storage is saved, the calculation amount is reduced, and the accuracy of a multi-dimensional information particle identification result is improved.

Description

Technical Field

[0001] The present application relates to the field of environmental radioactivity detection and particle physics technology, and in particular to a training method, an application method and a related system of a waveform data processing model. Background Art

[0002] Ionizing radiation is a type of radiation that can ionize atoms or molecules, including alpha rays and beta rays. Ionizing radiation is usually not directly perceived by the human body, but once the human body is exposed to radiation exceeding the safe dose or continues to be affected by internal radiation from the ingested radiation source, it will be seriously harmed. Therefore, it is very important to monitor the levels of alpha, beta and other radiation in various environments. The levels of alpha and beta radiation in daily life are very weak, which poses a challenge to the discrimination ability of measuring instruments and the realization of low background detection.

[0003] To solve this problem, a new measurement method is a low-background measurement method of α and β rays based on the TPC (Time Projection Chamber) detector. TPC uses the drift output waveform data of the electron cloud generated by ionizing radiation in the electric field to determine the trajectory and energy loss of particles. In many ionizing radiation detection application scenarios, the radioactivity of the samples being tested is relatively weak, and there is a similarity and overlap between background particles and particles of interest, so it is crucial to develop TPC data processing methods.

[0004] At present, the mainstream TPC data processing method is a multi-dimensional information particle identification method based on track detection. It mainly extracts features from waveform signals, measures physical information such as particle tracks, energy, and energy loss, and integrates multiple information to identify particles. However, in this multi-dimensional information particle identification method, the acquisition of multi-dimensional information usually requires a large amount of data collection and storage, and requires the processing of a large amount of data, which poses a challenge to the storage and processing capabilities of hardware data. In addition, the analysis and feature selection of multi-dimensional information are crucial to the effect and performance of the algorithm. In high-dimensional data space, feature selection and dimensionality curse problems will affect the accuracy and efficiency of the algorithm, and are easily affected by empirical factors. There will be large errors, resulting in inaccurate and unreliable multi-dimensional information particle identification results.

[0005] In summary, how to improve the accuracy of multi-dimensional information particle identification results while saving data storage and reducing the amount of calculation, so as to solve the problems of large data volume, strong influence of subjective factors and low accuracy in traditional multi-dimensional information particle identification methods, has become a technical problem that needs to be solved urgently in this field. Summary of the invention

[0006] The purpose of this application is to provide a training method, an application method and a related system for a waveform data processing model, which can improve the accuracy of multi-dimensional information particle identification results while saving data storage and reducing the amount of calculation.

[0007] To achieve the above objectives, this application provides the following solutions:

[0008] In a first aspect, the present application provides a method for training a waveform data processing model, the method for training a waveform data processing model comprising:

[0009] Obtain original waveform data and its corresponding particle real label data; the original waveform data is multi-channel particle waveform data collected by using a TPC detector, the TPC detector includes multiple channels, and the multiple channels synchronously collect particle waveform data to form the multi-channel particle waveform data; the particle real label data includes a real label for characterizing the actual category of the particle.

[0010] The original waveform data is subjected to downsampling processing and data compression processing in sequence to obtain preprocessed original waveform data.

[0011] The preprocessed raw waveform data is input into each channel of the waveform data processing model according to the spatial dimension of the channel, and the network loss is determined according to the predicted probability values ​​of each category of particles output by the waveform data processing model and the true label of the particles; the waveform data processing model is a model with a multi-channel network structure established based on an artificial neural network, and the network structure of each channel includes multiple convolution layers and multiple pooling layers. The waveform data processing model is built by first passing through all convolution layers and then passing through all pooling layers.

[0012] According to the network loss, the network parameters of the waveform data processing model are updated through back propagation to iterate the waveform data processing model multiple times to obtain a trained waveform data processing model.

[0013] Optionally, downsampling processing and data compression processing are sequentially performed on the original waveform data to obtain pre-processed original waveform data, specifically including:

[0014] The original waveform data is downsampled by adopting an interval resampling method to obtain the downsampled original waveform data.

[0015] According to the order of the channel numbers of the TPC detector, the downsampled original waveform data is compressed and saved in a sparse tensor format to obtain the preprocessed original waveform data.

[0016] Optionally, the particle categories include alpha particles, beta particles and muons.

[0017] In a second aspect, the present application provides an application method of a waveform data processing model, the application method of the waveform data processing model comprising:

[0018] The target waveform data is obtained, where the target waveform data is multi-channel particle waveform data of the particles to be measured collected by using a TPC detector.

[0019] The target waveform data is subjected to downsampling processing and data compression processing in sequence to obtain preprocessed target waveform data.

[0020] The preprocessed target waveform data is input into each channel of the waveform data processing model according to the spatial dimension of the channel to obtain the predicted probability value of each category of the particle to be measured; the waveform data processing model is a model trained according to the training method of the waveform data processing model described in the first aspect.

[0021] Optionally, after the step of inputting the preprocessed target waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel to obtain the predicted probability value of each category of the particle to be measured, the application method of the waveform data processing model further includes:

[0022] Set the confidence threshold.

[0023] The predicted probability value corresponding to the highest category among the predicted probability values ​​of each category of the particles to be tested is compared with the confidence threshold. The case where the predicted probability value is less than the confidence threshold is determined as a background case, and the case where the predicted probability value is greater than or equal to the confidence threshold is determined as a particle case. All background cases are eliminated and all particle cases are retained.

[0024] In a third aspect, the present application provides a training system for a waveform data processing model, the training system for the waveform data processing model comprising:

[0025] The first data acquisition module is used to acquire original waveform data and its corresponding particle real label data; the original waveform data is multi-channel particle waveform data collected by using a TPC detector, the TPC detector includes multiple channels, and the multiple channels synchronously collect particle waveform data to form the multi-channel particle waveform data; the particle real label data includes a real label for characterizing the actual category of the particle.

[0026] The first preprocessing module is used to perform downsampling processing and data compression processing on the original waveform data in sequence to obtain preprocessed original waveform data.

[0027] A network loss determination module is used to input the preprocessed original waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, and determine the network loss according to the predicted probability values ​​of each category of particles output by the waveform data processing model and the true label of the particles; the waveform data processing model is a model with a multi-channel network structure established based on an artificial neural network, and the network structure of each channel includes multiple convolution layers and multiple pooling layers. The waveform data processing model is built by first passing through all convolution layers and then passing through all pooling layers.

[0028] The training module is used to update the network parameters of the waveform data processing model through back propagation according to the network loss, so as to iterate the waveform data processing model multiple times to obtain a trained waveform data processing model.

[0029] Optionally, the first preprocessing module includes:

[0030] The downsampling unit is used to downsample the original waveform data by adopting an interval resampling method to obtain the downsampled original waveform data.

[0031] A data compression unit is used to compress and save the downsampled original waveform data into a sparse tensor format in the order of the channel number of the TPC detector to obtain the preprocessed original waveform data.

[0032] Optionally, the particle categories include alpha particles, beta particles and muons.

[0033] In a fourth aspect, the present application provides an application system of a waveform data processing model, the application system of the waveform data processing model comprising:

[0034] The second data acquisition module is used to acquire target waveform data, where the target waveform data is multi-channel particle waveform data of the particles to be measured acquired by using a TPC detector.

[0035] The second preprocessing module is used to perform downsampling processing and data compression processing on the target waveform data in sequence to obtain preprocessed target waveform data.

[0036] A target particle category prediction module is used to input the preprocessed target waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, so as to obtain the predicted probability value of each category of the particle to be measured; the waveform data processing model is a model trained according to the training method of the waveform data processing model described in the first aspect.

[0037] Optionally, the application system of the waveform data processing model further includes:

[0038] A confidence threshold setting module, used to set a confidence threshold;

[0039] The screening module is used to compare the predicted probability value corresponding to the highest category among the predicted probability values ​​of each category of the particles to be tested with the confidence threshold, determine the case where the predicted probability value is less than the confidence threshold as a background case, and determine the case where the predicted probability value is greater than or equal to the confidence threshold as a particle case, and eliminate all background cases and retain all particle cases.

[0040] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0041] The present application provides a training method, application method and related system of a waveform data processing model, which combines TPC detection technology, downsampling and data compression technology and neural network technology and applies them to specific scenarios of waveform processing. On the one hand, multi-channel particle waveform data is collected by TPC detector, and downsampling and data compression are preprocessed to obtain waveform data with high compression ratio and high resolution, which effectively saves data storage, reduces the amount of calculation, reduces the difficulty of operation, and improves the efficiency of waveform data processing and particle type prediction. On the other hand, in the context of multi-channel particle waveform data of TPC detector, based on machine learning algorithms such as artificial neural networks, multi-channel particle waveform data is input into each channel of the waveform data processing model according to the spatial dimension of the channel, and the corresponding particle identification result is output. By training the waveform data processing model, the trained waveform data processing model has high-precision particle identification ability, realizes high-precision particle identification, solves the problems of large data volume and low accuracy in traditional multi-dimensional information particle identification methods, and can be widely used in high-sensitivity detection scenarios of α and β radionuclides in ray environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 A flowchart of a method for training a waveform data processing model provided in Example 1 of the present application.

[0044] Figure 2 This is a schematic diagram of the structure of the TPC detector provided in Example 1 of the present application.

[0045] Figure 3This is a schematic diagram of the waveform data processing method provided in Example 1 of the present application.

[0046] Figure 4 A waveform diagram of the original waveform data provided in Example 1 of the present application.

[0047] Figure 5 A waveform diagram of the original waveform data provided in Example 1 of the present application after data compression processing.

[0048] Figure 6 A schematic diagram of the waveform after downsampling of the original waveform data provided in Example 1 of the present application.

[0049] Figure 7 This is a structural diagram of the waveform data processing model provided in Example 1 of the present application.

[0050] Figure 8 A schematic diagram of the structure of the convolution layer and pooling layer of each channel of the waveform data processing model provided in Example 1 of the present application.

[0051] Fig. 9 A schematic diagram of the structure of the fully connected output part of the waveform data processing model provided in Example 1 of the present application.

[0052] Fig.10 This is a distribution diagram of the starting points of the β sample cases provided in Example 1 of the present application.

[0053] Fig.11 This is a distribution diagram of the starting points of the remaining cases of the β samples after screening provided in Example 1 of the present application.

[0054] Fig.12 This is a starting point distribution diagram of the background sample example provided in Example 1 of the present application.

[0055] Fig.13 This is a distribution diagram of the starting points of the remaining cases of the background samples after screening provided in Example 1 of the present application.

[0056] Fig.14 A flowchart of the application method of the waveform data processing model provided in Example 2 of the present application.

[0057] Fig.15 This is a structural block diagram of the training system for the waveform data processing model provided in Example 3 of the present application.

[0058] Fig.16 This is a structural block diagram of the application system of the waveform data processing model provided in Example 4 of the present application. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0060] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0061] Example 1

[0062] like Figure 1 As shown, this embodiment proposes a training method for a waveform data processing model, and the training method for the waveform data processing model includes the following steps.

[0063] Step S1, obtaining original waveform data and its corresponding particle real label data. The original waveform data is multi-channel particle waveform data collected by using a TPC detector, the TPC detector includes multiple channels, and the multiple channels synchronously collect particle waveform data to form the multi-channel particle waveform data; the particle real label data includes a real label for characterizing the actual category of the particle.

[0064] Step S2: performing downsampling processing and data compression processing on the original waveform data in sequence to obtain pre-processed original waveform data.

[0065] Step S3, input the preprocessed raw waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, and determine the network loss according to the predicted probability values ​​of each category of particles output by the waveform data processing model and the true label of the particles. Wherein, the network loss refers to the network loss value of the loss function, and the waveform data processing model is a model with a network structure of multiple channels established based on an artificial neural network. The network structure of each channel corresponds to the input of the preprocessed raw waveform data of a spatial dimension. The network structure of each channel includes multiple convolutional layers and multiple pooling layers, and is constructed by first passing through all convolutional layers and then passing through all pooling layers. The convolution kernel sizes of different convolutional layers in the same channel are different.

[0066] Step S4: According to the network loss, the network parameters of the waveform data processing model are updated through back propagation to iterate the waveform data processing model multiple times to obtain a trained waveform data processing model.

[0067] In this embodiment, step S2 sequentially performs downsampling processing and data compression processing on the original waveform data to obtain pre-processed original waveform data, which specifically includes the following steps.

[0068] Step S21 : down-sample the original waveform data by using interval resampling to obtain down-sampled original waveform data.

[0069] Step S22: compress and save the downsampled original waveform data in a sparse tensor format in the order of the channel numbers of the TPC detector to obtain the preprocessed original waveform data.

[0070] In this embodiment, the particle categories mainly include α particles, β particles, muons, etc. That is to say, the particle category may be α particles, β particles, muons, etc.

[0071] like Figure 2 As shown, this embodiment uses a TPC detector, which consists of a drift region and an end cap readout detector. The end cap readout detector mainly includes a tank body and two end faces, which are arranged opposite to each other. The two end faces are a cathode plane and an anode plane, respectively, and the remaining faces are the sides of the tank body, as shown in FIG. Figure 2 As shown, a readout bar is provided on the anode plane, and the readout bar is used to read out the electronic signal. In the TPC detector, the drift region of the TPC detector is filled with working gases such as argon and isobutane, and a uniform drift electric field is applied. When charged particles pass through the drift region, these charged particles will ionize the working gas molecules to produce electron ion pairs. Under the action of the uniform drift electric field inside the drift region, electrons will drift to the anode region and produce avalanche amplification. Finally, the electronic signal is read out by the readout bar of the end cap readout detector. After pre-amplification and waveform digitization, the original waveform data of multiple channels can be obtained.

[0072] In this embodiment, the data of a single channel of the TPC detector is a one-dimensional array composed of a series of numbers. Since the TPC detector has multiple channels, the data obtained by the TPC detector is a multi-dimensional array, and the number of dimensions is equal to the number of channels. The multi-channel particle waveform data of the TPC detector refers to the particle waveform data of multiple channels obtained in the TPC detector after pre-amplification and waveform digitization.

[0073] The waveform data processing process based on the TPC detector in this embodiment mainly includes a data preprocessing algorithm, a neural network algorithm, an algorithm training method, and an algorithm deployment.

[0074] The data preprocessing algorithm includes resampling the multi-channel particle waveform data, i.e. the original waveform data, collected by the TPC detector through interval resampling, taking the peak position of the waveform as the center of the interval, selecting a suitable interval, resampling the interval while retaining the basic waveform information, reducing the sampling rate of the waveform, and compressing the data according to the channel number of the TPC detector and saving it in the format of a sparse tensor. The first column of the sparse tensor is the channel number of the triggered channel, and the second column is the corresponding waveform data. At the same time, the sparse tensor also saves the dimension size and data type of the data.

[0075] In this embodiment, interval resampling includes interval sampling and interval resampling. Interval sampling is performed first, and then interval resampling is performed on the basis of interval sampling. Since the characteristic of the sampled waveform is a pulse signal, the peak position of the pulse is taken as the center of the interval, and a suitable interval is selected as the sampling interval. This interval needs to contain the complete rising and falling edges of the pulse. Then, interval resampling is performed within the above-mentioned sampling interval. Specifically, sampling is performed once every several points within the sampling interval, and all sampled points are combined into a new waveform signal. The interval between the sampling points needs to ensure that the original waveform pulse signal is not distorted. At this point, the entire interval resampling is completed.

[0076] The neural network algorithm includes dividing the above-mentioned preprocessed raw waveform data into different dimensions, and sending the preprocessed raw waveform data of different dimensions as input into a multi-channel waveform data processing model. For the preprocessed raw waveform data, the waveform data of each dimension is used as the input of a channel. The waveform data of each dimension is first processed by multiple convolution layers and then by the pooling layer. The convolution layer uses a set of learnable filters (convolution kernels) to perform convolution operations on the input to capture features of different scales. The output of the convolution layer is downsampled by the pooling layer to reduce the dimension of the feature map. After the data output by the pooling layer of each channel is reduced in dimension, it is spliced ​​into a one-dimensional tensor form, and the spliced ​​data is input into a neural network module composed of multiple fully connected layers. In each fully connected layer, the input is weighted and summed, and nonlinear transformation is performed through the activation function, such as the ReLU function, to further extract features and classify, and map the input to the output space. The last layer activation function uses the sigmoid function to compress the output to the range of [0, 1]. The network finally normalizes the output particle identification results through the Softmax layer.

[0077] In this embodiment, the pre-processed waveform data is divided into dimensions (such as X dimension and Y dimension) and input into a multi-channel artificial neural network in parallel. The meaning of the dimension division is the same as the channel division, because when the TPC detector is collected, the number of dimensions is equal to the number of channels of the neural network and the number of types of detector channels, that is, the number of dimensions, the number of channels of the neural network and the number of types of detector channels are all equal. Each channel corresponds to waveform data of one dimension. The waveform data of each dimension is first processed by multiple convolutional layers and then by pooling layers. After the data output by the pooling layer of each channel is reduced in dimension, it is spliced ​​into a one-dimensional tensor form. The spliced ​​data is input into a neural network composed of multiple fully connected layers and output, and the network normalizes the output of the particle identification results through the Softmax layer. By comparing with the original true label, the network parameters are updated by back propagation and iterative training.

[0078] In the preprocessing process, since the multidimensional array has been converted into a multidimensional sparse tensor, the format of the data input to the waveform data processing model is a multidimensional sparse tensor format, and the output result can be regarded as the predicted probability value of each particle event being judged as each type of particle, that is, the confidence value. The categories of particles can include alpha particles, beta particles, and muons. Therefore, the output result of the waveform data processing model is the probability value of predicting that the current particle event is an alpha particle, the probability value of predicting that the current particle event is a beta particle, and the probability value of predicting that the current particle event is a muon.

[0079] In this embodiment, the training method of the algorithm includes dividing the collected particle waveform data and background waveform data into a training set and a test set according to a certain ratio. In this embodiment, the sample ratio of the training set and the test set is 8:2. The training set is used for the parameter update and training process of the neural network, and the test set is used to evaluate the performance of the network on unseen data. At the same time, in the training of the network, regularization methods such as a dynamic learning rate algorithm and a Dropout layer are used to further improve the accuracy of the network and prevent the occurrence of overfitting. In this embodiment, the deployment of the algorithm includes loading the particle case data and the background case data into the trained waveform data processing model and outputting the results. The output result is a predicted probability value, which can be regarded as the predicted probability value of each case being judged as each type of particle, that is, the confidence. By setting the confidence threshold, the output results are strictly screened, and only particles with consistent labels and predicted probability values ​​greater than or equal to the confidence threshold will be judged as correct category recognition, thereby obtaining the background screening rate and particle retention rate.

[0080] Among them, label consistency means that the predicted label of the particle predicted and output by the waveform data processing model is consistent with its actual label. In practical applications, the trained waveform data processing model is used to classify and screen particles. The waveform data of the particle can be input into the trained waveform data processing model, and the trained waveform data processing model can predict the probabilities of different particle categories corresponding to the waveform data of the particle. Select the predicted probability value A corresponding to the category Q with the largest predicted probability value, and compare it with the confidence threshold B. If the predicted probability value A is greater than or equal to the confidence threshold B, then it is determined that the category corresponding to the particle waveform data is Q, and Q can be one of the categories of α particles, β particles or muons. If the predicted probability value A is less than the confidence threshold B, then it is determined that the category corresponding to the particle waveform data is background. Here, background refers to all other signals or noises except the target signal during the measurement process. This case is a background case.

[0081] Figure 3 is a schematic diagram of a waveform data processing method based on a TPC detector, such as Figure 3 As shown, the specific process of waveform data processing mainly includes the following contents.

[0082] 1) Based on the original waveform data collected by the TPC detector, the original waveform data is multi-channel particle waveform data, and then the collected original waveform data is subjected to a pre-processing process of downsampling processing and data compression processing. Figure 4 It is a waveform diagram of the original waveform data; Figure 5 It is a waveform diagram of the original waveform data after data compression processing; Figure 6 The figure is a waveform diagram of the original waveform data after downsampling. During preprocessing, the fixed interval where the pulse peak is located is first selected for the original waveform data, and downsampling is performed by interval sampling to obtain the waveform data after downsampling. Then, the downsampled waveform data is combined into waveform data in sparse tensor format according to the channel number sequence. Only the position and value of non-zero elements need to be stored to obtain the preprocessed original waveform data. Among them, the channel number sequence is determined by the TPC detector. The channel number sequence of the TPC detector can be data of channel number 1, data of channel number 2, data of channel number 3, and so on.

[0083] 2) The pre-processed raw waveform data is input into the pre-built waveform data processing model in different dimensions for training. In this embodiment, the model is firstly trained through multiple convolution layers and then through pooling layers. Figure 7 and Figure 8As shown, the preprocessed raw waveform data obtained in 1), that is, the waveform data in sparse tensor format, is input into the waveform data processing model in dimensions, and each dimension corresponds to a neural network input channel in the model. In each input channel, the waveform data is first converted into a normal tensor format, and then the waveform data of each input channel is normalized. The waveform data after normalization is first passed through a plurality of connected convolutional layer networks, and then the data of each channel is output through the pooling layer.

[0084] 3) The data of each channel output by the pooling layer is processed by dimensionality reduction and splicing, and the final particle type identification result is output through the fully connected layer. Figure 7 and Fig. 9 As shown, the result of each channel in 2) is still a two-dimensional tensor. In order to facilitate the dimensionality reduction and feature extraction of the fully connected layer, the data output by each channel of the neural network is reduced in dimensionality and combined into a one-dimensional tensor form. The spliced ​​data is input into a neural network composed of multiple fully connected layers to output the identification result representing the particle type. The last layer activation function uses the sigmoid function. The network outputs the identification result through the Softmax layer. The identification result is a one-dimensional tensor with a length of M, where M is the number of categories of particles to be identified. Each value of the vector represents the predicted probability value of the input case corresponding to each particle, that is, the confidence value. By comparing with the real label representing the actual category of the particle, back propagation and iterative training are performed to update the network parameters of the waveform data processing model, and a trained waveform data processing model is obtained, thereby improving the performance of the waveform data processing model and improving the classification accuracy of the waveform data processing model for particle categories.

[0085] 4) The target waveform data is loaded into the trained waveform data processing model, and the output result of the network Softmax layer is used as the predicted probability value. The target waveform data is screened by setting the confidence threshold related to the probability.

[0086] Among them, the target waveform data is different from the original waveform data. The original waveform data is used to train the waveform data processing model to obtain a trained waveform data processing model. The target waveform data is the waveform data to be classified and screened that is actually processed by the trained waveform data processing model after the trained waveform data processing model is obtained. The trained waveform data processing model is used to perform particle classification and case screening on the target waveform data. When screening data, the cases whose output prediction probability values ​​are greater than or equal to the confidence threshold are retained, and the cases whose output prediction probability values ​​are less than the confidence threshold are screened out. By setting the confidence threshold, this embodiment increases the screening rate of background cases while retaining higher cases.

[0087] This embodiment takes 512 sampling point data of 240 channels collected by the TPC detector as an example to illustrate the specific implementation process of the technical solution of this embodiment, and the specific process includes the following contents.

[0088] (1) Down-sampling and data compression are performed on the acquired multi-channel sampling waveform.

[0089] The TPC detector used in this embodiment has a total of 240 data readout channels, with X dimension and Y dimension each occupying 120 channels, and each channel having a total of 512 sampling points.

[0090] The peak value of each channel waveform is roughly located near the 150th sampling point, such as Figures 4 to 6 As shown, the preprocessing method based on downsampling and data compression of this embodiment is used to select the interval from the 50th to the 250th sampling point, and the channel waveform is retained to 100 sampling points by sampling at one point intervals, and the above operation is performed on the waveforms of all channels. Then, the data obtained by downsampling is combined into a sparse tensor format according to the channel number sequence.

[0091] (2) The processed waveform data is input into the neural network in different dimensions, first through multiple convolutional layers, and then through pooling layers for processing and training.

[0092] In this embodiment, the waveform data in the sparse tensor format obtained by (1) is input into the dual-channel neural network in batches, and the X dimension and Y dimension correspond to a neural network input channel respectively. In each input channel, the waveform data is first converted into a normal tensor format, and then the waveforms of the 120 channels of the data are normalized. The processed waveform data first passes through 3 connected convolutional layer networks, and then passes through the pooling layer to output the channel results. Finally, each channel outputs a tensor with a shape of 15×4.

[0093] (3) After reducing the dimension of the data output by each channel and splicing them, the identification result is output through the fully connected layer.

[0094] In this embodiment, the data output by each channel of the neural network in (2) is flattened and dimensionally reduced, and each channel obtains a one-dimensional tensor of length 60. Then, each one-dimensional tensor is sequentially combined and spliced ​​into a one-dimensional tensor of length 120. The spliced ​​one-dimensional tensor is input into a neural network module composed of 3 fully connected layers to output the identification result. The output result of the network is a one-dimensional tensor of length 2, and each value represents the predicted probability value of the input case corresponding to each particle. By comparing with the label, back propagation is performed and iterative training is performed.

[0095] (4) Load the data onto the trained network, regard the output result of the network Softmax layer as the predicted probability value, and filter the data by setting the confidence threshold related to the probability.

[0096] In this embodiment, the cases whose predicted probability values ​​are greater than or equal to the confidence threshold are retained, and the cases whose predicted probability values ​​are less than the confidence threshold are screened out. After setting a reasonable confidence threshold, the changes in the distribution of the hit positions of the cases before and after the screening are finally screened out accurately.

[0097] This embodiment sets a reasonable confidence threshold, determines the cases whose output confidence values ​​are greater than or equal to the confidence threshold as the required case particles and retains them, and determines the cases whose output confidence values ​​are less than the confidence threshold as background data, and screens out the background data. This embodiment conducts experiments on β sample cases according to the above confidence threshold screening method. The experimental results show that the number of β sample cases before screening is 87340, the number of remaining β sample cases after screening is 48242, and the retention rate is 55.0%; the number of background sample cases before screening is 19161, the number of remaining background sample cases after screening is 302, and the screening rate is 98.4%. Figures 10 to 13 It is the distribution diagram of the starting points obtained by different sample data before and after the background cases are screened out using the above algorithm, where: Fig.10 is the distribution diagram of the starting points of the β sample cases, Fig.11 is the distribution diagram of the starting points of the remaining cases of the β sample after screening, Fig.12 is the distribution diagram of the starting points of the background sample cases, Fig.13 This is the distribution diagram of the starting points of the remaining cases of the background samples after screening. Figures 10 to 13 It can be seen intuitively that the starting point distribution of the β sample case is more concentrated, indicating that most of the background cases outside the incident window are screened out, while the particles regarded as real cases within the incident window are mostly retained, which illustrates the effectiveness of the technical solution of this embodiment in screening case particles. The starting point distribution of the background case is usually two prominent circular patterns, with fewer outer circle cases and more inner circle cases, which is caused by the mechanical structure of the incident window being a step shape. This illustrates the accuracy of the starting point distribution of the technical solution of this embodiment.

[0098] This embodiment establishes, trains and applies a waveform data processing model based on the TPC detector, which greatly compresses the data volume while retaining relatively complete waveform information, effectively reducing the amount and complexity of data calculation, and effectively reducing the consumption of computer memory and other hardware during the algorithm processing. In addition, this embodiment directly processes the original waveform data output by the TPC detector instead of extracting multi-dimensional physical features, so that the waveform information and time information in the waveform can be used as much as possible, avoiding the subjectivity and large errors in feature selection of traditional algorithms, and improving the accuracy of classification and screening of particle waveform data. Compared with traditional algorithms, the idea is simple, the program is highly portable, and the background resolution is higher.

[0099] Example 2

[0100] like Fig.14 As shown, this embodiment provides an application method of a waveform data processing model. The application method is actually a method for using the waveform data processing model trained in Example 1. The application method of the waveform data processing model includes the following steps.

[0101] 101. Obtain target waveform data, where the target waveform data is multi-channel particle waveform data of particles to be measured collected by using a TPC detector.

[0102] 102. Perform downsampling processing and data compression processing on the target waveform data in sequence to obtain pre-processed target waveform data.

[0103] 103. Input the preprocessed target waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel to obtain the predicted probability value of each category of the particle to be measured. The waveform data processing model is a model trained according to the training method of the waveform data processing model described in Example 1.

[0104] In this embodiment, after step 103, the application method of the waveform data processing model may further include the following steps.

[0105] 104. Set the confidence threshold.

[0106] 105. Compare the predicted probability value corresponding to the highest category among the predicted probability values ​​of each category of the particles to be tested with the confidence threshold to screen background cases and particle cases. Specifically, the case where the predicted probability value is less than the confidence threshold is determined as a background case, and the case where the predicted probability value is greater than or equal to the confidence threshold is determined as a particle case, and all background cases are eliminated, and all particle cases are retained.

[0107] Example 3

[0108] like Fig.15 As shown, this embodiment provides a training system for a waveform data processing model, which is actually a system corresponding to the training method of Embodiment 1. The training system for the waveform data processing model includes the following functional modules.

[0109] The first data acquisition module M1 is used to acquire original waveform data and its corresponding particle real label data; the original waveform data is multi-channel particle waveform data collected by using a TPC detector, and the TPC detector includes multiple channels, and the multiple channels synchronously collect particle waveform data to form the multi-channel particle waveform data; the particle real label data includes a real label for characterizing the actual category of the particle.

[0110] The first preprocessing module M2 is used to perform downsampling processing and data compression processing on the original waveform data in sequence to obtain preprocessed original waveform data.

[0111] The network loss determination module M3 is used to input the preprocessed original waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, and determine the network loss according to the predicted probability values ​​of each category of particles output by the waveform data processing model and the true label of the particles; the waveform data processing model is a model with a multi-channel network structure established based on an artificial neural network, and the network structure of each channel includes multiple convolution layers and multiple pooling layers. The waveform data processing model is built by first passing through all convolution layers and then passing through all pooling layers.

[0112] The training module M4 is used to update the network parameters of the waveform data processing model through back propagation according to the network loss, so as to iterate the waveform data processing model multiple times to obtain a trained waveform data processing model.

[0113] In this embodiment, the first preprocessing module M2 includes the following functional units.

[0114] The downsampling unit is used to downsample the original waveform data by adopting an interval resampling method to obtain the downsampled original waveform data.

[0115] A data compression unit is used to compress and save the downsampled original waveform data into a sparse tensor format in the order of the channel number of the TPC detector to obtain the preprocessed original waveform data.

[0116] Example 4

[0117] like Fig.16 As shown, this embodiment provides an application system of a waveform data processing model, which is actually a system corresponding to the application method of Embodiment 2. The application system of the waveform data processing model includes the following functional modules.

[0118] The second data acquisition module N1 is used to acquire target waveform data, where the target waveform data is multi-channel particle waveform data of the particles to be measured acquired by using a TPC detector.

[0119] The second preprocessing module N2 is used to perform downsampling processing and data compression processing on the target waveform data in sequence to obtain preprocessed target waveform data.

[0120] The target particle category prediction module N3 is used to input the preprocessed target waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, so as to obtain the predicted probability value of each category of the particle to be measured; the waveform data processing model is a model trained according to the training method of the waveform data processing model described in Example 1.

[0121] The confidence threshold setting module N4 is used to set the confidence threshold.

[0122] The screening module N5 is used to compare the predicted probability value corresponding to the highest category among the predicted probability values ​​of each category of the particles to be tested with the confidence threshold, determine the case where the predicted probability value is less than the confidence threshold as a background case, and determine the case where the predicted probability value is greater than or equal to the confidence threshold as a particle case, and eliminate all background cases and retain all particle cases.

[0123] This embodiment preprocesses the multi-channel signal waveform data of the TPC detector obtained after pre-amplification and waveform digitization to obtain waveform data with high compression ratio while maintaining high resolution. The waveform data of each channel is simultaneously input into the pre-built and trained multi-channel artificial neural network model according to the spatial distribution to obtain the particle identification result. In the context of using the original waveform data of the TPC detector, high-precision particle identification is achieved based on machine learning algorithms such as artificial neural networks. Under the premise of achieving a reasonable retention rate for the particles of interest, most of the background cases are screened out. In addition, making full use of the time and space information of the interaction between the incident particles and the TPC detector hidden in the signal waveform of the TPC detector can effectively increase the particle identification capability, which is particularly suitable for realizing high-sensitivity detection of α and β radioactive nuclides in a ray environment, and has broad application prospects in the fields of environmental radioactivity detection and radiopharmaceutical research.

[0124] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A training method for a waveform data processing model, characterized in that: The training method of the waveform data processing model includes: Acquire original waveform data and its corresponding particle real label data; the original waveform data is multi-channel particle waveform data acquired by using a TPC detector, the TPC detector includes multiple channels, and the multiple channels synchronously acquire particle waveform data to form the multi-channel particle waveform data; the particle real label data includes a real label for characterizing the actual category of the particle; The original waveform data is sequentially subjected to downsampling processing and data compression processing to obtain preprocessed original waveform data; The preprocessed raw waveform data is input into each channel of the waveform data processing model according to the spatial dimension of the channel, and the network loss is determined according to the predicted probability values ​​of each category of particles output by the waveform data processing model and the true label of the particles; the waveform data processing model is a model with a multi-channel network structure established based on an artificial neural network, and the network structure of each channel includes multiple convolution layers and multiple pooling layers. The waveform data processing model is built by first passing through all convolution layers and then passing through all pooling layers. According to the network loss, the network parameters of the waveform data processing model are updated through back propagation to iterate the waveform data processing model multiple times to obtain a trained waveform data processing model.

2. The training method of the waveform data processing model according to claim 1, characterized in that: The original waveform data is subjected to downsampling processing and data compression processing in sequence to obtain the preprocessed original waveform data, specifically including: Downsampling the original waveform data by using interval resampling to obtain downsampled original waveform data; According to the order of the channel numbers of the TPC detector, the downsampled original waveform data is compressed and saved in a sparse tensor format to obtain the preprocessed original waveform data.

3. The training method of the waveform data processing model according to claim 1, characterized in that: The particle classes include alpha particles, beta particles, and muons.

4. An application method of a waveform data processing model, characterized in that: The application method of the waveform data processing model includes: Acquire target waveform data, where the target waveform data is multi-channel particle waveform data of the particles to be measured collected by using a TPC detector; Performing downsampling processing and data compression processing on the target waveform data in sequence to obtain preprocessed target waveform data; The preprocessed target waveform data is input into each channel of the waveform data processing model according to the spatial dimension of the channel to obtain the predicted probability value of each category of the particle to be measured; the waveform data processing model is a model trained according to the training method of the waveform data processing model according to claim 1.

5. The application method of the waveform data processing model according to claim 4, characterized in that: After the step of inputting the preprocessed target waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel to obtain the predicted probability value of each category of the particle to be detected, the application method of the waveform data processing model also includes: Set confidence threshold; The predicted probability value corresponding to the highest category among the predicted probability values ​​of each category of the particles to be tested is compared with the confidence threshold. The case where the predicted probability value is less than the confidence threshold is determined as a background case, and the case where the predicted probability value is greater than or equal to the confidence threshold is determined as a particle case. All background cases are eliminated and all particle cases are retained.

6. A training system for a waveform data processing model, characterized in that: The training system of the waveform data processing model includes: A first data acquisition module is used to acquire original waveform data and its corresponding particle real label data; the original waveform data is multi-channel particle waveform data acquired by using a TPC detector, the TPC detector includes multiple channels, and the multiple channels synchronously acquire particle waveform data to form the multi-channel particle waveform data; the particle real label data includes a real label for characterizing the actual category of the particle; A first preprocessing module, used for sequentially performing downsampling processing and data compression processing on the original waveform data to obtain preprocessed original waveform data; A network loss determination module is used to input the preprocessed original waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, and determine the network loss according to the predicted probability values ​​of each category of particles output by the waveform data processing model and the real label of the particles; the waveform data processing model is a model with a multi-channel network structure established based on an artificial neural network, and the network structure of each channel includes multiple convolution layers and multiple pooling layers. The waveform data processing model is built by first passing through all convolution layers and then passing through all pooling layers. The training module is used to update the network parameters of the waveform data processing model through back propagation according to the network loss, so as to iterate the waveform data processing model multiple times to obtain a trained waveform data processing model.

7. The training system of the waveform data processing model according to claim 6, characterized in that: The first preprocessing module comprises: A downsampling unit, configured to downsample the original waveform data by adopting an interval resampling method to obtain the downsampled original waveform data; A data compression unit is used to compress and save the downsampled original waveform data into a sparse tensor format in the order of the channel number of the TPC detector to obtain the preprocessed original waveform data.

8. The training system of the waveform data processing model according to claim 6, characterized in that: The particle classes include alpha particles, beta particles, and muons.

9. An application system of a waveform data processing model, characterized in that: The application system of the waveform data processing model includes: A second data acquisition module is used to acquire target waveform data, where the target waveform data is multi-channel particle waveform data of the particles to be measured acquired by using a TPC detector; A second preprocessing module is used to perform downsampling processing and data compression processing on the target waveform data in sequence to obtain preprocessed target waveform data; A target particle category prediction module is used to input the preprocessed target waveform data into each channel of the waveform data processing model according to the spatial dimension of the channel, so as to obtain the predicted probability value of each category of the particle to be measured; the waveform data processing model is a model trained according to the training method of the waveform data processing model according to claim 1.

10. The application system of the waveform data processing model according to claim 9, characterized in that: The application system of the waveform data processing model also includes: A confidence threshold setting module, used to set a confidence threshold; The screening module is used to compare the predicted probability value corresponding to the highest category among the predicted probability values ​​of each category of the particles to be tested with the confidence threshold, determine the case where the predicted probability value is less than the confidence threshold as a background case, and determine the case where the predicted probability value is greater than or equal to the confidence threshold as a particle case, and eliminate all background cases and retain all particle cases.

Images