Pulse ionization chamber ionization signal and vibration signal distinguishing method based on neural network

Abstract

The invention discloses a pulse ionization chamber ionization signal and vibration signal distinguishing method based on a neural network, and the method comprises the steps: firstly, collecting a pulse ionization chamber ionization signal data set and a vibration signal data set; preprocessing and standardizing the ionization signal data set and the vibration signal data set, wherein the signal preprocessing and standardizing comprise differentiated length normalizing preprocessing, self-adaptive amplitude standardizing based on median and quartile distance, data set division and final standardizing processing; a model architecture based on the neural network is constructed, wherein the model architecture comprises a causal expansion convolution layer, five user-defined residual blocks, residual jump connection, an adaptive pooling classifier and updating optimization of model parameters; and finally, performing model evaluation and distinguishing effect test based on a neural network: selecting a cross entropy loss function, accuracy and an F1 score as model evaluation indexes. According to the invention, a vibration signal and an ionization signal can be effectively distinguished, the influence of vibration interference on the accuracy of a measurement result on an instrument is reduced, and more accurate radon concentration measurement is realized.

Description

Technical Field

[0001] This invention belongs to the field of radiation detection and intelligent signal processing, and in particular, it is a method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks. Background Technology

[0002] To effectively monitor radon concentration and assess its potential radiation risks, various radon measurement devices based on different detection principles have been developed both domestically and internationally. The pulse ionization chamber method has been widely used in radon concentration monitoring due to its advantages such as fast response speed, high detection efficiency, relatively low cost, and resistance to radiation damage. However, this method is sensitive to mechanical vibration in practical applications and is easily affected by environmental disturbances, thus impacting measurement accuracy and limiting its reliability and stability in complex field environments.

[0003] The pulsed ionization chamber works by detecting the ionization process initiated by alpha particles within the chamber to obtain radon concentration data. However, when the instrument is in a mechanically vibrating environment, such as near a construction site or subway line, external vibrations can cause micro-displacements of the electrodes within the chamber, leading to instantaneous changes in the electric field and charge redistribution. This process may generate vibration interference pulses with amplitudes similar to the alpha particle ionization pulses. These interference pulses are often misinterpreted by the instrument as valid signals, resulting in an increase in non-physical counts, thus causing significant measurement errors and reducing the accuracy and reliability of the data.

[0004] Currently, the ALPHAGUARD radon detector in Germany has successfully distinguished between vibration signals and ionization signals generated by high-energy alpha particles in a pulsed ionization chamber, effectively improving measurement accuracy. However, the core algorithm of this device has not been disclosed, causing domestic instrument manufacturers to face algorithmic bottlenecks when developing radon detectors with similar anti-vibration interference capabilities, and they have yet to find an efficient signal differentiation method specifically for pulsed ionization chambers. Therefore, how to reduce the interference of vibration signals on measurement results, reduce instrument deviation, and improve the measurement accuracy and stability of radon detectors has become a key technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for distinguishing ionization signals and vibration signals in a pulsed ionization chamber based on neural networks, comprising the following specific steps:

[0007] Step 1: Acquisition of ionization signal dataset and vibration signal dataset from the pulse ionization chamber;

[0008] Step 2: Preprocessing and standardization of ionization signal dataset and vibration signal dataset: The signal preprocessing and standardization include differential length normalization preprocessing, adaptive amplitude standardization based on median and interquartile range, dataset partitioning, and final standardization.

[0009] Step 3: Construct a neural network-based model architecture, including causal dilated convolutional layers, five custom residual blocks, residual skip connections, an adaptive pooling classifier, and optimization of model parameters.

[0010] Step 4: Model evaluation and discrimination test based on neural networks: Cross-entropy loss function, accuracy and F1 score are selected as model evaluation indicators.

[0011] A further technical solution of the present invention is as follows: the data acquisition process includes firstly, the sampling circuit captures the output signal of the pulse ionization chamber and converts the analog signal into a digital signal through an analog-to-digital converter (ADC); then, the sampling microcontroller processes the digital signal and sets a threshold to filter noise; finally, the acquired digital signal is transmitted to the main microcontroller after the measurement cycle ends, the main microcontroller stores the data in an SD card, and transmits it to the host computer via a serial port.

[0012] A further technical solution of the present invention is as follows: the specific operation of the noise filtering is to start data acquisition when the amplitude of the detected digital signal exceeds the threshold; and to stop acquisition when the amplitude of the digital signal drops below the threshold.

[0013] A further technical solution of the present invention is as follows: when collecting ionization signal data in step one, the radon source is placed at the cavity inlet of the pulse ionization chamber, so that the radon gas diffuses into the cavity. The radon gas decays in the cavity, releases alpha particles, induces air ionization and generates a weak ionization signal; when collecting vibration signal data, vibration signals are generated by striking with different forces.

[0014] A further technical solution of the present invention is: the specific steps of the differential length normalization preprocessing in step two are as follows: all input signals are uniformly adjusted to a preset target length L. target If the value is 512, then the following signal conditioning methods are included:

[0015] When the length of the original signal is less than L target When the value is 4, the original signal is compensated, and the formula is expressed as follows:

[0016]

[0017] Where, x target This represents the signal after length warping; x represents the original signal; tile(·) represents a periodic repetition operation, and [.] represents a floor operation; :Ltarget Indicates taking the first L target One sampling point.

[0018] When the length of the original signal is greater than L target When / 4 is less than 300, the preprocessing formula is expressed as:

[0019] x target =reflect_pad(x,L target (2)

[0020] Here, reflect_pad() represents the reflection padding operation.

[0021] For signal sequences of medium length, the Dynamic Time Warping (DTW) algorithm is first used for time alignment. The core operation of time alignment is as follows:

[0022] x target =DTW(x,t) ref (3)

[0023] Among them, t ref The constructed reference template, with a length of 512, is defined as a normalized time series sampled at equal intervals in the interval [0,1], used to provide a unified time anchor point to align multi-source signals;

[0024] Then, when the signal contains a lot of noise, extreme distortion, or the length differs too much from the template, causing DTW matching to fail or the alignment path to be unstable, linear interpolation is used for proportional resampling to adjust the original signal to the target length. The formula is:

[0025] x target =interp(x,L target ,L) (4)

[0026] Where interp(·) represents the length normalization function based on linear interpolation.

[0027] For ultra-long signals exceeding 1000, a main energy segment truncation strategy based on the sliding energy spectrum is adopted. Let the sampled long signal be represented as x = [x1, x2, ..., x]. L The sliding energy function E[i] is defined as follows:

[0028]

[0029] Where, x i+kLet h[k] represent the amplitude of the signal at the (i+k)th sampling point, and h[k] be the Hamming function with a window width w = 200, used to suppress edge effects and reduce spectral leakage. By traversing all sliding window positions, the global energy spectrum distribution is obtained, and then the time point of the main energy concentration is located using the maximum energy criterion.

[0030] i peak =argmax i E[i] (6)

[0031] Where, argmax i E[i] represents the index of the position where E[i] reaches its maximum value among all possible values ​​of i; peak This represents the core region with the most concentrated energy in the corresponding signal, which usually contains key dynamic patterns.

[0032] To further focus on this critical area and reduce redundant interference, i peak A local sub-segment of length 256 is extracted from the center as the main energy region. The specific extraction method is as follows:

[0033] x core =x[i peak -128:i peak +128] (7)

[0034] Where, x core This represents the most discriminative local core segment extracted from the original signal x; to meet the model's requirement for a uniform input length, the amplitude of the last sample point of this local segment is repeated until it is extended to the required length.

[0035] The specific steps of the adaptive amplitude normalization based on the median and interquartile range are as follows: for each length-normalized signal x... target First, calculate the median, then shift the signal to a distribution centered at zero to obtain the centered signal, as shown in the following formula:

[0036] x'=x target -median(x target (8)

[0037] Where x′ represents the signal after centering; median(·) represents the median calculation function, used to extract the median value of all values ​​in the input sequence.

[0038] To enhance robustness, interquartile range (IQR) is further used to scale the amplitude of the centered signal. Specifically, the 25th and 75th percentiles of x′, i.e., the first quartile Q1 and the third quartile Q3, are first calculated, and the interquartile range is obtained accordingly, as shown in the following formula:

[0039]

[0040] Among them, percentile 25 (·) and percentile 75 (·) represents the calculation functions for the 25th and 75th percentiles, respectively, used to measure the upper and lower quartile boundaries of a signal.

[0041] If the interquartile range is greater than the set fault tolerance threshold ε, then set ε to 1×10. -6 If the interquartile range is less than the fault tolerance threshold ε, then the centering operation is performed; otherwise, only the centering operation is retained. The specific normalization formula is as follows:

[0042]

[0043] Where, x norm This is the final normalized signal sequence.

[0044] The normalized signal sequence is augmented using a dataset signal enhancement method. The enhancement is achieved by randomly combining the following four strategies with independent probabilities of 0.5:

[0045] A, Amplitude perturbation: Applying random scaling to the signal amplitude, the specific expression is:

[0046] x aug =a·x train ,a~u(0.7,1.3) (11)

[0047] Where, x aug Indicates the enhanced signal, x train Let represent the original signal of the training set, and 'a' represent a random scaling factor that follows a uniform distribution in the interval [0.7, 1.3].

[0048] B, Timing Distortion: Nonlinear stretching or compression of the signal time axis, specifically expressed as follows:

[0049] x aug =interp(x train ,r),r~u(0.8,1.2) (12)

[0050] Where interp(·) represents the interpolation resampling operation, and r represents the random time scaling factor, which is used to control the degree of compression or expansion;

[0051] C, Additive Gaussian noise: Adding zero-mean Gaussian white noise to a signal, the specific expression is:

[0052] x aug =x train+σ·N(0,1),σ~u(0.01,0.05) (13)

[0053] Where N(0,1) represents a standard Gaussian distribution, and σ represents the noise intensity sampled uniformly from the interval [0.01,0.05].

[0054] D, Random Pruning and Padding: Randomly prune a continuous subsequence of length 256 from the signal and extend it to a uniform length of 512 using edge copying.

[0055] The enhanced signal is then subjected to final standardization processing. The specific formula for standardization is as follows:

[0056]

[0057] Where, x final This represents the signal after final standardization; μ = mean(x) represents the mean of the signal; ρ = std(x) represents the standard deviation of the signal; δ = 1 × 10⁻⁶. -6 This represents a small constant introduced to prevent division by zero errors when the standard deviation is zero;

[0058] Specifically, the dataset is divided into a test set, a training set, and a validation set.

[0059] A further technical solution of the present invention is: the causal dilated convolutional layer in step three is configured with a kernel size of 7 and a dilation rate of 1. The mathematical expression for the dilated convolution mechanism introduced in this layer is as follows:

[0060]

[0061] Where z(t) represents the convolution output at time step t; x(t) represents the input sequence, i.e., x final w(k) represents the k-th weight of the convolution kernel, k is the kernel size, and d is the inflation factor.

[0062] The custom residual consists of two layers of causal dilation convolutions. Each layer is followed by batch normalization, LeakyReLU activation function and regularization module. The input is adjusted by 1×1 convolution and then added to the output to form a residual jump connection.

[0063] The normalized mathematical expression is as follows:

[0064]

[0065] Where BatchNorm1d(z) represents the output layer result of the layer normalization; z represents the convolution output vector, which is the input feature of this layer; μ represents the mean of the input vector z; γ represents the standard deviation of the input vector z; γ represents the learned scaling parameter used to scale the normalized result; β represents the learned bias parameter used to shift the normalized result.

[0066] The specific formula for the LeakyReLU activation function is as follows:

[0067]

[0068] Where α represents the leakage coefficient of the negative half-shaft.

[0069] The residual jump connection is processed as follows: when the input and output dimensions are the same, an identity mapping is used; when the dimensions are different, a 1×1 convolution is introduced for linear mapping, and the residual block output is represented as:

[0070] y = z² + Skip(x) (18)

[0071] Where y represents the final output of the residual block, z2 represents the output value after two convolutional layers, and Skip(x) represents the residual path function, which is defined as follows:

[0072]

[0073] Among them, Conv 1×1 (x) represents a 1×1 convolution operation performed on the input x; C in and C out These represent the number of channels for the input and output features, respectively.

[0074] The adaptive pooling classifier includes an adaptive average pooling layer, a flattening layer, a first fully connected layer, a LeakyReLU activation function, a regularization module, and a second fully connected layer. The classifier output is represented as follows:

[0075] Y = Classifier(y) (20)

[0076] Where Y represents the binary classification result output after the second fully connected layer; Classifier(.) represents the classifier function that includes average pooling, flattening and fully connected layer structure.

[0077] The model parameters are updated and optimized using the cross-entropy loss function and learning rate scheduling; the cross-entropy loss function is defined as follows:

[0078]

[0079] Among them, L CE (θ) represents the cross-entropy loss function; θ represents the set of parameters of the model, including the weights and biases of the neural network; N represents the number of samples; Y represents the predicted probability output by the model; Y∈{0,1} represents the true data of sample i.

[0080] The learning rate scheduling employs a cosine annealing learning rate scheduler combined with a warm-up strategy; the learning rate increases gradually in a linear manner, with the following growth formula:

[0081]

[0082] Where, η t η represents the learning rate in the t-th period; min Represents the initial minimum learning rate; η0 represents the set initial learning rate; T warmup This represents the total number of cycles in the warm-up phase. After the warm-up phase is complete, the learning rate decays periodically according to a cosine function. The update formula for the cosine annealing learning rate scheduler is:

[0083]

[0084] Among them, T cosine This represents the total number of cycles in the cosine annealing stage.

[0085] A further technical solution of the present invention is: the expansion factors of the five residual blocks in the custom residual block are set to 1, 2, 4, 8, 16 respectively, and the number of channels increases layer by layer to 64, 128, 256, 256, 512 respectively.

[0086] A further technical solution of the present invention is as follows: the accuracy calculation expression in step four is as follows:

[0087]

[0088] Where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively.

[0089] The F1 score is the harmonic mean of precision and recall, calculated as follows:

[0090]

[0091] Accuracy and F1 score together reflect the ability to distinguish between ionized signals and vibration signals.

[0092] Compared with the prior art, the present invention has the following characteristics:

[0093] The deep learning algorithm of this invention can effectively distinguish between vibration signals and effective ionization signals, reduce the impact of vibration interference on the accuracy of measurement results, and achieve more accurate radon concentration measurement.

[0094] The detailed structure of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0095] Figure 1 This is a flowchart of the method for distinguishing ionization signals and vibration signals according to the present invention;

[0096] Figure 2 The flowchart shows the acquisition process for ionization signal datasets and vibration signal datasets.

[0097] Figure 3 This is a schematic diagram of data processing for a causal dilated convolutional layer.

[0098] Figure 4 A schematic diagram of data processing for a single custom residual block;

[0099] Figure 5 This is a schematic diagram of data processing for an adaptive pooling classifier.

[0100] Figure 6(a) , 6(b) Figures 6 and 6(c) show the curves of the cross-entropy loss function, accuracy, and F1 score as a function of the number of iterations on the training and validation sets, respectively. Detailed Implementation

[0101] Example 1, as Figure 1 As shown in Figure 6, the method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks aims to improve the accuracy of signal recognition and reduce measurement errors. The method includes the following specific steps:

[0102] Step 1: Acquisition of ionization signal data set and vibration signal data set from the pulse ionization chamber

[0103] To ensure the authenticity, accuracy, and diversity of the dataset, signal data from the pulse ionization chamber were collected under real-world measurement conditions to improve the generalization ability of the deep learning model. The dataset collection process is as follows: Figure 2 As shown: First, the sampling circuit captures the output signal of the pulse ionization chamber and converts the analog signal into a digital signal through an analog-to-digital converter (ADC). Then, the sampling microcontroller processes the digital signal, sets a threshold to effectively filter noise, and starts data acquisition when the amplitude of the digital signal exceeds the threshold. When the amplitude of the digital signal drops below the threshold, acquisition stops to reduce interference from invalid data. Finally, the acquired valid digital signal is transmitted to the main microcontroller after the measurement cycle ends. The main microcontroller is responsible for storing the data on the SD card and transmitting it to the host computer via serial port.

[0104] The acquisition of ionization signal datasets must be conducted in a vibration-free environment to ensure signal purity. When acquiring ionization signal datasets, a radon source is placed at the inlet of the pulse ionization chamber, allowing radon gas to diffuse into the chamber. The radon decays within the chamber, releasing alpha particles, which ionize the air and generate a weak ionization signal. The output signal of the pulse ionization chamber is this ionization signal. Because this signal is weak, it needs to be amplified by an amplification circuit before sampling. When acquiring vibration signal datasets, a radon source is not required, as the concentration of radon in the air is extremely low, and its influence on the vibration signal is negligible. To simulate different vibration conditions, the pulse ionization chamber is placed on a table, and different tools, such as metal and plastic, are used to strike it with varying forces to generate vibration signals. The output signal of the pulse ionization chamber under these different conditions represents the vibration signal. To ensure the sufficiency and diversity of the data, this embodiment acquires approximately 60,000 ionization signals and approximately 60,000 vibration signals to construct a complete dataset. The ionization signal dataset and the vibration signal dataset were amplified by analog circuits and converted from analog to digital before being sampled into a one-dimensional time series. Due to differences in the sampling trigger point and the signal duration, the length of the final raw data fluctuated between 150 and 1642 sampling points.

[0105] Step 2: Preprocessing and standardization of ionization signal datasets and vibration signal datasets

[0106] The ionization signal originates from the action of alpha particles released during radon decay in the pulsed ionization chamber. Due to differences in the decay time, emission position, energy, direction, and incident angle of the alpha particles, the resulting ionization trajectories are not uniform. This randomness causes significant differences in the amplitude, duration, and waveform characteristics of the amplified ionization signal, such as variations in rising and falling edges. Furthermore, vibration signals, primarily from external mechanical disturbances, simulate various typical mechanical vibrations depending on the material of the striking tool (e.g., metal, plastic, or other materials), the applied force, and the surface material. Due to the inconsistency of the vibration source, the waveform characteristics of the generated vibration signals also exhibit significant amplitude differences and morphological diversity. Therefore, to ensure the stability and generalization ability of the model training, systematic data preprocessing and standardization are necessary before inputting the original signals into the model.

[0107] Considering the physical characteristics of the signal source, the differences between ionized and vibration signals during the sampling process, and the format requirements of the input tensor in the deep learning model, the specific process of signal preprocessing and standardization in this embodiment is as follows:

[0108] 2.1 Differentiated Length Regularization Preprocessing

[0109] To meet the requirement of fixed input dimensions for deep learning models while preserving the original signal information to the maximum extent, differential length normalization is performed on the original signal, and all input signals are uniformly adjusted to a preset target length L. target If the value is 512, then the following signal curvature method exists.

[0110] 2.1.1 Short signal compensation

[0111] When the length of the original signal is less than L target When the length is 4, conventional interpolation or padding methods are insufficient to effectively preserve the original signal waveform characteristics. To avoid information dilution, a periodic repetitive expansion strategy is adopted, where the signal content is repeatedly spliced ​​in a periodic manner until the target length requirement is met. Any excess is truncated at the end to ensure consistent signal dimensions. The formula is as follows:

[0112]

[0113] Where, x target This represents the signal after length warping; x represents the original signal; tile(·) represents a periodic repetition operation, and [.] represents a floor operation; :L target Indicates taking the first L target This method uses multiple sampling points. It can extend the signal's time range while preserving the original signal's local structural features relatively well, and is particularly suitable for short-period signals with obvious repetitive patterns.

[0114] When the length of the original signal is greater than L target When / 4 is less than 300, although its information density is relatively sufficient, its length is still insufficient to directly meet the requirements of a fixed input dimension. A boundary reflection padding strategy is used to extend the length by mirroring the edge waveforms at both ends of the signal to fill in the missing parts, thus obtaining a normalized signal. The formula is as follows:

[0115] x target =reflect_pad(x,L target (2)

[0116] Here, `reflect_pad()` represents the reflection padding operation. This strategy can preserve the original signal boundary structure while avoiding non-physical disturbances introduced by zero padding or random completion, thus improving the fidelity of edge information. It is suitable for extended processing of most physical measurement signals.

[0117] 2.1.2 Timing regularization of medium-length signals

[0118] For signal sequences of moderate length, Dynamic Time Warping (DTW) is preferentially used for time alignment to achieve nonlinear stretching or compression of the signal shape in the time dimension, thereby maintaining good consistency with the standard reference template. The core operation of this step is as follows:

[0119] x target =DTW(x,t) ref (3)

[0120] Among them, t ref The constructed reference template, with a length of 512, is defined as a normalized time series sampled at equal intervals within the interval [0,1], used to provide a unified time anchor point for aligning multi-source signals. This template represents the distribution structure of an ideal signal in the standard time domain, enabling the elimination of time scale differences caused by varying sampling lengths.

[0121] When the signal contains a large amount of noise, extreme distortion, or a length that differs significantly from the template, DTW matching may fail or the alignment path may become unstable. To improve the overall robustness of the algorithm, linear interpolation is used to perform proportional resampling in such abnormal situations to adjust the original signal to the target length.

[0122] x target =interp(x,L target ,L) (4)

[0123] Here, interp(·) represents the length normalization function based on linear interpolation. This method quickly completes the uniform length normalization operation without introducing nonlinear deformation, and is suitable for signal samples that cannot be stably aligned by DTW.

[0124] 2.1.3 Extraction of the main energy segment of a long signal

[0125] For extremely long signals exceeding 1000, directly scaling them to the target length may lead to compression or even loss of crucial information. Considering that ionization and vibration signals typically exhibit localized energy concentration in the time domain, this example employs a main energy segment truncation strategy based on the sliding energy spectrum to extract the most representative core segments as model input, in order to improve the effectiveness of feature extraction and signal utilization.

[0126] Let the sampled long signal be represented as x = [x1, x2, ..., x...]. L The strategy first calculates the energy spectrum distribution of the entire signal using a sliding window method to identify the key regions where the energy is most concentrated. The sliding energy function E[i] is defined as follows:

[0127]

[0128] Where, x i+k Let h[k] represent the amplitude of the signal at the (i+k)th sampling point, and h[k] represent the Hamming function with a window width w = 200, used to suppress edge effects and reduce spectral leakage. The sliding window slides across the signal with a step size of 1, symmetrically covering 100 sampling points before and after the current point, thus obtaining a smooth energy spectrum estimate.

[0129] The global energy spectrum distribution is obtained by traversing all sliding window positions. Then, the maximum energy criterion is used to locate the time point of the principal energy concentration.

[0130] i peak =argmax i E[i] (6)

[0131] Where, argmax i E[i] represents the index of the position where E[i] reaches its maximum value among all possible values ​​of i; peak This represents the core region with the most concentrated energy in the corresponding signal, which usually contains key dynamic patterns.

[0132] To further focus on this critical area and reduce redundant interference, i peak A local sub-segment of length 256 is extracted from the center as the main energy region. The specific extraction method is as follows:

[0133] x core =x[i peak -128:i peak +128] (7)

[0134] Where, x core This represents the most discriminative local core segment extracted from the original signal x. To meet the model's requirement for a uniform input length, a constant edge padding strategy is further employed, which involves repeating the amplitude of the last sampling point of this local segment until it is expanded to the required length. This operation can regularize the signal dimension while preserving the main dynamic features, avoiding the introduction of spurious features that could negatively impact model performance.

[0135] By employing the aforementioned adjustment strategies, we can ensure that the most representative part of the ultra-long signal is extracted as the input to the model, while also taking into account the preservation of signal features and the uniformity of dimensions, thereby effectively improving the model's ability to model long time-series signals and its recognition accuracy.

[0136] 2.2 Adaptive Amplitude Standardization Based on Median and Interquartile Range

[0137] Raw signals often exhibit asymmetric distributions, baseline drift, and abrupt pulses during acquisition. Traditional mean-variance normalization can be susceptible to extreme value interference, leading to feature distortion. To improve the robustness of the normalization process to outliers, this example introduces an adaptive normalization strategy based on the median and interquartile range. The core idea is to center the signal using the median and then compress the signal scale using the interquartile range, thereby mitigating the impact of outlier fluctuations on the overall distribution. Specifically:

[0138] 2.2.1 Median Centering

[0139] For each length-normalized signal x target First, calculate the median, then shift the signal to a distribution centered at zero to obtain the centered signal, as shown in the following formula:

[0140] x'=x target -median(x target (8)

[0141] Where x′ represents the signal after centering; median(·) represents the median calculation function, used to extract the median value among all values ​​in the input sequence. This is achieved by subtracting the median(x′) from each sample point. target This allows the signal to be adjusted to a distribution centered at 0, which helps in the stable training of the subsequent model and feature extraction.

[0142] 2.2.2 Quartile Normalization

[0143] To enhance robustness, interquartile range (IQR) is further used to scale the amplitude of the centered signal. Specifically, the 25th and 75th percentiles of x′, i.e., the first quartile Q1 and the third quartile Q3, are first calculated, and the interquartile range is obtained accordingly, as shown in the following formula:

[0144]

[0145] Among them, percentile 25 (·) and percentile 75 (·) represents the calculation functions for the 25th and 75th percentiles, respectively, used to measure the upper and lower quartile boundaries of the signal; the interquartile range represents the dispersion of the signal within the middle 50% of the data range, and is a robust scaling estimation method with strong resistance to outliers.

[0146] To further standardize the signal amplitude scale, after median centering, an amplitude normalization strategy based on interquartile range is introduced. If the interquartile range is greater than a set fault tolerance threshold ε, this ε is set to 1 × 10⁻⁶. -6If the interquartile range is less than the tolerance threshold ε, then the centered signal x′ is normalized; otherwise, if the interquartile range is less than the tolerance threshold ε, then only the centered operation is retained to avoid the risk of numerical instability or noise amplification due to excessively small scale. The specific normalization formula is as follows:

[0147]

[0148] Where, x norm This represents the final normalized signal sequence.

[0149] 2.2.3 Dataset Signal Enhancement

[0150] To enhance the model's generalization ability and robustness to various perturbations, this example designs diverse signal enhancement strategies for the training set. By simulating noise and variations that may occur in real-world scenarios, the model's adaptability to complex environments is improved. Specific enhancement strategies include the following transformation methods, which are applied independently and randomly in combination with a probability of 0.5 during training:

[0151] A, Amplitude perturbation: By applying random scaling to the signal amplitude, it simulates fluctuations in sensor sensitivity or nonlinear responses of amplifiers. The specific expression is as follows:

[0152] x aug =a·x train ,a~u(0.7,1.3) (11)

[0153] Where, x aug Indicates the enhanced signal, x train Let be the original signal of the training set, and a be a random scaling factor that follows a uniform distribution in the interval [0.7, 1.3].

[0154] B, Timing Distortion: Nonlinear stretching or compression of the signal time axis to simulate changes in signal rhythm or speed, enhancing the model's adaptability to inputs at different time scales. The specific expression is:

[0155] x aug =interp(x train ,r), r~u(0.8,1.2) (12)

[0156] Where interp(·) represents the interpolation resampling operation, and r is a random time scaling factor used to control the degree of compression or expansion.

[0157] C, Additive Gaussian noise: Adding zero-mean Gaussian white noise to a signal to simulate external noise interference such as background interference, electromagnetic noise, or instrument errors.

[0158] x aug =x train+σ·N(0,1),σ~u(0.01,0.05) (13)

[0159] Where N(0,1) represents a standard Gaussian distribution, and σ represents the noise intensity sampled uniformly from the interval [0.01,0.05].

[0160] D. Random Pruning and Filling: A continuous subsequence of length 256 is randomly pruned from the signal and expanded to a uniform length of 512 using edge replication to simulate scenarios of local signal truncation, missing data, or incomplete measurements. This method not only preserves key local features of the signal but also improves the robustness of the model under conditions of missing information.

[0161] Each of the above enhancement operations A, B, C, and D is applied independently with a probability of 0.5. The enhanced signal retains the original features while introducing reasonable mutations, thereby effectively improving the robustness and generalization ability of the model under various perturbation conditions.

[0162] 2.2.4 Final Standardization Processing

[0163] After preprocessing and augmentation, to eliminate differences between samples due to varying amplitude scales and accelerate model training convergence, standardization is performed on each sample signal. Standardization converts the signal into a form with zero mean and unit variance, allowing signals at different scales to be processed at the same scale. The specific standardization formula is as follows:

[0164]

[0165] Where, x final This represents the signal after final standardization; μ = mean(x) represents the mean of the signal; ρ = std(x) represents the standard deviation of the signal; δ = 1 × 10⁻⁶. -6 This is a tiny constant introduced to prevent division-by-zero errors when the standard deviation is zero. After this operation, each signal sample will be converted into a standard normal distribution with zero mean and unit variance, thereby effectively improving the robustness and adaptability of the model when processing signals of different scales.

[0166] 2.3 Dataset Partitioning

[0167] To ensure the scientific rigor and reliability of the model evaluation, 10,000 ionization signals and 10,000 vibration signals were extracted from the original dataset to form a test set for final performance evaluation. The remaining 100,000 data points were divided into a training set and a validation set, with 80% and 20% respectively, for training the model and evaluating its generalization ability.

[0168] Step 3: Construct a model architecture based on neural networks

[0169] In this example, the neural network-based model architecture includes a single causal dilated convolutional layer, five custom residual blocks containing causal dilated convolutions, residual skip connections, an adaptive pooling classifier, and optimized model parameter updates. This structure has been specifically optimized in terms of parameter settings and module design to achieve efficient classification of ionization and vibration signals output from the pulsed ionization chamber. The following provides a detailed explanation of the model's network structure, the configuration parameters of each layer, and their functions.

[0170] 3.1 Causal Dilated Convolutional Layer

[0171] The causal dilated convolutional layer uses a kernel size of 7 and a dilation rate of 1, which can extract local temporal features while maintaining temporal consistency, such as... Figure 3 As shown, the mathematical expression for the dilated convolution mechanism introduced in this layer is as follows:

[0172]

[0173] Where z(t) represents the convolution output at time step t; x(t) is the input sequence, i.e., x final Here, w(k) represents the k-th weight of the convolution kernel, k is the kernel size, and d is the dilation factor. In this model, the first causal dilated convolutional layer has a small dilation rate d = 1, aiming to accurately capture the local variation features in ionization and vibration signals while strictly adhering to temporal causality constraints. This shallow structure provides a good feature foundation for the fusion of multi-scale features and high-order semantic abstraction in the subsequent residual module. Simultaneously, this structure possesses good parallel computing capabilities, effectively improving the model's efficiency during training and inference phases.

[0174] 3.2 Custom Residual Block

[0175] To further enhance the model's ability to extract multi-scale temporal features and alleviate the gradient vanishing and performance degradation problems that may occur during the training of deep networks, five custom residual modules are introduced after the causal dilated convolutional layers. Each custom residual consists of two causal dilated convolutional layers, followed by batch normalization, the LeakyReLU activation function, and a regularization module. The input is adjusted by a 1×1 convolution and then added to the output to form a residual skip connection. This structure aims to extract spatiotemporal features through multiple convolutions, enhance the model's expressive power by combining batch normalization and activation functions, and suppress overfitting using regularization. Figure 4 As shown.

[0176] 3.2.1 Causal Dilated Convolution

[0177] Each custom residual block contains two layers of one-dimensional causal dilated convolutions with a fixed kernel size of k=5, and its structure is consistent with the first causal dilated convolution layer mentioned above. To achieve layer-by-layer modeling of multi-scale temporal features, the dilation factors in the five residual blocks are set to d=1, 2, 4, 8, and 16 respectively, with the receptive field increasing exponentially with the number of network layers. Simultaneously, the number of channels increases layer by layer to 64, 128, 256, 256, and 512, respectively, enabling the network to have stronger feature extraction capabilities at different temporal granularities, focusing on local details and global dynamic changes, and enhancing its ability to model complex signal patterns and its classification performance.

[0178] 3.2.2 Batch Normalization

[0179] Batch normalization is introduced after each causal dilation convolution layer to alleviate internal covariate bias and improve the stability and convergence speed of model training. The mathematical expression of normalization is as follows:

[0180]

[0181] Where BatchNorm1d(z) represents the output layer result of the normalized layer; z represents the convolution output vector, which is the input feature of this layer; μ represents the mean of the input vector z; φ represents the standard deviation of the input vector z; γ represents the learned scale parameter, which is used to scale the normalized result; β represents the learned bias parameter, which is used to translate the normalized result.

[0182] 3.2.3 LeakyReLU activation function

[0183] Each batch normalization layer is followed by the LeakyReLU activation function, giving the network non-linear expressive capabilities. Compared to the traditional ReLU activation function, its output in the positive region is the same as the input, while in the negative region, the output is the input multiplied by a small positive number (0.2). This avoids the problem of the traditional ReLU output always being zero in the negative region, thus reducing the risk of "neuron death." The specific formula is shown below:

[0184]

[0185] Here, α represents the leakage coefficient of the negative half-axis, with a value of 0.2. This function can maintain gradient flow in the negative region, thereby improving the training performance of deep networks.

[0186] 3.2.4 Dropout Regularization Mechanism

[0187] To suppress model overfitting and enhance robustness to noise and perturbations, a Dropout regularization mechanism is introduced in each residual block. During training, Dropout randomly "masks" the output of a portion of neurons with a probability of 0.35, forcing the model to learn a sparser and more general feature representation.

[0188] 3.3 Residual Jump Connection

[0189] To achieve effective fusion of cross-layer features, a residual connection mechanism is introduced. When the input and output dimensions are the same, an identity mapping is used; when the dimensions are different, a 1×1 convolution is introduced for linear mapping. In this model, there are five residual blocks, four of which incorporate the residual connection mechanism. The output of the residual block is represented as follows:

[0190] y = z² + Skip(x) (18)

[0191] Where y represents the final output of the residual block, z2 represents the output value after two convolutional layers, and Skip(x) represents the residual path function, which is defined as follows:

[0192]

[0193] Among them, Conv 1×1 (x) represents a 1×1 convolution operation performed on the input x; C in and C out These represent the number of channels for the input and output features, respectively. This structure enables the fusion of the output after nonlinear transformation with the original input features, enhancing the model's ability to perceive details and change patterns.

[0194] 3.4 Adaptive Pooling Classifier

[0195] To further compress the time-series dimension to achieve binary classification, such as Figure 5 As shown, the adaptive pooling classifier includes an adaptive average pooling layer, a flattening layer, a first fully connected layer, a LeakyReLU activation function, a regularization module, and a second fully connected layer. Its workflow is as follows: First, the adaptive average pooling layer performs a global average operation on each channel in the time dimension; then, the flattening layer compresses the pooled temporal features; finally, through the combined action of the first fully connected layer, the LeakyReLU activation function, the regularization module, and the second fully connected layer, the temporal dimension of the signal features is reduced, completing the classification task. The classifier output is shown below:

[0196] Y = Classifier(y) (20)

[0197] Where Y represents the binary classification result output after the second fully connected layer; Classifier(.) represents the function of the entire adaptive pooling classifier, covering the adaptive average pooling, flattening and fully connected operations.

[0198] 3.5 Model Parameter Update and Optimization

[0199] To improve the accuracy, stability, and classification time of binary classification, the model parameters are updated and optimized using the cross-entropy loss function and learning rate scheduling.

[0200] The cross-entropy loss function can accurately measure the difference between the model's predicted probability distribution and the actual data distribution, and its definition is as follows:

[0201]

[0202] Among them, L CE (θ) represents the cross-entropy loss function; θ represents the set of parameters of the model, including the weights and biases of the neural network; N represents the number of samples; Let Y represent the predicted probability output by the model; Y∈{0,1} represents the true data of sample i. By minimizing the cross-entropy loss, the model can continuously adjust its parameters to improve the accuracy of class prediction and reduce misclassification.

[0203] To further improve training efficiency and model generalization ability, the optimizer chosen is AdamW, an enhanced version of the traditional Adam optimizer. Unlike the standard Adam, AdamW explicitly decouples the weight decay mechanism from gradient updates, adding a regularization term when updating parameters to effectively suppress excessively large model parameters and prevent overfitting. The update rule of AdamW is expressed as follows:

[0204]

[0205] Where η represents the learning rate; Let θ represent the bias correction terms for the first and second moment estimates, respectively; ∈ represent the numerical stability term; λ represent the weight decay factor; and θ represent the weight decay factor. t This represents the model parameters for the t-th iteration.

[0206] In each actual training iteration, the cross-entropy loss between the model output and the true label is first calculated. Then, the AdamW optimizer is used to update the model parameters based on this loss. The optimizer calculates the gradient through backpropagation and adjusts the parameters according to the update rules described above, so that the model reduces the loss while avoiding excessive parameter inflation. This approach allows the model to balance accuracy and generalization ability during training, ensuring good robustness and performance when dealing with complex ionization and vibration signal discrimination tasks. Through the AdamW optimizer, the model can not only converge quickly but also maintain stable discrimination ability when faced with diverse data.

[0207] The learning rate scheduling employs a cosine annealing learning rate scheduler combined with a warm-up strategy. This combined strategy optimizes the learning rate adjustment process, enabling the model to update parameters more efficiently at different training stages. The learning rate increases linearly, with the following growth formula:

[0208]

[0209] Where, η t η represents the learning rate in the t-th period; min Represents the initial minimum learning rate; η0 represents the set initial learning rate; T warmup This indicates the total number of warm-up periods. A warm-up strategy is used in the first few periods of training to gradually increase the learning rate from a small initial value to a preset initial learning rate. This phase aims to stabilize the model's parameter updates and prevent excessively large learning rates from causing training instability. By gradually increasing the learning rate, the model can smoothly enter a suitable parameter space.

[0210] After the warm-up phase, the learning rate decays periodically according to a cosine function, gradually decreasing from the initial learning rate η0 until it approaches the set minimum learning rate. This method, by gradually decreasing the learning rate, allows the model to finely adjust its parameters in the later stages of training. The update formula for the cosine annealing learning rate scheduler is:

[0211]

[0212] Among them, T cosine This indicates the total number of cycles in the cosine annealing stage.

[0213] By combining the learning rate adjustment process with the preheating and cosine annealing strategies, the model can be trained stably in the early stage, avoiding poor convergence caused by excessively rapid parameter updates. At the same time, the optimization accuracy of the model can be improved in the later stage by slowly decaying the learning rate, avoiding getting trapped in local optima, and ultimately improving the overall training effect and model performance.

[0214] Step 4: Model Evaluation and Discrimination Performance Testing Based on Neural Networks

[0215] In order to comprehensively evaluate its performance in classification tasks during model training and testing, cross-entropy loss function, accuracy and F1 score were selected as model evaluation metrics.

[0216] The accuracy rate, which measures the proportion of correctly predicted samples out of the total number of samples, is the most intuitive performance metric, and is defined as follows:

[0217]

[0218] Where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively.

[0219] The F1 score is the harmonic mean of precision and recall, and is particularly suitable for scenarios with imbalanced class distributions. Its definition is as follows:

[0220]

[0221] The accuracy rate provides a macro-level assessment of the overall prediction performance, while the F1 score focuses more on the model's balanced performance between the positive and negative classes. Combining the two provides a more comprehensive reflection of the model's practical application capability in binary classification tasks involving ionized and vibration signals.

[0222] The constructed neural network-based model was trained iteratively for 600 rounds on the training set, and its classification performance was evaluated on both the training and validation sets. The specific results of each evaluation metric are shown in Table 1.

[0223] Table 1 Performance metrics of the hybrid model after 600 iterations

[0224]

[0225] As shown in the table, after 600 rounds of training, the model achieved a cross-entropy loss of 0.0823 on the training set, with an accuracy of 0.9682 and an F1 score of 0.9687, indicating that the model has excellent classification performance on the training set and can effectively distinguish between ionization and vibration signals. Meanwhile, on the validation set, the cross-entropy loss further decreased to 0.0387, and both the accuracy and F1 score reached 0.9915, demonstrating the model's extremely strong generalization ability on unseen data.

[0226] Meanwhile, the curves showing the changes in the model's cross-entropy loss function, accuracy, and F1 score with the number of iterations on the training and validation sets during training are as follows: Figure 6(a) , 6(b)As shown in Figure 6(c), the model exhibits good fitting performance on both the training and validation sets, with a consistent overall convergence trend and no obvious overfitting. During training, the loss function value shows a typical "fast at first, slow later" decreasing trend: in the early stages of training, the loss value decreases rapidly, indicating that the model can quickly capture the main discriminative features in the data; as training progresses, the loss decreases more gradually, entering a fine-tuning phase where the model begins to learn more complex and subtle features, thereby further improving overall performance. Furthermore, the model's accuracy and F1 score on the validation set remain highly consistent, indicating that the model not only performs well in overall classification accuracy but also demonstrates good robustness and balance when handling class imbalance or complex signal features. The stability of the F1 score further confirms that the model has achieved a good trade-off between precision and recall, possessing high generalization ability and practical application potential.

[0227] To further verify the classification performance of the model, a secondary test was designed on a large-scale test set completely independent of the training and validation sets. This test set includes 10,000 ionization signals and 10,000 vibration signals. Since these test samples have never appeared during the model training and validation phases, they can truly reflect the model's generalization ability. The test results are shown in Table 2.

[0228] Table 2. Test results of neural network-based models

[0229]

[0230]

[0231] As shown in Table 2, the classification accuracy and F1 score for vibration and ionization signals on this test set both exceeded 0.99. This indicates that the model not only possesses extremely high classification accuracy but also demonstrates strong stability and robustness when handling complex patterns such as ionization and vibration signals. Therefore, this neural network-based model can maintain excellent recognition capabilities even when faced with unseen data in real-world scenarios, demonstrating promising deployment prospects and application value.

Claims

1. A method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks, characterized by: The specific steps include the following: Step 1: Acquisition of ionization signal dataset and vibration signal dataset from the pulse ionization chamber; Step 2: Preprocessing and standardization of ionization signal dataset and vibration signal dataset: The signal preprocessing and standardization include differential length normalization preprocessing, adaptive amplitude standardization based on median and interquartile range, dataset partitioning, and final standardization. Step 3: Construct a neural network-based model architecture, including causal dilated convolutional layers, five custom residual blocks, residual skip connections, an adaptive pooling classifier, and optimization of model parameters. Step 4: Model evaluation and discrimination test based on neural networks: Cross-entropy loss function, accuracy and F1 score are selected as model evaluation indicators.

2. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks as described in claim 1, characterized in that: The data acquisition process includes: first, the sampling circuit captures the output signal of the pulse ionization chamber and converts the analog signal into a digital signal through an analog-to-digital converter (ADC); then, the sampling microcontroller processes the digital signal and sets a threshold to filter noise; finally, the acquired digital signal is transmitted to the main microcontroller after the measurement cycle ends, the main microcontroller stores the data in an SD card, and transmits it to the host computer via a serial port.

3. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks as described in claim 2, characterized in that: The specific operation of the noise filtering is to start data acquisition when the amplitude of the detected digital signal exceeds the threshold, and to stop acquisition when the amplitude of the digital signal drops below the threshold.

4. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks as described in claim 1, characterized in that: In step one, when collecting the ionization signal dataset, the radon source is placed at the cavity inlet of the pulse ionization chamber, allowing the radon gas to diffuse into the cavity. The radon gas decays in the cavity, releasing alpha particles, which trigger air ionization and generate a weak ionization signal. When collecting the vibration signal dataset, vibration signals are generated by striking the object with different forces.

5. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks as described in claim 1, characterized in that: The specific steps of the differential length normalization preprocessing in step two are as follows: All input signals are uniformly adjusted to a preset target length L. target If the value is 512, then the following signal conditioning methods are included: When the length of the original signal is less than L target When the value is 4, the original signal is compensated, and the formula is expressed as follows: Where, x target This represents the signal after length warping; x represents the original signal; tile(·) represents a periodic repetition operation, and [.] represents a floor operation; :L target Indicates taking the first L target One sampling point; When the length of the original signal is greater than L target When / 4 is less than 300, the preprocessing formula is expressed as: x target =reflect_pad(x,L target ) (2) Where reflect_pad(,) represents the reflection padding operation; For signal sequences of medium length, the Dynamic Time Warping (DTW) algorithm is first used for time alignment. The core operation of time alignment is as follows: x target =DTW(x,t ref ) (3) Among them, t ref The constructed reference template, with a length of 512, is defined as a normalized time series sampled at equal intervals within the interval [0,1], used to provide a unified time anchor point for aligning multi-source signals; Then, when the signal contains a lot of noise, extreme distortion, or the length differs too much from the template, causing DTW matching to fail or the alignment path to be unstable, linear interpolation is used for proportional resampling to adjust the original signal to the target length. The formula is: x target =interp(x,L target (4) Where interp(·) denotes the length normalization function based on linear interpolation; For ultra-long signals exceeding 1000, a main energy segment truncation strategy based on the sliding energy spectrum is adopted. Let the sampled long signal be represented as x = [x1, x2, ..., x]. L The sliding energy function E[i] is defined as follows: Where, x i+k Let h[k] represent the amplitude of the signal at the (i+k)th sampling point, and h[k] be the Hamming function with a window width w = 200, used to suppress edge effects and reduce spectral leakage. By traversing all sliding window positions, the global energy spectrum distribution is obtained, and then the time point of the main energy concentration is located using the maximum energy criterion. i peak =argmax i E[i] (6) Where, arg max i E[i] represents the index of the position where E[i] reaches its maximum value among all possible values ​​of i; peak This represents the core region with the most concentrated energy in the corresponding signal, which usually contains key dynamic patterns; To further focus on this critical area and reduce redundant interference, i peak A local sub-segment of length 256 is extracted from the center as the main energy region. The specific extraction method is as follows: x core =x[i peak -128:i peak +128] (7) Where, x core This represents the most discriminative local core segment extracted from the original signal x; to meet the model's requirement for a uniform input length, the amplitude of the last sampling point of this local segment is repeated until it is extended to the required length; The specific steps of the adaptive amplitude normalization based on the median and interquartile range are as follows: for each length-normalized signal x... target First, calculate the median, then shift the signal to a distribution centered at zero to obtain the centered signal, as shown in the following formula: x'=x target -median(x target ) (8) Where x′ represents the signal after centering; median(·) represents the median calculation function, used to extract the median value of all values ​​in the input sequence; To enhance robustness, interquartile range (IQR) is further used to scale the amplitude of the centered signal. Specifically, the 25th and 75th percentiles of x′, i.e., the first quartile Q1 and the third quartile Q3, are first calculated, and the interquartile range is obtained accordingly, as shown in the following formula: Among them, percentile 25 (·) and percentile 75 (·) represent the calculation functions for the 25th and 75th percentiles, respectively, used to measure the upper and lower quartile boundaries of a signal; If the interquartile range is greater than the set fault tolerance threshold ε, then set ε to 1×10. -6 If the interquartile range is less than the fault tolerance threshold ε, then the centering operation is performed; otherwise, only the centering operation is retained. The specific normalization formula is as follows: Where, x norm This is the final normalized signal sequence; The normalized signal sequence is augmented using a dataset signal enhancement method. The enhancement is achieved by randomly combining the following four strategies with independent probabilities of 0.5: A, Amplitude perturbation: Applying random scaling to the signal amplitude, the specific expression is: x aug =a·x train ,a~u(0.7,1.3) (11) Where, x aug Indicates the enhanced signal, x train Let represent the original signal of the training set, and 'a' represent a random scaling factor that follows a uniform distribution in the interval [0.7, 1.3]. B, Timing Distortion: Nonlinear stretching or compression of the signal time axis, specifically expressed as follows: x aug =interp(x train ,r),r~u(0.8,1.2) (12) Where interp(·) represents the interpolation resampling operation, and r represents the random time scaling factor, which is used to control the degree of compression or expansion; C, Additive Gaussian noise: Adding zero-mean Gaussian white noise to a signal, the specific expression is: x aug =x train +σ·N(0,1),σ~u(0.01,0.05) (13) Where N(0,1) represents a standard Gaussian distribution, and σ represents the noise intensity sampled uniformly from the interval [0.01,0.05]. D, Random cropping and padding: Randomly crop a continuous subsequence of length 256 from the signal and expand it to a uniform length of 512 using edge copying; The enhanced signal is then subjected to final standardization processing. The specific formula for standardization is as follows: Where, x final This represents the signal after final standardization; μ = mean(x) represents the mean of the signal; ρ = std(x) represents the standard deviation of the signal; δ = 1 × 10⁻⁶. -6 This represents a small constant introduced to prevent division by zero errors when the standard deviation is zero; Specifically, the dataset is divided into a test set, a training set, and a validation set.

6. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks as described in claim 1, characterized in that: In step three, the causal dilated convolutional layer uses a kernel size of 7 and a dilation rate of 1. The mathematical expression for the dilated convolution mechanism introduced in this layer is as follows: Where z(t) represents the convolution output at time step t; x(t) represents the input sequence, i.e., x final w(k) represents the k-th weight of the convolution kernel, k is the kernel size, and d is the inflation factor; The custom residual consists of two layers of causal dilation convolutions. Each layer is followed by batch normalization, LeakyReLU activation function and regularization module. The input is adjusted by 1×1 convolution and then added to the output to form a residual jump connection. The normalized mathematical expression is as follows: Where BatchNorm1d(z) represents the output layer result of the layer normalization; z represents the convolution output vector, which is the input feature of this layer; μ represents the mean of the input vector z; γ represents the standard deviation of the input vector z; γ represents the learned scaling parameter, used to scale the normalized result; β represents the learned bias parameter, used to shift the normalized result. The specific formula for the LeakyReLU activation function is as follows: Where α represents the leakage coefficient of the negative half-shaft; The residual jump connection is processed as follows: when the input and output dimensions are the same, an identity mapping is used; when the dimensions are different, a 1×1 convolution is introduced for linear mapping, and the residual block output is represented as: y = z² + Skip(x) (18) Where y represents the final output of the residual block, z2 represents the output value after two convolutional layers, and Skip(x) represents the residual path function, which is defined as follows: Among them, Conv 1×1 (x) represents a 1×1 convolution operation performed on the input x; C in and C out These represent the number of channels for the input and output features, respectively. The adaptive pooling classifier includes an adaptive average pooling layer, a flattening layer, a first fully connected layer, a LeakyReLU activation function, a regularization module, and a second fully connected layer. The classifier output is represented as follows: Y = Classifier(y) (20) Where Y represents the binary classification result output after the second fully connected layer; Classifier(.) represents the classifier function that includes average pooling, flattening, and a fully connected layer structure; The model parameters are updated and optimized using the cross-entropy loss function and learning rate scheduling; the cross-entropy loss function is defined as follows: Among them, L CE (θ) represents the cross-entropy loss function; θ represents the set of parameters of the model, including the weights and biases of the neural network; N represents the number of samples; Y represents the predicted probability output by the model; Y∈{0,1} represents the true data of sample i. The learning rate scheduling employs a cosine annealing learning rate scheduler combined with a warm-up strategy; the learning rate increases gradually in a linear manner, with the following growth formula: Where, η t η represents the learning rate in the t-th period; min Represents the initial minimum learning rate; η0 represents the set initial learning rate; T warmup This represents the total number of cycles in the warm-up phase. After the warm-up phase is complete, the learning rate decays periodically according to a cosine function. The update formula for the cosine annealing learning rate scheduler is: Among them, T cosine This represents the total number of cycles in the cosine annealing stage.

7. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on a neural network as described in claim 6, characterized in that: The expansion factors of the five residual blocks in the custom residual block are set to 1, 2, 4, 8, and 16 respectively, and the number of channels increases layer by layer to 64, 128, 256, 256, and 512 respectively.

8. The method for distinguishing ionization signals and vibration signals in a pulse ionization chamber based on neural networks as described in claim 1, characterized in that: The accuracy calculation expression in step four is as follows: Wherein, TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively; The F1 score is the harmonic mean of precision and recall, calculated as follows: Accuracy and F1 score together reflect the ability to distinguish between ionized signals and vibration signals.

Images