Energy spectrum noise reduction method and system of direct radiation detector
By processing the radiation signal of the direct radiation detector through an end-to-end convolutional neural network, the problem of reduced energy spectrum resolution caused by noise and interference was solved, achieving high-precision energy spectrum measurement and improved anti-interference capability.
Patent Information
- Application Number
- CN202510887431.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-31
AI Technical Summary
Existing direct radiation detectors are susceptible to electronic noise, pulse accumulation, Compton scattering, charge sharing, and defects during operation, which significantly reduces the signal-to-noise ratio and measurement accuracy of the energy spectrum detection results.
An end-to-end convolutional neural network is used to process the radiation signal, including high-pass filtering, threshold screening, transient current calculation and smoothing noise reduction. Combined with a non-passing waveform discrimination mechanism, time jitter is eliminated by peak alignment, and the noise-reduced energy spectrum data is reconstructed.
It significantly improves the measurement accuracy and anti-interference capability of the particle energy spectrum of the radiation field, achieves a waveform classification accuracy of 97.2%, improves energy resolution, and maintains network topology simplicity in millisecond-level real-time signal processing.
Smart Images
Figure CN120871218A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radiometric measurement technology, and in particular to a method and system for reducing the energy spectrum noise of a direct radiation detector. Background Technology
[0002] Direct radiation detectors (also known as spectral detectors, photon counting detectors, or multi-energy X-ray detectors) are advanced sensor devices that directly convert radiation energy into electrical signals based on semiconductor sensitive materials (such as cadmium telluride, zinc cadmium telluride, silicon-based materials, etc.). Their core feature is that by combining an array-type pixel structure design with a high-speed ASIC readout circuit, they can simultaneously acquire photon energy deposition information over a wide energy spectrum range during a single scan and generate digital pulse signals containing energy level distribution characteristics. This energy resolution capability allows them to be widely applied in multiple technical fields such as low-dose photon counting computed tomography (PCCT), positron emission tomography (PET-CT), K-boundary feature imaging, radioactive isotope identification, and material composition analysis, demonstrating unique advantages in medical imaging diagnosis, industrial non-destructive testing, and radioactive monitoring scenarios.
[0003] Despite their theoretically high spectral sensitivity, current direct radiation detectors still face multiple physical interference challenges in practical applications. First, the background noise of the detector's front-end electronics system (including thermal noise, dark current noise, and readout circuit noise) can superimpose with weak X-ray photon signals, leading to a significant degradation in the low-energy spectral resolution. Second, under high count rate operation, the time interval between adjacent photon events may be shorter than the detector's response time, causing pulse stacking and resulting in nonlinear distortion of the energy deposition signal. Furthermore, secondary electrons generated by Compton scattering events can induce cross-pixel charge diffusion, and the charge sharing phenomenon between micrometer-level pixels can cause single-photon events to be misjudged as low-energy photon events by multiple pixels. Finally, defects such as surface, grain boundaries, and twins introduced during the detector's manufacturing process act as charge traps, capturing charges or accelerating electron-hole recombination, interfering with the detector's collection of photon signals and causing signal amplitude attenuation. The coupling effect of these interference factors causes peak position shifts, spectral line broadening, and characteristic peak distortion in the energy spectrum curve output by the detector, which severely restricts its energy resolution and material identification accuracy, especially under low radiation dose conditions.
[0004] Therefore, how to reduce the impact of noise and improve the accuracy of the detection results obtained by the detector during the energy spectrum detection process remains the core bottleneck problem restricting the large-scale application of this technology in clinical and industrial scenarios.
[0005] Therefore, there is a need in the art for an energy spectrum noise reduction method and system for direct radiation detectors to solve the above problems. Summary of the Invention
[0006] To address the aforementioned technical problems, namely the susceptibility of existing direct radiation detectors to multiple interference factors such as electronic noise, pulse accumulation, Compton scattering, charge sharing, and defect effects during operation, which leads to a significant reduction in the signal-to-noise ratio and measurement accuracy of their energy spectrum detection results.
[0007] In one aspect, the present invention provides an energy spectrum noise reduction method for a direct radiation detector, the method comprising:
[0008] S1: Collect radiation signals and convert incident photon energy into electron-hole pairs, generate induced current pulses under a bias electric field, amplify the induced current pulses into voltage pulse signals, and perform digital processing;
[0009] S2: Perform high-pass filtering and threshold screening on the digitized voltage pulse signal, retain the effective signal with an amplitude higher than the set threshold, and eliminate time jitter by peak alignment;
[0010] S3: Calculate the transient current based on the aligned voltage pulse signal, and perform smoothing, noise reduction, and normalization processing on the transient current waveform;
[0011] S4: Input the normalized transient current data into an end-to-end convolutional neural network for classification and determine whether it is a complete transit event;
[0012] S5: Based on the classification results, exclude non-crossing events and reconstruct the denoised energy spectrum data.
[0013] In some preferred embodiments, in step S1, a preamplifier is used to convert the charge signal output by the direct radiation detector into a voltage pulse signal, as shown in the formula:
[0014]
[0015] Among them, u out This is the output voltage signal of the preamplifier, K is the open-loop gain of the preamplifier op-amp, Q is the input charge, and C... f This is the feedback capacitor for the preamplifier.
[0016] In some preferred embodiments, in step S3, the formula for calculating the transient current is:
[0017]
[0018] Where i is the calculated transient current and t is time.
[0019] In some preferred embodiments, in step S3, the formula used to smooth and reduce noise in the transient current waveform is:
[0020]
[0021] Where x is the original transient current data sequence, y n The output data sequence is defined by M, which is the dimensionless length of the sliding window, n, which is the sequence number of the current data point, and k, which is the summation index used to traverse all samples within the window.
[0022] In some preferred embodiments, the end-to-end convolutional neural network includes an input layer, a first set of convolutional pooling layers, a second set of convolutional pooling layers, a flattening layer, and an output layer arranged sequentially, wherein the output layer is a fully connected layer.
[0023] In some preferred embodiments, the activation functions of the first group of convolutional pooling layers and the second group of convolutional pooling layers are both ReLU functions;
[0024] The formula for the ReLU function of the first group of convolutional pooling layers is:
[0025]
[0026] Where x1 is the input value of the first group of convolutional pooling layers, and f(x1) is the output value of the first group of convolutional pooling layers;
[0027] The formula for the ReLU function of the second group of convolutional pooling layers is:
[0028]
[0029] Where x2 is the input value of the second convolutional pooling layer, and f(x2) is the output value of the second convolutional pooling layer.
[0030] In some preferred embodiments, the activation function of the output layer is the Sigmoid activation function, with the following formula:
[0031]
[0032] Where x3 is the input value of the output layer, and f(x3) is the output value of the output layer.
[0033] In some preferred embodiments, the output layer outputs two types of probability values through a sigmoid activation function:
[0034] The output vector is [p1, p2], where p1 is the complete transit probability and p2 is the incomplete transit probability.
[0035] In step S4, the rule for determining whether it is a complete crossing event is as follows:
[0036] If p1 > p2, it is determined to be a complete crossing event;
[0037] If p1 < p2, it is determined to be an incomplete crossing event.
[0038] In another aspect, the present invention also provides an energy spectrum noise reduction system for a direct radiation detector, the system comprising:
[0039] A radiation signal acquisition module is used to acquire radiation signals and convert incident photon energy into electron-hole pairs, generate induced current pulses under a bias electric field, amplify the induced current pulses into voltage pulse signals, and perform digital processing.
[0040] The pulse signal processing module is used to perform high-pass filtering and threshold screening on the digitized voltage pulse signal, retain the effective signal with an amplitude higher than the set threshold, and eliminate time jitter by peak alignment;
[0041] A current shaping module is used to calculate the transient current based on the aligned voltage pulse signal, and to smooth, reduce noise, and normalize the transient current waveform.
[0042] An end-to-end convolutional neural network module is used to receive normalized transient current data and classify it.
[0043] The energy spectrum data reconstruction module is used to determine whether a complete transit event is based on the output of the end-to-end convolutional neural network module, exclude non-transit events based on the classification results, and reconstruct the noise-reduced energy spectrum data.
[0044] In some preferred embodiments, the radiation signal acquisition module includes a direct radiation detector, a preamplifier, and an analog-to-digital converter;
[0045] The pulse signal processing module includes RC-CR 2 Filters, threshold comparators, and software peak alignment modules;
[0046] The current shaping module includes a transient current calculation module and a moving average filtering module.
[0047] The energy spectrum noise reduction method and system for direct radiation detectors of the present invention have the following beneficial effects:
[0048] This invention achieves temporal feature analysis and waveform classification of transient current signals through an end-to-end convolutional network architecture. Combined with a non-passing waveform discrimination mechanism, it suppresses energy spectrum noise. Compared with existing technologies, it significantly improves the measurement accuracy and anti-interference capability of the radiation field particle energy spectrum. The technical solution of this invention adopts a lightweight processing architecture to achieve millisecond-level real-time signal processing. While ensuring the simplicity of the network topology, it achieves a waveform classification accuracy of 97.2%. The network processes a single radiation event in about 135 μs. It improves detection efficiency by relying on the end-to-end feature extraction mechanism and introduces a non-passing waveform exclusion algorithm to improve the energy resolution of the detector. Attached Figure Description
[0049] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:
[0050] Figure 1 This is a flowchart of the energy spectrum noise reduction method for the direct radiation detector of the present invention;
[0051] Figure 2 This is an architectural diagram of the energy spectrum noise reduction system for the direct radiation detector of the present invention;
[0052] Figure 3 This is a network architecture diagram of an end-to-end convolutional neural network according to an embodiment of the present invention;
[0053] Figure 4 This is a schematic diagram of the energy spectrum noise reduction results according to an embodiment of the present invention;
[0054] Figure 5 This is a test graph showing the robustness of the end-to-end convolutional neural network model according to an embodiment of the present invention. Detailed Implementation
[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0056] Based on the background art, existing direct radiation detectors are susceptible to multiple interference factors such as electronic noise, pulse accumulation, Compton scattering, and charge sharing during operation, which leads to a significant reduction in the signal-to-noise ratio and measurement accuracy of their energy spectrum detection results. This invention provides an energy spectrum noise reduction method and system for direct radiation detectors, aiming to improve detection efficiency by relying on an end-to-end feature extraction mechanism and to significantly improve the energy resolution of the detector by introducing a non-passing waveform exclusion algorithm.
[0057] like Figure 1 As shown, the energy spectrum noise reduction method for the direct radiation detector of the present invention includes:
[0058] S1: Collects radiation signals and converts incident photon energy into electron-hole pairs, generates induced current pulses under a bias electric field, amplifies the induced current pulses into voltage pulse signals, and performs digital processing.
[0059] Specifically, when a direct radiation detector is irradiated by radiation photons, it generates electron-hole pairs proportional to the energy deposited by the radiation photons, which are collected under a bias electric field to generate induced current pulses.
[0060] Preferably, in step S1, a preamplifier is used to convert the charge signal output by the direct radiation detector into a voltage pulse signal, as shown in the formula:
[0061]
[0062] Among them, u out This is the output voltage signal of the preamplifier, K is the open-loop gain of the preamplifier op-amp, Q is the input charge, and C... f This is the feedback capacitor for the preamplifier;
[0063] Preferably, an analog-to-digital converter (ADC) is used to convert the continuously transmitted analog signal into a digital signal, which facilitates the processing and analysis of the analog signal by the digital system.
[0064] S2: Perform high-pass filtering and threshold screening on the digitized voltage pulse signal, retain the effective signal with an amplitude higher than the set threshold, and eliminate time jitter by peak alignment;
[0065] Preferably, RC-CR is used. 2 The filter filters the voltage pulse signal. The high-pass filter circuit performs two differentiating stages to subtract the effects of baseline and low-frequency fluctuations. After discrimination by a threshold comparator, the trigger time is obtained. The bipolar output signal ensures that the zero-crossing trigger signal is independent of the pulse amplitude. The threshold comparator compares the RC-CR... 2 If the output amplitude of the filter is higher than the set threshold, the voltage signal will be retained; if it is lower than or equal to the threshold, the signal will be discarded. The software peak alignment module will align the valid data according to the trigger time and save the voltage pulse data.
[0066] S3: Calculate the transient current based on the aligned voltage pulse signal, and perform smoothing, noise reduction, and normalization on the transient current waveform;
[0067] Preferably, the transient current is obtained by differentiating the voltage pulse data and multiplying it by a constant related to the preamplifier using a transient current calculation module. The formula for calculating the transient current is as follows:
[0068]
[0069] Where i is the calculated transient current and t is time;
[0070] Preferably, a rectangular window averaging filter is used with a moving average module. The sum of all data points within the window is divided by the window length to obtain the first output value. The window is then moved one data point to the right, discarding the leftmost data point and incorporating the new data point. This process is repeated until all data points have been traversed. The sliding step size is set to 600, and the window width is 10. The transient current waveform is smoothed and denoised using the following formula:
[0071]
[0072] Where x is the original transient current data sequence, y n To output the data sequence, M is the dimensionless length of the sliding window, n is the sequence number of the current data point, and k is the summation index, used to traverse all samples within the window;
[0073] Preferably, the moving average module is processed using the following normalization formula:
[0074]
[0075] Among them, y n,min It is y n The minimum value of y n,max It is y n The maximum value is y. n ', scales the data to the range [0,1];
[0076] The value of the data point refers to the value of the vertical axis in the transient current waveform, i.e., the current amplitude.
[0077] S4: Input the normalized transient current data into an end-to-end convolutional neural network for classification and determine whether it is a complete transit event;
[0078] Preferably, the end-to-end convolutional neural network includes an input layer, a first set of convolutional pooling layers, a second set of convolutional pooling layers, a flattening layer, and an output layer arranged sequentially, wherein the output layer is a fully connected layer.
[0079] The input layer receives the output of a one-dimensional normalized moving average module with a length of 300. The convolutional pooling layer consists of two sets of convolutional-pooling layers. Both sets of convolutional layers use convolutional kernels of length 3 for feature extraction. The first set of convolutional layers has 1 input channel, 32 output channels, and a stride of 1, with padding at both ends to handle the boundary parts of the input data. The second set of convolutional layers has 32 input channels, 64 output channels, and a stride of 1, with padding at both ends. The first and second sets of convolutional layers are followed by max pooling layers with a kernel size of 2 and a stride of 2, without additional padding. After the second set of max pooling layers, the output is flattened into a one-dimensional vector and connected to a fully connected layer with 128 nodes. A Dropout process is introduced to randomly drop some neurons during training. The activation function of the convolutional pooling layers is the ReLU function, and the activation function of both the first and second sets of convolutional pooling layers is the ReLU function.
[0080] The formula for the ReLU function of the first group of convolutional pooling layers is:
[0081]
[0082] Where x1 is the input value of the first group of convolutional pooling layers, and f(x1) is the output value of the first group of convolutional pooling layers;
[0083] The formula for the ReLU function of the second group of convolutional pooling layers is:
[0084]
[0085] Where x2 is the input value of the second convolutional pooling layer, and f(x2) is the output value of the second convolutional pooling layer.
[0086] The specific end-to-end convolutional neural network architecture is shown in the table below:
[0087]
[0088] The output layer is a fully connected layer with 2 nodes. It receives the output from the convolutional pooling layer, passes it through a sigmoid activation function, and outputs a vector of length 2, representing the probability of the transient current present or absent from the complete transit process. The activation function of the output layer is the sigmoid activation function, and the formula is:
[0089]
[0090] Where x3 is the input value of the output layer, and f(x3) is the output value of the output layer;
[0091] The output layer outputs two types of probability values through the Sigmoid activation function:
[0092] The output vector is [p1, p2], where p1 is the complete transit probability and p2 is the incomplete transit probability.
[0093] In step S4, the rule for determining whether it is a complete crossing event is as follows:
[0094] If p1 > p2, it is determined to be a complete crossing event;
[0095] If p1 < p2, it is determined to be an incomplete crossing event.
[0096] S5: Based on the classification results, exclude non-crossing events and reconstruct the denoised energy spectrum data.
[0097] It should be noted that in practical applications, there is only a very low probability that p1 = p2, a probability so low that it can be ignored. If p1 = p2 occurs, the uncertain event can be recorded separately and not included in the energy spectrum reconstruction to avoid introducing noise.
[0098] In the above, the convolutional and pooling layers of the backbone network are used for feature extraction and dimensionality reduction to prevent overfitting, respectively; the fully connected layer network is used to flatten the feature maps obtained by the backbone network, so as to aggregate global information and map it to the output space.
[0099] In some embodiments of the present invention, the noise sources of energy spectrum detection may come from internal noise, external noise, and counts caused by low charge collection efficiency; internal noise includes thermal noise, shot noise, leakage current noise, etc.; the thermal noise is caused by the thermal motion of charge carriers inside the detector; shot noise is caused by the random generation and recombination process of charge carriers and belongs to statistical fluctuations; leakage current noise is the leakage current fluctuation inside the detector; external noise includes electromagnetic interference and power supply noise, etc.; electromagnetic interference comes from radiation interference from surrounding electronic devices, power lines, or wireless signals; power supply noise comes from power supply ripple, voltage fluctuations, etc.; counts caused by low charge collection efficiency refer to the erroneous counts (low energy tails) caused by charge loss during the charge collection process of the direct radiation detector.
[0100] In some embodiments of the present invention, the end-to-end convolutional neural network module is obtained in the following manner:
[0101] Using a CsPbBr3 detector at 1000-2000 Vcm -1The 5.5 Me alpha particle characteristic spectral lines of a 241Am natural radioactive source were resolved under a bias electric field. The distance between the radioactive source and the detector was kept as close as possible. Transient current data were collected and stored. The sampling frequency of the ADC was 100 MHz, the data bit depth was 14 bits, and the model was CAEN DT5781. Radiation events near the full-energy peak were selected and labeled to indicate whether the transient current had a complete transit process as training data. The number of samples was 2000, and the ratio of complete transit events to non-transit events was 1:1. The training data was input into a convolutional neural network with the cross-entropy loss function, the Adam optimizer, the batch size per round was 64, the learning rate was 0.0001, and the number of training rounds was 80.
[0102] Optionally, in one embodiment of the present invention, an early stopping mechanism is introduced, in which training is stopped if the performance of the validation set does not improve within a set number of iterations.
[0103] In addition, refer to Figure 2 The present invention also provides an energy spectrum noise reduction system for a direct radiation detector, the system comprising:
[0104] The radiation signal acquisition module is used to acquire radiation signals and convert incident photon energy into electron-hole pairs, generate induced current pulses under a bias electric field, amplify the induced current pulses into voltage pulse signals, and perform digital processing.
[0105] The pulse signal processing module is used to perform high-pass filtering and threshold screening on the digitized voltage pulse signal, retain the effective signal with an amplitude higher than the set threshold, and eliminate time jitter by peak alignment;
[0106] The current shaping module is used to calculate the transient current based on the aligned voltage pulse signal and to smooth, reduce noise, and normalize the transient current waveform.
[0107] An end-to-end convolutional neural network module is used to receive normalized transient current data and classify it.
[0108] The energy spectrum data reconstruction module is used to determine whether a complete transit event is based on the output of the end-to-end convolutional neural network module, and to exclude non-transit events based on the classification results, thereby reconstructing the denoised energy spectrum data.
[0109] Preferably, the radiation signal acquisition module includes a direct radiation detector, a preamplifier, and an analog-to-digital converter; the pulse signal processing module includes an RC-CR. 2 The system includes filters, threshold comparators, and software peak alignment modules; the current shaping module includes a transient current calculation module and a moving average filtering module.
[0110] The radiation signal acquisition module acquires the charge signal of the detector, converts it into a voltage pulse output through a charge-sensitive preamplifier, and is sampled and stored by the ADC. In some embodiments of the present invention, the charge-sensitive preamplifier has a sensitivity of 400mV / MeV(Si), a decay time constant of 27μs, a rise time of less than 110ns, and the ADC has a sampling rate of 100MHz and a bit depth of 14 bits.
[0111] Furthermore, in the pulse signal processing module, the traditional nuclear detector readout electronics will be processed via RC-CR 2 After the filter and threshold comparator screen the valid signal, the signal from the preamplifier is further shaped and input into the digital multichannel analyzer (MCA). It has a certain noise reduction performance for small-amplitude input noise. However, the inventors found that the collection process of the direct radiation detector has internal defects and recombination centers can delay or recombine charge carriers. Therefore, considering that the transient current waveform reflects the charge carrier collection process of the direct radiation detector and that semi-quantitative transient current calculation is convenient, the transient current waveform is used to eliminate low charge collection efficiency events.
[0112] Furthermore, the inventors discovered that, before calculating transient currents, due to the randomness of radiation events and RC-CR... 2 The position of the zero-crossing point of the filter may have a certain deviation. Direct application may lead to misjudgment and the generation of spurious peaks. In order to achieve better noise reduction effect, it is necessary to align the peak position of the transient current waveform. In the preferred embodiment of the present invention, the peak position of the transient current waveform is aligned at 200ns.
[0113] As an example, labeled transient current data is input into the network for training. The network identifies the transient current waveform and determines whether the transient current waveform contains a complete transit process. The judgment result is compared with the input label value. Incorrect network outputs will have higher losses. The network is trained by minimizing the loss function.
[0114] As an example, the network functionality of an end-to-end convolutional neural network is as follows: Figure 3 As shown, labeled transient current data is input into the network for training. The network identifies the transient current waveform and determines whether the transient current waveform contains a complete transit process. The judgment result is compared with the input label value. Incorrect network outputs will have higher losses. The network is trained by minimizing the loss function.
[0115] As an example, Figure 4 The results of the optimization of the 5.5 MeV α particle energy spectrum of the CsPbBr3 direct radiation detector under a bias electric field of 2000 V cm⁻¹ are presented. The energy spectrum data contains 137,017 radiation events. The end-to-end convolutional neural network took a total of 18.50 s to process the above data, with an average time of 135 μs per event.
[0116] As an example, Figure 5 Robustness tests of the end-to-end convolutional neural network model are presented. Gaussian white noise is added to the transient current before inputting it into the model, with noise standard deviations of 5%, 10%, 20%, and 50% of the peak value of the transient current. The network output results are shown in the table below. In scenarios with a noise standard deviation ≤ 20%, the model maintains stable resolution of the feature spectral lines. When the noise intensity reaches 50%, the model performance deteriorates, but it still possesses noise reduction capabilities. The model exhibits strong robustness and is suitable for practical scenarios with low to moderate noise levels. However, in environments with strong noise interference, further optimization using other techniques is required. The table below shows the robustness test results of the end-to-end convolutional neural network model:
[0117] Noise Standard Deviation All-energy peak position Full-range peak half-width Energy resolution 0％ 286.14 62.33 21.78％ 5％ 287.01 60.61 21.12％ 10％ 285.29 61.9 21.70％ 20％ 283.88 60.65 21.36％ 50％ 279.46 71.98 25.76％
[0118] The beneficial effects of this invention are:
[0119] The calculation and identification of transient current waveforms of direct radiation detectors based on end-to-end convolutional neural networks provided in this invention, compared with the prior art, accurately measures the particle energy spectrum of the radiation field to be measured by denoising the energy spectrum by excluding non-passing waveforms, thereby obtaining accurate information on the radiation situation of the radiation field to be measured. In addition, this invention also has the advantages of fast response speed, flexible adjustment of the transient current waveform to be received according to different radiation conditions, high measurement accuracy, simple and reliable network structure, and easy deployment and promotion.
[0120] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0121] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.
[0122] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present invention as described above, which are not provided in detail for the sake of brevity.
[0123] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description.
[0124] One or more embodiments of the present invention are intended to cover all such substitutions, modifications, and variations that fall within the scope of protection of the present invention. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of the present invention should be included within the scope of protection of this disclosure.
Claims
1. A method for energy spectrum noise reduction of a direct radiation detector, characterized in that, The method includes: S1: Collect radiation signals and convert incident photon energy into electron-hole pairs, generate induced current pulses under a bias electric field, amplify the induced current pulses into voltage pulse signals, and perform digital processing; S2: Perform high-pass filtering and threshold screening on the digitized voltage pulse signal, retain the effective signal with an amplitude higher than the set threshold, and eliminate time jitter by peak alignment; S3: Calculate the transient current based on the aligned voltage pulse signal, and perform smoothing, noise reduction, and normalization processing on the transient current waveform; S4: Input the normalized transient current data into an end-to-end convolutional neural network for classification and determine whether it is a complete transit event; S5: Based on the classification results, exclude non-crossing events and reconstruct the denoised energy spectrum data.
2. The energy spectrum noise reduction method for a direct radiation detector according to claim 1, characterized in that, In step S1, a preamplifier is used to convert the charge signal output by the direct radiation detector into a voltage pulse signal, as shown in the formula: Among them, u out This is the output voltage signal of the preamplifier, K is the open-loop gain of the preamplifier op-amp, Q is the input charge, and C... f This is the feedback capacitor for the preamplifier.
3. The energy spectrum noise reduction method for a direct radiation detector according to claim 2, characterized in that, In step S3, the formula for calculating the transient current is: Where i is the calculated transient current and t is time.
4. The energy spectrum noise reduction method for a direct radiation detector according to claim 1, characterized in that, In step S3, the formula used to smooth and reduce noise in the transient current waveform is: Where x is the original transient current data sequence, y n The output data sequence is defined by M, which is the dimensionless length of the sliding window, n, which is the sequence number of the current data point, and k, which is the summation index used to traverse all samples within the window.
5. The energy spectrum noise reduction method for a direct radiation detector according to claim 1, characterized in that, The end-to-end convolutional neural network includes an input layer, a first set of convolutional pooling layers, a second set of convolutional pooling layers, a flattening layer, and an output layer arranged sequentially, wherein the output layer is a fully connected layer.
6. The energy spectrum noise reduction method for a direct radiation detector according to claim 5, characterized in that, The activation functions for both the first group of convolutional pooling layers and the second group of convolutional pooling layers are the ReLU function; The formula for the ReLU function of the first group of convolutional pooling layers is: Where x1 is the input value of the first group of convolutional pooling layers, and f(x1) is the output value of the first group of convolutional pooling layers; The formula for the ReLU function of the second group of convolutional pooling layers is: Where x2 is the input value of the second convolutional pooling layer, and f(x2) is the output value of the second convolutional pooling layer.
7. The energy spectrum noise reduction method for a direct radiation detector according to claim 5, characterized in that, The activation function of the output layer is the Sigmoid activation function, and the formula is: Where x3 is the input value of the output layer, and f(x3) is the output value of the output layer.
8. The energy spectrum noise reduction method for a direct radiation detector according to claim 7, characterized in that, The output layer outputs two types of probability values through the Sigmoid activation function: The output vector is [p1, p2], where p1 is the complete transit probability and p2 is the incomplete transit probability. In step S4, the rule for determining whether it is a complete crossing event is as follows: If p1 > p2, it is determined to be a complete crossing event; If p1 < p2, it is determined to be an incomplete crossing event.
9. A spectral noise reduction system for a direct radiation detector, characterized in that, The system includes: A radiation signal acquisition module is used to acquire radiation signals and convert incident photon energy into electron-hole pairs, generate induced current pulses under a bias electric field, amplify the induced current pulses into voltage pulse signals, and perform digital processing. The pulse signal processing module is used to perform high-pass filtering and threshold screening on the digitized voltage pulse signal, retain the effective signal with an amplitude higher than the set threshold, and eliminate time jitter by peak alignment; A current shaping module is used to calculate the transient current based on the aligned voltage pulse signal, and to smooth, reduce noise, and normalize the transient current waveform. An end-to-end convolutional neural network module is used to receive normalized transient current data and classify it. The energy spectrum data reconstruction module is used to determine whether a complete transit event is based on the output of the end-to-end convolutional neural network module, exclude non-transit events based on the classification results, and reconstruct the noise-reduced energy spectrum data.
10. The energy spectrum noise reduction system for a direct radiation detector according to claim 9, characterized in that, The radiation signal acquisition module includes a direct radiation detector, a preamplifier, and an analog-to-digital converter; The pulse signal processing module includes RC-CR 2 Filters, threshold comparators, and software peak alignment modules; The current shaping module includes a transient current calculation module and a moving average filtering module.