An adaptive signal denoising and baseline correction system and method

Abstract

This invention discloses an adaptive signal denoising and baseline correction system and method. It includes: an environmental parameter sensing module for real-time acquisition of sensor operating environment parameters; a multimodal feature extraction module for fusing environmental parameters with time-domain, frequency-domain, and wavelet-domain features of the original electrical signal to form a high-dimensional feature tensor; a signal decomposition module for mapping the feature tensor to a structured latent space via an encoder, using a dynamic persistence constraint layer to generate an adaptive mask based on environmental parameters to apply to noise features, and then reconstructing the purified signal via a decoder; a training and self-evolution module for pre-training and online fine-tuning of the model; and a hardware-software co-engineering architecture that utilizes a front-end FPGA and a core GPU / NPU, dynamically allocating computing resources based on environmental interference levels. This invention effectively solves the problem of decreased sensor signal denoising accuracy caused by environmental changes, significantly improving signal quality and system stability.

Description

Technical Field

[0001] This invention relates to the fields of signal processing, machine learning and adaptive control technology, and in particular to an adaptive signal denoising and baseline correction system and method. Background Technology

[0002] Sensor arrays are widely used to acquire weak electrical signals in fields such as industrial detection, biomedical signal acquisition, environmental monitoring, and intelligent sensing. However, in practical applications, the raw signals output by sensors are highly susceptible to the influence of complex working environments, such as temperature changes, electromagnetic interference, mechanical vibration, humidity fluctuations, and power supply noise. These environmental factors introduce random noise, systematic interference, and baseline drift into the signal, significantly reducing the signal-to-noise ratio and stability, and affecting the accuracy of subsequent feature analysis and decision-making results.

[0003] Existing signal denoising and baseline correction techniques mainly include traditional digital filtering methods (such as low-pass filtering, band-pass filtering, and adaptive filtering), statistical model-based methods (such as Kalman filtering and empirical mode decomposition), and signal processing methods based on machine learning or deep learning developed in recent years. While these methods can improve signal quality under certain conditions, they generally suffer from the following shortcomings:

[0004] On the one hand, existing technologies mostly use single-domain features or fixed model structures to process signals, and usually only perform analysis in the time domain or frequency domain. This makes it difficult to fully characterize the complex characteristics of signals at different scales and frequency bands, resulting in limited ability to distinguish multi-source noise.

[0005] On the other hand, many noise reduction methods assume a relatively stable working environment during their design, lacking awareness and utilization of real-time environmental parameters, making it difficult to dynamically adjust noise reduction strategies when environmental conditions change rapidly. When the type or intensity of environmental interference changes, model performance is prone to significant degradation, and may even introduce new distortions.

[0006] Furthermore, most existing deep learning-based signal processing solutions rely on statically trained offline models, which are difficult to adaptively update based on real-world data after deployment, resulting in insufficient model generalization ability and long-term performance degradation. At the same time, these solutions typically have high computational resource requirements, lack hardware-software co-design mechanisms, and are difficult to operate efficiently in resource-constrained or real-time-critical embedded systems.

[0007] Therefore, there is an urgent need for an adaptive signal denoising and baseline correction system that can integrate multimodal signal features, sense changes in the working environment, dynamically constrain noise components, and have continuous self-evolution capabilities, in order to overcome the shortcomings of existing technologies in terms of adaptability to complex environments, noise separation accuracy, and long-term system stability. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide an adaptive signal denoising and baseline correction system and method for achieving high-precision, adaptive purification and baseline restoration of sensor array signals under complex and changing environments, thereby significantly improving signal quality, system stability and long-term operational reliability.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an adaptive signal denoising and baseline correction system, comprising:

[0010] An environmental parameter sensing module is used to collect the working environment parameters of the sensor array in real time, including at least the working temperature and cumulative radiation dose.

[0011] The multimodal feature extraction module is connected to the environmental parameter sensing module. It is used to receive the working environment parameters and preprocess the raw electrical signal output by the sensor to extract time-domain features, frequency-domain features and wavelet-domain features, and fuse them to form a four-dimensional feature tensor.

[0012] The signal decomposition module, connected to the multimodal feature extraction module, is used to receive the four-dimensional feature tensor and perform signal decomposition, including:

[0013] The encoder maps the four-dimensional feature tensor to a latent space divided into useful signal features and various noise features;

[0014] A dynamic persistence constraint layer, connected to the encoder, takes the working environment parameters and signal temporal attenuation characteristics as input, dynamically generates noise waveforms through a time-series prediction sub-model and converts them into inverse masks, which are then applied to the corresponding noise features in the latent space.

[0015] The decoder, connected to the dynamic afterglow constraint layer, extracts useful signal features from the latent space and reconstructs the purified signal;

[0016] The training and self-evolution module is connected to the signal decomposition module. It pre-trains the signal decomposition module based on a multi-environment simulation dataset and fine-tunes it online based on high-confidence samples of the purified signal after deployment.

[0017] The hardware and software co-engineering architecture includes a front-end programmable logic device that implements the multimodal feature extraction module, and a core processor that implements the signal decomposition module and the training and self-evolution modules. The core processor integrates a resource scheduler for dynamically allocating computing resources according to the level of environmental interference.

[0018] Furthermore, the multimodal feature extraction module preprocesses the original electrical signal and extracts multidimensional features, specifically including:

[0019] The time-domain feature extraction unit is used to perform baseline recovery and peak detection on the original electrical signal, and extract pulse amplitude, rise time, full width at half maximum, pulse decay time constant and afterglow intensity ratio coefficient as time-domain features;

[0020] The frequency domain feature extraction unit is connected to the time domain feature extraction unit. It calculates the power spectral density of the original electrical signal through the FFT accelerator built into the FPGA and extracts the peak frequency and 3dB bandwidth as frequency domain features.

[0021] The wavelet domain feature extraction unit is connected to the time domain feature extraction unit. It decomposes the original electrical signal into multiple layers through wavelet transform based on the db4 wavelet basis and calculates the energy ratio of each layer as wavelet domain features.

[0022] The feature fusion unit is connected to the time-domain feature extraction unit, the frequency-domain feature extraction unit, and the wavelet-domain feature extraction unit, respectively, and is used to splice the time-domain features, the frequency-domain features, the wavelet-domain features with the operating temperature and the cumulative radiation dose to generate the four-dimensional feature tensor.

[0023] Furthermore, the encoder is specifically a three-layer convolutional neural network that receives the four-dimensional feature tensor as input and maps it to a latent space with a dimension of 64; the latent space is configured to include the following three functionally independent feature subspaces:

[0024] Dimensions 1 to 16 are configured as a real scintillation photon feature subspace to characterize the real signal excited by 511 keV gamma photons.

[0025] Dimensions 17 to 32 are configured as a permaluminance noise feature subspace to characterize delayed luminescence noise caused by crystal lattice defects;

[0026] Dimensions 33 to 64 are configured as other noise feature subspaces to characterize dark noise and inter-crystal crosstalk noise.

[0027] Furthermore, the encoder's three-layer convolutional neural network sequentially includes:

[0028] The first convolutional layer uses 64 3×3×3×2 convolutional kernels to convolve the four-dimensional feature tensor with a stride of 2, and uses the ReLU activation function to output the first feature map.

[0029] The second convolutional layer, connected to the first convolutional layer, uses 128 3×3×3×64 convolutional kernels to convolve the first feature map with a stride of 1, and uses the ReLU activation function to output the second feature map.

[0030] The third convolutional layer, connected to the second convolutional layer, uses 64 1×1×1×128 convolutional kernels to map the second feature map to the latent space of dimension 64.

[0031] Furthermore, the dynamic afterglow constraint layer includes an LSTM sub-model for constraining afterglow noise features. This LSTM sub-model is a two-layer long short-term memory network, and its data processing specifically includes:

[0032] Input steps: The current operating temperature, cumulative radiation dose, and signal time-domain attenuation feature vector extracted from the four-dimensional feature tensor are used as the input parameter set, wherein the signal time-domain attenuation feature vector includes the pulse attenuation time constant and the afterglow intensity scaling factor.

[0033] Hidden state calculation steps: The first LSTM layer receives the input parameter set and calculates the first hidden state; the second LSTM layer receives the first hidden state and calculates the second hidden state.

[0034] Environment-adaptive output steps: Map the second hidden state through a fully connected layer to output a personalized set of afterglow waveform parameters that matches the current environment;

[0035] Constraint implementation steps: Generate the inverse mask based on the personalized afterglow waveform parameter set, and perform a dot product operation between the inverse mask and the afterglow noise feature subspace in the latent space to achieve environmentally adaptive noise feature constraints.

[0036] Furthermore, the dynamic afterglow constraint layer also includes a dynamic weighting mechanism for the loss function, configured as follows:

[0037] Calculate the absolute difference between the current operating temperature and the reference temperature;

[0038] Calculate the dynamic weighting factor based on the absolute difference;

[0039] The dynamic weighting factor is used as the coefficient of the afterglow constraint loss term of the dynamic afterglow constraint layer to form the environment-weighted total loss function.

[0040] When environmental interference increases, the constraint strength on the afterglow noise characteristics is automatically strengthened by increasing the dynamic weighting factor.

[0041] Furthermore, the decoder specifically includes:

[0042] The feature selection subunit, connected to the dynamic afterglow constraint layer, is used to extract only the real scintillation photon feature subspace from the latent space, and to shield the afterglow noise feature subspace and other noise feature subspaces.

[0043] A three-layer deconvolutional network, connected to the feature selection subunit, is used to reconstruct the extracted real scintillation photon features into a purified signal waveform, wherein:

[0044] The first deconvolutional layer uses 128 3×3×3×16 convolutional kernels to upsample the real scintillation photon feature subspace and output the first reconstructed feature map.

[0045] The second deconvolutional layer uses 64 3×3×3×128 convolutional kernels to upsample the first reconstructed feature map and output the second reconstructed feature map.

[0046] The third deconvolutional layer uses a 1×1×1×64 convolutional kernel to reconstruct the second reconstructed feature map into a purified signal waveform that matches the dimensions of the original electrical signal.

[0047] The output subunit is connected to the three-layer deconvolution network and is used to output the purified signal waveform to the image reconstruction module.

[0048] The decoder also includes a residual connection structure, specifically:

[0049] The four-dimensional feature tensor generated by the multimodal feature extraction module is processed through a 1×1×1×1 convolution kernel for channel adjustment, and then connected to the second reconstructed feature map output by the second deconvolution layer via a skip connection.

[0050] The contributions of the original features and the reconstructed features are balanced by learnable weight parameters;

[0051] The weighted summed features are input into the third deconvolution layer to preserve the details of the original signal and improve the reconstruction accuracy.

[0052] Furthermore, the training and self-evolution module includes:

[0053] The pre-training unit is used to pre-train the signal decomposition module based on a multi-environment simulation dataset, wherein the multi-environment simulation dataset covers a preset temperature range and a preset dose range, and contains original signals and clean signal labels corresponding to each working environment parameter.

[0054] An online self-evolving unit, connected to the pre-training unit, is used to achieve online model optimization after system deployment, including:

[0055] The high-confidence sample screening subunit is used to calculate the quality evaluation index of the purified signal and the reference image after image reconstruction. When the quality evaluation index is greater than the preset quality threshold, the sample is marked as a valid sample and associated with its working environment parameters.

[0056] A scenario-specific fine-tuning trigger subunit is connected to the high-confidence sample screening subunit. When the number of effective samples in a specific environment reaches a preset threshold, it triggers fine-tuning of the LSTM sub-model in the dynamic persistence constraint layer. This fine-tuning process freezes other parameters in the signal decomposition module except for the specific layer of the LSTM sub-model.

[0057] Furthermore, the quality evaluation index in the high-confidence sample screening subunit adopts the structural similarity index;

[0058] The reference image is a high signal-to-noise ratio image acquired and confirmed under baseline environmental conditions;

[0059] The working environment parameters associated with the valid sample include the working temperature and cumulative radiation dose recorded by the environmental parameter sensing module when the sample was collected.

[0060] An adaptive signal denoising and baseline correction method, applied to the adaptive signal denoising and baseline correction system described above, includes:

[0061] Step S1: The environmental parameter sensing module collects the working environmental parameters of the sensor array in real time.

[0062] In step S2, the multimodal feature extraction module receives the working environment parameters and preprocesses the raw electrical signal output by the sensor to extract time-domain features, frequency-domain features and wavelet-domain features, and fuses them to form a four-dimensional feature tensor.

[0063] Step S3: The encoder maps the four-dimensional feature tensor to a latent space divided into useful signal features and various noise features; the dynamic persistence constraint layer takes the working environment parameters and signal temporal attenuation features as input, dynamically generates noise waveforms through a time-series prediction sub-model and converts them into inverse masks, which are then applied to the corresponding noise features in the latent space; the decoder extracts useful signal features from the latent space and reconstructs the cleaned signal.

[0064] Step S4: The training and self-evolution module pre-trains the signal decomposition module based on a multi-environment simulation dataset, and fine-tunes the signal decomposition module online based on high-confidence samples selected after image reconstruction and quality evaluation.

[0065] Step S5, the hardware and software co-architecture includes a front-end programmable logic device that implements the multimodal feature extraction module, and a core processor that implements the signal decomposition module and the training and self-evolution module. The core processor integrates a resource scheduler for dynamically allocating computing resources according to the level of environmental interference.

[0066] The beneficial effects of this invention are:

[0067] 1. Improve signal noise reduction and baseline correction in complex environments: This invention introduces environmental parameter sensing and participation in the signal processing process, enabling the system to adaptively adjust the noise reduction strategy according to changes in the working environment, significantly improving the stability and accuracy of signal purification in complex and variable environments.

[0068] 2. Enhanced ability to distinguish and suppress multi-source noise: By fusing time-domain, frequency-domain, and wavelet-domain features to form a four-dimensional feature tensor, and distinguishing useful signal features from noise features in the latent space, the system can achieve fine separation and suppression of different types of noise, thereby reducing useful signal distortion.

[0069] 3. Achieve dynamic modeling and adaptive constraints for noise: Utilize a dynamic persistence constraint layer combined with environmental parameters and a time-series prediction model to generate a noise inverse mask, enabling the system to dynamically suppress noise components as environmental interference changes, thereby improving the consistency and robustness of noise reduction.

[0070] 4. Supports continuous optimization and self-evolution of system performance: Through a training and self-evolution mechanism that combines pre-training and online fine-tuning, the system can continue to optimize model performance based on high-confidence signal samples after actual deployment, avoiding performance degradation during long-term operation.

[0071] 5. Engineering implementation advantages that balance real-time performance and computational efficiency: It adopts a hardware and software co-architecture and dynamically allocates computing resources according to the level of environmental interference, which reduces system latency and resource consumption while ensuring signal processing accuracy, making it suitable for engineering and embedded applications. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the adaptive signal denoising and baseline correction system in this invention;

[0073] Figure 2 This is a schematic diagram of the decoder structure in this invention;

[0074] Figure 3 This is a schematic diagram of the structure of the online self-evolution unit in this invention;

[0075] Figure 4 This is a flowchart of the adaptive signal denoising and baseline correction system in this invention.

[0076] Figure labels: 1. Environmental parameter perception module; 2. Multimodal feature extraction module; 21. Temporal feature extraction unit; 22. Frequency domain feature extraction unit; 23. Wavelet domain feature extraction unit; 24. Feature fusion unit; 3. Signal decomposition module; 31. Encoder; 32. Dynamic persistence constraint layer; 33. Decoder; 331. Feature selection subunit; 332. Three-layer deconvolutional network; 333. Output subunit; 334. Residual connection structure; 4. Training and self-evolution module; 41. Pre-training unit; 42. Online self-evolution unit; 421. High-confidence sample selection subunit; 422. Scene-specific fine-tuning trigger subunit. Detailed Implementation

[0077] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0078] Example 1, refer to Figure 1 This is the first embodiment of the present invention. This embodiment provides an adaptive signal denoising and baseline correction system, deployed within a signal acquisition and processing scenario, suitable for applications such as industrial detection, bioelectrical signal acquisition, and environmental monitoring. The system uses a sensor array as the signal source input, connected to a backend signal processing and computing platform, to achieve adaptive denoising and baseline correction of the original electrical signal.

[0079] The system includes:

[0080] The system comprises an environmental parameter perception module 1, a multimodal feature extraction module 2, a signal decomposition module 3, a training and self-evolution module 4, and a hardware / software collaborative architecture. These modules are interconnected according to... Figure 1 Data interaction and collaborative work are carried out through the defined connection relationships.

[0081] In this embodiment, the system operating environment includes:

[0082] Multi-channel analog or digital sensor array;

[0083] Interference sources such as temperature, electromagnetic fields, and vibration in industrial or laboratory environments;

[0084] An embedded computing platform equipped with programmable logic devices and general-purpose processors;

[0085] A software runtime environment that supports model training and online updates.

[0086] The specific implementation methods of each module:

[0087] (a) Environmental parameter sensing module 1;

[0088] The environmental parameter sensing module 1 is used to collect the operating environmental parameters of the sensor array in real time. In this embodiment, this module can be implemented by or in combination with the following devices:

[0089] Temperature sensor, used to collect changes in ambient temperature;

[0090] Humidity sensor, used to collect ambient humidity information;

[0091] Electromagnetic interference detection unit, used to sense the intensity of external electromagnetic noise;

[0092] Vibration or acceleration sensors are used to detect mechanical vibration disturbances.

[0093] The collected working environment parameters are output in a time-synchronized manner and serve as an important auxiliary input for subsequent signal processing, reflecting the interference characteristics of the current signal acquisition environment.

[0094] (ii) Multimodal feature extraction module 2;

[0095] The multimodal feature extraction module 2 is connected to the environmental parameter sensing module 1 and is used to receive working environment parameters and preprocess and extract features from the raw electrical signals output by the sensor.

[0096] In this embodiment, the module is preferably deployed in a front-end programmable logic device (such as an FPGA) to meet real-time requirements, and its specific functions include:

[0097] Signal preprocessing: The original electrical signal is amplified, normalized, and de-DC biased to ensure that the signals from different acquisition channels and at different times are comparable.

[0098] Time-domain feature extraction: Extracting features such as amplitude, energy, and decay trend of a signal from the time domain perspective to characterize the basic form of signal changes over time.

[0099] Frequency domain feature extraction: Extracting the energy distribution characteristics of the signal in the frequency domain through methods such as fast Fourier transform, which is used to distinguish interference components in different frequency bands.

[0100] Wavelet domain feature extraction: Multi-scale wavelet transform is used to decompose the signal and extract local features at different scales to characterize non-stationary noise and transient interference.

[0101] The aforementioned time-domain features, frequency-domain features, and wavelet-domain features are fused to form a unified four-dimensional feature tensor, which serves as the input to signal decomposition module 3.

[0102] (iii) Signal decomposition module 3;

[0103] The signal decomposition module 3 is connected to the multimodal feature extraction module 2 and is used to receive the four-dimensional feature tensor and perform signal decomposition processing. This module includes an encoder 31, a dynamic persistence constraint layer 32, and a decoder 33.

[0104] 1. Encoder 31;

[0105] The encoder 31 is used to map a four-dimensional feature tensor to a latent space. In this embodiment, the encoder 31 can be implemented using a deep neural network structure, and its output latent space is divided into: a useful signal feature subspace; and multiple noise feature subspaces.

[0106] This mapping process separates the signals and noise that were originally mixed in the observation space at the feature level, providing a basis for subsequent noise suppression.

[0107] 2. Dynamic afterglow constraint layer 32;

[0108] The dynamic afterglow constraint layer 32 is connected to the encoder 31, and its inputs include: operating environment parameters; and the time-domain attenuation characteristics of the signal.

[0109] In this embodiment, the dynamic persistence constraint layer 32 incorporates a time-series prediction sub-model to predict the waveform characteristics of the corresponding noise based on the changing trends of environmental parameters over time. The predicted noise waveform is converted into an inverse mask and applied to the corresponding noise feature subspace in the latent space.

[0110] This method enables dynamic suppression of noise characteristics without unnecessarily weakening useful signal characteristics, thereby improving the accuracy and stability of signal purification.

[0111] 3. Decoder 33;

[0112] Decoder 33 is connected to dynamic persistence constraint layer 32 to extract constrained useful signal features from the latent space, reconstruct them, and output the purified signal.

[0113] The reconstruction process of decoder 33 focuses on preserving the true shape of the signal and baseline stability, thereby achieving baseline correction while completing noise reduction.

[0114] (iv) Training and Self-Evolution Module 4;

[0115] The training and self-evolution module 4 is connected to the signal decomposition module 3 and is used for model training and parameter updates.

[0116] In this embodiment:

[0117] Before deployment, the system pre-trained the signal decomposition module 3 based on a multi-environment simulation dataset, enabling the model to have basic adaptability to various typical noise scenarios.

[0118] After system deployment, the model is fine-tuned online based on samples identified as having high confidence levels in the purified signals, thereby achieving continuous optimization of model parameters.

[0119] Through the above mechanism, the system can continuously evolve adaptively as the operating environment and equipment status change, avoiding performance degradation caused by long-term operation.

[0120] (v) Hardware and software co-engineering architecture;

[0121] This embodiment uses a hardware-software co-architecture to implement the above modules, wherein:

[0122] The multimodal feature extraction module 2 is implemented by a front-end programmable logic device to meet the requirements of high-speed parallel processing;

[0123] The signal decomposition module 3 and the training and self-evolution module 4 are implemented by a core processor, which can be a CPU, GPU or a dedicated AI processing unit.

[0124] The core processor integrates a resource scheduler to dynamically allocate computing resources based on the level of environmental interference.

[0125] When the level of environmental interference is high, the resource scheduler increases the allocation ratio of computing resources related to signal decomposition and noise constraints; when the environment is relatively stable, it reduces the computing load to achieve a balance between performance and power consumption.

[0126] III. Working principle and technical effects of Example 1:

[0127] During system operation, the environmental parameter sensing module 1 first collects the working environmental parameters of the sensor array in real time; while preprocessing the original electrical signal, the multimodal feature extraction module 2 extracts and fuses multi-domain features in combination with environmental parameters to form a four-dimensional feature tensor.

[0128] Subsequently, the signal decomposition module 3 maps the four-dimensional feature tensor to the latent space through the encoder 31 to separate the useful signal features from the noise features; the dynamic persistence constraint layer 32 dynamically suppresses the noise features based on environmental parameters and time-series prediction results; and the decoder 33 reconstructs the purified signal based on this.

[0129] Through continuous optimization via training and self-evolution module 4, the system can operate stably for extended periods, maintaining consistent noise reduction and baseline correction performance under varying environmental conditions. The hardware-software co-architecture further ensures that the system achieves good real-time performance and engineering applicability while maintaining processing accuracy.

[0130] Example 2, refer to Figure 2 and Figure 3This is the second embodiment of the present invention, which is an application example of the present invention in the field of medical imaging. The system still consists of the following modules:

[0131] The system includes an environmental parameter perception module 1, a multimodal feature extraction module 2, a signal decomposition module 3 (including an encoder 31, a dynamic persistence constraint layer 32, and a decoder 33), a training and self-evolution module 4, and a hardware and software co-engineering architecture.

[0132] The system operates within the PET detector signal processing chain and is deployed in an imaging system consisting of a LYSO scintillator array, a photodetector (such as a silicon photomultiplier SiPM), a signal preamplifier circuit, a data acquisition unit (DAQ), and an image reconstruction server.

[0133] Module structure and implementation:

[0134] (1) Environmental parameter sensing module 1: This module is used to sense the operating status parameters of the LYSO crystal array in real time.

[0135] Hardware configuration: Digital temperature sensors (such as PT1000 or DS18B20) and radiation dose integration probes (such as CsI:TI detector subunits or semiconductor radiation monitoring modules) are deployed around the crystal array.

[0136] Data acquisition: Temperature and dose data are synchronously recorded using an analog-to-digital converter (ADC) chip at high frequency and transmitted to the central processing unit.

[0137] Function: To enable dynamic monitoring of the crystal's working environment and provide environmentally dependent information for subsequent signal purification.

[0138] Technical benefits: Enables quantitative perception of the cumulative effects of temperature and radiation, provides a physical basis for dynamic modeling of noise sources, and avoids the accumulation of signal distortion.

[0139] (2) Multimodal feature extraction module 2: This module receives temperature and dose information from environmental parameter sensing module 1, as well as raw electrical signal data output by the detector.

[0140] Preprocessing: Amplitude normalization, DC component removal, and time synchronization of the electrical signal.

[0141] Feature extraction: Through FPGA or high-speed DSP hardware, time-domain features (pulse rising edge, peak value, attenuation constant, etc.), frequency-domain features (main frequency, bandwidth, power spectral density), and wavelet-domain features (multi-scale sub-band energy distribution) are extracted.

[0142] Feature fusion: The above three types of features are superimposed on time frames and combined with environmental parameters to construct a four-dimensional feature tensor (time × frequency × wavelet scale × environmental state).

[0143] Technical effect: Multimodal feature fusion enhances the ability to distinguish between signals and noise, providing high information density input for subsequent deep decomposition, thereby improving purification accuracy.

[0144] (3) Signal decomposition module 3: This module is the core part of the system and is used to separate the real scintillation photon features from complex aliased signals.

[0145] ① Encoder 31: It adopts a convolutional neural network (CNN) structure to map the input four-dimensional feature tensor to the latent space;

[0146] Latent space features are classified into three categories: real scintillation photon features, afterglow noise features, and other noise features.

[0147] Technical effect: The latent space mapping enables the structured decomposition of signals, providing a foundation for accurate noise identification.

[0148] ② Dynamic afterglow constraint layer 32: This layer is constructed based on the Long Short-Term Memory (LSTM) network model, with real-time ambient temperature, cumulative radiation dose and signal attenuation characteristics as input.

[0149] The model outputs a predicted value of the afterglow waveform and generates a corresponding inverse mask, which is used to adaptively suppress the afterglow noise features in the latent space.

[0150] Working principle: LSTM can learn the time-dependent characteristics of afterglow decay under different temperatures and radiation, and dynamically adjust the mask weights to achieve targeted noise reduction.

[0151] Technical effect: Effectively compensates for the drift in afterglow time domain characteristics caused by environmental changes, and maintains the robustness of signal purification.

[0152] ③Decoder 33: Receives the constrained latent space features and reconstructs the purified signal waveform using a deconvolution network.

[0153] Technical effect: The output purified signal is significantly better than the traditional static noise reduction result in terms of time resolution and energy measurement accuracy, and truly reflects the response to flicker events.

[0154] (4) Training and self-evolution module 4: This module is responsible for the system's model optimization and long-term adaptability.

[0155] Initial stage: The signal decomposition module 3 is pre-trained offline using a multi-environment simulation dataset (covering different temperatures and radiation levels);

[0156] Deployment phase: After the system is running, it automatically selects high-confidence cleansing samples for online fine-tuning to achieve continuous self-learning.

[0157] Technical benefits: Maintains the model's adaptability to environmental drift during long-term use, reduces the frequency of manual calibration, and improves system operational stability.

[0158] (5) Hardware and software co-architecture:

[0159] The system adopts a heterogeneous computing architecture: front-end FPGA: performs high-speed feature extraction and signal preprocessing to ensure real-time response;

[0160] Core GPU or NPU: Performs deep neural network inference, LSTM dynamic modeling, and online learning;

[0161] Environmental parameter scheduler: Integrated into the core GPU or NPU, it dynamically allocates computing resources based on the level of environmental interference.

[0162] Technical benefits: It achieves a balance between power consumption and performance while ensuring real-time performance, making it suitable for clinical high-throughput imaging tasks.

[0163] Overall working principle of Example 2:

[0164] When the LYSO crystal array is running in a PET detector, the incident light... Photons trigger scintillation events, and the resulting optical signals are converted into electrical signals by photodetectors and input into the system processing chain.

[0165] Environmental perception stage: The environmental parameter perception module 1 measures the crystal's operating temperature and radiation dose in real time, providing environmental labels for signal characteristics.

[0166] Feature fusion stage: Multimodal feature extraction module 2 simultaneously performs multi-domain feature analysis on the original signal and constructs a four-dimensional feature tensor containing environmental information.

[0167] Signal decomposition stage: The encoder 31 of the signal decomposition module 3 maps the features to the latent space, the dynamic afterglow constraint layer 32 generates an inverse mask based on the environmental information, modulates and suppresses the afterglow noise features, and finally the purified signal is reconstructed by the decoder 33.

[0168] Model optimization phase: The training and self-evolution module 4 fine-tunes the model based on high-confidence samples of the purified signal, so that the model parameters are updated adaptively as the environment evolves.

[0169] Collaborative execution phase: The hardware and software architecture enables parallel computation of multiple modules. The FPGA is responsible for data preprocessing and feature extraction, while the GPU or NPU is responsible for deep inference and online learning, ensuring that the system completes the purification task within millisecond response time.

[0170] Overall technical effects of Example 2:

[0171] It realizes a closed-loop signal purification process of environmental perception, feature fusion, cognitive decomposition, and dynamic optimization;

[0172] This improved the signal stability of the LYSO-PET detector under temperature fluctuations and radiation aging conditions.

[0173] This improved the time and energy resolution of the detector's output signal;

[0174] This reduces the number of manual calibrations required during system operation;

[0175] It implements a real-time adaptive signal purification mechanism that integrates software and hardware, and possesses long-term self-learning capabilities and excellent clinical adaptability.

[0176] Preferably, the hardware and software co-processing architecture in this embodiment consists of a front-end FPGA processing unit, a core computing unit (GPU or NPU), and an environment parameter scheduler. The three communicate with each other via a high-speed bus (e.g., PCIe or AXI interconnect).

[0177] Front-end FPGA processing unit: used for high-speed sampling, filtering, baseline recovery, and preliminary extraction of multimodal features from the raw electrical signal output by the detector;

[0178] Core GPU or NPU unit: used to perform deep neural network inference operations of signal decomposition module 3 (encoder 31, dynamic persistence constraint layer 32, decoder 33), as well as the online fine-tuning process of training and self-evolution module 4;

[0179] Environmental parameter scheduler: Embedded in the GPU or NPU driver layer or scheduling management module, it is used to adjust the system's computing resource allocation strategy in real time according to the level of environmental interference.

[0180] Definition and calculation of environmental interference level:

[0181] The environmental interference level is calculated in real time by the environmental parameter sensing module 1 and is used to describe the environmental stability under the current working state of the crystal array.

[0182] The system uses the following formula to determine the level of environmental interference. Quantification:

[0183] ;

[0184] in: Environmental interference level, The current operating temperature. As the reference temperature, For cumulative radiation dose, The preset upper limit of dosage, These are all weighting coefficients, used to reflect the sensitivity of temperature fluctuations and radiation dose to noise characteristics, respectively.

[0185] Based on the calculation results, the system will Divided into three levels:

[0186] Level 1 (Stable Region): ;

[0187] Level 2 (Mild Interference Zone): ;

[0188] Level 3 (Strong Interference Zone): .

[0189] Preferably, dynamic allocation strategies for computing resources include:

[0190] The environment parameter scheduler dynamically allocates computing resources for GPUs, NPUs, and FPGAs based on the interference level. The specific strategy is as follows:

[0191] Level 1 (Stable Region): The system is in a stable environmental phase with minimal signal characteristic fluctuations. The GPU retains only the inference cores and disables the online fine-tuning process; the FPGA performs routine feature extraction tasks and maintains low-power operation; the dynamic persistence constraint layer 32 runs in standard mode and does not trigger high-frequency adaptive updates.

[0192] Level 2 (Mild Interference Zone): The environment experiences slight fluctuations, and the afterglow features show a shifting trend. The GPU or NPU is dynamically increased to a medium computing load, and some LSTM weight updates are enabled; the FPGA opens its cache channel to extract features at a higher sampling rate; the environment scheduler increases the time step sampling density of the dynamic afterglow constraint layer 32 to compensate for slight noise changes.

[0193] Level 3 (Strong Interference Zone): Temperature or radiation fluctuations are severe, and noise characteristics change significantly. The GPU or NPU enters a high-load mode, activating all computing cores; an online fine-tuning mechanism is enabled, and the training and self-evolution module 4 performs rapid micro-updates based on high-confidence samples; the FPGA enables multi-channel feature extraction paths in parallel to ensure the real-time performance of the input signal; the environmental parameter scheduler calls a priority control algorithm to prioritize the allocation of computing bandwidth to the dynamic persistence constraint layer 32 to enhance persistence noise suppression.

[0194] Preferred implementation of the scheduling algorithm:

[0195] The environmental parameter scheduler adopts a hierarchical control architecture, including: a monitoring layer for real-time data acquisition. ; Decision layer: Determines the current interference level based on thresholds; Execution layer: Calls the computing resource management API through the CUDA / NPU driver interface to realize the allocation of computing cores and adjustment of clock frequency.

[0196] Preferably, when the system is running, the environmental parameter sensing module 1 monitors temperature and dosage information in real time. The environmental parameter scheduler continuously receives this data and calculates the real-time environmental interference level. .

[0197] when When the value is in the low range, the system maintains low power consumption and only performs routine signal purification; when it detects... Upon escalation, the scheduler immediately issues resource adjustment instructions to the GPU or NPU: increase the priority of inference threads; allocate more computing cores; and allow the LSTM module to perform short-term online fine-tuning.

[0198] At the same time, the FPGA's sampling rate and data buffering strategy will also be adjusted accordingly to maintain the real-time performance of the signal input chain.

[0199] The entire scheduling process can be completed in tens of milliseconds, ensuring that there is no delay or frame loss in signal purification during clinical scanning of the PET detection system.

[0200] Technical effects:

[0201] Achieve adaptive matching between computing power and environmental conditions: Through real-time scheduling, the system automatically reduces power consumption in low-interference environments and automatically increases computing power output in high-interference scenarios, realizing a "distribution on demand" computing power management mode.

[0202] Enhanced system stability and signal purification consistency: The environmental interference level and computing resource allocation form a closed-loop control, which enables the dynamic afterglow constraint layer 32 to maintain stable performance under different temperature and radiation conditions, ensuring consistent signal purification output.

[0203] Significantly improves real-time processing capabilities and energy efficiency: Under the same computing power conditions, the system's energy efficiency is improved by about 25%, and the real-time response speed is improved by about 15%, ensuring stable processing of high-throughput data during clinical imaging.

[0204] Extend hardware lifespan and reduce maintenance costs: Intelligent scheduling reduces the long-term high-load operation of GPUs or NPUs, slows down device aging, and thus extends the overall lifespan of the system.

[0205] Through the aforementioned hardware and software collaborative architecture and dynamic computing power scheduling mechanism, this embodiment achieves closed-loop linkage among "environmental status - computing power allocation - signal purification effect", enabling the system to maintain high precision, high efficiency and low energy consumption signal purification performance under different working conditions, which meets the real-time and stability requirements of PET clinical imaging systems.

[0206] The preferred structure and working principle of the multimodal feature extraction module 2:

[0207] This module includes a time-domain feature extraction unit 21, a frequency-domain feature extraction unit 22, a wavelet domain feature extraction unit 23, and a feature fusion unit 24, which are used to extract and fuse multi-dimensional features of the original electrical signal output by the detector.

[0208] (1) Time-domain feature extraction unit 21: First, the baseline of the input electrical signal is restored to eliminate the DC offset caused by the amplifier circuit and temperature drift; the peak detection algorithm is used to identify the flicker pulse and calculate the pulse amplitude, rise time and half width at half maximum; these parameters are important indicators reflecting the time shape of the signal and constitute the time-domain feature vector.

[0209] Technical effect: By baseline correction and time-domain feature extraction, circuit noise and drift error are effectively suppressed, providing stable input for subsequent frequency domain and wavelet domain analysis.

[0210] (2) Frequency domain feature extraction unit 22: Based on the Fast Fourier Transform (FFT) accelerator inside the FPGA, the original signal is converted to the frequency domain; the power spectral density (PSD) is calculated, and the peak frequency and 3dB bandwidth are extracted as frequency domain features.

[0211] Technical effect: Millisecond-level spectrum extraction is achieved through hardware-accelerated FFT operations, which can distinguish between real flicker signals (narrowband high-frequency components) and afterglow and dark noise (low-frequency smooth components).

[0212] (3) Wavelet domain feature extraction unit 23: Use db4 wavelet basis to perform multi-layer wavelet decomposition on the original electrical signal; calculate the energy ratio of each decomposition layer to form wavelet domain features.

[0213] Technical effect: Wavelet decomposition can capture the local time-frequency characteristics of a signal, enhancing the ability to distinguish transient noise and afterglow components.

[0214] (4) Feature fusion unit 24: After splicing the time domain, frequency domain and wavelet domain features respectively, it is fused with the real-time collected working temperature and cumulative radiation dose to form a four-dimensional feature tensor (time domain × frequency domain × wavelet scale × environmental dimension).

[0215] Technical effect: This fusion method introduces environmental variables while maintaining signal integrity, realizing synchronous mapping between signal features and environmental state, and providing high-dimensional input support for subsequent cognitive decomposition.

[0216] Structure and working principle of signal decomposition module 3:

[0217] This module includes an encoder 31 and a dynamic afterglow constraint layer 32, which together constitute the core of signal cognitive decomposition.

[0218] (1) Structure and principle of encoder 31:

[0219] Encoder 31 is a three-layer convolutional neural network used to map four-dimensional feature tensors to a latent space of dimension 64.

[0220] First convolutional layer: 64 3×3×3×2 convolutional kernels are used for convolution extraction with a stride of 2, and ReLU activation is used;

[0221] The second convolutional layer uses 128 3×3×3×64 convolutional kernels with a stride of 1 to output deeper features.

[0222] The third convolutional layer uses 64 1×1×1×128 convolutional kernels to compress and map the second feature map into latent space feature vectors.

[0223] The latent space is structured into three independent functional subspaces:

[0224] Dimensions 1–16: Real scintillation photon feature subspace;

[0225] Dimensions 17–32: Peripheral noise feature subspace;

[0226] Dimensions 33–64: Other noise feature subspaces (including dark noise and crystal crosstalk noise).

[0227] Technical effects:

[0228] Latent space structured mapping enables physical separation of signal features at the representation level; it improves the distinguishability between the afterglow and the real signal, and provides a clear boundary for subsequent dynamic constraints.

[0229] (2) Structure and principle of dynamic afterglow constraint layer 32:

[0230] This layer is a two-layer Long Short-Term Memory (LSTM) network structure, used to adaptively generate persistent waveform constraints based on real-time environmental conditions. The specific process is as follows:

[0231] Input steps: Enter the current time Operating temperature Cumulative radiation dose and signal time-domain attenuation eigenvector ,in The attenuation constant is The proportion of afterglow intensity;

[0232] Hidden state calculation: The first LSTM layer calculates the first hidden state. The second LSTM layer calculates the second hidden state. ;

[0233] Environment-adaptive output: Mapping h2t to a set of persistent waveform parameters via a fully connected layer. ,in This represents the environmental adaptive afterglow decay time constant. Indicates the ambient adaptive afterglow amplitude. Indicates the environmental adaptive strength coefficient;

[0234] Constraint Enforcement: Generate Inverse Mask and the characteristic subspace of afterglow noise in the latent space. Perform dot product operations to achieve dynamic noise suppression;

[0235] Loss weighting mechanism: Calculate the current operating temperature Compared with reference temperature absolute difference According to the absolute difference Calculate dynamic weighting factors ,in This refers to the temperature sensitivity coefficient.

[0236] Dynamic weighting factor As a residual light constraint loss item The coefficients form the environmentally weighted total loss function. ,in This represents the signal reconstruction loss. When environmental interference increases, the dynamic weighting factor is increased. Automatically strengthen the constraint strength on the afterglow noise characteristics.

[0237] Technical effects:

[0238] Two-layer LSTM can capture the evolution of afterglow over long time scales and achieve refined dynamic compensation;

[0239] Dynamic loss weights ensure that the model automatically enhances afterglow suppression under high temperature or high dose perturbation environments;

[0240] This improves the system's adaptability to environmental fluctuations and makes the purification signal more stable in both the time and energy domains.

[0241] Preferably, the structure and working principle of decoder 33 are as follows:

[0242] The decoder 33 includes a feature selection subunit 331, a three-layer deconvolutional network 332, an output subunit 333, and a residual connection structure 334.

[0243] Feature selection subunit 331: Extracts only the real scintillation photon feature subspace from the latent space, and masks afterglow and other noise features. This ensures that subsequent reconstruction is based only on physically real signals.

[0244] A three-layer deconvolutional network 332 includes:

[0245] First deconvolutional layer: 128 3×3×3×16 convolutional kernels, for initial upsampling;

[0246] The second deconvolutional layer consists of 64 3×3×3×128 convolutional kernels, generating refined features.

[0247] The third deconvolutional layer consists of a 1×1×1×64 convolutional kernel, which outputs a cleaned waveform with the same dimensions as the original signal.

[0248] Residual connection structure 334: The four-dimensional feature tensor output by the multimodal feature extraction module 2 is subjected to dimensionality reduction convolution and then connected to the output of the second deconvolution layer via a skip connection; simultaneously, learnable weight parameters are used. The original and reconstructed features are balanced; finally, they are input into the third deconvolution layer for fusion output.

[0249] Technical effects:

[0250] The residual mechanism avoids the loss of details during deep reconstruction, ensuring that the cleaned signal retains the key timing characteristics of the original waveform; the output signal noise suppression rate is significantly improved (up to 10-15dB), the time resolution is improved by about 15%, and the energy resolution is improved by about 10%.

[0251] Preferred structure and function of training and self-evolution module 4:

[0252] (1) Pre-training unit 41, which is executed before system deployment, is used to perform offline training of the signal decomposition module 3 (including encoder 31, dynamic persistence constraint layer 32 and decoder 33) under multiple environmental conditions.

[0253] Multi-environment simulation dataset: The dataset covers a preset temperature range (e.g., 15℃–45℃) and radiation dose range (e.g., 0–2000mGy), and includes the original signals collected or simulated under various environmental conditions and the corresponding clean signal labels.

[0254] Training objective: To optimize model parameters through supervised learning, enabling encoder 31 to accurately separate real scintillation photon features from noise features, and dynamic afterglow constraint layer 32 to generate the correct suppression mask based on environmental parameters.

[0255] Training strategy: A phased training approach is adopted. First, the reconstruction capabilities of encoder 31 and decoder 33 are trained independently. Then, a dynamic persistence constraint layer 32 is introduced for joint optimization to ensure overall convergence.

[0256] Working principle: The pre-training phase enables the system to establish an initial cognitive model under a wide range of environmental conditions, giving the latent space features the ability to generalize across temperature and dose, providing a stable foundation for subsequent online optimization.

[0257] Technical benefits: Learning noise distribution patterns in multiple environments in advance reduces the training burden after on-site deployment; improving the initial purification accuracy of the model and ensuring reliable performance in different PET scanning environments; avoiding the accumulation of purification errors caused by sudden environmental changes.

[0258] (2) Online self-evolution unit 42, which is executed after the system is deployed and running, is used to realize dynamic adaptive optimization of the model based on actual collected data, and includes the following two core sub-units:

[0259] (2.1) High-confidence sample screening subunit 421 receives the purified signal output from decoder 33 and sends it to image reconstruction module to generate PET image. Subsequently, quality assessment is performed between the reconstructed image and the reference image.

[0260] Quality evaluation index: The structural similarity index (SSIM) is adopted. This index comprehensively considers brightness, contrast and structural similarity, and is a commonly used objective index for evaluating image quality.

[0261] Reference image: A high signal-to-noise ratio PET image acquired under standard environmental conditions (reference temperature Tref and low dose conditions) and verified by experts.

[0262] Screening process: When the structural similarity index (SSIM) of the reconstructed image is greater than a preset threshold (e.g., 0.95), the sample corresponding to the clean signal is considered to be of reliable quality and is marked as a valid sample.

[0263] Parameter association: The system automatically associates and stores the valid sample with the working temperature and cumulative radiation dose information recorded by the environmental parameter sensing module 1 at the time of collection.

[0264] Working principle: By introducing image quality evaluation as a feedback signal, the system can filter high-confidence samples from real-time running data without human intervention, providing an accurate sample set for subsequent self-evolving learning.

[0265] Technical effects: Avoid interference from low-quality samples to model updates and ensure the reliability of online learning; realize a closed-loop mechanism for reverse optimization of the signal layer model from the "image level"; improve the long-term self-stability of the system in complex environments.

[0266] (2.2) Scene-specific fine-tuning trigger subunit 422, which receives the high-confidence sample screening results and is used to determine whether to trigger model fine-tuning operation. When the number of valid samples accumulated under specific environmental conditions (such as temperature range or radiation dose range) reaches a preset threshold (such as 1000 samples), the system automatically triggers fine-tuning operation;

[0267] The fine-tuning target is limited to the LSTM sub-model in the dynamic persistence constraint layer 32. The system freezes all parameters in the signal decomposition module 3 except for the specific layer of the LSTM sub-model. Through local backpropagation, only the time-related parameters of the LSTM are updated to better adapt it to the current environmental characteristics. The updated model version is cached and automatically numbered. The system dynamically selects the best matching version for inference based on the real-time environmental parameters.

[0268] Working principle: The scenario-specific fine-tuning trigger mechanism enables the system to automatically adjust model parameters according to different working environments (such as cold start, high-dose continuous scanning, or equipment aging stage) during long-term operation, achieving "scenario-specific adaptive" evolution.

[0269] Technical effects: Enhances the model's self-learning and anti-drift capabilities in non-standard environments; avoids overfitting through parameter freezing strategies, ensuring stable model evolution; achieves self-maintenance throughout the system's lifecycle, significantly reducing the costs of manual annotation and retraining.

[0270] A PET detector signal denoising method based on LYSO scintillator is applied to the above-mentioned adaptive signal denoising and baseline correction system, with reference to... Figure 4 ,include:

[0271] Step S1: The environmental parameter sensing module 1 collects the working environmental parameters of the sensor array in real time.

[0272] In step S2, the multimodal feature extraction module 2 receives the working environment parameters and preprocesses the raw electrical signal output by the sensor to extract time-domain features, frequency-domain features and wavelet-domain features, and fuses them to form a four-dimensional feature tensor.

[0273] In step S3, encoder 31 maps the four-dimensional feature tensor to a latent space divided into useful signal features and various noise features; dynamic persistence constraint layer 32 takes working environment parameters and signal temporal attenuation features as input, dynamically generates noise waveforms through time-series prediction sub-model and converts them into inverse masks, which are then applied to the corresponding noise features in the latent space; decoder 33 extracts useful signal features from the latent space and reconstructs the cleaned signal.

[0274] Step S4: The training and self-evolution module 4 pre-trains the signal decomposition module 3 based on the multi-environment simulation dataset, and fine-tunes it online based on the high-confidence samples of the purified signal after deployment.

[0275] Step S5, the hardware and software co-architecture includes a front-end programmable logic device that implements the multimodal feature extraction module 2, and a core processor that implements the signal decomposition module 3 and the training and self-evolution module 4. The core processor integrates a resource scheduler to dynamically allocate computing resources according to the level of environmental interference.

[0276] It is understood that the above systems and methods are not limited to the application examples described above. Those skilled in the art can apply them to any scenario requiring high-precision processing of weak pulse signals affected by environmental interference, such as ultrasonic sensor signal processing in industrial non-destructive testing, optical signal demodulation in fiber optic sensing systems, and particle detector signal readout in high-energy physics experiments. As long as the sensor signals have similar noise aliasing characteristics and environmental dependence, performance improvements can be achieved by adapting the system of this invention.

[0277] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive signal noise reduction and baseline correction system, characterized in that, include: An environmental parameter sensing module is used to collect the working environment parameters of the sensor array in real time, including at least the working temperature and cumulative radiation dose. The multimodal feature extraction module is connected to the environmental parameter sensing module. It is used to receive the working environment parameters and preprocess the raw electrical signal output by the sensor to extract time-domain features, frequency-domain features and wavelet-domain features, and fuse them to form a four-dimensional feature tensor. The signal decomposition module, connected to the multimodal feature extraction module, is used to receive the four-dimensional feature tensor and perform signal decomposition, including: The encoder maps the four-dimensional feature tensor to a latent space divided into useful signal features and various noise features; A dynamic persistence constraint layer, connected to the encoder, takes the working environment parameters and signal temporal attenuation characteristics as input, dynamically generates noise waveforms through a time-series prediction sub-model and converts them into inverse masks, which are then applied to the corresponding noise features in the latent space. The decoder, connected to the dynamic afterglow constraint layer, extracts useful signal features from the latent space and reconstructs the purified signal; The training and self-evolution module is connected to the signal decomposition module. It pre-trains the signal decomposition module based on a multi-environment simulation dataset and fine-tunes it online based on high-confidence samples of the purified signal after deployment. The hardware and software co-engineering architecture includes a front-end programmable logic device that implements the multimodal feature extraction module, and a core processor that implements the signal decomposition module and the training and self-evolution modules. The core processor integrates a resource scheduler for dynamically allocating computing resources according to the level of environmental interference.

2. The adaptive signal denoising and baseline correction system according to claim 1, characterized in that, The multimodal feature extraction module preprocesses the original electrical signal and extracts multidimensional features, specifically including: The time-domain feature extraction unit is used to perform baseline recovery and peak detection on the original electrical signal, and extract pulse amplitude, rise time, full width at half maximum, pulse decay time constant and afterglow intensity ratio coefficient as time-domain features; The frequency domain feature extraction unit is connected to the time domain feature extraction unit. It calculates the power spectral density of the original electrical signal through the FFT accelerator built into the FPGA and extracts the peak frequency and 3dB bandwidth as frequency domain features. The wavelet domain feature extraction unit is connected to the time domain feature extraction unit. It decomposes the original electrical signal into multiple layers through wavelet transform based on the db4 wavelet basis and calculates the energy ratio of each layer as wavelet domain features. The feature fusion unit is connected to the time-domain feature extraction unit, the frequency-domain feature extraction unit, and the wavelet-domain feature extraction unit, respectively, and is used to splice the time-domain features, the frequency-domain features, the wavelet-domain features with the operating temperature and the cumulative radiation dose to generate the four-dimensional feature tensor.

3. The adaptive signal denoising and baseline correction system according to claim 2, characterized in that, The encoder is specifically a three-layer convolutional neural network that receives the four-dimensional feature tensor as input and maps it to a latent space with a dimension of 64. The latent space is configured to contain the following three functionally independent feature subspaces: Dimensions 1 to 16 are configured as a real scintillation photon feature subspace to characterize the real signal excited by 511 keV gamma photons. Dimensions 17 to 32 are configured as a permaluminance noise feature subspace to characterize delayed luminescence noise caused by crystal lattice defects; Dimensions 33 to 64 are configured as other noise feature subspaces to characterize dark noise and inter-crystal crosstalk noise.

4. The adaptive signal denoising and baseline correction system according to claim 3, characterized in that, The encoder's three-layer convolutional neural network comprises, in sequence: The first convolutional layer uses 64 3×3×3×2 convolutional kernels to convolve the four-dimensional feature tensor with a stride of 2, and uses the ReLU activation function to output the first feature map. The second convolutional layer, connected to the first convolutional layer, uses 128 3×3×3×64 convolutional kernels to convolve the first feature map with a stride of 1, and uses the ReLU activation function to output the second feature map. The third convolutional layer, connected to the second convolutional layer, uses 64 1×1×1×128 convolutional kernels to map the second feature map to the latent space of dimension 64.

5. The adaptive signal denoising and baseline correction system according to claim 1, characterized in that, The dynamic afterglow constraint layer includes an LSTM sub-model for constraining afterglow noise features. This LSTM sub-model is a two-layer long short-term memory network, and its data processing specifically includes: Input steps: The current operating temperature, cumulative radiation dose, and signal time-domain attenuation feature vector extracted from the four-dimensional feature tensor are used as the input parameter set, wherein the signal time-domain attenuation feature vector includes the pulse attenuation time constant and the afterglow intensity scaling factor. Hidden state calculation steps: The first LSTM layer receives the input parameter set and calculates the first hidden state; the second LSTM layer receives the first hidden state and calculates the second hidden state. Environment-adaptive output steps: Map the second hidden state through a fully connected layer to output a personalized set of afterglow waveform parameters that matches the current environment; Constraint implementation steps: Generate the inverse mask based on the personalized afterglow waveform parameter set, and perform a dot product operation between the inverse mask and the afterglow noise feature subspace in the latent space to achieve environmentally adaptive noise feature constraints.

6. The adaptive signal denoising and baseline correction system according to claim 5, characterized in that, The dynamic afterglow constraint layer also includes a dynamic weighting mechanism for the loss function, configured as follows: Calculate the absolute difference between the current operating temperature and the reference temperature; Calculate the dynamic weighting factor based on the absolute difference; The dynamic weighting factor is used as the coefficient of the afterglow constraint loss term of the dynamic afterglow constraint layer to form the environment-weighted total loss function. When environmental interference increases, the constraint strength on the afterglow noise characteristics is automatically strengthened by increasing the dynamic weighting factor.

7. The adaptive signal denoising and baseline correction system according to claim 3, characterized in that, The decoder specifically includes: The feature selection subunit, connected to the dynamic afterglow constraint layer, is used to extract only the real scintillation photon feature subspace from the latent space, and to shield the afterglow noise feature subspace and other noise feature subspaces. A three-layer deconvolutional network, connected to the feature selection subunit, is used to reconstruct the extracted real scintillation photon features into a purified signal waveform, wherein: The first deconvolutional layer uses 128 3×3×3×16 convolutional kernels to upsample the real scintillation photon feature subspace and output the first reconstructed feature map. The second deconvolutional layer uses 64 3×3×3×128 convolutional kernels to upsample the first reconstructed feature map and output the second reconstructed feature map. The third deconvolutional layer uses a 1×1×1×64 convolutional kernel to reconstruct the second reconstructed feature map into a purified signal waveform that matches the dimensions of the original electrical signal. The output subunit is connected to the three-layer deconvolution network and is used to output the purified signal waveform to the image reconstruction module. The decoder also includes a residual connection structure, specifically: The four-dimensional feature tensor generated by the multimodal feature extraction module is processed through a 1×1×1×1 convolution kernel for channel adjustment, and then connected to the second reconstructed feature map output by the second deconvolution layer via a skip connection. The contributions of the original features and the reconstructed features are balanced by learnable weight parameters; The weighted summed features are input into the third deconvolution layer to preserve the details of the original signal and improve the reconstruction accuracy.

8. The adaptive signal denoising and baseline correction system according to claim 5, characterized in that, The training and self-evolution module includes: The pre-training unit is used to pre-train the signal decomposition module based on a multi-environment simulation dataset, wherein the multi-environment simulation dataset covers a preset temperature range and a preset dose range, and contains original signals and clean signal labels corresponding to each working environment parameter. An online self-evolving unit, connected to the pre-training unit, is used to achieve online model optimization after system deployment, including: The high-confidence sample screening subunit is used to calculate the quality evaluation index of the purified signal and the reference image after image reconstruction. When the quality evaluation index is greater than the preset quality threshold, the sample is marked as a valid sample and associated with its working environment parameters. A scenario-specific fine-tuning trigger subunit is connected to the high-confidence sample screening subunit. When the number of effective samples in a specific environment reaches a preset threshold, it triggers fine-tuning of the LSTM sub-model in the dynamic persistence constraint layer. This fine-tuning process freezes other parameters in the signal decomposition module except for the specific layer of the LSTM sub-model.

9. The adaptive signal denoising and baseline correction system according to claim 8, characterized in that, The quality evaluation index in the high-confidence sample screening subunit adopts the structural similarity index; The reference image is a high signal-to-noise ratio image acquired and confirmed under baseline environmental conditions; The working environment parameters associated with the valid sample include the working temperature and cumulative radiation dose recorded by the environmental parameter sensing module when the sample was collected.

10. An adaptive signal denoising and baseline correction method, applied to the adaptive signal denoising and baseline correction system according to any one of claims 1-9, characterized in that, include: Step S1: The environmental parameter sensing module collects the working environmental parameters of the sensor array in real time. In step S2, the multimodal feature extraction module receives the working environment parameters and preprocesses the raw electrical signal output by the sensor to extract time-domain features, frequency-domain features and wavelet-domain features, and fuses them to form a four-dimensional feature tensor. Step S3: The encoder maps the four-dimensional feature tensor to a latent space divided into useful signal features and various noise features; the dynamic persistence constraint layer takes the working environment parameters and signal temporal attenuation features as input, dynamically generates noise waveforms through the time-series prediction sub-model and converts them into inverse masks, which are then applied to the corresponding noise features in the latent space. The decoder extracts useful signal features from the latent space and reconstructs the cleaned signal; Step S4: The training and self-evolution module pre-trains the signal decomposition module based on a multi-environment simulation dataset, and fine-tunes the signal decomposition module online based on high-confidence samples selected after image reconstruction and quality evaluation. Step S5, the hardware and software co-architecture includes a front-end programmable logic device that implements the multimodal feature extraction module, and a core processor that implements the signal decomposition module and the training and self-evolution module. The core processor integrates a resource scheduler for dynamically allocating computing resources according to the level of environmental interference.

Images