A deep learning-based analog front-end sensor noise filtering system and method

Abstract

This invention discloses a deep learning-based analog front-end sensor noise filtering system and method, belonging to the field of sensor signal processing technology. It includes an analog front-end sensor, a signal preprocessing module, a differentiable noise coupling feature detection module, a programmable RC filter network, an ADC module, a deep learning fine filtering module, and a signal output module. By first preprocessing the sensor output signal, detecting noise coupling features, and then driving the programmable RC filter network to achieve analog domain coupling noise pre-suppression, followed by analog-to-digital conversion and digital domain fine filtering by a lightweight residual CNN model, a closed-loop linkage mechanism between the analog and digital domains is constructed. This invention achieves precise end-to-end suppression of coupling noise, solves the problem of low-frequency noise residue, adapts to a wide temperature range environment, improves system filtering stability and operational reliability, and is suitable for high-precision sensor signal acquisition scenarios.

Description

Technical Field

[0001] This invention relates to the field of sensor signal processing technology, specifically to a deep learning-based analog front-end sensor noise filtering system and method. Background Technology

[0002] Analog front-end sensors are widely used in industrial inspection, medical equipment, automotive electronics, aerospace, and other fields. Their core function is to acquire external physical quantities and convert them into processable electrical signals. The signals output by analog front-end sensors are typically weak signals ranging from a few microvolts to tens of millivolts, making them highly susceptible to various types of noise interference during acquisition and transmission. I / f noise and thermal noise are the main inherent noise types of analog front-end sensors, and their coupling leads to a decrease in the signal-to-noise ratio and signal distortion, severely affecting the accuracy and reliability of subsequent signal processing. With the increasing demands for sensor detection accuracy across various fields, higher requirements are being placed on the performance of noise filtering technology for analog front-end sensors. Developing efficient and stable noise filtering methods and systems has become a key research focus for those skilled in the art.

[0003] Currently, noise filtering techniques for analog front-end sensors are mainly divided into two categories: analog domain-based filtering methods and digital domain-based deep learning filtering methods. Analog domain filtering methods often employ fixed-parameter RC filter networks and low-noise amplifier circuits to suppress noise. For example, the invention patent with publication number CN101599767A discloses a fourth-order single-loop local negative feedback Sigma-Delta modulator, which suppresses analog front-end noise through a local negative feedback structure and a fixed-parameter filter circuit. However, this type of method cannot adapt to changes in the coupling characteristics of 1 / f noise and thermal noise in real time, resulting in limited noise suppression accuracy, especially in effectively suppressing low-frequency coupling noise. Another type of digital domain deep learning filtering method uses a CNN model to process noisy signals in the digital domain, improving the signal denoising effect. However, this type of method only filters digital domain signals and does not consider the coupling characteristics of 1 / f noise and thermal noise in the analog domain. It does not perform targeted noise pre-suppression in the analog domain, resulting in obvious low-frequency noise residue after filtering, which cannot meet the requirements of high-precision analog front-end sensors. In addition, some existing filtering schemes that integrate analog and digital domains do not build a closed-loop linkage mechanism between the analog and digital domains, and cannot dynamically adjust the filtering parameters according to the noise coupling characteristics, making it difficult to achieve accurate and efficient suppression of coupled noise.

[0004] In view of the shortcomings of the existing technology, there is an urgent need for a technical solution to solve the problem of low-frequency noise residue caused by the coupling of 1 / f noise and thermal noise in analog front-end sensors. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a noise filtering system and method for analog front-end sensors based on deep learning. This addresses the problems of existing analog domain filtering methods being unable to adapt to changes in the coupling characteristics of 1 / f noise and thermal noise in real time, having limited noise suppression accuracy, existing digital domain deep learning filtering methods only processing digital domain signals without targeted noise pre-suppression in the analog domain, and existing fusion filtering schemes failing to construct a closed-loop linkage mechanism between the analog and digital domains. These issues result in significant low-frequency noise residue in the output signal of the analog front-end sensor, failing to meet the requirements for high-precision signal acquisition.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a deep learning-based analog front-end sensor noise filtering system, comprising an analog front-end sensor, a signal preprocessing module, a differentiable noise coupling feature detection module, a programmable RC filter network, an ADC module, a deep learning fine filtering module, and a signal output module; the output terminal of the analog front-end sensor is connected to the input terminal of the signal preprocessing module, the output terminal of the signal preprocessing module is connected to the input terminal of the differentiable noise coupling feature detection module, the output terminal of the differentiable noise coupling feature detection module is connected to the control terminal of the programmable RC filter network and the first input terminal of the deep learning fine filtering module, and the input terminal of the programmable RC filter network is connected to the signal preprocessing module. The output of the module is connected to the input of the programmable RC filter network, the output of the ADC module is connected to the second input of the deep learning fine filtering module, and the output of the deep learning fine filtering module is connected to the input of the signal output module. The deep learning fine filtering module and the programmable RC filter network form a closed-loop control. The differentiable noise coupling feature detection module is used to detect the coupling characteristics of 1 / f noise and thermal noise in the output signal of the analog front-end sensor and transmit it to the deep learning fine filtering module. The deep learning fine filtering module outputs adjustment instructions to the programmable RC filter network according to the coupling characteristics to achieve pre-suppression of analog domain coupling noise, and then performs fine filtering on the digital signal converted by the ADC module.

[0007] Furthermore, the signal preprocessing module includes a low-noise operational amplifier (LNP) and a buffer. The input of the LNP is connected to the output of the analog front-end sensor, and the output of the LNP is connected to the input of the buffer. The output of the buffer is connected to the input of the differentiable noise coupling feature detection module and the input of the programmable RC filter network, respectively. The LNP is used to amplify the weak signal output by the analog front-end sensor with low noise, and the buffer is used to isolate the influence of subsequent modules on the output signal of the LNP, ensuring signal transmission stability. The differentiable noise coupling feature detection module includes a differentiable detection circuit and an anti-interference unit. The differentiable detection circuit is used to detect the corner frequency of 1 / f noise, the power spectral density of thermal noise, and the coupling coefficient between the two. The anti-interference unit uses a copper foil shielding layer and a differential input method to suppress external electromagnetic interference and common-mode noise. The coupling coefficient is calculated using the following formula: ,in, The coupling coefficient between 1 / f noise and thermal noise is given. The power value of 1 / f noise. This represents the power value of the thermal noise. This is the combined power value after the 1 / f noise and thermal noise are coupled.

[0008] Furthermore, the programmable RC filter network adopts a second-order low-pass RC filter structure, including a programmable resistor and a programmable capacitor, which are interconnected to form a filter circuit. The programmable RC filter network receives adjustment commands from the deep learning fine filtering module and dynamically adjusts the parameters of the programmable resistor and programmable capacitor based on the coupling coefficient detected by the differentiable noise coupling feature detection module, thereby adjusting the cutoff frequency of the filter network. When the coupling coefficient is greater than or equal to 0.6, the time constant of the filter network is decreased to lower the cutoff frequency; when the coupling coefficient is less than or equal to 0.4, the time constant of the filter network is increased to raise the cutoff frequency; when the coupling coefficient is greater than 0.4 and less than 0.6, the cutoff frequency is adjusted to a preset intermediate value to achieve balanced suppression of 1 / f noise and thermal noise. The time constant of the filter network is calculated using the following formula: ,in The time constant of the programmable RC filter network. This is the resistance value of the programmable resistor. This is the capacitance value of the programmable capacitor.

[0009] Furthermore, the ADC module includes an anti-aliasing filter and an analog-to-digital converter (ADC). The input of the anti-aliasing filter is connected to the output of a programmable RC filter network, and the output of the anti-aliasing filter is connected to the input of the ADC. The output of the ADC is connected to the second input of the deep learning fine filtering module. The anti-aliasing filter is used to suppress noise interference caused by high-frequency signal aliasing. The ADC is used to convert the analog signal after analog domain pre-suppression into a digital signal. The sampling interval of the ADC is synchronized with the feature acquisition interval of the differentiable noise coupling feature detection module to ensure the timing consistency of the digital signal and noise coupling features, providing a synchronous input signal for the fine filtering processing of the deep learning fine filtering module.

[0010] Furthermore, the deep learning fine filtering module employs a lightweight residual CNN model, which includes an input layer, a convolutional layer, and an output layer. Each convolutional layer comprises at least three sequentially connected convolutional units. The model parameters are controlled within a preset range, allowing for embedded deployment on an FPGA chip. The first input of the deep learning fine filtering module receives coupling features output from the differentiable noise coupling feature detection module, and the second input receives the digital signal output from the ADC module. Shallow convolutional units extract residual coupling noise features from the digital signal, while deep convolutional units restore signal details, outputting a clean digital signal. Simultaneously, the deep learning fine filtering module generates adjustment instructions in real-time based on the coupling features, transmitting them to a programmable RC filter network to achieve closed-loop linkage filtering between the analog and digital domains, ensuring effective suppression of coupling noise.

[0011] Furthermore, the lightweight residual CNN model is trained using an unsupervised training method based on physical constraints. The training samples consist of a combination of simulated coupling noise and clean signals. The simulated coupling noise is generated according to the Shockley-Reid Hall noise theory model, covering different coupling coefficient ranges. During training, the output impedance of the simulated front-end sensor and the ambient temperature are introduced as physical constraints, allowing the model to learn the correlation between noise and physical parameters, thereby improving the model's adaptability to different noise scenarios. The objective function of the model training is to minimize the residual coupling noise, ensuring that the low-frequency noise residue after filtering is controlled within a preset range. After training, the model can receive coupling features and digital signals in real time, and complete fine filtering and adjustment command generation.

[0012] Furthermore, the signal output module includes a digital-to-analog converter (DAC) unit and an amplitude detection unit. The input of the DAC unit is connected to the output of the deep learning fine filtering module, and the output of the DAC unit is connected to the input of the amplitude detection unit. The DAC unit is used to convert the pure digital signal output by the deep learning fine filtering module into an analog signal, or directly output a digital signal to adapt to different back-end application scenarios. The amplitude detection unit is used to monitor the amplitude of the output signal in real time to ensure that the amplitude error of the output signal is controlled within a preset range, avoid signal amplitude distortion, and ensure the accuracy of the output signal.

[0013] Furthermore, it also includes a temperature adaptive adjustment module. The input of the temperature adaptive adjustment module is used to collect ambient temperature data from the analog front-end sensor, and the output of the temperature adaptive adjustment module is connected to the deep learning fine filtering module. The temperature adaptive adjustment module combines the correlation model of temperature and noise to transmit the ambient temperature data to the deep learning fine filtering module. The deep learning fine filtering module dynamically adjusts the weights of the model loss function according to the ambient temperature data. At low temperatures, it focuses on suppressing 1 / f noise, and at high temperatures, it focuses on suppressing thermal noise. At the same time, it optimizes the activation function of the model to ensure that the signal-to-noise ratio fluctuation of the filtered signal is controlled within a preset range over a wide temperature range, thereby improving the filtering stability of the system in different temperature scenarios.

[0014] Furthermore, it also includes a fault monitoring and self-repair module. The input of the fault monitoring and self-repair module is connected to the deep learning fine filtering module, and the output of the fault monitoring and self-repair module is connected to the programmable RC filter network and the signal output module, respectively. The fault monitoring and self-repair module uses two indicators, noise suppression residual and signal distortion, output by the deep learning fine filtering module to evaluate the filtering effect in real time. When the indicators exceed the preset threshold, it automatically switches to the backup filtering mode. At the same time, it uses the model to reverse locate the fault link and outputs a fault prompt. The backup filtering mode adopts an improved Kalman filter algorithm to ensure that the signal is not lost in the fault state and improve the reliability of the system.

[0015] A deep learning-based method for noise filtering in analog front-end sensors includes the following steps: The first step involves simulating the weak signal output from the front-end sensor. After low-noise amplification and isolation by the signal preprocessing module, the signal is transmitted to the differentiable noise coupling feature detection module and the programmable RC filter network. The second step involves the differentiable noise coupling feature detection module acquiring signals in real time, detecting the corner frequency of 1 / f noise, the power spectral density of thermal noise, and the coupling coefficient between the two, and transmitting the coupling features to the deep learning fine filtering module. The third step involves the deep learning fine filtering module outputting adjustment instructions based on the coupling characteristics to adjust the parameters of the programmable RC filter network and pre-suppress coupling noise in the analog signal. The fourth step involves the pre-suppressed analog signal undergoing anti-aliasing filtering and analog-to-digital conversion via the ADC module, converting it into a digital signal, and then transmitting it to the deep learning fine filtering module. The fifth step involves the deep learning fine filtering module performing fine filtering on the digital signal, extracting and eliminating residual coupling noise, and outputting a clean digital signal. The sixth step involves the output of the pure digital signal after processing by the signal output module. Simultaneously, the deep learning fine filtering module continuously receives new coupling features and dynamically adjusts the parameters of the programmable RC filter network to achieve adaptive closed-loop filtering.

[0016] Compared with existing technologies, this deep learning-based analog front-end sensor noise filtering system and method has the following advantages: I. This invention constructs a closed-loop linkage filtering architecture between the analog and digital domains. First, a differentiable noise coupling feature detection module accurately identifies the coupling characteristics of 1 / f noise and thermal noise. Then, a deep learning fine filtering module drives a programmable RC filter network to achieve dynamic pre-suppression of coupling noise in the analog domain. Subsequently, a lightweight residual CNN model performs fine filtering on the digital signal, achieving end-to-end suppression of coupling noise. It can adapt to changes in the coupling characteristics of 1 / f noise and thermal noise in real time, effectively solving the problem of significant low-frequency noise residue after filtering in existing technologies. It significantly improves the signal-to-noise ratio of the analog front-end sensor output signal and meets the requirements of high-precision signal acquisition.

[0017] Second, this invention optimizes the deep learning model by adding a temperature adaptive adjustment module and a fault monitoring and self-repair module, combined with an unsupervised training method constrained by physical laws. This enables the filtering system to dynamically adjust the filtering strategy according to the ambient temperature, while also evaluating the filtering effect in real time and automatically switching to a backup filtering mode in case of a fault. It can maintain stable filtering performance over a wide temperature range and significantly improve the system's fault tolerance and operational reliability. It is suitable for complex usage environments in various scenarios such as industrial testing and automotive electronics, reduces system maintenance costs, and enhances engineering practicality.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] like Figure 1 and Figure 2 As shown, a deep learning-based analog front-end sensor noise filtering system includes an analog front-end sensor, a signal preprocessing module, a differentiable noise coupling feature detection module, a programmable RC filter network, an ADC module, a deep learning fine filtering module, a signal output module, a temperature adaptive adjustment module, and a fault monitoring and self-repair module.

[0023] The output of the analog front-end sensor is electrically connected to the input of the signal preprocessing module. The output of the signal preprocessing module is electrically connected to the input of the differentiable noise coupling feature detection module. The output of the differentiable noise coupling feature detection module is electrically connected to the control terminal of the programmable RC filter network and the first input of the deep learning fine filtering module. The input of the programmable RC filter network is electrically connected to the output of the signal preprocessing module. The output of the programmable RC filter network is electrically connected to the input of the ADC module. The output of the ADC module is electrically connected to the second input of the deep learning fine filtering module. The output of the deep learning fine filtering module is electrically connected to the input of the signal output module.

[0024] The temperature adaptive adjustment module's input is used to collect ambient temperature data, and its output is electrically connected to the deep learning fine filtering module. The fault monitoring and self-repair module's input is electrically connected to the deep learning fine filtering module, and its output is electrically connected to the programmable RC filter network and the signal output module, respectively. The deep learning fine filtering module and the programmable RC filter network form a closed-loop control of the electrical signal. After the noise coupling characteristics detected by the differentiable noise coupling feature detection module are transmitted to the deep learning fine filtering module, the adjustment commands generated by the deep learning fine filtering module are fed back to the programmable RC filter network in real time, realizing dynamic pre-suppression of coupled noise in the analog domain. Then, the digital signal converted by the ADC module is finely filtered to complete the full-link suppression of coupled noise.

[0025] The analog front-end sensor is a conventional physical quantity acquisition sensor that can collect external physical quantities such as pressure, temperature, displacement, and bioelectricity and convert them into electrical signals. The output electrical signal is a weak analog signal with an amplitude range of a few microvolts to tens of millivolts. This signal is prone to 1 / f noise and thermal noise during acquisition and transmission, and the two types of noise are coupled to each other in the signal transmission link, which is the object of processing by subsequent modules.

[0026] The signal preprocessing module includes a low-noise operational amplifier and a buffer. The input of the low-noise operational amplifier is electrically connected to the output of the analog front-end sensor, and the output of the low-noise operational amplifier is electrically connected to the input of the buffer. The output of the buffer is electrically connected to the input of the differentiable noise coupling feature detection module and the input of the programmable RC filter network, respectively.

[0027] Specifically, the low-noise operational amplifier (op-amp) uses an operational amplifier with low input offset voltage and low noise figure to amplify the weak signal output from the analog front-end sensor with low noise. The amplification factor is set according to the amplitude of the sensor output signal and the signal processing requirements of subsequent modules, thereby increasing the amplitude of the weak signal while avoiding the introduction of additional noise during the amplification process. The buffer adopts a voltage follower structure with high input impedance and low output impedance. It is used to isolate the influence of subsequent differentiable noise coupling characteristic detection modules and programmable RC filter networks on the output signal of the low-noise op-amp, preventing impedance matching problems of subsequent modules from causing signal attenuation or distortion, and ensuring the stability and integrity of weak signal transmission.

[0028] The differentiable noise coupling feature detection module includes a differentiable detection circuit and an anti-interference unit. The differentiable detection circuit is a detection circuit based on analog integrated circuit design, which has continuous signal detection and feature extraction capabilities. It is used to detect the corner frequency of 1 / f noise, the power spectral density of thermal noise, and the coupling coefficient between 1 / f noise and thermal noise in real time. The anti-interference unit adopts a copper foil shielding layer and a differential input method. The copper foil shielding layer covers the periphery of the differentiable detection circuit to shield external electromagnetic interference signals. The differential input method is used to suppress common-mode noise during transmission, ensuring that the noise feature signals collected by the differentiable detection circuit are real and effective, without the introduction of external interference.

[0029] The corner frequency of 1 / f noise is the frequency point at which the power spectral density of 1 / f noise decreases with increasing frequency to the same level as the power spectral density of thermal noise. The differentiable detection circuit extracts this corner frequency feature by performing frequency domain analysis on the acquired noisy signal. The power spectral density of thermal noise is the power value of thermal noise per unit frequency. The differentiable detection circuit calculates this parameter by acquiring the voltage signal of thermal noise and combining it with the physical characteristics of thermal noise.

[0030] The coupling coefficient between 1 / f noise and thermal noise is calculated using the following formula:

[0031] In the formula, The coupling coefficient between 1 / f noise and thermal noise is given. The power value of 1 / f noise. This represents the power value of the thermal noise. This is the combined power value after the 1 / f noise and thermal noise are coupled.

[0032] Specifically, The 1 / f noise voltage signal acquired by the differentiable detection circuit is obtained through power calculation. The thermal noise voltage signal acquired by the differentiable detection circuit is combined with the thermal noise power calculation method to obtain the result. The overall signal containing coupling noise, acquired by a differentiable detection circuit, is obtained through power calculation, and the coupling coefficient is... The value ranges from 0 to 1, and its magnitude reflects the degree of coupling between 1 / f noise and thermal noise. The larger the value, the higher the degree of coupling.

[0033] The programmable RC filter network adopts a second-order low-pass RC filter structure, which includes a programmable resistor and a programmable capacitor. The programmable resistor and the programmable capacitor are electrically connected to form a filter circuit. The programmable resistor is a digital potentiometer-type programmable resistor device, and the programmable capacitor is a programmable capacitor device based on a capacitor array. Both can receive electrical signal commands and adjust their own parameters.

[0034] The control terminal of the programmable RC filter network is electrically connected to the output terminal of the deep learning fine filtering module, receiving adjustment commands from the deep learning fine filtering module and determining the coupling coefficient detected by the differentiable noise coupling feature detection module. By dynamically adjusting the resistance value of the programmable resistor and the capacitance value of the programmable capacitor, the cutoff frequency of the filter network can be adjusted to achieve pre-suppression of coupled noise in the analog domain.

[0035] The time constant of a programmable RC filter network is calculated using the following formula:

[0036] In the formula, The time constant of the programmable RC filter network. This is the resistance value of the programmable resistor. This is the capacitance value of the programmable capacitor.

[0037] time constant The magnitude of the time constant directly determines the cutoff frequency of the filter network. The cutoff frequency of the second-order low-pass RC filter network is negatively correlated with the time constant. A decrease in the time constant leads to a decrease in the cutoff frequency, while an increase in the time constant leads to an increase in the cutoff frequency. The resistance value of the programmable resistor and the capacitance value of the programmable capacitor are continuously or incrementally adjusted according to the adjustment instructions of the deep learning fine filtering module to achieve dynamic changes in the time constant.

[0038] Specifically, the adjustment logic for the coupling coefficient is as follows: when the coupling coefficient... When the coefficient of friction is ≥0.6, it indicates a high degree of coupling between 1 / f noise and thermal noise. In this case, reducing the time constant of the filter network lowers the cutoff frequency and enhances the suppression of low-frequency coupled noise. When the frequency is ≤0.4, it indicates a low coupling degree between 1 / f noise and thermal noise. In this case, increasing the time constant of the filter network raises the cutoff frequency, ensuring that the high-frequency components of the effective signal are not excessively filtered out while suppressing noise. When 0.4 < When the value is less than 0.6, it indicates that the 1 / f noise and thermal noise are moderately coupled. At this time, the cutoff frequency is adjusted to the preset intermediate value to achieve balanced suppression of 1 / f noise and thermal noise, taking into account both the noise suppression effect and the integrity of the effective signal.

[0039] The ADC module includes an anti-aliasing filter and an analog-to-digital converter (ADC). The input of the anti-aliasing filter is electrically connected to the output of the programmable RC filter network. The output of the anti-aliasing filter is electrically connected to the input of the ADC. The output of the ADC is electrically connected to the second input of the deep learning fine filtering module.

[0040] Specifically, the anti-aliasing filter is a low-pass filter with fixed parameters. Its cutoff frequency is set to half of the sampling frequency of the analog-to-digital converter unit, which conforms to the Nyquist sampling theorem. It is used to suppress noise interference caused by high-frequency signal aliasing and prevent high-frequency unwanted signals from folding into the effective signal frequency band during analog-to-digital conversion, thus introducing additional noise. The analog-to-digital converter unit is a high-speed, high-precision analog-to-digital converter that converts the analog signal after analog domain pre-suppression into a digital signal. The converted digital signal is a binary digital sequence, which is transmitted to the deep learning fine filtering module for subsequent digital domain fine filtering processing.

[0041] The sampling interval of the analog-to-digital conversion unit is strictly synchronized with the feature acquisition interval of the differentiable noise coupling feature detection module. The clock signals of the two are output from the same clock source, ensuring the timing consistency of the digital signal and the noise coupling feature. This provides a synchronous input signal for the fine filtering process of the deep learning fine filtering module, ensuring that the deep learning fine filtering module can perform targeted filtering on the digital signal based on the noise coupling features of the same time dimension.

[0042] The deep learning fine filtering module employs a lightweight residual CNN model, which is embedded in an FPGA chip for deployment. This model meets the real-time processing requirements of industrial and automotive applications, and its parameter count is controlled within a preset range, ensuring filtering performance while reducing hardware resource consumption. The lightweight residual CNN model consists of an input layer, a convolutional layer, and an output layer. Each convolutional layer includes at least three sequentially connected convolutional units, with residual connections between them to avoid the vanishing gradient problem during model training and improve the model's feature extraction and signal reconstruction capabilities.

[0043] The first input of the deep learning fine filtering module receives the coupling features output by the differentiable noise coupling feature detection module, including the corner frequency of 1 / f noise, the power spectral density of thermal noise, and the coupling coefficient. The second input receives the digital signal output by the ADC module. Both input signals are normalized by the input layer and then input to the convolutional layer. Specifically, the shallow convolutional units use small-sized convolutional kernels to extract residual coupling noise features in the digital signal, achieving accurate identification of noise features. The deep convolutional units use multi-scale convolutional kernels combined with residual connection structures to restore the detailed features of the signal. While eliminating residual coupling noise, the waveform and amplitude of the effective signal are not distorted to the greatest extent possible, and finally, a clean digital signal is output through the output layer.

[0044] Meanwhile, the deep learning fine filtering module generates adjustment instructions in real time based on the noise coupling characteristics of the input through built-in logic operations and parameter mapping relationships. These instructions are in the form of electrical signals and are transmitted to the control terminal of the programmable RC filter network to realize closed-loop linkage filtering between the analog and digital domains, ensuring that coupled noise is effectively suppressed throughout the entire link of analog domain pre-suppression and digital domain fine filtering.

[0045] The lightweight residual CNN model is trained using an unsupervised training method based on physical constraints, eliminating the need for manual labeling of training samples, reducing training costs, and improving the model's engineering applicability. The training samples consist of a combination of simulated coupling noise and clean signals. The simulated coupling noise is generated based on the Shockley-Reid Hall noise theory model, covering different coupling coefficient ranges from 0 to 1 to ensure the comprehensiveness of the training samples. The clean signal is a noise-free ideal signal collected by a simulated front-end sensor, generated according to the sensor's signal type and amplitude range.

[0046] During training, the output impedance of a simulated front-end sensor and ambient temperature are introduced as physical constraints. These physical parameters are used as auxiliary inputs to the model, allowing it to learn the correlation between noise characteristics and physical parameters, thus improving its adaptability to different noise scenarios and physical environments. The objective function for model training is to minimize the residual coupled noise, and the objective function expression is:

[0047] In the formula, This represents the model's loss value. The number of sampling points for the signal. Let i be the pure signal value at the i-th sampling point. Let be the model filtered output signal value at the i-th sampling point.

[0048] During training, the convolution kernel parameters of the model are continuously updated using the gradient descent algorithm until the loss value L converges to the preset threshold, ensuring that the low-frequency noise residue after filtering is controlled within the preset range. After training, the model can receive noise coupling features and digital signals in real time, and complete fine filtering and real-time generation of adjustment instructions.

[0049] The temperature adaptive adjustment linkage realizes the electrical connection between the deep learning fine filtering module and the temperature adaptive adjustment module. The temperature adaptive adjustment module is a signal processing module with an integrated temperature sensor. Its input terminal collects the ambient temperature data of the analog front-end sensor through the temperature sensor, converts the temperature data into an electrical signal, and then transmits it to the deep learning fine filtering module.

[0050] The temperature adaptive adjustment module incorporates a temperature-noise correlation model, which is fitted based on experimental data and reflects the relationship between ambient temperature and 1 / f noise and thermal noise power values: 1 / f noise is the main noise component at low temperatures, while thermal noise is the main noise component at high temperatures. The deep learning fine filtering module dynamically adjusts the weights of the model's loss function based on the received ambient temperature data. Specifically, at low temperatures, it increases the weight of the residual 1 / f noise in the loss function, focusing on suppressing 1 / f noise; at high temperatures, it increases the weight of the residual thermal noise in the loss function, focusing on suppressing thermal noise. Simultaneously, the deep learning fine filtering module optimizes the model's activation function based on temperature data, adjusting the slope and threshold of the activation function to ensure that the signal-to-noise ratio fluctuation of the filtered signal is controlled within a preset range over a wide temperature range, improving the system's filtering stability in different temperature scenarios.

[0051] The signal output module includes a digital-to-analog converter unit and an amplitude detection unit. The input terminal of the digital-to-analog converter unit is electrically connected to the output terminal of the deep learning fine filtering module, and the output terminal of the digital-to-analog converter unit is electrically connected to the input terminal of the amplitude detection unit.

[0052] Specifically, the digital-to-analog conversion unit is a high-precision digital-to-analog converter that can convert the pure digital signal output by the deep learning fine filtering module into an analog signal, or directly output a digital signal. The two output modes can be switched via a hardware switch to adapt to the signal type requirements of different back-end application scenarios. The amplitude detection unit is a detection module based on a voltage comparator and amplitude sampling circuit, used to monitor the amplitude of the output signal in real time. It compares the monitored amplitude with a preset amplitude reference value. When the amplitude deviation exceeds the preset range, it makes a small correction through the internal feedback circuit to ensure that the amplitude error of the output signal is controlled within the preset range, avoids signal amplitude distortion, and ensures the accuracy of the output signal.

[0053] Fault monitoring and self-repair module The fault monitoring and self-repair module is a logic control module based on digital signal processing. Its input is electrically connected to the deep learning fine filtering module. It collects two indicators in real time: noise suppression residual and signal distortion. The noise suppression residual is the difference in noise power between the input signal and the output signal of the deep learning fine filtering module, and the signal distortion is the waveform similarity deviation between the output signal and the clean signal. These two indicators are the core quantitative parameters for evaluating the filtering effect.

[0054] The fault monitoring and self-repair module has built-in indicator thresholds. It compares the collected noise suppression residual and signal distortion with preset thresholds in real time. When any indicator exceeds the preset threshold, it determines that the current filtering link of the system has failed. At this time, the fault monitoring and self-repair module automatically sends a switching command to the programmable RC filter network and signal output module to switch the system to the backup filtering mode. At the same time, it uses a model to reverse locate the fault link: based on the abnormal characteristics of the noise suppression residual and signal distortion, it matches the preset fault link mapping table to determine the module where the fault occurred, and outputs fault prompts through external display or communication modules to facilitate staff to troubleshoot and repair.

[0055] Specifically, the backup filtering mode adopts an improved Kalman filter algorithm, which is integrated into the fault monitoring and self-repair module. The state equation and observation equation of the Kalman filter are optimized for the noisy signal characteristics of the analog front-end sensor. In the fault state, the digital signal output by the ADC module is filtered in real time to ensure that the signal is not lost and the effective signal characteristics are not destroyed, thereby improving the reliability and fault tolerance of the system.

[0056] The deep learning-based analog front-end sensor noise filtering method of the present invention is applied to the above-mentioned filtering system. This method realizes closed-loop adaptive filtering in both the analog and digital domains, and specifically includes the following steps: The first step is signal acquisition and preprocessing: the analog front-end sensor acquires external physical quantities and converts them into weak analog electrical signals. The signal is then output to the signal preprocessing module, which amplifies the weak signal with low noise using a low-noise operational amplifier and then isolates it using a buffer. The isolated signal is then transmitted to both the differentiable noise coupling feature detection module and the programmable RC filter network.

[0057] The second step is noise coupling feature detection: the differentiable noise coupling feature detection module acquires the preprocessed signal in real time through the differentiable detection circuit, performs frequency domain and power analysis on the signal, detects the corner frequency of 1 / f noise and the power spectral density of thermal noise, and calculates the coupling coefficient between 1 / f noise and thermal noise through the coupling coefficient calculation formula. The above coupling features are then transmitted to the deep learning fine filtering module in real time.

[0058] The third step is analog domain coupling noise pre-suppression: The deep learning fine filtering module generates adjustment instructions based on the received noise coupling characteristics through the built-in parameter mapping relationship, and transmits the adjustment instructions to the control terminal of the programmable RC filter network. The programmable RC filter network dynamically adjusts the resistance value of the programmable resistor and the capacitance value of the programmable capacitor according to the adjustment instructions, changes the time constant and cutoff frequency of the filter network, and performs targeted coupling noise pre-suppression on the preprocessed analog signal.

[0059] The fourth step is analog-to-digital conversion: the analog signal after analog domain pre-suppression is transmitted to the ADC module. First, the high-frequency aliasing noise is suppressed by the anti-aliasing filter, and then the analog signal is converted into a digital signal by the analog-to-digital conversion unit. The sampling interval of the analog-to-digital conversion unit is synchronized with the acquisition interval of the noise coupling characteristics. The converted digital signal is transmitted to the deep learning fine filtering module.

[0060] Step 5, Digital Domain Fine Filtering: The deep learning fine filtering module synchronously inputs the received noise coupling features and digital signal into the lightweight residual CNN model. The model extracts the residual coupling noise features in the digital signal through shallow convolutional units, restores the effective signal details through deep convolutional units, and outputs a clean digital signal after eliminating residual coupling noise.

[0061] Step 6, Signal Output and Closed-Loop Adjustment: The pure digital signal is transmitted to the signal output module. The signal output module selects to output either a digital signal or an analog signal after digital-to-analog conversion according to the requirements of the backend application. The amplitude detection unit monitors the amplitude of the output signal in real time and ensures the amplitude accuracy. At the same time, the deep learning fine filtering module continuously receives new coupling features output by the differentiable noise coupling feature detection module. Based on the changes in coupling features, it dynamically updates the adjustment instructions, adjusts the parameters of the programmable RC filter network, realizes adaptive closed-loop filtering, and continuously and effectively suppresses coupling noise.

[0062] Throughout the system's operation, the temperature adaptive adjustment module collects ambient temperature data in real time and transmits it to the deep learning fine filtering module. The deep learning fine filtering module dynamically adjusts model parameters to ensure filtering stability over a wide temperature range. The fault monitoring and self-repair module evaluates the filtering effect in real time and automatically switches to the backup filtering mode when a system fault occurs, locating and alerting the faulty component to ensure reliable system operation.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based analog front-end sensor noise filtering system, characterized in that, The system includes an analog front-end sensor, a signal preprocessing module, a differentiable noise coupling feature detection module, a programmable RC filter network, an ADC module, a deep learning fine filtering module, and a signal output module. The output of the analog front-end sensor is connected to the input of the signal preprocessing module. The output of the signal preprocessing module is connected to the input of the differentiable noise coupling feature detection module. The output of the differentiable noise coupling feature detection module is connected to the control terminal of the programmable RC filter network and the first input of the deep learning fine filtering module. The input of the programmable RC filter network is connected to the output of the signal preprocessing module. The output of the programmable RC filter network... The input terminal of the ADC module is connected to the input terminal of the deep learning fine filtering module, and the output terminal of the deep learning fine filtering module is connected to the input terminal of the signal output module. The deep learning fine filtering module and the programmable RC filter network form a closed-loop control. The differentiable noise coupling feature detection module is used to detect the coupling characteristics of 1 / f noise and thermal noise in the output signal of the analog front-end sensor and transmit it to the deep learning fine filtering module. The deep learning fine filtering module outputs adjustment instructions to the programmable RC filter network according to the coupling characteristics to achieve pre-suppression of analog domain coupling noise, and then performs fine filtering on the digital signal converted by the ADC module.

2. The deep learning-based analog front-end sensor noise filtering system according to claim 1, characterized in that, The signal preprocessing module includes a low-noise operational amplifier (LNP) and a buffer. The input of the LNP is connected to the output of the analog front-end sensor, and the output of the LNP is connected to the input of the buffer. The output of the buffer is connected to the input of the differentiable noise coupling feature detection module and the input of the programmable RC filter network. The LNP is used to amplify the weak signal output by the analog front-end sensor with low noise. The buffer is used to isolate the influence of subsequent modules on the output signal of the LNP, ensuring signal transmission stability. The differentiable noise coupling feature detection module includes a differentiable detection circuit and an anti-interference unit. The differentiable detection circuit is used to detect the corner frequency of 1 / f noise, the power spectral density of thermal noise, and the coupling coefficient between the two. The anti-interference unit uses a copper foil shielding layer and differential input to suppress external electromagnetic interference and common-mode noise. The coupling coefficient is calculated using the following formula: ,in, The coupling coefficient between 1 / f noise and thermal noise is given. The power value of 1 / f noise. This represents the power value of the thermal noise. This is the combined power value after the 1 / f noise and thermal noise are coupled.

3. The deep learning-based analog front-end sensor noise filtering system according to claim 1, characterized in that, The programmable RC filter network adopts a second-order low-pass RC filter structure, including a programmable resistor and a programmable capacitor, which are interconnected to form a filter circuit. The programmable RC filter network receives adjustment instructions output by the deep learning fine filtering module and dynamically adjusts the parameters of the programmable resistor and programmable capacitor according to the coupling coefficient detected by the differentiable noise coupling feature detection module, thereby adjusting the cutoff frequency of the filter network. When the coupling coefficient is greater than or equal to 0.6, the time constant of the filter network is decreased to lower the cutoff frequency; when the coupling coefficient is less than or equal to 0.4, the time constant of the filter network is increased to raise the cutoff frequency; when the coupling coefficient is greater than 0.4 and less than 0.6, the cutoff frequency is adjusted to a preset intermediate value to achieve balanced suppression of 1 / f noise and thermal noise; the time constant of the filter network is calculated using the following formula: ,in The time constant of the programmable RC filter network. This is the resistance value of the programmable resistor. This is the capacitance value of the programmable capacitor.

4. The deep learning-based analog front-end sensor noise filtering system according to claim 1, characterized in that, The ADC module includes an anti-aliasing filter and an analog-to-digital converter (ADC). The input of the anti-aliasing filter is connected to the output of a programmable RC filter network, and the output of the anti-aliasing filter is connected to the input of the ADC. The output of the ADC is connected to the second input of the deep learning fine filtering module. The anti-aliasing filter is used to suppress noise interference caused by high-frequency signal aliasing. The ADC is used to convert the analog signal, after pre-suppression in the analog domain, into a digital signal. The sampling interval of the ADC is synchronized with the feature acquisition interval of the differentiable noise coupling feature detection module to ensure the timing consistency between the digital signal and the noise coupling feature, providing a synchronous input signal for the fine filtering processing of the deep learning fine filtering module.

5. The deep learning-based analog front-end sensor noise filtering system according to claim 1, characterized in that, The deep learning fine filtering module employs a lightweight residual CNN model, which includes an input layer, a convolutional layer, and an output layer. Each convolutional layer comprises at least three sequentially connected convolutional units. The model's parameter count is controlled within a preset range, allowing for embedded deployment on an FPGA chip. The first input of the deep learning fine filtering module receives coupling features output from a differentiable noise coupling feature detection module, and the second input receives the digital signal output from an ADC module. Shallow convolutional units extract residual coupling noise features from the digital signal, while deep convolutional units restore signal details, outputting a clean digital signal. Simultaneously, the deep learning fine filtering module generates adjustment instructions in real-time based on the coupling features, transmitting them to a programmable RC filter network to achieve closed-loop linkage filtering between the analog and digital domains, ensuring effective suppression of coupling noise.

6. The deep learning-based analog front-end sensor noise filtering system according to claim 5, characterized in that, The lightweight residual CNN model is trained using an unsupervised training method based on physical constraints. The training samples consist of a combination of simulated coupling noise and clean signals. The simulated coupling noise is generated according to the Shockley-Reid Hall noise theory model and covers different coupling coefficient ranges. During training, the output impedance of the simulated front-end sensor and the ambient temperature are introduced as physical constraints to allow the model to learn the correlation between noise and physical parameters, thereby improving the model's adaptability to different noise scenarios. The objective function of the model training is to minimize the amount of residual coupling noise, ensuring that the low-frequency noise residue after filtering is controlled within a preset range. After training, the model can receive coupling features and digital signals in real time to complete fine filtering and generate adjustment instructions.

7. The deep learning-based analog front-end sensor noise filtering system according to claim 1, characterized in that, The signal output module includes a digital-to-analog converter unit and an amplitude detection unit. The input terminal of the digital-to-analog converter unit is connected to the output terminal of the deep learning fine filtering module, and the output terminal of the digital-to-analog converter unit is connected to the input terminal of the amplitude detection unit. The digital-to-analog converter is used to convert the pure digital signal output by the deep learning fine filtering module into an analog signal, or directly output a digital signal to adapt to different back-end application scenarios; the amplitude detection unit is used to monitor the amplitude of the output signal in real time to ensure that the amplitude error of the output signal is controlled within a preset range, avoid signal amplitude distortion, and ensure the accuracy of the output signal.

8. The deep learning-based analog front-end sensor noise filtering system according to claim 1, characterized in that, It also includes a temperature adaptive adjustment module. The input of the temperature adaptive adjustment module is used to collect ambient temperature data from the analog front-end sensor, and the output of the temperature adaptive adjustment module is connected to the deep learning fine filtering module. The temperature adaptive adjustment module combines the correlation model of temperature and noise to transmit the ambient temperature data to the deep learning fine filtering module. The deep learning fine filtering module dynamically adjusts the weights of the model loss function according to the ambient temperature data. At low temperatures, it focuses on suppressing 1 / f noise, and at high temperatures, it focuses on suppressing thermal noise. At the same time, it optimizes the activation function of the model to ensure that the signal-to-noise ratio fluctuation of the filtered signal is controlled within a preset range over a wide temperature range, thereby improving the filtering stability of the system in different temperature scenarios.

9. The analog front-end sensor noise filtering system based on deep learning according to claim 1, characterized in that, It also includes a fault monitoring and self-repair module. The input of the fault monitoring and self-repair module is connected to the deep learning fine filtering module, and the output of the fault monitoring and self-repair module is connected to the programmable RC filter network and the signal output module, respectively. The fault monitoring and self-repair module uses two indicators, noise suppression residual and signal distortion, output by the deep learning fine filtering module to evaluate the filtering effect in real time. When the indicators exceed the preset threshold, it automatically switches to the backup filtering mode. At the same time, it uses the model to reverse locate the fault link and outputs a fault prompt. The backup filtering mode adopts an improved Kalman filter algorithm to ensure that the signal is not lost under fault conditions and improve the reliability of the system.

10. A method for noise filtering in analog front-end sensors based on deep learning, characterized in that, Applied to the system according to any one of claims 1 to 9, the method Includes the following steps: The first step involves simulating the weak signal output from the front-end sensor. After low-noise amplification and isolation by the signal preprocessing module, the signal is transmitted to the differentiable noise coupling feature detection module and the programmable RC filter network. The second step involves the differentiable noise coupling feature detection module acquiring signals in real time, detecting the corner frequency of 1 / f noise, the power spectral density of thermal noise, and the coupling coefficient between the two, and transmitting the coupling features to the deep learning fine filtering module. The third step involves the deep learning fine filtering module outputting adjustment instructions based on the coupling characteristics to adjust the parameters of the programmable RC filter network and pre-suppress coupling noise in the analog signal. The fourth step involves the pre-suppressed analog signal undergoing anti-aliasing filtering and analog-to-digital conversion via the ADC module, converting it into a digital signal, and then transmitting it to the deep learning fine filtering module. The fifth step involves the deep learning fine filtering module performing fine filtering on the digital signal, extracting and eliminating residual coupling noise, and outputting a clean digital signal. The sixth step involves the output of the pure digital signal after processing by the signal output module. Simultaneously, the deep learning fine filtering module continuously receives new coupling features and dynamically adjusts the parameters of the programmable RC filter network to achieve adaptive closed-loop filtering.

Images