An adaptive weak signal extraction method under strong background interference

Abstract

The application provides a weak signal extraction method under strong background interference, which is based on a signal processing system, the signal processing system is based on a precision measurement experiment platform and comprises a main measurement sensor, an auxiliary sensor array and a data processing unit; specifically comprising: (1) data acquisition and quality control preprocessing; (2) adaptive frequency domain parameter generation; (3) rough noise reduction and signal automatic peak searching based on unprotected greedy iteration; (4) multi-order mask protection band generation and radius calculation; (5) fine noise reduction based on global dynamic greedy iteration; (6) frequency domain mask interpolation operation based on complex domain decomposition; (7) standard physical quantity conversion and output. The application has reasonable concept, is widely applied to complex measurement scenes with extremely high signal-to-noise ratio requirement and existing reference noise source, such as gravitational wave detection, weak electromagnetic signal detection, high-precision optical interference measurement and general high-precision weak signal extraction system.

Description

Technical Field

[0001] This invention relates to the field of precision measurement and signal processing technology, specifically to an adaptive weak signal extraction method under strong background interference. Background Technology

[0002] In the field of contemporary precision measurement and signal detection technology, particularly in gravitational wave detection, weak electromagnetic signal detection, high-precision optical interferometry, and general-purpose high-precision weak signal extraction systems, the weak signals of the target signals typically have extremely low energy amplitudes as the requirements for system detection sensitivity continue to increase. Their intensity is often several orders of magnitude lower than the ambient background noise, and the sources of ambient noise are extremely complex, encompassing multi-physics effects such as vibration coupling, electromagnetic interference, temperature drift, and common-mode interference from electronic readout systems. Current technological trends in this field indicate that, as the performance of traditional physical shielding and hardware isolation systems gradually reaches its technological bottleneck, relying solely on hardware means is no longer sufficient to effectively suppress low-frequency broadband noise caused by complex environments. Therefore, data post-processing and software denoising techniques based on multi-sensor fusion have become core methods for improving the final system's signal-to-noise ratio. In particular, utilizing environmental monitoring auxiliary sensors for "coherent noise cancellation" has become an indispensable standard data processing procedure for high-end precision measurement systems.

[0003] Against the aforementioned technical background, the closest existing technology to this invention is the Wiener filtering or least mean square noise cancellation method based on multi-channel auxiliary sensors. This method, based on the coherence of environmental noise in a multidimensional spatial distribution, aims to accurately remove background noise from the main measurement data by establishing a mathematical model. The specific steps of this general method are typically as follows: First, while constructing the main measurement channel, several auxiliary sensors are deployed at key nodes of the system specifically for monitoring environmental interference and minimizing response to target signals; second, synchronous data is collected and converted to the frequency domain, and based on the correlation assumption, the self-power spectral density of each auxiliary channel and the cross-power spectral density between the auxiliary and main channels are calculated; third, according to the statistical error minimization criterion, the transfer function of environmental noise coupling from the auxiliary sensors to the main measurement sensor is calculated using wideband data, thereby constructing a linear mapping model from environmental noise to the main channel; finally, the calculated system transfer function is used to reconstruct the real-time noise data recorded by the auxiliary channels in the frequency or time domain, estimating the background interference coupled into the main channel and subtracting it from the original main measurement data. Ideally, this method can maximize the use of auxiliary reference information to reduce the broadband noise floor of the system, thereby improving the overall data quality and measurement accuracy.

[0004] The aforementioned closest existing technologies, as well as other mainstream signal extraction methods, all have significant technical shortcomings when processing weak or bursty signals at specific frequencies. First, the core flaw of the closest multi-channel Wiener filtering technology and similar adaptive filtering technologies lies in the "signal self-eating" effect: because they use the minimum energy criterion across the entire frequency band when calculating the transfer function, when the reference sensor inevitably senses a small amount of the target signal due to crosstalk, or when the target signal itself is strong, the algorithm incorrectly identifies the target signal as a related part of noise and cancels it out, resulting in the accidental deletion of valuable physical signals; adaptive filtering, through continuous iteration, gradually "learns" to eliminate constant target signals. Second, while notch filtering or band-stop filtering avoids accidental deletion, it causes permanent loss of frequency domain information by directly cutting off energy in a specific frequency band, making it impossible to observe the actual physical processes at that frequency. Third, although matched filtering technology has high detection sensitivity, it heavily relies on accurate prior waveform models, making it completely ineffective for physical experiments with unknown waveforms or those in the exploratory stage. Finally, although the deep learning denoising methods that have emerged in recent years have performed well in the field of image processing, they have problems such as "black box" uninterpretability, lack of sufficient signal-containing samples for training, and easy generation of "model illusion" spurious signals at low signal-to-noise ratios in precision measurements.

[0005] In summary, none of the existing technologies can effectively protect weak signals at specific frequency points and accurately remove background noise without relying on prior models, losing frequency domain information, or accidentally deleting target signals. Summary of the Invention

[0006] To address the technical problems existing in the background art mentioned above, this invention proposes an adaptive weak signal extraction method under strong background interference. Its concept is reasonable and it can be widely used in complex measurement scenarios with extremely high signal-to-noise ratio requirements and the presence of reference noise sources, such as gravitational wave detection, weak electromagnetic signal detection, high-precision optical interferometry, and general high-precision weak signal extraction systems.

[0007] To address the aforementioned technical problems, this invention provides an adaptive weak signal extraction method under strong background interference. This method is based on a signal processing system, which relies on a precision measurement experimental platform and includes a main measurement sensor, an auxiliary sensor array, a data processing unit, a signal conditioning module, a multi-channel synchronous acquisition module, a data storage module, and a result output module. The main measurement sensor is used to acquire a mixed signal containing the target signal and environmental background interference; The auxiliary sensor array is used to acquire reference signals related to environmental background interference. It includes several sensors that are sensitive to environmental interference but theoretically insensitive to the target physical signal. Each auxiliary sensor in the auxiliary sensor array is a reference sensor that is insensitive to or only weakly responds to the target physical signal. It is arranged at a position that can reflect environmental disturbances, structural coupling, common-mode interference, or changes in the external background. Its output signal is used to characterize the reference background component in the main measurement signal. The auxiliary sensor array responds to environmental background interference but does not respond or only weakly responds to the target signal. By using the reference signals acquired by the auxiliary sensor array, the data processing unit estimates the correlation between background interference and the main measurement signal, and thereby reduces coherent background interference from the main measurement signal. The data processing unit is a computer, server, embedded processor, or other hardware device with data processing capabilities that executes data processing and noise reduction algorithms. The data processing unit is configured with a data reading and preprocessing module, a parameter generation module, a preliminary noise reduction and target frequency search module, a guard band generation module, a fine noise reduction module, a frequency domain mapping processing module, and a result conversion and output module. The data reading and preprocessing module is responsible for reading multi-channel synchronous data, step detection, effective segment extraction, detrending, and DC removal. The parameter generation module is responsible for generating frequency-resolution... The algorithm defines the following parameters: rate, search range, iteration upper limit, and convergence threshold. The initial denoising and target frequency search module performs a finite number of rounds of unprotected coarse denoising and automatically locks the target fundamental frequency. The guard band generation module generates a list of protected frequencies and a mask interval based on the fundamental frequency and its harmonics. The fine denoising module performs multiple rounds of greedy denoising under guard band constraints. The frequency domain mapping processing module calculates the complex frequency domain mapping from the reference signal to the current residual signal and interpolates and reconstructs it within the mask region. The result conversion and output module converts the final residual into the target physical quantity and outputs the time-domain and frequency-domain results. The signal conditioning module is used to amplify, filter, isolate, range match, or calibrate the output signals of each sensor. The multi-channel synchronous acquisition module is used to synchronously sample the main measurement channel and the auxiliary measurement channel; The data storage module is used to store the original data, intermediate processed data, and noise reduction results; The result output module is used to output the processed time-domain signal, frequency-domain signal, and corresponding physical quantity results. The main measuring sensor and the auxiliary sensor array are respectively connected to the signal conditioning module, the signal conditioning module is connected to the multi-channel synchronous acquisition module, the multi-channel synchronous acquisition module is connected to the data processing unit, and the data processing unit is respectively connected to the data storage module and the result output module; The adaptive weak signal extraction method mainly includes the following steps: (1) Data acquisition and quality control preprocessing The signal processing system simultaneously reads the output data of the main measurement sensor and the auxiliary sensor array through the multi-channel synchronous acquisition module. The main measurement sensor outputs a main channel signal, which includes the target signal and environmental background interference. The auxiliary sensor array outputs several auxiliary channel signals, which are used to characterize environmental background interference or reference information related to the background interference. To prevent non-stationary abrupt changes or steps in the raw time-domain data from compromising the accuracy of frequency-domain analysis, the data reading and preprocessing module performs a robust step detection algorithm on each channel. (2) Adaptive frequency domain parameter generation The data processing unit dynamically matches and generates the processing parameters required for frequency domain analysis and noise reduction based on the characteristics of the input data, so as to ensure that the frequency resolution of the mapping relationship, i.e. the frequency domain correlation between the auxiliary channel reference signal and the main channel signal or the current residual signal, can automatically adapt to different observation durations. The data processing unit estimates the background interference components in the main channel based on the mapping relationship and subtracts the estimated background interference components from the main channel signal. (3) Coarse noise reduction and automatic peak search based on unprotected greedy iteration Before entering the protected fine noise reduction, the preliminary noise reduction and target frequency search module in the data processing unit first reuses the core framework of "full-domain dynamic greedy iteration" to perform a round of "blind search" operation. Without pre-specifying the protection frequency, the preliminary noise reduction and target frequency search module performs preliminary background interference suppression on multi-channel data to make the target frequency characteristics covered by strong background interference more obvious, and achieves rapid stripping of the dominant strong background interference to reveal the true frequency characteristics of the weak signal buried deep within. (4) Generation and radius calculation of multi-level mask protection zone The base frequency locked according to step (3) above The protection band generation module in the data processing unit automatically generates a list of protected frequencies containing the fundamental frequency and its associated harmonics. For the list of protected frequencies For each frequency point, the data processing unit dynamically matches and generates the mask protection radius based on the characteristics of the input data to construct the optimal protection boundary; (5) Fine-grained noise reduction based on global dynamic greedy iteration The fine noise reduction module in the data processing unit initializes a dynamic data pool containing all candidate reference signals and initializes the main channel signal as the current residual. Then, a fine iterative loop is performed, which is a fine noise reduction process that suppresses background interference in the main channel signal in multiple rounds based on the determined guard band. (6) Frequency domain mapping processing module performs frequency domain mask interpolation operation based on complex domain decomposition; (7) Standard physical quantity conversion and output After the iteration is completed, the result conversion output module in the data processing unit multiplies the final residual signal by the sensitivity coefficient corresponding to the device, converts it into the standard physical units required by the target system, and outputs the time-domain waveform and spectrum data after noise reduction.

[0008] As a preferred embodiment of the present invention, the specific process of the signal processing system performing the robust step detection algorithm for each channel in step (1) is as follows: (1.1) Signal conditioning: Low-pass filtering is performed on the original time-domain signal of each channel, and the first-order difference is calculated to highlight the abrupt change characteristics; (1.2) Statistical threshold determination: Calculate the median absolute deviation of the differential signal, set a statistical threshold, and time points exceeding the threshold are marked as "step anomalies"; (1.3) Optimal Segment Extraction: The data processing unit performs a merged scan of the step anomaly markers for all channels. When any channel is marked as an anomaly at a certain time point or time period, that time point or time period is considered unavailable; when none of the channels are marked as anomaly, that time point or time period is considered available. The data processing unit traverses the entire time series, finds the interval with the longest continuous available time, and synchronously extracts the data of all channels within that interval as the effective dataset. Subsequently, for the valid dataset Detrending and DC removal processes are performed to reduce the impact of slow drift and fixed bias on subsequent frequency domain analysis.

[0009] As a preferred embodiment of the present invention, the low-pass filtering in step (1.1) is used to suppress high-frequency glitches or acquisition noise, and the differential processing is used to highlight abrupt changes, steps, or stair anomalies in the signal; the differential processing includes differential sampling between adjacent sampling points, differential sampling between segmented averages, differential sampling between sliding windows, or other equivalent differential methods that can reflect abrupt changes between adjacent time periods. The specific process of step (1.2) is as follows: robust statistical analysis is performed on the differential signal obtained by differential analysis of adjacent sampling points, or on the segmented differential results obtained by segmented mean differential analysis or sliding window differential analysis, to obtain a robust statistic that can reflect the normal fluctuation range, and an anomaly judgment threshold is set according to the robust statistic; when the differential amplitude at a certain time point or a certain time period exceeds the anomaly judgment threshold, it is marked as a step anomaly point; the robust statistic can be the absolute deviation of the median, quantile statistics, sliding window statistics, or other statistics that are not sensitive to outliers; the anomaly judgment threshold is preset or adaptively adjusted according to the noise level and stability requirements of the specific measurement system.

[0010] As a preferred embodiment of the present invention, the specific process of step (3) is as follows: (3.1) Unprotected global exploration and elimination: Initialize the dynamic data pool of temporary residual signal and candidate reference channel; within a limited number of preset iterations, traverse the entire domain and calculate the noise reduction gain of all remaining reference channels for the current residual; since this stage is in a "blind search" state, the signal processing system does not enable frequency domain mask interpolation, but directly uses the original response relationship of the entire frequency band to lock the "dominant interference source" with the strongest correlation and eliminate it; (3.2) Dynamic iteration: Repeatedly execute the above steps (3.1) to continuously strip away the outermost high-energy background interference until the iteration limit is reached or the noise reduction gain converges; (3.3) Spectrum Analysis and Automatic Peak Locking: The amplitude spectral density of the temporary residual signal obtained after several rounds of preliminary background suppression in steps (3.1) and (3.2) is analyzed; the maximum signal-to-noise ratio peak on the ASD spectrum is searched within the preset wideband search range; the frequency corresponding to the peak is automatically locked by the preliminary noise reduction and target frequency search module in the data processing unit as the target fundamental frequency that needs to be strictly protected in the future. The preliminary noise reduction and target frequency search module automatically determines the target fundamental frequency within the target search range based on the spectrum analysis results after coarse noise reduction, and transmits the target fundamental frequency to the guard band generation module to generate the guard band required for subsequent fine noise reduction.

[0011] As a preferred embodiment of the present invention, the specific process of initializing the temporary residual signal and the dynamic data pool of the candidate reference channel in step (3.1) is as follows: In the preliminary noise reduction stage, the data processing unit takes the main channel signal after effective segment truncation, detrending, DC removal and necessary filtering / calibration as the temporary residual signal, and takes each auxiliary channel signal as the candidate reference signal set; the candidate reference signal set is screened according to channel quality, effectiveness, synchronization and noise level, so as to quickly reduce the background interference that dominates the main channel, rather than to obtain the final noise reduction result; The specific process of calculating the noise reduction gain of all remaining reference channels for the current residual in step (3.1) is as follows: In each round of preliminary noise reduction, the data processing unit calculates the noise reduction contribution of each candidate reference signal to the current temporary residual signal; the noise reduction contribution is obtained by the residual energy reduction amount. The preliminary noise reduction and target frequency search module in the data processing unit selects the candidate reference signal with the largest noise reduction contribution as the dominant background reference signal for this round, and estimates the background interference component based on its correlation with the current temporary residual signal, and then subtracts it from the current temporary residual signal. In step (3.1), the full-band original response relationship refers to the frequency domain correlation between the auxiliary reference signal and the current residual signal when no guard band is set. This full-band original response relationship is used to initially estimate the manifestation of background interference in the main channel and to quickly remove the main background interference in the initial noise reduction stage. The step (3.2) of stripping the outermost high-energy background interference refers to gradually reducing the high-energy background components that can be characterized by the auxiliary reference signal from the current temporary residual signal of the main channel; and these high-energy background components are related to environmental disturbances, structural transmission, common-mode interference or external noise sources. In step (3.3), the search for the maximum signal-to-noise ratio peak on the ASD spectrum is specifically as follows: the data processing unit scans the amplitude spectral density curve within a preset wideband search range and calculates the relative prominence of each frequency point in combination with the local background level; the frequency point with the highest relative prominence and that meets the preset significance condition is determined as the target fundamental frequency candidate point.

[0012] As a preferred embodiment of the present invention, the specific process of performing fine iterative loops in step (5) is as follows: (5.1) Full domain probing: Traverse all remaining reference channels in the data pool and calculate the cross-correlation noise reduction of each channel with the current residual; during this calculation process, the frequency domain mapping processing module is forcibly invoked to ensure that the protected signal energy is not accidentally deleted; (5.2) Optimal locking: The data processing unit compares the noise reduction amount corresponding to all candidate reference signals and selects the candidate reference signal with the largest noise reduction amount as the dominant background reference signal for this round; then, the data processing unit constructs the background interference estimate for this round based on the dominant background reference signal and subtracts it from the current residual signal; if the maximum noise reduction amount corresponding to all candidate reference signals is lower than the preset convergence threshold, or the number of iterations reaches the preset upper limit, it is considered that continuing to process will have limited effect on improving the noise reduction effect, and the fine noise reduction module in the data processing unit automatically terminates the fine noise reduction process; (5.3) Residual update: After the dominant background reference signal of the current round is determined, the frequency domain mapping processing module in the data processing unit generates a background interference estimate based on its corresponding frequency domain mapping result, and subtracts the background interference estimate from the current residual signal to obtain the updated residual signal; then, the frequency domain mapping processing module in the data processing unit updates the candidate reference signal set or its participation status, and enters the next round of loop.

[0013] As a preferred embodiment of the present invention, the initialization process of the fine noise reduction module in the data processing unit in step (5) is as follows: the preprocessed main channel signal is used as the initial residual signal; the reference signals corresponding to the auxiliary sensor array are used as the candidate reference signal set; and a protected frequency list containing the fundamental frequency and its associated harmonics is automatically generated by the protection band generation module in step (4). The mask protection radius and protection interval corresponding to each frequency point are set; the upper limit of the number of iterations and the convergence judgment condition are set; then, in each round of processing, the fine noise reduction module selects the reference signal that contributes the most to the current residual signal from the candidate reference signal set, estimates its corresponding background interference component, and subtracts it from the current residual signal until the stopping condition is met.

[0014] As a preferred embodiment of the present invention, the specific process of calculating the cross-correlation noise reduction amount of each channel and the current residual in step (5.1) is as follows: In the fine noise reduction stage, the data processing unit traverses the candidate reference signal set, estimates the interpretable background component of each candidate reference signal to the current residual signal, attempts to subtract the background component and calculates the residual change amount; the residual change amount is represented by the residual energy reduction index; during the process of estimating the background component, attempting to subtract and calculating the residual change amount for each candidate reference signal, the data processing unit calls the frequency domain mapping processing module and performs protection processing within the target guard band to avoid signal energy near the target frequency from participating in the background mapping estimation.

[0015] As a preferred embodiment of the present invention, the specific process of step (6) is as follows: (6.1) Power spectrum calculation: The cross power spectral density and self power spectral density of the input signal and the reference signal are calculated using discrete Fourier transform; (6.2) Original response estimation: Based on the above power spectrum results, calculate the basic mapping relationship of the original full frequency band; (6.3) Mask area locking: Based on the mask protection radius generated in step (4), mark the corresponding interval on the frequency axis as the mask area, and mark the remaining points as safety anchor points; (6.4) Independent interpolation of complex components: As an abstract mathematical operation, the original response function is decomposed into two independent components, the real part and the imaginary part; using the data of the safety anchor point, mathematical fitting and interpolation operations are performed independently on the real part curve and the imaginary part curve respectively to reconstruct the values ​​in the mask area.

[0016] As a preferred embodiment of the present invention, the specific process of calculating the cross-power spectral density and self-power spectral density of the input signal and the reference signal using discrete Fourier transform in step (6.1) is as follows: the data processing unit converts the input signal and the reference signal to the frequency domain, and calculates the cross-power spectrum between the input signal and the reference signal and the self-power spectrum of the reference signal itself according to conventional spectral estimation methods; the cross-power spectrum is used to characterize the correlation between the input signal and the reference signal, and the self-power spectrum is used to characterize the energy distribution of the reference signal at each frequency; The specific process of step (6.2) is as follows: Based on the cross power spectrum and the self power spectrum, the data processing unit calculates the basic frequency domain mapping relationship from the reference signal to the current residual signal; the basic frequency domain mapping relationship is used to describe the corresponding performance of the background interference component in the reference signal in the current residual signal; the basic frequency domain mapping relationship can be in complex form, with its amplitude part representing the gain relationship and its phase part representing the phase delay relationship. The specific process of step (6.3) is as follows: Based on the guard band generated in step (4), the data processing unit marks the guard interval near the target fundamental frequency and its related frequencies on the frequency axis; the guard interval is the mask area, and the frequency points outside the mask area are used as the safety reference area to estimate the frequency domain mapping relationship of the background interference. The specific implementation process of step (6.4) is as follows: the data processing unit decomposes the basic frequency domain mapping relationship into two components that can be processed separately, and uses the safe reference area data outside the mask area to smoothly reconstruct the mapping relationship within the mask area; the reconstructed mapping relationship remains continuous on the frequency axis and is used for subsequent background interference estimation.

[0017] As a preferred embodiment of the present invention, the specific process of step (7) is as follows: after the fine noise reduction is completed, the data processing unit uses the final residual signal as the target signal extraction result; the result conversion output module converts the target signal extraction result into the target physical quantity according to the sensitivity coefficient of the main measurement sensor, the gain of the amplifier circuit, the quantization coefficient of the acquisition system or the calibration coefficient of the experimental platform; The converted target physical quantity is displacement, velocity, acceleration, force, electromagnetic signal intensity, phase change, or other physical quantities corresponding to the specific measurement object; the result conversion output module in the data processing unit finally outputs the denoised time-domain waveform, frequency-domain spectrum, and related data files.

[0018] By adopting the above technical solution, the present invention has the following beneficial effects: The adaptive weak signal extraction method under strong background interference of this invention is reasonably conceived and can be widely used in complex measurement scenarios with extremely high signal-to-noise ratio requirements and the presence of reference noise sources, such as gravitational wave detection, weak electromagnetic signal detection, high-precision optical interferometry, and general high-precision weak signal extraction systems.

[0019] Based on the multi-channel noise cancellation method using auxiliary sensors, this invention proposes the following four core technological innovations by improving the response estimation model, noise iteration process, and spectrum protection range: (1) This invention constructs a “frequency domain mask interpolation model with independent complex domain components.” Unlike existing technologies that directly process signal magnitude or simply cut off signal frequency bands, the specific technical means of this invention is as follows: First, the interval where the target signal is located is locked as the mask area; then, response reconstruction based on complex domain decomposition of the signal is performed, and effective data from the sidebands on both sides of the mask area are used as anchor points to independently perform mathematical interpolation operations to fit the background interference substrate within the mask area. This step, by mathematically achieving feature decoupling, ensures that while forcibly skipping the target frequency point to prevent “signal self-eating,” it effectively solves the response distortion problem caused by conventional noise reduction methods.

[0020] (2) This invention employs a noise stripping process based on "global dynamic greedy iteration with maximum noise reduction gain". Unlike existing technologies that process reference channels in a fixed, pre-defined order, the specific technical means of this invention is as follows: a dynamic data pool containing all candidate reference signals is constructed. In each round of noise reduction iteration, all remaining reference channels in the data pool are traversed across the entire domain, and the cross-correlation noise reduction amount between each channel and the current main signal residual is calculated. The system automatically selects the channel with the largest current noise reduction amount as the "dominant interference source" for this round for subtraction, and simultaneously updates the main signal residual of the system. This method dynamically determines the processing path by calculating the global correlation score in real time, eliminating model bias caused by manual sorting, and realizing dynamic decoupling and global stripping of multi-source interference.

[0021] (3) This invention establishes a "synchronous mask protection mechanism for multi-harmonic correlation". For periodic weak signals with complex spectral structures, existing technologies that only protect the fundamental frequency will cause the signal energy in the high-frequency band to be mistakenly deleted as noise, resulting in time-domain waveform distortion. The specific technical means of this invention is: the system automatically generates a harmonic sequence containing the fundamental frequency and multiple harmonics for protection; when constructing the mapping model, all frequency points in the sequence are synchronously mapped as protected mask regions, and the above-mentioned interpolation fitting operation is performed simultaneously at all harmonic frequency points. This means ensures that the complete spectral structure of the target signal in the entire frequency band is not destroyed, so that the real target waveform with a specific shape can be restored without distortion in the output result after noise reduction.

[0022] (4) This invention introduces a "two-stage guided adaptive frequency positioning" mechanism. In view of the problem that traditional band-stop filtering or masking technology relies heavily on manual experience to blindly set a fixed protection frequency, when the initial frequency of the target signal is unknown or there is a slight frequency shift due to non-stationary environment, the protection is easily misaligned or the real signal is eliminated. The specific technical means of this invention is as follows: First, the first stage of maskless coarse noise reduction preprocessing is performed to initially remove the dominant broadband background interference, thereby exposing the local energy peak in the residual spectrum; then, the system extracts the coordinates of the dynamic peak as the "reference frequency"; finally, the reference frequency is automatically fed back to the second stage of precise noise reduction iteration to directly guide the accurate generation and anchoring of the center position of the aforementioned protected mask. This method breaks the vulnerability of static parameters to frequency drift by constructing a dynamic detection closed loop of "exposure first, then fixed-point protection". It completely eliminates the heavy reliance on precise prior waveform templates and complicated manual parameter tuning, and realizes the automatic location and protection of unknown sudden transient signals and weak signals from strong background interference. It greatly improves the cold start capability and robustness of the algorithm system in complex non-stationary environments. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the adaptive weak signal extraction method under strong background interference according to the present invention. Detailed Implementation

[0025] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The present invention will be further explained below with reference to specific embodiments.

[0027] like Figure 1 As shown in the figure, this embodiment provides an adaptive weak signal extraction method under strong background interference, which is based on a signal processing system that relies on a precision measurement experimental platform.

[0028] The aforementioned signal processing system mainly includes a main measurement sensor, an auxiliary sensor array, and a data processing unit.

[0029] The main measurement sensor (main channel) is used to acquire a mixed signal containing the target signal and environmental background interference, denoted as... .

[0030] This auxiliary sensor array (reference channel) is used to acquire reference signals related to environmental background interference. It includes several sensors that are sensitive to environmental interference but theoretically insensitive to target physical signals, specifically including but not limited to: environmental condition monitoring sensors covering multiple physical fields such as electricity, magnetism, light, and vibration. These reference channel signals are denoted as the set. .

[0031] Each auxiliary sensor in the auxiliary sensor array is a reference sensor that is insensitive to or only weakly responds to the physical signal of the target under test. They are arranged in positions that can reflect environmental disturbances, structural coupling, common-mode interference, or changes in the external background. Their output signals are used to characterize the reference background components in the main measurement signal.

[0032] The auxiliary sensor array primarily responds to environmental background interference, but does not respond or responds weakly to the target signal. By utilizing the reference signal collected by the auxiliary sensor array, the data processing unit can estimate the correlation between background interference and the main measurement signal, and thereby reduce coherent background interference from the main measurement signal.

[0033] The data processing unit is a computer industrial control computer, server, embedded processor, or other hardware device with data computing capabilities that executes data processing and noise reduction algorithms. The data processing unit is equipped with software functional modules, including a data reading and preprocessing module, a parameter generation module, a preliminary noise reduction and target frequency search module, a guard band generation module, a fine noise reduction module, a frequency domain mapping processing module, and a result conversion and output module. The software functional modules are implemented using programs, algorithms, firmware, or a combination thereof, and are uniformly scheduled and executed by the data processing unit.

[0034] The data reading and preprocessing module is responsible for reading multi-channel synchronous data, step detection, effective segment extraction, detrending, and DC removal. The parameter generation module generates the frequency resolution, search range, iteration upper limit, and convergence threshold. The preliminary denoising and target frequency search module performs finite-round unprotected coarse denoising and automatically locks the target fundamental frequency. The guard band generation module generates a list of protected frequencies and mask intervals based on the fundamental frequency and its harmonics. The fine denoising module performs multi-round greedy denoising under guard band constraints. The frequency domain mapping processing module calculates the complex frequency domain mapping from the reference signal to the current residual signal and interpolates and reconstructs it within the mask area. The result conversion and output module converts the final residual into the target physical quantity and outputs the time-domain and frequency-domain results.

[0035] This signal conditioning module is used to amplify, filter, isolate, range match, or calibrate the output signals of each sensor.

[0036] This multi-channel synchronous acquisition module is used to synchronously sample the main measurement channel and the auxiliary measurement channel.

[0037] This data storage module is used to save the original data, intermediate processed data, and noise reduction results.

[0038] The output module is used to output the processed time-domain signal, frequency-domain signal, and corresponding physical quantity results.

[0039] The main measuring sensor and the auxiliary sensor array are respectively connected to the signal conditioning module, which is connected to the multi-channel synchronous acquisition module. The multi-channel synchronous acquisition module is connected to the data processing unit, which is connected to the data storage module and the result output module.

[0040] The adaptive weak signal extraction method under strong background interference of the present invention mainly includes the following steps: Step S1: Data Acquisition and Quality Control Preprocessing The signal processing system simultaneously reads the output data of the main measurement sensor and the auxiliary sensor array through the multi-channel synchronous acquisition module. The main measurement sensor outputs a main channel signal, which includes the target signal and environmental background interference. The auxiliary sensor array outputs several auxiliary channel signals, which are used to characterize environmental background interference or reference information related to the background interference. To prevent non-stationary abrupt changes or steps in the raw time-domain data from compromising the accuracy of frequency-domain analysis, the data reading and preprocessing module in the data processing unit performs a robust step detection algorithm on each channel. S1.1 Signal Conditioning: Low-pass filtering is applied to the original signal, and first-order difference is calculated to highlight abrupt changes. Specifically, the original time-domain signal of each channel is preprocessed, including low-pass filtering and differential processing. Low-pass filtering is used to suppress high-frequency glitches or acquisition noise, while differential processing is used to highlight abrupt changes, steps, or staggered anomalies in the signal. Differential processing can employ adjacent sampling point differentials or segmented mean differentials; the differential processing includes adjacent sampling point differentials, segmented mean differentials, sliding window mean differentials, or other equivalent differential methods that can reflect abrupt changes in adjacent time periods.

[0041] S1.2 Statistical Threshold Determination: Calculate the median absolute deviation of the differential signal, set a statistical threshold, and mark time points exceeding this threshold as "step anomalies." The specific process is as follows: Perform robust statistical analysis on the differential signals obtained by differentiating adjacent sampling points, or on the piecewise differential results obtained by piecewise mean difference or sliding window difference, to obtain a statistic that reflects the normal fluctuation range, and set an anomaly determination threshold based on this statistic; when the differential amplitude at a certain time point or over a certain time period exceeds the anomaly determination threshold, it is marked as a step anomaly. This robust statistic can be the median absolute deviation, quantile statistics, sliding window statistics, or other statistics that are insensitive to outliers; the anomaly determination threshold can be preset or adaptively adjusted according to the noise level and stability requirements of the specific measurement system.

[0042] S1.3 Optimal Segment Extraction: The data processing unit merges and scans the step anomaly markers of all channels. When any channel is marked as an anomaly at a certain time point or time period, that time point or time period is considered unavailable; when none of the channels are marked as anomaly, that time point or time period is considered available. The data processing unit traverses the entire time series, finds the interval with the longest continuous available time, and simultaneously extracts the data from all channels within that interval as the valid dataset. Subsequently, for the valid dataset Detrending and DC removal processes are performed to reduce the impact of slow drift and fixed bias on subsequent frequency domain analysis.

[0043] Step S2: Adaptive Frequency Domain Parameter Generation To ensure that the frequency resolution of the mapping relationship can automatically adapt to different observation durations, the main control unit (which can be a control program module or a control hardware module within the data processing unit; the main control unit is used to generate the operating parameters required for subsequent frequency domain processing and noise reduction based on the basic characteristics of the acquired data, and to control each functional module to execute according to a preset process) dynamically matches and generates core parameters (core parameters refer to the processing parameters required for frequency domain analysis and noise reduction, including frequency resolution, segment processing length, target search range, etc.) based on the characteristics of the input data (input data characteristics may include sampling rate, data length, observation duration, number of channels, frequency range, noise level, and target search frequency band, etc.). This process includes setting protection bandwidth, iteration limits, and convergence criteria to ensure that the frequency resolution of the mapping relationship (which refers to the frequency domain correlation between the auxiliary channel reference signal and the main channel signal or the current residual signal; the data processing unit estimates the background interference components in the main channel based on this mapping relationship and subtracts the estimated background interference components from the main channel signal) can automatically adapt to different observation durations. As a general mathematical processing module (which can be understood as the frequency domain calculation program module in the data processing unit, used to complete operations such as spectrum analysis, correlation estimation, bandwidth protection, and signal reconstruction), this step ensures that the granularity of the noise reduction algorithm in capturing interference features remains constant regardless of the data length.

[0044] Step S3: Coarse noise reduction and automatic signal peak search based on unprotected greedy iteration Before entering the protected fine noise reduction stage, the preliminary noise reduction and target frequency search module in the data processing unit first reuses the core framework of "global dynamic greedy iteration" to perform a round of "blind search." This preliminary noise reduction and target frequency search module performs preliminary background interference suppression on multi-channel data without pre-specifying the protection frequency, making the target frequency characteristics masked by strong background interference more apparent. This achieves rapid removal of dominant strong background interference to reveal the true frequency characteristics of the deeply buried weak signal. This stage is completely consistent with the subsequent fine noise reduction algorithm, differing only in the upper limit of the number of iterations (set to a lower limit) and frequency domain mask protection (interpolation protection is not enabled in this stage). Its purpose is to quickly remove dominant strong background interference to reveal the true frequency characteristics of the deeply buried weak signal. Specific steps include: S3.1 Unprotected Global Probe and Reduction: Initialize the temporary residual signal and the dynamic data pool of candidate reference channels (in the initial denoising stage, the data processing unit uses the main channel signal after effective segment truncation, detrending, DC removal, and necessary filtering / calibration as the temporary residual signal; differential processing is only used as an intermediate calculation process for step anomaly detection and is not used as the input signal for subsequent denoising, and each auxiliary channel signal is used as a candidate reference signal set; the candidate reference signal set can be screened according to channel quality, effectiveness, synchronization, and noise level; the purpose of this stage is to quickly reduce the dominant background interference in the main channel, rather than to obtain the final denoising result); within a limited preset number of iterations (the preset number of iterations refers to the maximum number of loops allowed to be executed in the initial denoising stage; this maximum number of loops is preset by the system to limit the processing depth of the initial denoising and avoid over-processing the main channel signal before the target frequency is determined), the entire domain is traversed and the denoising gain of all remaining reference channels on the current residual is calculated (in each round of initial denoising, the data processing unit... The processing unit calculates the noise reduction contribution of each candidate reference signal to the current temporary residual signal. This noise reduction contribution is measured by the decrease in residual energy (the noise reduction contribution is the decrease in energy or percentage of the current temporary residual signal after the background interference component estimated by the candidate reference signal through frequency domain mapping is subtracted). The preliminary noise reduction and target frequency search module in the data processing unit selects the candidate reference signal with the largest noise reduction contribution as the dominant background reference signal for this round, and estimates the background interference component based on its correlation with the current temporary residual signal, and then subtracts it from the current temporary residual signal. Since this stage is in a "blind search" state, the signal processing system does not enable frequency domain mask interpolation, but directly uses the full-band original response relationship (the full-band original response relationship refers to the frequency domain correlation between the auxiliary reference signal and the current residual signal without setting a guard band; this frequency domain correlation relationship is used to initially estimate the manifestation of background interference in the main channel and to quickly remove the main background interference in the preliminary noise reduction stage) to lock in the "dominant interference source" with the strongest correlation and subtract it.

[0045] S3.2 Dynamic Iteration: Repeatedly execute step S3.1 above, continuously stripping away the outermost high-energy background interference (this stripping away of the outermost high-energy background interference refers to gradually reducing the high-energy background components that can be characterized by the auxiliary reference signal from the current temporary residual signal of the main channel; these background components are usually related to environmental disturbances, structural transmission, common-mode interference, or external noise sources; this step does not directly delete a signal of a certain frequency band, but estimates and subtracts the referenced background interference part based on the correlation between the reference channel and the main channel), until the iteration limit is reached or the noise reduction gain converges.

[0046] S3.3, Spectrum Analysis and Automatic Peak Locking: Amplitude Spectral Density (ASD) analysis is performed on the residual signal obtained after coarse denoising (the residual signal after coarse denoising refers to the temporary residual signal obtained after several rounds of preliminary background suppression in steps S3.1 and S3.2; the strong background interference in this signal has been reduced to a certain extent, so the frequency characteristics of the target signal are relatively easier to identify). (Amplitude spectral density analysis is a commonly used frequency domain analysis method in signal processing, used to characterize the amplitude distribution of a signal at different frequencies; the data processing unit can use conventional spectrum analysis methods to calculate the amplitude spectrum of the residual signal after coarse denoising.) The system searches for the maximum signal-to-noise ratio (SNR) peak value on the ASD spectral line within a preset wideband search range (the data processing unit scans the amplitude spectral density curve within the preset wideband search range and calculates the relative prominence of each frequency point in conjunction with the local background level; the frequency point with the highest relative prominence and meeting the preset significance condition is identified as the candidate point of the target fundamental frequency; to improve reliability, it can also be confirmed by combining the spectral line continuity, peak width, octave relationship, or consistency of repeated measurements near the target frequency). The system automatically locks the frequency corresponding to this peak value as the target fundamental frequency that needs to be strictly protected in the future. (The target fundamental frequency is locked by the preliminary noise reduction and target frequency search module in the data processing unit; the preliminary noise reduction and target frequency search module automatically determines the target fundamental frequency within the target search range based on the spectrum analysis results after coarse noise reduction, and transmits the target fundamental frequency to the guard band generation module to generate the guard band required for subsequent fine noise reduction).

[0047] Step S4: Generation and radius calculation of multi-level mask protection zone The signal processing system locks the fundamental frequency according to step (3) above. Automatically generate a list of protected frequencies containing the fundamental frequency and its associated harmonics. The specific process is as follows: The guard band generation module in the data processing unit automatically generates a list of protected frequencies. Based on the target fundamental frequency locked in step S3, this module generates a list of protected frequencies containing the target fundamental frequency and its related frequency components according to preset rules. Related frequency components may include harmonics of the target fundamental frequency, adjacent characteristic frequencies, or other frequencies to be protected determined based on the specific measurement object. The guard band generation module generates a corresponding guard band for each frequency to be protected to avoid mistakenly deleting effective signals near the target frequency during subsequent noise reduction processing.

[0048] For the list of protected frequencies For each frequency point in the signal processing system, in order to prevent the subsequent frequency domain smoothing operation from spreading the "contaminated area" (target signal energy) to the "safe area" (anchor point), the data processing unit dynamically matches and generates the mask protection radius based on the characteristics of the input data to construct the optimal protection boundary.

[0049] Step S5: Fine-grained noise reduction based on global dynamic greedy iteration Initialization is performed by the signal processing system (initialization is executed by the fine noise reduction module in the data processing unit, and the initialization process includes: using the preprocessed main channel signal as the initial residual signal; using each reference signal corresponding to the auxiliary sensor array as a set of candidate reference signals; and reading a list of protected frequencies containing the fundamental frequency and its associated harmonics automatically generated by the guard band generation module in step S4). The system sets the mask protection radius / protection zone for each frequency point; sets an upper limit for the number of iterations and convergence criteria; then, in each round of processing, the fine noise reduction module selects the reference signal that contributes the most to the current residual signal from the candidate reference signal set, estimates its corresponding background interference component, and subtracts it from the current residual signal until the stopping condition is met. A dynamic data pool containing all candidate reference signals is created, and the main channel signal is initialized to the current residual, followed by a fine iterative loop. This fine iterative loop refers to a fine noise reduction process that performs multiple rounds of background interference suppression on the main channel signal based on the determined protection frequency band. This process differs from the preliminary noise reduction in step S3, and its characteristic is that the target frequency and its related frequency components are protected during the noise reduction process to reduce the risk of the target signal being mistakenly deleted.

[0050] The system initializes a dynamic data pool containing all candidate reference signals and initializes the main channel signal to the current residual; then it enters a fine-grained iterative loop: S5.1, Full-Domain Probe: Traverse all remaining reference channels in the data pool and calculate the cross-correlation noise reduction of each channel with the current residual (the specific process is as follows: in the fine noise reduction stage, the data processing unit traverses the candidate reference signal set, estimates the interpretable background component of each candidate reference signal for the current residual signal, attempts to subtract the background component and calculates the residual change; the residual change is represented by the residual energy reduction index; during the background component estimation, trial subtraction and residual change calculation for each candidate reference signal, the data processing unit calls the frequency domain mapping processing module and performs protection processing within the target protection frequency band to avoid signal energy near the target frequency from participating in the background mapping estimation); during this calculation process, the frequency domain mapping processing module described in step S6 is forcibly called to ensure that the protected signal energy is not accidentally deleted.

[0051] S5.2 Optimal Locking: Compare the noise reduction amounts of all channels and select the channel with the maximum noise reduction amount as the "dominant interference source" for this round; use the protected mapping relationship calculated by the selected channel to construct an interference model and subtract it from the current residual; if the maximum noise reduction amount is lower than the preset convergence threshold, the iteration is automatically terminated. Specifically, the data processing unit compares the noise reduction amounts corresponding to all candidate reference signals and selects the candidate reference signal with the largest noise reduction amount as the dominant background reference signal for this round; subsequently, the data processing unit constructs the background interference estimate for this round based on the dominant background reference signal and subtracts it from the current residual signal; if the maximum noise reduction amount corresponding to all candidate reference signals is lower than the preset convergence threshold, or the number of iterations reaches the preset upper limit, it is considered that continuing processing will have limited improvement on the noise reduction effect, and the fine noise reduction module in the data processing unit automatically terminates the fine noise reduction process.

[0052] S5.3 Residual Update: Using the protected mapping relationship calculated from the selected channel, an interference model is constructed and subtracted from the current residual. Specifically, step S5.2 is used to filter the dominant background reference signal for this round and determine whether the termination condition is met; step S5.3 is used to perform residual update if the termination condition is not met.

[0053] The specific process of step S5.3 is as follows: After the dominant background reference signal of this round is determined, the frequency domain mapping processing module in the data processing unit generates a background interference estimate based on its corresponding frequency domain mapping result, and subtracts the background interference estimate from the current residual signal to obtain the updated residual signal; subsequently, the frequency domain mapping processing module in the data processing unit updates the candidate reference signal set or its participation status, and enters the next round of loop.

[0054] Step S6: The frequency domain mapping processing module performs frequency domain mask interpolation based on complex domain decomposition. This frequency domain mapping processing module can be a software function module within the data processing unit, or it can be implemented by firmware, a hardware processing unit, or a combination thereof. This module processes the mapping relationship between the auxiliary reference signal and the current residual signal in the frequency domain, and generates continuous frequency domain mapping results within the guard band using a smooth reconstruction method.

[0055] This step is the core data processing subroutine that is repeatedly called in step S5, and it is mainly implemented through the general mathematical processing module: S6.1 Power Spectrum Calculation: The cross-power spectral density and auto-power spectral density of the input signal and the reference signal are calculated using Discrete Fourier Transform (DFT). Specifically, the data processing unit converts the input and reference signals to the frequency domain and calculates the cross-power spectrum between the input and reference signals, as well as the auto-power spectrum of the reference signal itself, using conventional spectral estimation methods (e.g., periodogram method obtained from Discrete Fourier Transform or Fast Fourier Transform, Welch piecewise averaging method, Bartlett piecewise averaging method, windowed periodogram method, multi-cone window spectral estimation method, etc. In practice, the Welch piecewise averaging method is preferred to improve the stability of power spectral estimation). The cross-power spectrum characterizes the correlation between the input and reference signals, while the auto-power spectrum characterizes the energy distribution of the reference signal at various frequencies. The above power spectral calculation can be implemented using conventional Discrete Fourier Transform, piecewise averaging, or other frequency domain spectral estimation methods.

[0056] S6.2 Original Response Estimation: Based on the power spectrum results above, the basic mapping relationship of the original full-band is calculated; the specific process is as follows: based on the cross power spectrum and the self power spectrum, the data processing unit calculates the basic frequency domain mapping relationship from the reference signal to the current residual signal; this basic frequency domain mapping relationship is used to describe the corresponding performance of the background interference component in the reference signal in the current residual signal; this basic frequency domain mapping relationship can be in complex form, with its amplitude part representing the gain relationship and its phase part representing the phase delay relationship.

[0057] S6.3 Mask Region Locking: Based on the mask protection radius generated in step S4, the corresponding interval on the frequency axis is marked as the mask region, and the remaining points are marked as safety anchor points. Specifically, based on the guard band generated in step S4, the data processing unit marks the target fundamental frequency and the guard interval near its related frequencies on the frequency axis. This guard interval is the mask region, and the frequency points outside the mask region serve as a safety reference area for estimating the frequency domain mapping relationship of background interference.

[0058] S6.4, Independent Interpolation of Complex Components: As an abstract mathematical operation, the original response function is decomposed into two independent components: the real part and the imaginary part. Using data from the safety anchor point, mathematical fitting and interpolation operations are performed independently on the real and imaginary curves to reconstruct the values ​​within the mask area. Specifically, the data processing unit decomposes the basic frequency domain mapping relationship into two separately processable components and uses data from the safety reference area outside the mask area to smoothly reconstruct the mapping relationship within the mask area. The reconstructed mapping relationship remains continuous on the frequency axis and is used for subsequent background interference estimation. The purpose of this processing is to avoid directly using mapping results within the target guard band that may be affected by the target signal, thereby reducing the possibility of the target signal being accidentally deleted during noise reduction.

[0059] Step S7: Standard Physical Quantity Conversion and Output After the iteration, the result conversion output module in the data processing unit multiplies the final residual signal by the corresponding sensitivity coefficient of the device, converts it into the standard physical units required by the target system, and outputs the denoised time-domain waveform and spectrum data. The specific process is as follows: After fine denoising, the data processing unit obtains the final residual signal; the result conversion output module converts the final residual signal into the standard physical quantity required by the target system according to the device calibration coefficient, sensor sensitivity coefficient, or system conversion coefficient; the converted result can include the time-domain waveform, frequency-domain spectrum, amplitude information, phase information, comparison results before and after denoising, and data files of the target physical quantity; the aforementioned results can be output through a display interface, data interface, or storage file; each precision measurement experimental platform has its own series of physical quantity conversions because the dimensions need to be unified during data processing, but not necessarily the dimensions of the final evaluation index; the converted target physical quantity can be displacement, velocity, acceleration, force, electromagnetic signal intensity, phase change, or other physical quantities corresponding to the specific measurement object; the result conversion output module in the data processing unit finally outputs the denoised time-domain waveform, frequency-domain spectrum, and related data files.

[0060] This invention aims to solve the problems faced by existing precision measurement data processing technologies when extracting weak signals at specific frequencies, such as "signal self-eating," lack of frequency domain information, and strong model dependence. Specifically, the technical problems to be solved by this invention include: (1) When using a reference channel for multi-channel correlated noise cancellation, how to avoid the "signal self-eating" phenomenon caused by signal crosstalk in the reference channel or the high strength of the target signal itself. That is, how to establish a frequency-selective protection mechanism while using auxiliary sensor data to significantly reduce the background noise floor, so as to ensure that the energy and waveform characteristics of the target physical signal (especially narrowband or burst signals) are not mistakenly identified as noise and subtracted by the algorithm.

[0061] (2) How to accurately estimate and remove the background noise component superimposed on the target signal frequency point without directly cutting out specific frequency band data (such as notch filtering). That is, how to "fill in" the background noise model of the signal position by mining the noise correlation law on both sides of the signal frequency point and using interpolation fitting methods, so as to effectively suppress the deep background noise at the frequency point while completely preserving the frequency domain information of the target signal, and significantly improve the final detection signal-to-noise ratio.

[0062] (3) How to achieve accurate decoupling and extraction of complex multi-source environmental noise in the absence of a priori template of the target signal waveform. That is, how to construct a general preprocessing method that does not depend on the specific shape of the signal and is based only on the assumptions of statistical independence and frequency domain continuity, so as to solve the problem that traditional matched filtering technology cannot be applied to unknown signal search or waveform restoration applications.

[0063] Compared with existing technologies, the present invention provides a multi-channel weak signal denoising method and system based on complex domain frequency domain mask interpolation and global dynamic greedy iteration, offering a general weak signal extraction paradigm applicable to various complex industrial and scientific research scenarios, and possessing the following significant technical advantages: (1) Breaking through the bottleneck of "signal self-eating" in traditional coherent denoising, lossless protection of complex signal spectrum is achieved: Traditional multi-channel Wiener filtering, under the global energy minimization criterion, is prone to misinterpreting weak target signals interfering with the reference channel as coherent interference and over-attenuating them, leading to the "self-eating" of the true waveform. This invention, in the step of constructing the system mapping relationship, employs a "complex domain component-independent mask interpolation model," locking the target frequency band as a mask region and using clean data from the sidebands on both sides of the mask region to independently perform interpolation fitting on the real and imaginary parts of the mathematical model. Because this technique mathematically achieves precise decoupling of the response basis amplitude and phase, the technical solution of this invention achieves lossless removal of interference at specific frequency points under low signal-to-noise ratio without directly truncating the target frequency band or distorting the signal phase delay, thus maximizing the preservation of the intrinsic energy of the target signal.

[0064] (2) Overcome the limitations of static prior models and achieve adaptive global optimal decoupling of multi-source strong coupling interference. In complex industrial and scientific monitoring scenarios, strong multidimensional cross-coupling often exists between multi-physics interferences. Traditional static models that rely on manual experience to pre-set fixed denoising sequences are prone to error accumulation. This invention employs a "global dynamic greedy iteration based on maximum denoising gain" technique. In each iteration, the cross-correlation denoising amount of each channel and residual is calculated globally, dynamically locking the current dominant interference source and subtracting and updating it. Because this iteration mechanism has the characteristics of eliminating manual sorting bias and automatically terminating based on the marginal gain of denoising, it improves the system's generalization ability under strong background interference. Therefore, the technical solution of this invention completely eliminates the risk of ill-conditioned inversion of large covariance matrices and significantly reduces the background noise limit across the entire frequency band.

[0065] (3) Solve the time-domain distortion problem of complex broadband signals and achieve lossless restoration of the target waveform. Real target signals in complex measurement scenarios often possess complex spectral structures and are accompanied by multiple associated frequency bands. Existing technologies, if only protecting a single dominant frequency point, can lead to the signal energy of associated frequency bands being mistakenly deleted as noise, resulting in smoothing or severe distortion of the time-domain waveform edges. This invention establishes a "multi-order associated frequency point synchronous mask protection" mechanism. The system automatically generates a protection sequence containing the target fundamental frequency and its multiple associated frequencies, and synchronously performs complex domain interpolation fitting operations at all protected frequency points. Because this technology features coordinated locking and decoupling protection of the target signal's dominant frequency and higher-order associated components across the entire frequency band, the technical solution of this invention ensures that the complete spectral structure of the target signal across the entire frequency band is not destroyed, thereby enabling the distortion-free reconstruction of the real target waveform with a specific shape in the noise-reduced output.

[0066] (4) Improve the algorithm's cold start capability and system robustness, and construct a general adaptive signal extraction closed loop. Traditional band-stop filtering or masking techniques heavily rely on manual experience to blindly set fixed protection parameters. When faced with unknown initial states of the target signal or dynamic frequency shifts due to non-stationary environments, protection misalignment or rejection of the true signal is highly likely. This invention introduces a "two-stage guided adaptive frequency point positioning" mechanism. First, unprotected coarse noise reduction preprocessing exposes local energy peaks in the residual spectrum. Then, the coordinates of these dynamic peaks are extracted as a reference frequency point and automatically fed back to the second stage to guide the precise generation of the mask center position. Because this technique features a dynamic detection closed-loop characteristic of "exposure first, then fixed-point protection," it overcomes the vulnerability of static parameters to environmental drift. Therefore, the technical solution of this invention completely eliminates the heavy reliance on precise prior waveform templates and complex manual parameter tuning, achieving automated positioning and protection of unknown weak signals from strong background interference, and significantly improving the cold-start capability and robustness of the algorithm system in complex non-stationary environments.

[0067] Among them, the alternative to mapping relationship decomposition is: Besides decomposing the complex mapping model (which refers to the frequency domain mapping relationship between the auxiliary reference signal and the main channel signal or the current residual signal. This mapping relationship contains both amplitude and phase information, thus more accurately describing the transmission characteristics of background interference from the reference channel to the main channel) into real and imaginary parts, it can also be transformed to other mathematical spaces such as polar coordinates for processing. The algorithm performs mathematical interpolation on the two independent feature components at the safety anchor point, and finally reconstructs the mapping model within the mask area using the transformation formula. This scheme also achieves decoupling in the mathematical dimension, enabling the reconstruction of a continuous response basis. (The alternative here is used to...) An alternative to the mask region interpolation algorithm (because a guard band is set, the target signal region does not participate in the calculation of the transfer function, and the transfer function of this guard band is obtained by complex-domain linear interpolation of the transfer function values ​​on both sides of the guard band. The replacement method is this linear interpolation replacement method): Besides linear interpolation, missing data within the mask region can also be fitted using higher-order fitting algorithms such as cubic spline interpolation, polynomial fitting, or Gaussian process regression. By extracting anchor point data from a wider range on both sides of the mask region for curve optimization, it is possible to fit the higher-order evolution trend and complex fluctuation characteristics of background interference over a wide frequency band.

[0068] This invention is well-conceived and can be widely applied in complex measurement scenarios where the signal-to-noise ratio is extremely high and there are reference noise sources, such as gravitational wave detection, weak electromagnetic signal detection, high-precision optical interferometry, and general-purpose high-precision weak signal extraction systems.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive weak signal extraction method under strong background interference, characterized in that: The adaptive weak signal extraction method is based on a signal processing system, which relies on a precision measurement experimental platform and includes a main measurement sensor, an auxiliary sensor array, a data processing unit, a signal conditioning module, a multi-channel synchronous acquisition module, a data storage module, and a result output module. The main measurement sensor is used to acquire a mixed signal containing the target signal and environmental background interference; The auxiliary sensor array is used to acquire reference signals related to environmental background interference. It includes several sensors that are sensitive to environmental interference but theoretically insensitive to the target physical signal. Each auxiliary sensor in the auxiliary sensor array is a reference sensor that is insensitive to or only weakly responds to the target physical signal. It is arranged at a position that can reflect environmental disturbances, structural coupling, common-mode interference, or changes in the external background. Its output signal is used to characterize the reference background component in the main measurement signal. The auxiliary sensor array responds to environmental background interference but does not respond or only weakly responds to the target signal. By using the reference signals acquired by the auxiliary sensor array, the data processing unit estimates the correlation between background interference and the main measurement signal, and thereby reduces coherent background interference from the main measurement signal. The data processing unit is a computer, server, embedded processor, or other hardware device with data processing capabilities that executes data processing and noise reduction algorithms. The data processing unit is configured with a data reading and preprocessing module, a parameter generation module, a preliminary noise reduction and target frequency search module, a guard band generation module, a fine noise reduction module, a frequency domain mapping processing module, and a result conversion and output module. The data reading and preprocessing module is responsible for reading multi-channel synchronous data, step detection, effective segment extraction, detrending, and DC removal. The parameter generation module is responsible for generating frequency-resolution... The algorithm defines the following parameters: rate, search range, iteration upper limit, and convergence threshold. The initial denoising and target frequency search module performs a finite number of rounds of unprotected coarse denoising and automatically locks the target fundamental frequency. The guard band generation module generates a list of protected frequencies and a mask interval based on the fundamental frequency and its harmonics. The fine denoising module performs multiple rounds of greedy denoising under guard band constraints. The frequency domain mapping processing module calculates the complex frequency domain mapping from the reference signal to the current residual signal and interpolates and reconstructs it within the mask region. The result conversion and output module converts the final residual into the target physical quantity and outputs the time-domain and frequency-domain results. The signal conditioning module is used to amplify, filter, isolate, range match, or calibrate the output signals of each sensor. The multi-channel synchronous acquisition module is used to synchronously sample the main measurement channel and the auxiliary measurement channel; The data storage module is used to store the original data, intermediate processed data, and noise reduction results; The result output module is used to output the processed time-domain signal, frequency-domain signal, and corresponding physical quantity results. The main measuring sensor and the auxiliary sensor array are respectively connected to the signal conditioning module, the signal conditioning module is connected to the multi-channel synchronous acquisition module, the multi-channel synchronous acquisition module is connected to the data processing unit, and the data processing unit is respectively connected to the data storage module and the result output module; The adaptive weak signal extraction method mainly includes the following steps: (1) Data acquisition and quality control preprocessing The signal processing system simultaneously reads the output data of the main measurement sensor and the auxiliary sensor array through the multi-channel synchronous acquisition module. The main measurement sensor outputs a main channel signal, which includes the target signal and environmental background interference. The auxiliary sensor array outputs several auxiliary channel signals, which are used to characterize environmental background interference or reference information related to the background interference. To prevent non-stationary abrupt changes or steps in the raw time-domain data from compromising the accuracy of frequency-domain analysis, the data reading and preprocessing module performs a robust step detection algorithm on each channel. (2) Adaptive frequency domain parameter generation The data processing unit dynamically matches and generates the processing parameters required for frequency domain analysis and noise reduction based on the characteristics of the input data, so as to ensure that the frequency resolution of the mapping relationship, i.e. the frequency domain correlation between the auxiliary channel reference signal and the main channel signal or the current residual signal, can automatically adapt to different observation durations. The data processing unit estimates the background interference components in the main channel based on the mapping relationship and subtracts the estimated background interference components from the main channel signal. (3) Coarse noise reduction and automatic peak search based on unprotected greedy iteration Before entering the protected fine noise reduction, the preliminary noise reduction and target frequency search module in the data processing unit first reuses the core framework of "full-domain dynamic greedy iteration" to perform a round of "blind search" operation. Without pre-specifying the protection frequency, the preliminary noise reduction and target frequency search module performs preliminary background interference suppression on multi-channel data to make the target frequency characteristics that are covered by strong background interference more obvious, and realizes the rapid stripping of the dominant strong background interference to reveal the true frequency characteristics of the weak signal buried deep within. (4) Generation and radius calculation of multi-level mask protection zone The base frequency locked according to step (3) above The protection band generation module in the data processing unit automatically generates a list of protected frequencies containing the fundamental frequency and its associated harmonics. For the list of protected frequencies For each frequency point, the data processing unit dynamically matches and generates the mask protection radius based on the characteristics of the input data to construct the optimal protection boundary; (5) Fine-grained noise reduction based on global dynamic greedy iteration The fine noise reduction module in the data processing unit initializes a dynamic data pool containing all candidate reference signals and initializes the main channel signal as the current residual. Then, a fine iterative loop is performed, which is a fine noise reduction process that suppresses background interference in the main channel signal in multiple rounds based on the determined guard band. (6) Frequency domain mapping processing module performs frequency domain mask interpolation operation based on complex domain decomposition; (7) Standard physical quantity conversion and output After the iteration is completed, the result conversion output module in the data processing unit multiplies the final residual signal by the sensitivity coefficient corresponding to the device, converts it into the standard physical units required by the target system, and outputs the time-domain waveform and spectrum data after noise reduction.

2. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The specific process of the signal processing system performing the robust step detection algorithm for each channel in step (1) is as follows: (1.1) Signal conditioning: Low-pass filtering is performed on the original time-domain signal of each channel, and the first-order difference is calculated to highlight the abrupt change characteristics; (1.2) Statistical threshold determination: Calculate the median absolute deviation of the differential signal, set a statistical threshold, and time points exceeding the threshold are marked as "step anomalies"; (1.3) Optimal Segment Extraction: The data processing unit performs a merged scan of the step anomaly markers for all channels. When any channel is marked as an anomaly at a certain time point or time period, that time point or time period is considered unavailable; when none of the channels are marked as anomaly, that time point or time period is considered available. The data processing unit traverses the entire time series, finds the interval with the longest continuous available time, and synchronously extracts the data of all channels within that interval as the effective dataset. Subsequently, for the valid dataset Detrending and DC removal processes are performed to reduce the impact of slow drift and fixed bias on subsequent frequency domain analysis.

3. The adaptive weak signal extraction method under strong background interference as described in claim 2, characterized in that, The low-pass filtering in step (1.1) is used to suppress high-frequency glitches or acquisition noise, and the differential processing is used to highlight abrupt changes, steps, or stair anomalies in the signal; the differential processing includes differential sampling between adjacent sampling points, differential sampling between segmented averages, differential sampling between sliding windows, or other equivalent differential methods that can reflect abrupt changes between adjacent time periods. The specific process of step (1.2) is as follows: perform robust statistical analysis on the differential signal obtained by differential analysis of adjacent sampling points, or on the segmented differential result obtained by segmented mean differential analysis or sliding window differential analysis, to obtain a robust statistical quantity that can reflect the normal fluctuation range, and set an anomaly judgment threshold based on the robust statistical quantity. When the difference amplitude at a certain time point or during a certain time period exceeds the anomaly detection threshold, it is marked as a step anomaly point; the robust statistic can be the absolute deviation of the median, quantile statistic, sliding window statistic, or other statistics that are not sensitive to outliers; the anomaly detection threshold is preset or adaptively adjusted according to the noise level and stability requirements of the specific measurement system.

4. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The specific process of step (3) is as follows: (3.1) Unprotected global probing and elimination: Initialize the dynamic data pool of temporary residual signal and candidate reference channel; within a limited number of preset iterations, traverse the entire domain and calculate the noise reduction gain of all remaining reference channels for the current residual; since this stage is in a "blind search" state, the signal processing system does not enable frequency domain mask interpolation, but directly uses the original response relationship of the entire frequency band to lock the "dominant interference source" with the strongest correlation and eliminate it; (3.2) Dynamic iteration: Repeatedly execute the above steps (3.1) to continuously strip away the outermost high-energy background interference until the iteration limit is reached or the noise reduction gain converges; (3.3) Spectrum Analysis and Automatic Peak Locking: The amplitude spectral density of the temporary residual signal obtained after several rounds of preliminary background suppression in steps (3.1) and (3.2) is analyzed; the maximum signal-to-noise ratio peak on the ASD spectrum is searched within the preset wideband search range; the frequency corresponding to the peak is automatically locked by the preliminary noise reduction and target frequency search module in the data processing unit as the target fundamental frequency that needs to be strictly protected in the future. The preliminary noise reduction and target frequency search module automatically determines the target fundamental frequency within the target search range based on the spectrum analysis results after coarse noise reduction, and transmits the target fundamental frequency to the guard band generation module to generate the guard band required for subsequent fine noise reduction.

5. The adaptive weak signal extraction method under strong background interference as described in claim 4, characterized in that, The specific process of initializing the temporary residual signal and the dynamic data pool of the candidate reference channel in step (3.1) is as follows: In the initial noise reduction stage, the data processing unit takes the main channel signal after effective segment truncation, detrending, DC removal and necessary filtering / calibration as the temporary residual signal, and takes each auxiliary channel signal as the candidate reference signal set. The candidate reference signal set is selected based on channel quality, effectiveness, synchronization and noise level in order to quickly reduce the dominant background interference in the main channel, rather than to obtain the final noise reduction result; The specific process of calculating the noise reduction gain of all remaining reference channels for the current residual in step (3.1) is as follows: In each round of preliminary noise reduction, the data processing unit calculates the noise reduction contribution of each candidate reference signal to the current temporary residual signal; the noise reduction contribution is obtained by the residual energy reduction amount. The preliminary noise reduction and target frequency search module in the data processing unit selects the candidate reference signal with the largest noise reduction contribution as the dominant background reference signal for this round, and estimates the background interference component based on its correlation with the current temporary residual signal, and then subtracts it from the current temporary residual signal. In step (3.1), the full-band original response relationship refers to the frequency domain correlation between the auxiliary reference signal and the current residual signal when no guard band is set. This full-band original response relationship is used to initially estimate the manifestation of background interference in the main channel and to quickly remove the main background interference in the initial noise reduction stage. The step (3.2) of stripping the outermost high-energy background interference refers to gradually reducing the high-energy background components that can be characterized by the auxiliary reference signal from the current temporary residual signal of the main channel; and these high-energy background components are related to environmental disturbances, structural transmission, common-mode interference or external noise sources. In step (3.3), the search for the maximum signal-to-noise ratio peak on the ASD spectrum is specifically as follows: the data processing unit scans the amplitude spectral density curve within a preset wideband search range and calculates the relative prominence of each frequency point in combination with the local background level; the frequency point with the highest relative prominence and that meets the preset significance condition is determined as the target fundamental frequency candidate point.

6. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The specific process of performing the fine iterative loop in step (5) is as follows: (5.1) Full domain probing: Traverse all remaining reference channels in the data pool and calculate the cross-correlation noise reduction of each channel with the current residual; during this calculation process, the frequency domain mapping processing module is forcibly invoked to ensure that the protected signal energy is not accidentally deleted; (5.2) Optimal locking: The data processing unit compares the noise reduction amount corresponding to all candidate reference signals and selects the candidate reference signal with the largest noise reduction amount as the dominant background reference signal for this round; Subsequently, the data processing unit constructs the current background interference estimate based on the dominant background reference signal and subtracts it from the current residual signal; If the maximum noise reduction amount corresponding to all candidate reference signals is lower than the preset convergence threshold, or the number of iterations reaches the preset upper limit, it is considered that continuing to process will have limited effect on improving the noise reduction effect, and the fine noise reduction module in the data processing unit will automatically terminate the fine noise reduction process. (5.3) Residual update: After the dominant background reference signal of the current round is determined, the frequency domain mapping processing module in the data processing unit generates a background interference estimate based on its corresponding frequency domain mapping result, and subtracts the background interference estimate from the current residual signal to obtain the updated residual signal; then, the frequency domain mapping processing module in the data processing unit updates the candidate reference signal set or its participation status, and enters the next round of loop.

7. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The initialization process of the fine noise reduction module in the data processing unit in step (5) is as follows: the preprocessed main channel signal is used as the initial residual signal; the reference signals corresponding to the auxiliary sensor array are used as the candidate reference signal set; and a protected frequency list containing the fundamental frequency and its associated harmonics is automatically generated by the protection band generation module in step (4). And the mask protection radius and protection range corresponding to each frequency point; Set an upper limit on the number of iterations and convergence criteria; Subsequently, in each round of processing, the fine noise reduction module selects the reference signal that contributes the most to the current residual signal from the candidate reference signal set, estimates its corresponding background interference component, and subtracts it from the current residual signal until the stopping condition is met.

8. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The specific process of calculating the cross-correlation noise reduction of each channel with the current residual in step (5.1) is as follows: In the fine noise reduction stage, the data processing unit traverses the candidate reference signal set, estimates the interpretable background component of each candidate reference signal for the current residual signal, attempts to subtract the background component, and calculates the residual change after the background component is subtracted; the residual change is represented by the residual energy reduction index; during the process of estimating the background component, attempting to subtract the background component, and calculating the residual change for each candidate reference signal, the data processing unit calls the frequency domain mapping processing module and performs protection processing within the target protection frequency band to avoid signal energy near the target frequency from participating in the background mapping estimation.

9. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The specific process of step (6) is as follows: (6.1) Power spectrum calculation: The cross power spectral density and self power spectral density of the input signal and the reference signal are calculated using discrete Fourier transform; (6.2) Original response estimation: Based on the above power spectrum results, calculate the basic mapping relationship of the original full frequency band; (6.3) Mask area locking: Based on the mask protection radius generated in step (4), mark the corresponding interval on the frequency axis as the mask area, and mark the remaining points as safety anchor points; (6.4) Independent interpolation of complex components: As an abstract mathematical operation, the original response function is decomposed into two independent components, the real part and the imaginary part; using the data of the safety anchor point, mathematical fitting and interpolation operations are performed independently on the real part curve and the imaginary part curve respectively to reconstruct the values ​​in the mask area.

10. The adaptive weak signal extraction method under strong background interference as described in claim 9, characterized in that, The specific process of calculating the cross-power spectral density and auto-power spectral density of the input signal and the reference signal using discrete Fourier transform in step (6.1) is as follows: The data processing unit converts the input signal and the reference signal to the frequency domain, and calculates the cross-power spectrum between the input signal and the reference signal and the auto-power spectrum of the reference signal itself according to conventional spectral estimation methods; the cross-power spectrum is used to characterize the correlation between the input signal and the reference signal, and the auto-power spectrum is used to characterize the energy distribution of the reference signal at each frequency; The specific process of step (6.2) is as follows: Based on the cross power spectrum and the self power spectrum, the data processing unit calculates the basic frequency domain mapping relationship from the reference signal to the current residual signal; the basic frequency domain mapping relationship is used to describe the corresponding performance of the background interference components in the reference signal in the current residual signal; The fundamental frequency domain mapping relationship can be in complex form, where the magnitude part represents the gain relationship and the phase part represents the phase delay relationship. The specific process of step (6.3) is as follows: Based on the guard band generated in step (4), the data processing unit marks the guard interval near the target fundamental frequency and its related frequencies on the frequency axis; the guard interval is the mask area, and the frequency points outside the mask area are used as the safety reference area to estimate the frequency domain mapping relationship of the background interference. The specific implementation process of step (6.4) is as follows: the data processing unit decomposes the basic frequency domain mapping relationship into two components that can be processed separately, and uses the safe reference area data outside the mask area to smoothly reconstruct the mapping relationship within the mask area; the reconstructed mapping relationship remains continuous on the frequency axis and is used for subsequent background interference estimation.

11. The adaptive weak signal extraction method under strong background interference as described in claim 1, characterized in that, The specific process of step (7) is as follows: After the fine noise reduction is completed, the data processing unit uses the final residual signal as the target signal extraction result; the result conversion output module converts the target signal extraction result into the target physical quantity according to the sensitivity coefficient of the main measurement sensor, the gain of the amplifier circuit, the quantization coefficient of the acquisition system or the calibration coefficient of the experimental platform. The converted target physical quantity is displacement, velocity, acceleration, force, electromagnetic signal intensity, phase change, or other physical quantities corresponding to the specific measurement object; the result conversion output module in the data processing unit finally outputs the denoised time-domain waveform, frequency-domain spectrum, and related data files.

Images