A real-time suppression method of signal noise of high performance liquid chromatography detector

Abstract

The application discloses a real-time suppression method for signal noise of a high performance liquid chromatography detector, and belongs to the field of analytical chemistry. The method realizes real-time collection of a detector output signal through sliding window segmentation, evaluates noise characteristics on line by using time-frequency joint analysis, realizes signal and noise separation based on an adaptive variational mode decomposition framework, dynamically adjusts filter parameters according to signal-to-noise ratio and peak shape retention index by using a fuzzy logic controller, and reconstructs the output signal after denoising treatment by using an adaptive soft threshold function. The application can accurately maintain the chromatographic peak shape while ensuring good noise reduction effect, the processing delay is controlled within 2 sampling periods, and the application requirements of on-line quality monitoring and real-time feedback control are met.

Description

Technical Field

[0001] This application belongs to the field of analytical chemistry, and specifically relates to a real-time method for suppressing signal noise in a high-performance liquid chromatography detector.

[0002] High-performance liquid chromatography (HPLC), one of the most important separation and detection techniques in modern analytical chemistry, works by using a high-pressure pumping system to inject the mobile phase into a chromatographic column packed with a stationary phase. This allows the components in the sample to repeatedly partition between the two phases, achieving separation based on differences in retention time. The detector then converts the concentration of each component into a measurable electrical signal. This technology is widely used in pharmaceutical analysis, environmental monitoring, food safety, clinical diagnostics, and life science research, serving as a core tool for qualitative and quantitative analysis of substances. In an HPLC system, the detector, as the terminal component of the entire analytical process, plays a crucial role in converting the eluted components separated by the chromatographic column into physical signals. Its performance directly determines the detection limit, sensitivity, and quantitative accuracy of the analytical method.

[0003] Among various detection methods, such as ultraviolet-visible detectors, electrochemical detectors, fluorescence detectors, and mass spectrometry detectors, a common challenge in trace analysis applications is how to effectively extract target component information from weak analytical signals. The raw electrical signal output by the detector typically contains noise interference from multiple frequency components. This noise originates from thermal and shot noise of the internal photoelectric conversion elements of the detector, as well as pulsation noise generated by the mobile phase delivery system. Furthermore, temperature fluctuations in the column oven and detection cell can cause baseline drift. In actual analysis, these noise components often overlap with the analytical signal in the frequency domain, forming a so-called noise-signal coupling phenomenon. This makes it difficult for traditional fixed-parameter filters to achieve an ideal balance between noise reduction and signal fidelity.

[0004] In existing technologies, noise reduction of high-performance liquid chromatography (HPLC) detector output signals mainly employs offline post-processing, meaning that denoising is performed after data acquisition for the entire chromatographic run. While this method can utilize computationally complex algorithms to achieve good noise reduction, it suffers from severe lag, failing to meet the requirements of online quality monitoring and real-time feedback control. Meanwhile, traditional time-domain moving average filtering or simple low-pass filtering, although capable of real-time processing, typically requires manual pre-setting of filter parameters and cannot be adaptively adjusted. This leads to a significant decrease in denoising effectiveness when processing real-world samples with severe baseline drift or complex and variable noise characteristics. Furthermore, existing adaptive filtering methods often struggle to accurately identify the boundary between noise and effective signals when dealing with chromatographic peaks with low signal-to-noise ratios and high overlap between the analytical signal and noise interference spectra, resulting in distortion of the analytical signal peak shape and quantitative information. These technical deficiencies severely restrict the widespread application of HPLC detectors in scenarios requiring both real-time performance and accuracy, such as trace analysis, process analysis, and rapid detection. Therefore, a signal-noise suppression method that balances real-time performance with noise reduction accuracy is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time method for suppressing signal noise in high-performance liquid chromatography detectors, which can effectively solve the problems of offline processing lag, inability to adaptively adjust filter parameters, difficulty in identifying signal boundaries under low signal-to-noise ratio conditions, and baseline drift affecting analytical accuracy in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A real-time method for suppressing signal noise in a high-performance liquid chromatography detector includes the following specific steps: Step 1: Real-time acquisition and sliding window segmentation of the raw electrical signal output by the detector: The output signal of the high performance liquid chromatography detector is sampled at equal intervals with a preset sampling period to obtain a continuous signal sequence containing multiple time points, and the continuous signal sequence is slid segmented according to a preset window length to obtain several signal windows that are processed sequentially. Step 2: Online evaluation of noise characteristics based on joint time-frequency analysis: Perform short-time Fourier transform on the time-domain signal within each signal window to extract the energy distribution characteristics of the signal in the time-frequency domain, calculate the power spectral density of noise in each frequency band, and dynamically evaluate the frequency distribution characteristics and intensity change trend of noise within the current signal window. Step 3: Construct an adaptive variational mode decomposition framework for signal separation: Based on the noise frequency distribution characteristics evaluated in Step 2, adaptively determine the number of intrinsic mode components and the initial center frequency range of the variational mode decomposition, perform variational mode decomposition on the signal window, separate the original signal into multiple intrinsic mode components, and identify and extract the effective component containing the chromatographic analysis signal and the residual component containing noise interference based on the center frequency and bandwidth of each component. Step 4: Dynamically update adaptive filter parameters based on signal quality assessment: Calculate the signal-to-noise ratio estimate and peak preservation index of the effective component signal. Based on the signal-to-noise ratio estimate and peak preservation index, dynamically adjust the control parameters of the adaptive filter through a preset fuzzy logic controller, including the convergence factor and regularization coefficient of the filter, to achieve adaptive adjustment of the filter strength. Step 5 uses an adaptive soft threshold function to denoise the effective components and reconstruct the output signal: The extracted effective components are denoised using the adaptive filtering parameters updated in Step 4 to remove noise components in the overlapping spectral regions. Then, the denoised effective components are reconstructed to obtain the detector output signal after noise suppression. Step 6: Output the noise-suppressed signal and update the noise characteristic evaluation model: The denoised signal obtained in step 5 is used as the final output of the method of the present invention. At the same time, the noise evaluation results and adaptive parameter update records of this processing are fed back to step 2 to optimize the noise characteristic evaluation model parameters during subsequent signal window processing.

[0007] Preferably, in step 1, the preset sampling period is adaptively set according to the detector type and the chromatographic peak width. For conventional ultraviolet detectors, the sampling period is set to 50 milliseconds, and for fast scan detectors, the sampling period is set to 10 milliseconds. The window length is set to 20 to 50 times the sampling period, and the sliding step size is set to 1 to 5 times the sampling period.

[0008] Preferably, in step 2, the short-time Fourier transform uses the Hanning window function, the window length is one-quarter of the signal window length, the frequency resolution is set to the reciprocal of the sampling period divided by the window length, and the result of the time-frequency joint analysis is used to construct the noise power spectral density matrix, the dimension of which is the product of the number of sampling points in the signal window and the number of frequency resolution points.

[0009] Preferably, in step 3, when adaptively determining the number of intrinsic mode components, the boundary of the energy concentration region in the noise power spectral density matrix in step 2 is used as the basis for setting the initial center frequency. The number of intrinsic mode components is set to 2 to 3 times the number of the main peaks of the noise frequency distribution. The iteration termination condition of variational mode decomposition is that the relative difference between the decomposition results of two adjacent iterations is less than 0.001 or the number of iterations reaches 200.

[0010] Preferably, in step 4, the input variables of the fuzzy logic controller include the signal-to-noise ratio estimate of the effective components extracted in step 3, the peak preservation index, and the rate of change of the noise power spectral density. The output variables are the adjustment amount of the convergence factor and the adjustment amount of the regularization coefficient. The domain range of the signal-to-noise ratio estimate is 0 to 40 dB, the domain range of the peak preservation index is 0 to 1, the adjustment range of the convergence factor is 0.001 to 0.1, and the adjustment range of the regularization coefficient is 0.01 to 1.

[0011] Preferably, the threshold setting rule for the adaptive soft thresholding function in step 5 is as follows: when the signal-to-noise ratio estimate is below 15 dB, a hard thresholding strategy is adopted; when the signal-to-noise ratio estimate is between 15 dB and 25 dB, a semi-soft thresholding strategy is adopted; and when the signal-to-noise ratio estimate is above 25 dB, a soft thresholding strategy is adopted. The threshold value is dynamically calculated and determined based on the convergence factor and regularization coefficient updated in step 4.

[0012] Preferably, the method further includes step 7, baseline drift correction: after signal reconstruction in step 5, the baseline of the reconstructed signal is estimated using a linear or piecewise linear fitting method based on robust estimation. The baseline estimate is subtracted from the original denoised signal to obtain the final output signal after baseline correction. During the baseline fitting process, data points in the chromatographic peak region are automatically identified and removed to avoid peak shape distortion.

[0013] Preferably, the process also includes step 8, peak shape quality assessment and adaptive adjustment of processing strategy: the peak shape quality of the baseline-corrected signal is assessed, and the peak width, symmetry factor and separation index are calculated. When any index exceeds the preset quality threshold range, the processing strategy adjustment mechanism is automatically triggered, and the process returns to step 3 to re-perform variational mode decomposition and optimize the selection of intrinsic mode components until the peak shape quality assessment result meets the requirements.

[0014] Preferably, when the method is applied to an online quality monitoring system for high-performance liquid chromatography, the processing delay of steps 1 to 6 is controlled within two sampling cycles, meeting the time requirements of real-time feedback control, and the sampling rate of the output signal remains consistent with the original sampling rate of the input signal. Compared with the prior art, the present invention, using the above technical solution, has the following technical effects: 1. This invention uses a technical solution that combines sliding window segmentation with time-frequency joint analysis to complete the entire process from noise characteristic assessment to signal denoising and reconstruction within each signal window processing cycle. The processing delay is controlled within 2 sampling cycles, which completely solves the lag problem of existing offline post-processing methods and meets the application requirements of online quality monitoring and real-time feedback control.

[0015] 2. The adaptive variational mode decomposition framework constructed in this invention can dynamically adjust the decomposition parameters according to the real-time frequency distribution characteristics of noise. Combined with the dynamic update mechanism of filter parameters based on fuzzy logic control, it realizes the automatic adaptation of filter intensity to different noise characteristics and signal states, and solves the technical problem that the noise reduction effect of traditional fixed parameter filters is significantly reduced when the noise characteristics are complex and varied.

[0016] 3. This invention effectively solves the problems of peak distortion and quantitative information distortion in the analysis signal caused by existing adaptive filtering methods under low signal-to-noise ratio conditions through a multi-level adaptive processing strategy, including adaptive soft threshold function selection based on signal-to-noise ratio, filter parameter adjustment based on peak shape preservation, and robust estimation and correction of baseline drift. It achieves accurate peak shape preservation while ensuring good noise reduction effect, providing a reliable data foundation for subsequent qualitative and quantitative analysis. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical architecture of a real-time suppression method for high-performance liquid chromatography detector signal noise proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of signal denoising based on the combination of adaptive variational mode decomposition and fuzzy logic control in this invention; Figure 3 This is a flowchart illustrating the logic flow of the real-time acquisition and sliding window segmentation stage of the high-performance liquid chromatography detector output signal in this invention. Figure 4 This is a flowchart illustrating the logical framework of the online noise characteristic evaluation and adaptive filter parameter dynamic update mechanism based on time-frequency joint analysis in this invention. Figure 5 This is a schematic diagram illustrating the multi-level data flow and feedback adjustment relationship between the original signal, decomposition processing, threshold denoising, and reconstructed output in the entire signal noise suppression process of this invention; Figure 6 This diagram illustrates a comparison of the effectiveness of the adaptive soft threshold denoising method, which is the core of this invention, in terms of signal-to-noise ratio improvement and peak shape fidelity with existing fixed parameter filtering techniques. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] Example 1: As Figures 1 to 6As shown, the present invention proposes a real-time noise suppression method for high-performance liquid chromatography (HPLC) detector signals. The overall technical architecture includes the following core processing modules: a real-time signal acquisition and sliding window segmentation module, a time-frequency joint analysis noise characteristic evaluation module, an adaptive variational mode decomposition signal separation module, a fuzzy logic adaptive filtering parameter update module, an adaptive soft threshold denoising and signal reconstruction module, a denoised signal output and evaluation model feedback update module, a baseline drift correction module, and a peak shape quality evaluation and strategy adaptive adjustment module. The specific implementation methods of each module are described in detail below.

[0020] Step 1: Real-time Acquisition and Sliding Window Segmentation of the Raw Electrical Signal Output from the HPLC Detector. The output signal of the HPLC detector first enters the real-time signal acquisition and sliding window segmentation module. In this module, the output signal of the HPLC detector is sampled at equal intervals according to a preset sampling period. Specifically, the sampling period is adaptively set according to the detector type and chromatographic peak width: for conventional UV-Vis detectors, whose response time is typically on the order of seconds, the sampling period is set to 50 milliseconds, at which point sufficient data points can be obtained for each chromatographic peak for subsequent analysis; for fast scanning detectors such as diode array detectors or mass spectrometers, the response speed is even faster, and the sampling period can be set to 10 milliseconds to capture the peak shape details of rapidly eluting components.

[0021] Sampling is implemented using an analog-to-digital converter (ADC) with a conversion precision set to 16-bit to 24-bit to ensure quantization resolution for weak signals. During sampling, the timestamp information for each data point is recorded synchronously, with a timestamp precision better than one-hundredth of the sampling period. The acquired raw sampled data is stored in double-precision floating-point format, forming a continuous signal sequence containing multiple time points. The data structure of the continuous signal sequence includes three core fields: a timestamp field recording the absolute or relative time of each sampling point, an amplitude field recording the voltage or current value of the detector output signal, and a status field indicating whether the data point is a valid sample or a missing point due to a brief detector offline.

[0022] Continuous signal sequences are segmented using a sliding window of a preset length. The window length is set to 20 to 50 times the sampling period: when the window length is 20 times the sampling period, the processing window contains data from 20 consecutive sampling points, suitable for online monitoring scenarios with relatively stable noise characteristics and limited computational resources; when the window length is 50 times the sampling period, the processing window contains data from 50 consecutive sampling points, obtaining richer statistical characteristics for noise modeling, suitable for complex sample analysis scenarios with large noise fluctuations. The sliding step size is set to 1 to 5 times the sampling period. A smaller step size results in a larger overlap area between adjacent windows and better continuity of signal feature extraction, but the computational load increases accordingly.

[0023] The specific implementation process of sliding window segmentation is as follows: First, initialize the window start position pointer, pointing it to the first data point of the continuous signal sequence; then extract N consecutive data points starting from the start position to form the first signal window, where N is the preset window length; next, move the start position pointer forward by a step of S data points, extracting the (S+1)th to (S+N)th data points to form the second signal window; repeat the above process until the start position pointer reaches the end of the continuous signal sequence. Each extracted signal window data, along with the window's start timestamp and window number, is packaged and passed to the subsequent time-frequency joint analysis module for noise characteristic evaluation.

[0024] Step 2: Online evaluation of noise characteristics based on joint time-frequency analysis: Each signal window enters the time-frequency joint analysis noise characteristic evaluation module. The core algorithm of this module uses short-time Fourier transform to perform time-frequency characterization of the time-domain signal within the window. The Hanning window is selected as the window function for the short-time Fourier transform because its spectral characteristics include good sidelobe suppression and moderate main lobe width, achieving a good balance between frequency resolution and time resolution. The window length is set to one-quarter of the signal window length: when the signal window length is 20 times the sampling period, the window length is 5 times the sampling period; when the signal window length is 50 times the sampling period, the window length is 12.5 times the sampling period. The sliding step size of the window function is set to half the window length to achieve a smooth transition between adjacent analysis frames.

[0025] The mathematical expression for the short-time Fourier transform is: in, Given the input time-domain signal sequence, For the Hanning window function, Angular frequency, This is the time frame index. The frequency resolution is set by dividing the reciprocal of the sampling period by the window length: for a sampling period of 50 milliseconds and a window length of 1000 milliseconds, the frequency resolution is 1 Hz; for a sampling period of 10 milliseconds and a window length of 500 milliseconds, the frequency resolution is 2 Hz.

[0026] The time-frequency joint analysis module extracts the energy distribution characteristics of the signal in the time-frequency domain. Specifically, it calculates the squared amplitude of the short-time Fourier transform result in each time frame to obtain the time-frequency energy distribution matrix. The row indices of this matrix correspond to the frequency resolution points, and the column indices correspond to the time frame indices. By statistically averaging the time-frequency energy distribution matrix along the time dimension, the power spectral density of the signal can be estimated. Furthermore, the frequency ranges of energy-concentrated regions in the power spectrum are identified as the principal component frequency bands of the signal, and the regions of dispersed and persistent energy in the power spectrum are identified as noise frequency bands.

[0027] The noise power spectral density was calculated using a modified Welch method. The signal window was divided into several overlapping segments, and a periodogram was calculated for each segment. The periodograms of all segments were then averaged to reduce variance. The segment length was set to half the signal window length, and the overlap between adjacent segments was 50%. The dimension of the noise power spectral density matrix was the product of the number of sampling points within the signal window and the number of frequency-resolved points. This matrix served as a quantitative representation of the noise frequency distribution characteristics.

[0028] The evaluation of noise intensity trends employs a recursive update strategy. A historical noise power spectral density queue of fixed frame length is maintained. After a new signal window is processed, its corresponding noise power spectral density vector is appended to the end of the historical queue, and the oldest historical record at the head of the queue is removed. By comparing the current noise power spectral density with the mean and standard deviation of the noise power spectral density in the historical queue, the change in current noise intensity relative to the historical average level can be determined. If the current noise power spectral density is generally more than two standard deviations above the historical mean, it is considered that the noise intensity has significantly increased; if it is less than two standard deviations below the historical mean, it is considered that the noise intensity has significantly decreased; if it is within the range of the mean plus or minus two standard deviations, it is considered that the noise intensity is relatively stable.

[0029] Step 3: Construct an adaptive variational mode decomposition framework for signal separation: Based on the noise frequency distribution characteristics output in step 2, the signal enters the adaptive variational mode decomposition signal separation module. The core function of this module is to adaptively determine the number of intrinsic mode components and the initial center frequency range of the variational mode decomposition according to the real-time evaluated noise characteristics, and to decompose the signal window to achieve effective separation of signal and noise.

[0030] Variational mode decomposition (VMD) is a non-recursive signal decomposition method. Its basic idea is to decompose the input signal into several eigenmode components with sparse characteristics, each corresponding to a signal component within a specific frequency band. The mathematical optimization model for VMD is constructed as follows: in, For the first Each intrinsic mode component For the first The center frequencies of the intrinsic mode components For the Dirac function, This represents the convolution operation. The input signal is denoted as . The optimization variables include all intrinsic mode components and their corresponding center frequencies.

[0031] Adaptive determination of the number of intrinsic mode components The specific strategy is as follows: The boundary of the energy concentration region in the noise power spectral density matrix in step 2 is used as the basis for setting the initial center frequency. Specifically, the main energy peak and the energy valleys on both sides of the noise power spectral density curve are identified, and the frequencies corresponding to the valleys are used as the boundary points of the center frequencies of adjacent intrinsic mode components. (Number of main peaks in the noise frequency distribution) The number of intrinsic mode components is determined by a peak detection algorithm. Set the number of intrinsic mode components to 2 to 3 times the number of main peaks in the noise frequency distribution. For example, if two obvious main peaks are detected in the noise power spectral density curve, set the number of intrinsic mode components to 4 to 6; if three main peaks are detected, set the number to 6 to 9.

[0032] The initial center frequency range is determined based on the noise power spectral density matrix analysis results in step 2. First, the frequency positions corresponding to all local maxima in the noise power spectral density curve are identified; these frequency points represent the frequency bands where noise energy is most concentrated. The overall effective frequency range of the noise power spectral density curve is divided into K sub-bands, and the center frequency of each sub-band is used as the initial center frequency estimate for the corresponding intrinsic mode component.

[0033] The iterative solution of variational mode decomposition employs the alternating direction multiplier method. Each iteration includes the following sub-steps: First, for each eigenmode component... With other components and all center frequencies fixed, solve for the relationship between... The optimization subproblem, whose analytical solution is obtained by calculating the Hilbert transform in the frequency domain; secondly, for each center frequency... Fixing all intrinsic modal components, solve for the relationship between the intrinsic modal components and the eigenmode components. The optimization subproblem, which is solved by... The spectral energy centroid is used as the center frequency update value; third, a quadratic penalty term and Lagrange multipliers are introduced to enhance the numerical stability of constraint satisfaction.

[0034] The iteration termination condition for variational mode decomposition is set to one of the following two: the relative difference between the decomposition results of two adjacent iterations is less than 0.001, or the number of iterations reaches 200. The formula for calculating the relative difference is: in, For the first The result of the iteration is the first Each intrinsic mode component. When the relative difference is less than a preset threshold, the decomposition result is considered to have converged, and the iteration can be terminated early to save computation time.

[0035] Signal separation is performed based on feature identification of the individual intrinsic modal components obtained from the decomposition. Specifically, two characteristic parameters are calculated for each intrinsic modal component: center frequency and bandwidth. The center frequency is defined as the first-order origin moment of the component's spectrum, and the bandwidth is defined as the standard deviation of the component's spectrum. The effective components containing the chromatographic analysis signal are identified according to the following criteria: if the center frequency of a component is within the bandwidth of the principal component of the chromatographic signal obtained in step 2, and the bandwidth matches the theoretical bandwidth of the chromatographic peak, it is determined to be an effective component; if the center frequency of a component is within the noise bandwidth, or the bandwidth is significantly greater than the theoretical bandwidth of the chromatographic peak, it is determined to be a residual component containing noise interference. The bandwidth of the principal component of the chromatographic signal is dynamically determined by the power spectrum analysis results output in step 2, and the theoretical bandwidth is estimated based on the detector type and chromatographic conditions.

[0036] Step 4: Implement dynamic updates of adaptive filtering parameters based on signal quality assessment: The extracted effective components are fed into the adaptive filter parameter update module. The core function of this module is to calculate the quality assessment index of the effective component signal and dynamically adjust the control parameters of the adaptive filter through a fuzzy logic controller based on the quality assessment results.

[0037] The signal-to-noise ratio (SNR) estimate is calculated as the ratio of signal power to noise power. The signal power is approximated by the variance of the effective component in the peakless region (baseline region) as the noise power, and by subtracting the noise power from the overall variance of the effective component as the signal power. The SNR estimate is expressed in decibels (dB), and the calculation formula is: in, The population variance of the effective components. This is an estimate of the noise power. The domain of the signal-to-noise ratio (SNR) estimate ranges from 0 to 40 dB: when the SNR estimate is close to 0 dB, it indicates that the signal and noise power are comparable and difficult to distinguish effectively; when the SNR estimate reaches 40 dB, it indicates that the signal is much stronger than the noise, and there is more room for noise reduction processing.

[0038] The peak shape retention index is used to evaluate the degree to which denoising treatment preserves the shape of chromatographic peaks. Specifically, firstly, the position and boundaries of the chromatographic peaks are detected from the effective components, identifying characteristic points at the peak apex, leading edge, and trailing edge. Then, three original characteristic parameters—peak width, symmetry factor, and peak height—are calculated. Next, these parameters are compared with the corresponding chromatographic peak parameters identified within the original signal window in step 3, and the normalized values ​​of the parameter deviations are calculated. Finally, the normalized deviations of the three parameters are weighted and averaged to obtain the peak shape retention index. The domain of discourse for the peak shape retention index ranges from 0 to 1: when the index is close to 1, it indicates that denoising treatment has almost no effect on the peak shape; when it is close to 0, it indicates that denoising treatment has caused severe peak shape distortion.

[0039] The input variables of the fuzzy logic controller include three: the signal-to-noise ratio estimate of the effective components extracted in step 3, the peak preservation index, and the rate of change of the noise power spectral density. The rate of change of the noise power spectral density is defined as the absolute value of the relative deviation between the noise power spectral density vector of the current processing window and the historical mean vector, and is calculated using the following formula: in, This is the noise power spectral density vector of the current processing window. It is the average vector of historical noise power spectral density.

[0040] The fuzzy logic controller comprises three processing stages: input fuzzification, fuzzy rule inference, and output defuzzification. Input fuzzification maps precise input variable values ​​to membership degrees corresponding to fuzzy linguistic variables. The membership function for the signal-to-noise ratio (SNR) estimate is defined as follows: a Gaussian membership function is used for low SNR (0 to 10 dB), with the membership peak at 0 dB; a triangular membership function is used for medium SNR (10 to 30 dB), covering the entire range; and a Gaussian membership function is used for high SNR (30 to 40 dB), with the membership peak at 40 dB. The membership function for the kurtosis retention index is defined as follows: a Gaussian membership function is used for low retention (0 to 0.4), a triangular membership function for medium retention (0.4 to 0.7), and a Gaussian membership function for high retention (0.7 to 1). The membership function of the noise power spectral density change rate is defined as follows: small changes (0 to 0.1) use a Gaussian membership function, medium changes (0.1 to 0.3) use a triangular membership function, and large changes (0.3 to 1) use a Gaussian membership function.

[0041] The fuzzy rule-based inference is built upon expert knowledge and includes the following typical rules: When the input has a low signal-to-noise ratio (SNR), low peak preservation, and large noise variation, the output significantly increases the convergence factor and regularization coefficient to enhance denoising strength; when the input has a high SNR, high peak preservation, and small noise variation, the output appropriately decreases the convergence factor and regularization coefficient to retain more signal details; when the input has a moderate SNR and moderate peak preservation, the output slightly adjusts the convergence factor and regularization coefficient to maintain the current processing effect. The number of fuzzy rules is set to 27, covering all combinations of the three input variables.

[0042] Output deblurring employs the centroid method, converting the fuzzy output into precise convergence factor and regularization coefficient adjustments. The convergence factor is adjusted from 0.001 to 0.1, and the regularization coefficient is adjusted from 0.01 to 1. The updated convergence factor and regularization coefficient replace the original parameter values ​​and are used in the subsequent step 5 for calculating the adaptive soft threshold function.

[0043] Step 5: Use an adaptive soft thresholding function to denoise the effective components and reconstruct the output signal. The updated adaptive filtering parameters from step 4 are passed to the adaptive soft thresholding denoising module. This module applies adaptive soft thresholding to the extracted effective components to remove noise components in the overlapping spectral regions, and then reconstructs the signal from the denoised effective components.

[0044] The threshold setting rule of the adaptive soft thresholding function is dynamically selected based on the signal-to-noise ratio (SNR) estimate. The specific strategy is as follows: When the SNR estimate is below 15 dB, the noise and signal spectra highly overlap. Using a hard thresholding strategy can maximize the preservation of signal abrupt changes and avoid peak blurring caused by excessive smoothing. The formula for calculating the hard threshold is... in Here, N is the noise standard deviation estimate, and N is the signal length. When the signal-to-noise ratio estimate is between 15 dB and 25 dB, a semi-soft thresholding strategy is used to balance denoising intensity and signal fidelity. The semi-soft thresholding requires setting an upper threshold. and lower threshold And satisfy Threshold processing function in and The transition is linear; when the signal-to-noise ratio estimate is higher than 25 dB, a soft thresholding strategy is adopted to obtain a smooth denoising effect. The calculation formula of the soft threshold introduces a regularization coefficient correction term on the basis of the hard threshold.

[0045] The threshold value is dynamically calculated and determined based on the convergence factor and regularization coefficient updated in step 4. The comprehensive formula for threshold calculation is: in, The convergence factor is The regularization coefficient is . This is an adjustment factor based on the rate of change of noise power spectral density. When the noise intensity increases, A positive value is used to increase the threshold, thereby enhancing the noise reduction effect; when the noise intensity decreases... Negative values ​​are used to lower the threshold, thereby reducing signal loss.

[0046] The soft-threshold denoising process is specifically applied to the variational mode decomposition coefficients of the effective components. For each effective component obtained from the decomposition... First, the wavelet transform coefficients or equivalent frequency domain representation are calculated, and then the coefficients are thresholded: for a hard thresholding strategy, coefficients with absolute values ​​less than the threshold are set to zero, and coefficients with absolute values ​​greater than or equal to the threshold remain unchanged; for a soft thresholding strategy, coefficients with absolute values ​​less than the threshold are set to zero, and coefficients with absolute values ​​greater than or equal to the threshold are shrunk towards zero by the amount of the threshold; for a semi-soft thresholding strategy, linear interpolation is performed between the two thresholds.

[0047] The denoised effective components, along with the residual components identified as noise in step 3, are used for signal reconstruction. The reconstruction process employs the inverse transform of variational mode decomposition, summing the processed intrinsic mode components in the frequency domain to obtain the spectrum of the reconstructed signal, and then converting it back to the time domain using an inverse Fourier transform to obtain the reconstructed signal. The reconstructed signal is then used as the output of step 5 and passed to step 6 for final output and model update.

[0048] Step 6: Output the noise-suppressed signal and update the noise characteristic evaluation model: The denoised signal obtained in step 5 is used as the final output of the method of this invention. Simultaneously, the noise evaluation results and adaptive parameter update records from this process are fed back to step 2 to optimize the noise characteristic evaluation model parameters during subsequent signal window processing.

[0049] The output format of the denoised signal is consistent with the original sampled data format in step 1, including three fields: timestamp, amplitude, and status. The sampling rate of the output signal is consistent with the original sampling rate of the input signal, ensuring that the output signal can be directly used for subsequent applications such as chromatographic peak identification, quantitative analysis, or online quality monitoring without additional resampling processing.

[0050] The noise characteristic assessment model is updated using a recursive learning strategy. Specifically, a noise characteristic parameter database is maintained, recording the noise power spectral density vector, noise intensity change rate, and corresponding adaptive filtering parameter configurations for the most recently processed batch of signal windows (e.g., the last 100 windows). After a new window is processed, the noise assessment result for that window is appended to the database, and the long-term mean and variance estimates of the noise power spectral density are updated using an exponentially weighted moving average method. The update weight coefficient is set to 0.05, meaning each new sample contributes five percent to the long-term estimate. The updated long-term estimate serves as prior knowledge of noise characteristics for subsequent signal window processing, improving the accuracy of noise assessment.

[0051] Historical records of adaptive filtering parameters are also included in the database. These records include the signal-to-noise ratio estimate, peak preservation index, noise power spectral density change rate, and the adjustment amounts of the convergence factor and regularization coefficients of the fuzzy logic controller output for each processing step. Statistical analysis of the relationship between historical parameter configurations and processing results can identify the optimal parameter configuration patterns for different noise characteristics. When a certain combination of noise characteristics is detected to occur repeatedly, the membership function parameters and rule weights of the fuzzy logic controller can be fine-tuned to further improve the adaptive control effect.

[0052] Step 7: Baseline drift correction: After signal reconstruction in step 5, a baseline drift correction module is used to perform baseline estimation and correction on the reconstructed signal. This module uses a robust estimation-based linear or piecewise linear fitting method to estimate the baseline of the reconstructed signal, and subtracts the baseline estimate from the original denoised signal to obtain the final output signal after baseline correction.

[0053] The robust fitting method for baseline estimation employs iterative reweighted least squares. Specifically, the initial estimated baseline for the reconstructed signal is first set to the global length mean of the signal; then, the residual from each data point to the current baseline is calculated; next, the weight of each data point is calculated based on the residual, with data points having larger residuals assigned smaller weights, typically using the bisquare weighting function; subsequently, the baseline is refitted based on these weights; the above process is repeated until the change in baseline parameters is less than a preset convergence threshold or the number of iterations reaches a preset upper limit.

[0054] During baseline fitting, data points in chromatographic peak regions need to be automatically identified and removed to avoid peak shape distortion. The identification of chromatographic peak regions employs the following strategy: First, the first derivative of the signal is calculated; locations where the absolute value of the derivative significantly deviates from zero are identified as candidate peak boundaries. Then, the second derivative of the signal is calculated; locations where the second derivative changes from positive to negative correspond to inflection points at the leading edge of the peak, and locations where the second derivative changes from negative to positive correspond to inflection points at the trailing edge of the peak. The precise boundaries of the chromatographic peak regions are determined by combining the information from the first and second derivatives. Finally, data points falling into the chromatographic peak regions are removed from baseline fitting, and only data points from the baseline region are used for fitting.

[0055] The baseline-corrected signal peak shape retains the position, width, and height information of the chromatographic peaks in the original signal, eliminating only slowly changing baseline drift components. The output of the baseline drift correction module is an optional final output. This module can be enabled when the user is focused on accurate peak area quantitative analysis or needs to eliminate the influence of long-term drift on the analytical results; when the user is focused on rapid response and real-time monitoring, this module can be skipped and the output of step 5 can be used directly.

[0056] Step 8: Peak shape quality assessment and adaptive adjustment of processing strategies The baseline-corrected signal is evaluated for peak shape quality by calculating peak width, symmetry factor and separation index. When any index exceeds the preset quality threshold range, the processing strategy adjustment mechanism is automatically triggered, returning to step 3 to re-perform variational mode decomposition and optimize the selection of intrinsic mode components until the peak shape quality evaluation result meets the requirements.

[0057] Peak width is calculated using the half-peak width definition, which is the peak width at 50% of the chromatographic peak height. The preset peak width quality threshold range is determined based on chromatographic conditions and detector type: for routine analysis, the peak width threshold range is set to 80% to 120% of the target peak width; for rapid analysis, the peak width threshold range can be appropriately widened to 70% to 130% of the target peak width.

[0058] The symmetry factor index was calculated using the method specified in the pharmacopoeia. The symmetry factor was defined as the ratio of the peak leading edge width to the peak trailing edge width, with the width measured at 10% of the peak height. The preset symmetry factor quality threshold range was 0.8 to 1.5: a symmetry factor close to 1 indicated a perfectly symmetrical peak; a symmetry factor less than 0.8 indicated a forward-protruding peak (steep leading edge, trailing edge); and a symmetry factor greater than 1.5 indicated a trailing peak (tailing leading edge, steep trailing edge).

[0059] The resolution index is used to evaluate the separation effect between adjacent chromatographic peaks. The formula for resolution is the difference in retention time between two adjacent peaks divided by half the sum of their peak widths. The preset resolution quality threshold range is 1.5 to 3.0: when the resolution is greater than 3.0, it indicates that the two peaks are completely separated; when the resolution is between 1.5 and 3.0, it indicates that the two peaks are partially separated but still overlap; when the resolution is less than 1.5, it indicates that the two peaks are severely overlapped and cannot be effectively distinguished.

[0060] When any of the above indicators exceeds the preset threshold range, the system automatically triggers the processing strategy adjustment mechanism. The specific process of the adjustment mechanism is as follows: First, record the key parameters of the current processing strategy, including the number of intrinsic mode components in variational mode decomposition, the center frequency range setting, and the threshold strategy selection for adaptive soft thresholding; then, determine the adjustment direction based on the type of quality assessment failure: if the peak width exceeds the standard and the actual peak width is greater than the target range, appropriately increase the number of intrinsic mode components to improve frequency resolution; if the symmetry factor exceeds the standard and the value is too large, adjust the initial setting of the center frequency to better match the principal components of the signal; if the separation degree does not meet the standard, re-evaluate the division of the noise frequency band to reduce excessive suppression of the effective signal; finally, return to step 3, and re-perform variational mode decomposition and signal separation according to the adjusted parameters. Repeat the above iterative process until all quality assessment indicators meet the requirements or the number of iterations reaches the preset upper limit.

[0061] Specific Implementation of the Method in an Online Quality Monitoring System for High-Performance Liquid Chromatography When the method of this invention is applied to an online quality monitoring system for high performance liquid chromatography, the processing delay of steps 1 to 6 is controlled within 2 sampling cycles, which meets the time requirements of real-time feedback control, and the sampling rate of the output signal is consistent with the original sampling rate of the input signal.

[0062] The hardware architecture of the online quality monitoring system includes the following core components: an analog-to-digital converter (ADC) is responsible for converting the analog output of the detector into a digital signal. The sampling rate of the converter is set according to the detector type and chromatographic analysis requirements, and the conversion accuracy is not less than 16 bits to ensure the quantization resolution of weak signals; a digital signal processor (DSP) or field-programmable gate array (FPGA) is responsible for executing the core algorithms of steps 1 to 6. The processor's clock frequency is selected according to the processing delay requirements. When the processing delay requirement is 2 sampling cycles, the processor must complete the entire process of processing the signal window of a conventional ultraviolet detector within 100 milliseconds; the memory module adopts a double-buffered structure, with one buffer used to receive real-time sampled data and the other buffer used to store the signal window to be processed, realizing the parallel execution of data acquisition and signal processing; the communication interface adopts Ethernet or serial communication protocols to transmit the denoised signal to the host computer monitoring system in real time.

[0063] The software architecture employs a pipelined parallel processing model, organizing data acquisition, signal processing, and result output into three parallel processing stages. The data acquisition stage continuously reads raw sampled data from the analog-to-digital converter and writes it to the acquisition buffer. The signal processing stage extracts the complete signal window from the acquisition buffer, executes all processing steps 1 to 6, and writes the processing result to the output buffer. The result output stage reads the denoised signal from the output buffer and sends it to the host computer via the communication interface. When the acquisition buffer is full, it automatically switches to another buffer to continue writing, while the signal processing stage switches to the newly filled buffer for reading and processing. This dual-buffering alternation mechanism ensures the decoupling of the timing of data acquisition and signal processing, avoiding the accumulation of processing delays caused by mutual waiting.

[0064] Specific application example: Taking the detection of the content of multiple active ingredients in a pharmaceutical preparation using a conventional ultraviolet detector as an application scenario, the actual processing flow of the method of the present invention is described in detail.

[0065] The high-performance liquid chromatography (HPLC) system was equipped with a conventional ultraviolet detector, a wavelength set at 254 nm, a mobile phase flow rate set at 1.0 mL / min, and a column temperature set at 30 °C. The sample contained three active pharmaceutical ingredients, and the theoretical peak width of the target chromatographic peak was approximately 30 seconds. Based on the detector response characteristics and peak width, the preset sampling period in step 1 was set to 50 ms, the window length was set to 40 times the sampling period (2 seconds), and the sliding step size was set to twice the sampling period (100 ms). The sampled data were converted from analog to digital with 24-bit precision to form a continuous signal sequence.

[0066] Suppose that during a certain chromatographic run, the system begins to detect the elution signal of the first active ingredient at the second minute. The chromatographic peak signal corresponding to this active ingredient is a positive peak lasting approximately 30 seconds in the time domain, and a principal component signal with energy concentrated in the 0.05 Hz to 0.5 Hz frequency band in the frequency domain, while also being superimposed with high-frequency thermal noise and low-frequency baseline drift.

[0067] Step 2 involves performing a time-frequency joint analysis of the signal window corresponding to that moment. A Hanning window is used for short-time Fourier transform, with the window length set to one-quarter of the signal window length, i.e., 500 milliseconds, and the frequency resolution set to 2 Hz. The analysis results show that there is a significant energy concentration region in the 0.1 Hz to 0.3 Hz frequency band, which corresponds to the main component frequency band of the chromatographic peak signal; there is a slowly changing low-frequency component in the 0 Hz to 0.05 Hz frequency band, which corresponds to the baseline drift component; and there is a uniformly distributed low-amplitude component in the 0.5 Hz to 10 Hz frequency band, which corresponds to high-frequency electronic noise.

[0068] Step 3 involves adaptive variational mode decomposition based on the noise characteristic assessment results. Two main peak regions are identified in the noise power spectral density curve, and the number of intrinsic mode components is set to five, with initial center frequencies of 0.05 Hz, 0.15 Hz, 0.3 Hz, 1.5 Hz, and 5.0 Hz, corresponding to low-frequency drift components, the main component frequency band of the chromatographic peak and its higher harmonics, and high-frequency noise components, respectively. The variational mode decomposition converges after 15 iterations, with a relative difference of 0.0008. The decomposition results identify the two components with center frequencies of 0.15 Hz and 0.3 Hz as effective components, the component with a center frequency of 0.05 Hz as baseline drift residual, and the two components with center frequencies of 1.5 Hz and 5.0 Hz as high-frequency noise residual.

[0069] Step 4 calculates the quality assessment index of the effective components. The signal power is estimated to be 0.85 volts squared, the noise power is estimated to be 0.02 volts squared, and the signal-to-noise ratio is estimated to be 16.3 dB. The peak shape preservation index is calculated to be 0.92, indicating that the current decomposition result maintains the peak shape well. The noise power spectral density change rate is calculated to be 0.08, indicating that the current noise intensity is basically stable relative to the previous processing window. Based on the above inputs, the fuzzy logic controller determines that the current processing state is medium signal-to-noise ratio, good peak shape preservation, and small noise change. Accordingly, the output convergence factor adjustment is 0.005, the regularization coefficient adjustment is 0.1, the updated convergence factor is 0.015, and the regularization coefficient is 0.35.

[0070] Step 5 performs adaptive soft-threshold denoising based on the updated parameters. With a signal-to-noise ratio estimate of 16.3 dB, falling between 15 and 25 dB, a semi-soft thresholding strategy is selected for denoising. The lower threshold is set to 0.12, and the upper threshold to 0.25. The thresholding function performs linear interpolation between the two thresholds. The effective component after thresholding, along with baseline drift remnants and high-frequency noise remnants, is used to reconstruct the signal, obtaining the denoised detector output signal.

[0071] Step 6 outputs the denoised signal and updates the noise assessment model. The signal-to-noise ratio of the denoised signal is improved to 26.5 dB, the peak shape remains intact, the peak width is 31 seconds, the symmetry factor is 1.05, and the separation is 2.8. All indicators meet the quality requirements. The noise assessment results and parameter configurations of the current window are appended to the historical database for reference in subsequent window processing.

[0072] In the third and fourth minutes, the system detected elution signals for the second and third active ingredients, respectively. The method of this invention followed the same procedure as described above. During this period, at the fourth minute, a slight baseline drift occurred due to a brief fluctuation in the column oven temperature. Step 2 detected that the noise power in the 0.02 Hz to 0.08 Hz frequency band increased by approximately 25% compared to the historical average, with a noise change rate reaching 0.25. Step 4: The fuzzy logic controller identified the significantly enhanced noise intensity, adjusting the convergence factor to 0.08 and the regularization coefficient to 0.5, significantly enhancing the denoising strength. Step 7: The baseline drift correction module identified the drift region and performed effective correction. Ultimately, the signal-to-noise ratio of the detection results for all three active ingredients reached over 25 dB, and the peak shape quality assessment all met the standards, satisfying the application requirements for online quality monitoring.

[0073] Example 2 This example uses a fast scanning diode array detector combined with an ultra-high performance liquid chromatography system as an application scenario to describe in detail the implementation of the method of the present invention under higher sampling rate requirements.

[0074] The mobile phase flow rate of the ultra-high performance liquid chromatography (UHPLC) system was increased to 5.0 mL / min, and the column packing particle size was reduced to 1.7 μm, resulting in a significant reduction in the retention time of components and a peak width of approximately 8 seconds. To accurately capture the peak shape details of rapidly eluting components, the preset sampling period in step 1 was adjusted to 10 ms, the window length was set to 30 times the sampling period (300 ms), and the sliding step size was set to 3 times the sampling period (30 ms). The analog-to-digital converter (ADC) sampling rate was set to 100 kHz to meet the accuracy requirements of the short sampling period, and the conversion accuracy was 24 bits.

[0075] The noise characteristics of the fast-scan detector differ significantly from those of a conventional ultraviolet detector. Weak optical crosstalk exists between the array sensor units within the detector, resulting in periodic, weak modulation components in the noise spectrum. These modulation frequencies are related to the array scanning frequency and are approximately 0.8 Hz or integer multiples thereof. The time-frequency joint analysis in step 2 requires accurate identification of these periodic modulation components for targeted separation in step 3.

[0076] The parameters for variational mode decomposition were adjusted accordingly: the number of intrinsic modal components was set to 6 based on the number of noise peaks; the initial center frequency covered the low-frequency drift band of 0.01 Hz, the chromatographic peak main component band of 0.1 Hz to 0.5 Hz, the array scan modulation band of 0.8 Hz, and the high-frequency noise band above 2.0 Hz. The relative difference threshold in the iteration termination condition was kept at 0.001, and the maximum number of iterations was set to 200. In actual operation, convergence usually occurred after 50 to 80 iterations.

[0077] The domain of discourse for the fuzzy logic controller is adapted and adjusted: the domain of discourse for the signal-to-noise ratio estimate is adjusted to 0 to 50 dB to cover the higher signal-to-noise ratio that the fast detector may achieve, the domain of discourse for the peak preservation index remains unchanged, and the convergence factor is adjusted to 0.002 to 0.2 to adapt to the dynamic response requirements at higher sampling rates.

[0078] The baseline drift correction module in this embodiment plays a more crucial role. While ultra-high performance liquid chromatography (UHPLC) significantly reduces analysis time, rapid changes in column pressure still introduce additional baseline drift components, typically within the range of 0.02 Hz to 0.1 Hz. The robust fitting algorithm in step 7 adjusts the baseline region identification threshold to 30% of the derivative change to accommodate the steeper peak boundaries under rapid elution conditions.

[0079] The threshold ranges for peak shape quality assessment are also tightened accordingly: the peak width threshold range is set to 90% to 110% of the target peak width, the symmetry factor threshold range is set to 0.9 to 1.2, and the resolution threshold range is set to 2.0 to 4.0. When the quality assessment triggers a strategy adjustment, the maximum number of iterations to return to step 3 for re-decomposition is set to 5. If the target is still not met after 5 iterations, a warning message is output to prompt the operator to check the status of the chromatographic system.

[0080] This embodiment is applied to the screening of multiple pesticide residues in complex systems of traditional Chinese medicine. After separation by ultra-high performance liquid chromatography (UHPLC), 12 chromatographic peaks of the target pesticides appeared within 2.5 minutes, with peak widths ranging from 6 to 12 seconds and a minimum peak interval of 5 seconds. The method of this invention processes this complex chromatogram, and the peak shape quality assessment of all 12 chromatographic peaks passes on the first attempt. The signal-to-noise ratio is improved to 28 to 35 dB, and the resolution is greater than 2.5 for all peaks, meeting the qualitative and quantitative requirements for multiple residue screening. The actual processing delay is 18 milliseconds, approximately 1.8 sampling cycles, meeting the response speed requirements for real-time monitoring. Example

[0081] This embodiment uses a mass spectrometry detector combined with a liquid chromatography-mass spectrometry system as an application scenario to provide a detailed description of the application of the method of the present invention in trace analysis at extremely low concentrations.

[0082] The mass spectrometer detector in a liquid chromatography-mass spectrometry (LC-MS) system employs an electrospray ionization source and a triple quadrupole mass analyzer. The signal output by the mass spectrometer detector is an ion current intensity signal in selective ion monitoring mode, characterized by extremely weak signal amplitude, typically only one-thousandth to one-hundredth of the detector's range, while the noise level is relatively high, resulting in a signal-to-noise ratio often below 10 dB. This places higher demands on the performance of noise suppression methods.

[0083] In step 1, the sampling period setting needs to balance time resolution and signal-to-noise ratio requirements. For selective ion monitoring mode of triple quadrupole mass spectrometry, the scan time for a single ion is typically set to 50 to 100 milliseconds, and the sampling period is set to 10 milliseconds to obtain sufficient signal smoothing data points. The window length is set to 50 times the sampling period, i.e., 500 milliseconds, and the sliding step size is set to 1 times the sampling period, i.e., 10 milliseconds.

[0084] Step 2, the joint time-frequency analysis, employs an improved spectral estimation method to adapt to noise characteristic assessment under extremely low signal-to-noise ratio (SNR) conditions. Due to the extremely low SNR, the traditional periodogram method struggles to accurately estimate the noise power spectral density. This embodiment uses a multi-window spectral estimation method, averaging the spectral estimation results from multiple orthogonal window functions, effectively reducing the variance of the spectral estimation and improving the reliability of the noise characteristic assessment. The frequency resolution is set to the reciprocal of the sampling period divided by twice the window length; that is, for a sampling period of 10 milliseconds and a window length of 500 milliseconds, the frequency resolution is 1 Hz.

[0085] The parameters of variational mode decomposition are optimized in a targeted manner: the number of intrinsic mode components is set to three times the number of main peaks in the noise frequency distribution to more finely separate the signal from the noise; the determination of the initial center frequency introduces an adaptive learning mechanism of noise power spectral density, and a prior model of the noise spectrum is established by analyzing the pure noise signal segment in the early stage of processing. Then, the prior model is fused with the real-time analysis results to improve the accuracy of the center frequency setting; the relative difference threshold in the iteration termination condition is set to 0.0005 to improve the decomposition accuracy.

[0086] The fuzzy logic controller in step 4 is specially designed for extremely low signal-to-noise ratio (SNR) conditions. The domain of discourse for the SNR estimate is adjusted to negative values ​​up to 30 dB: when the actual SNR is negative dB, it indicates that the noise power is greater than the signal power. The adjustment strategy for the fuzzy rules is as follows: when the SNR is below 10 dB, regardless of other input conditions, the convergence factor and regularization coefficient are significantly increased to implement strong denoising; when the SNR is between 10 dB and 20 dB, fine-tuning is performed based on peak preservation and noise change rate; when the SNR is above 20 dB, the denoising intensity is appropriately reduced to retain more details of weak signals.

[0087] Step 5, the adaptive soft thresholding process, employs a hard thresholding strategy as the primary denoising method under extremely low signal-to-noise ratio conditions. A conservative threshold setting strategy is adopted, incorporating a safety factor of 1.5 into the threshold calculation formula to prevent excessive signal loss due to exceeding the threshold. The thresholding range is limited to the effective components whose center frequencies obtained from variational mode decomposition fall within the main component frequency band of the chromatographic peak. Components whose center frequencies deviate from the main component frequency band are directly identified as noise and completely removed.

[0088] Step 7, baseline drift correction, is particularly important in this embodiment. Fluctuations in the ion source temperature and minute changes in vacuum level of the mass spectrometer detector can introduce low-frequency baseline drift, with drift periods potentially lasting several minutes. Baseline estimation employs a piecewise linear fitting method, dividing the signal window into several sub-intervals. Each sub-interval is independently linearly fitted, and its length is dynamically adjusted based on signal characteristics. Peak region identification uses the second-derivative zero-crossing detection method, which is more robust than the first-derivative method under low signal-to-noise ratio conditions.

[0089] This embodiment is applied to the quantitative analysis of trace drug metabolites in blood samples. The concentration of the metabolite to be measured is only 10 nanograms per milliliter, and the amplitude of the ion current signal generated in selective ion monitoring mode is only 0.5% of the full scale. The signal-to-noise ratio of the raw acquired signal is approximately 5 dB, and the noise is mainly composed of shot noise from the mass spectrometry ion source and thermal noise from the electron multiplier, exhibiting an approximately Gaussian distribution.

[0090] The method of this invention processes this trace signal. First, step 2 accurately identifies the frequency distribution characteristics of the noise through multi-window spectral estimation. The noise power spectral density exhibits an approximately uniform distribution in the range of 0 to 20 Hz, and the ion current signal energy of the chromatographic peak is concentrated in the frequency band of 0.2 Hz to 1.5 Hz. Step 3 decomposes the signal into 8 intrinsic mode components, accurately identifying the two components with center frequencies of 0.5 Hz and 1.0 Hz as effective components. Step 4 evaluates and obtains a signal-to-noise ratio estimate of 5.2 dB and a peak shape retention index of 0.85. Based on this, the fuzzy logic controller outputs a strong denoising configuration with a convergence factor of 0.08 and a regularization coefficient of 0.6. Step 5's hard thresholding completely removes components with center frequencies higher than 2 Hz, and applies a hard threshold of 0.08 to the two effective components at 0.5 Hz and 1.0 Hz. Step 7's baseline correction eliminates slowly changing drift components.

[0091] The signal-to-noise ratio of the processed ion current signal was improved to 28 dB, the peak shape remained intact, the peak area was 92% of the original integral value, and the relative standard deviation of the quantitative analysis was 2.8%, which met the methodological requirements for trace analysis of biological samples.

[0092] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above, which are merely preferred embodiments. The scope of the present invention is defined by the claims rather than by the specific embodiments. Any modifications, equivalent substitutions, etc., made within the spirit and scope defined by the claims are within the protection scope of the present invention.

Claims

1. A method for real-time suppression of signal noise in a high-performance liquid chromatography detector, characterized in that, The specific steps include the following: Step 1: Real-time acquisition and sliding window segmentation of the raw electrical signal output by the detector: The output signal of the high performance liquid chromatography detector is sampled at equal intervals with a preset sampling period to obtain a continuous signal sequence containing multiple time points, and the continuous signal sequence is slid segmented according to a preset window length to obtain several signal windows that are processed sequentially. Step 2, online evaluation of noise characteristics based on joint time-frequency analysis: perform short-time Fourier transform on the time-domain signal in each signal window, extract the energy distribution characteristics of the signal in the time-frequency domain, calculate the power spectral density of noise in each frequency band, and dynamically evaluate the frequency distribution characteristics and intensity change trend of noise in the current signal window accordingly. Step 3: Construct an adaptive variational mode decomposition framework for signal separation: Based on the noise frequency distribution characteristics evaluated in Step 2, adaptively determine the number of intrinsic mode components and the initial center frequency range of variational mode decomposition, perform variational mode decomposition on the signal window, separate the original signal into multiple intrinsic mode components, and identify and extract the effective component containing the chromatographic analysis signal and the residual component containing noise interference based on the center frequency and bandwidth of each component. Step 4, Dynamic update of adaptive filter parameters based on signal quality assessment: Calculate the signal-to-noise ratio estimate and peak preservation index of the effective component signal. Based on the signal-to-noise ratio estimate and peak preservation index, dynamically adjust the control parameters of the adaptive filter through a preset fuzzy logic controller, including the convergence factor and regularization coefficient of the filter, to achieve adaptive adjustment of the filter strength. Step 5: Use an adaptive soft threshold function to denoise the effective components and reconstruct the output signal: Apply the updated adaptive filtering parameters from Step 4 to the extracted effective components for soft threshold denoising to remove noise components in the overlapping spectral regions. Then, reconstruct the signal from the denoised effective components to obtain the noise-suppressed detector output signal. Step 6: Output the noise-suppressed signal and update the noise characteristic evaluation model: The denoised signal obtained in step 5 is used as the final output of the method of the present invention. At the same time, the noise evaluation results and adaptive parameter update records of this processing are fed back to step 2 to optimize the noise characteristic evaluation model parameters during subsequent signal window processing.

2. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 1, characterized in that, In step 1, the preset sampling period is adaptively set according to the detector type and chromatographic peak width, the window length is set to twenty to fifty times the sampling period, and the sliding step size is set to one to five times the sampling period.

3. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 1, characterized in that, In step 2, a short-time Fourier transform is performed on each signal window using the Hanning window function. The window length is one-quarter of the signal window length. The results of the time-frequency joint analysis are used to construct the noise power spectral density matrix. The dimension of this matrix is ​​the product of the number of sampling points in the signal window and the number of frequency resolution points. The mathematical expression for the short-time Fourier transform is: in, Given the input time-domain signal sequence, For the Hanning window function, Angular frequency, For time frame index.

4. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 1, characterized in that, When adaptively determining the number of intrinsic mode components in step 3, the boundary of the energy concentration region in the noise power spectral density matrix in step 2 is used as the basis for setting the initial center frequency. The number of intrinsic mode components is set to two to three times the number of the main peaks of the noise frequency distribution. The iteration termination condition of variational mode decomposition is that the relative difference between the decomposition results of two adjacent iterations is less than a preset threshold or the number of iterations reaches a preset upper limit. Variational mode decomposition (VMD) is a non-recursive signal decomposition method. Its basic idea is to decompose the input signal into several eigenmode components with sparse characteristics, each corresponding to a signal component within a specific frequency band. The mathematical optimization model for VMD is constructed as follows: in, For the first Each intrinsic mode component For the first The center frequencies of the intrinsic mode components For the Dirac function, This represents the convolution operation. This is the input signal.

5. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 1, characterized in that, In step 4, the input variables of the fuzzy logic controller include the signal-to-noise ratio estimate of the effective components extracted in step 3, the peak preservation index, and the rate of change of the noise power spectral density. The output variables are the adjustment amount of the convergence factor and the adjustment amount of the regularization coefficient. The signal-to-noise ratio estimate is expressed in decibels and is calculated using the following formula: in, The population variance of the effective components. This is an estimate of the noise power; The input variables of the fuzzy logic controller include three: the signal-to-noise ratio estimate of the effective components extracted in step 3, the peak preservation index, and the rate of change of the noise power spectral density; the rate of change of the noise power spectral density is defined as the absolute value of the relative deviation between the noise power spectral density vector of the current processing window and the historical mean vector, and the calculation formula is as follows: in, This is the noise power spectral density vector of the current processing window. It is the average vector of historical noise power spectral density.

6. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 1, characterized in that, In step 5, the threshold setting rule of the adaptive soft thresholding function is as follows: when the signal-to-noise ratio estimate is lower than the preset first threshold, a hard thresholding strategy is adopted; when the signal-to-noise ratio estimate is between the preset first threshold and the preset second threshold, a semi-soft thresholding strategy is adopted; and when the signal-to-noise ratio estimate is higher than the preset second threshold, a soft thresholding strategy is adopted. The size of the threshold is dynamically calculated and determined based on the convergence factor and regularization coefficient updated in step 4. The threshold value is dynamically calculated and determined based on the convergence factor and regularization coefficient updated in step 4; the comprehensive formula for threshold calculation is: in, The convergence factor is The regularization coefficient is . This is an adjustment factor based on the rate of change of noise power spectral density; when the noise intensity increases, A positive value is used to increase the threshold, thereby enhancing the noise reduction effect; when the noise intensity decreases... Negative values ​​are used to lower the threshold, thereby reducing signal loss.

7. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 6, characterized in that, In step 5, the adaptive soft threshold function is implemented as follows: the effective components are subjected to threshold processing of variational mode decomposition coefficients. When a hard threshold strategy is adopted, coefficients with absolute values ​​less than the threshold are set to zero. When a soft threshold strategy is adopted, coefficients with absolute values ​​less than the threshold are shrunk towards zero. When a semi-soft threshold strategy is adopted, linear interpolation is performed between the lower threshold and the upper threshold.

8. The real-time suppression method for high-performance liquid chromatography detector signal noise according to claim 1, characterized in that, After signal reconstruction in step 5, a robust estimation-based fitting method is used to estimate the baseline of the reconstructed signal. The baseline estimate is subtracted from the original denoised signal to obtain the final output signal after baseline correction. During the baseline fitting process, data points in the chromatographic peak region are automatically identified and removed to avoid peak shape distortion.

9. A real-time noise suppression method for a high-performance liquid chromatography detector according to claim 8, characterized in that, It also includes steps to evaluate the peak quality of the baseline-corrected signal, calculate peak width, symmetry factor and separation index. When any index exceeds the preset quality threshold range, the processing strategy adjustment mechanism is automatically triggered, returning to step 3 to re-perform variational mode decomposition and optimize the selection of intrinsic mode components until the peak quality evaluation result meets the requirements.

Images