An adaptive residual-driven meshless power harmonic detection method

Abstract

The application discloses a self-adaptive residual driving meshless power harmonic detection method, and belongs to the technical field of power quality detection. In view of the problem that the existing discrete compressed sensing algorithm is mismatched with the grid and the noise is amplified under a low compression ratio, leading to reconstruction divergence, the application first adopts a Kaiser optimization window to window dimension reduction sampling on the original signal, and suppresses spectrum leakage; secondly, a joint continuous dictionary is constructed, a first-order Taylor expansion and a damped Gauss-Newton iteration mechanism are utilized for continuous frequency domain meshless approximation, signal characteristics are extracted, and a compressed domain physical fitting residual is calculated; finally, the observation noise covariance of the filter is dynamically updated based on the fitting residual, and a single frame result is adaptively smoothed. The application breaks the lower limit of the compression ratio theory of the traditional algorithm, realizes high-precision and high-robustness measurement of harmonic parameters under 9% low compression ratio and a strong noise environment, and has high engineering practicability.

Description

Technical Field

[0001] This application relates to the field of power system power quality monitoring technology, specifically to an adaptive residual-driven meshless power harmonic detection method. Background Technology

[0002] Currently, new power systems dominated by new energy sources have incorporated massive amounts of distributed photovoltaic, wind power, and nonlinear power electronic equipment, resulting in severe distortion of grid voltage and current waveforms. The spectral composition exhibits wide bandwidth, time-varying characteristics, and complex interharmonic interleaving. To achieve comprehensive observability of the distribution network, high-frequency, densely deployed intelligent sensing terminals generate massive amounts of sampling data, posing significant challenges to the transmission bandwidth and underlying storage and computing power of existing distribution communication networks.

[0003] Compressed sensing (CS) theory can perform non-adaptive sampling at rates far below the Nyquist frequency, combining data compression and sampling into one process, greatly alleviating communication and storage pressures, and showing broad application prospects in the field of power quality data compression. However, traditional compressed sensing power harmonic detection methods suffer from the following long-standing technical bottlenecks:

[0004] First, traditional discrete compressed sensing reconstruction models (such as orthogonal matching pursuit and sparse adaptive matching pursuit algorithms) are all based on a pre-defined discrete frequency dictionary. However, due to frequency fluctuations and non-integer interharmonics in real power grids, the actual signal frequency often cannot accurately fall within the pre-defined discrete grid points. This grid mismatch leads to severe spectral leakage and the "pick-up fence effect," causing energy diffusion in sparse vectors, resulting in a sharp increase in reconstruction residuals and a shift in frequency estimation.

[0005] Secondly, existing technologies employ a strategy of "windowing preprocessing combined with spectral interpolation" to mitigate errors caused by mesh mismatch. However, spectral interpolation is highly dependent on the leakage energy distribution of adjacent frequencies, requiring the reconstruction algorithm to retain a large number of effective observations. According to the Restricted Isometry Property (RIP), as the compression ratio (CR) decreases, the available observations cannot meet the sparse atom reconstruction requirements of the interpolation algorithm. Therefore, the effective compression ratio of existing discrete compressed sensing algorithms typically struggles to exceed the theoretical lower limit, failing to meet the evolving demand for extremely lightweight new IoT terminals.

[0006] Finally, under harsh conditions of low compression ratios (e.g., below 10%), the RIP of the observation matrix degrades, and the matrix condition number deteriorates drastically. This ill-conditioned matrix leads to a strong noise amplification effect, causing even extremely weak background white noise from the power distribution network to be greatly amplified when mapped back to the high-dimensional signal space, resulting in severe distortion and transient divergence in the single-frame reconstruction results. At this point, conventional time-domain adaptive Kalman filtering, which assumes a constant observation noise covariance, cannot adapt to the strong time-varying randomness of single-frame reconstruction errors under low compression ratios. It not only fails to effectively filter out noise but also easily induces filter divergence.

[0007] In summary, existing power harmonic detection technologies generally face problems such as grid mismatch, matrix ill-conditioning, and noise amplification at low sampling rates. There is an urgent need for a new gridless power harmonic detection mechanism that can achieve high-precision and robust parameter estimation under extremely low compression ratios and strong electromagnetic interference environments. Summary of the Invention

[0008] The technical problem this application aims to solve is: addressing the issues of mesh mismatch, difficulty in exceeding the theoretical lower limit of effective compression ratio, and severe noise amplification and filter divergence caused by matrix ill-conditioning in existing discrete compressed sensing harmonic detection methods. This application provides an adaptive residual-driven meshless power harmonic detection method. This method achieves high-precision and robust measurement of power harmonics and interharmonic parameters under extremely low compression ratios and strong electromagnetic interference environments through deep fusion of continuous frequency domain Newton approximation and residual-driven post-filtering. The method includes the following steps:

[0009] S101, Time-domain sampling is performed on the signal to be measured in the power system to obtain a sampling sequence. Construct a Gaussian random measurement matrix And combined with shape parameters as The Kaiser optimized window function is diagonalized to generate a windowed measurement matrix that incorporates the characteristics of the window function. Using the windowed measurement matrix For the sampled sequence Compressed sampling and dimensionality reduction projection are performed to obtain low-dimensional compressed observation vectors. ;

[0010] S102, construct an initial sensing matrix based on a preset discrete Fourier basis, and use a sparse adaptive matching pursuit algorithm to process the compressed observation vector. Iterative calculations are performed to extract the discrete frequency points corresponding to the peak values ​​of the spectral lines in the positive frequency domain, which serve as the initial frequency point set for continuous domain optimization.

[0011] S103, Define the complex exponential basis vectors for the corresponding frequency points in the continuous frequency domain. and the partial derivative vector of the basis vector with respect to frequency At the initial frequency point set, the complex exponential basis vectors are linearized using a first-order Taylor joint expansion, and the linearized basis atoms and derivative atoms are mapped to the compressed observation domain to construct an enhanced joint continuous sensing matrix. ;

[0012] S104, the enhanced joint continuous sensing matrix is ​​decoupled and solved using the least squares method to calculate the complex amplitude and frequency deviation. Introducing a damping limiting factor A damped Gaussian-Newton update rule is constructed with the maximum jump bandwidth to iteratively update the frequency step size, driving the frequency points to adaptively approximate within the continuous real domain until convergence is achieved in outputting single-frame harmonic parameters. Simultaneously, based on the joint continuous sensing matrix and sparse complex coefficient vector after the final iteration convergence, the physical fitting residual of the current frame data in the compressed observation domain is calculated. ;

[0013] S105, establish a physical state-space model of harmonic characteristic parameters including frequency, amplitude, and phase, and use the single-frame harmonic parameters as filter input; use the compressed observation domain physical fitting residual calculated in step S104. The observation noise covariance matrix of the post-filter is dynamically adjusted in real time. The weight of single-frame observations of sudden strong electromagnetic interference is actively reduced, and the filter gain is calculated and the state posterior estimation is performed based on this. In this way, the harmonic parameters of the single frame are subjected to time-domain inertial smoothing and transient divergence shielding, and the final harmonic and interharmonic characteristic parameters are output.

[0014] Optionally, the windowed measurement matrix that fuses the window function characteristics described in step S101... Represented as ;in, The Gaussian random measurement matrix, To obtain the diagonal matrix of the Kaiser optimization window function after diagonalization, the discrete-time expression of the Kaiser optimization window function is as follows:

[0015] ,

[0016] in For the first kind of zeroth order modified Bessel function, The number of sampling points. The shape parameter has a value of .

[0017] Optionally, the complex exponential basis vector described in step S103 And the partial derivative vector of the complex exponential basis vector with respect to frequency Specifically, it is expressed as follows:

[0018] ,

[0019] ,

[0020] in, For the frequency points to be solved, The sampling time interval, Represents vector transpose; the enhanced joint continuous sensing matrix Base atoms mapped to the compressed observation domain With derivative atoms It is constructed by splicing.

[0021] Optionally, the specific formula for decoupling and solving using the least squares method in step S104 is as follows:

[0022] ,

[0023] in, Let the complex amplitude be the unknown. It is an auxiliary variable and satisfies , This represents the conjugate transpose of a matrix. For an enhanced joint continuous sensing matrix, This is a low-dimensional compressed observation vector.

[0024] Optionally, the damped Gaussian-Newton update rule described in step S104 satisfies the following iterative relationship:

[0025] ,

[0026]

[0027] In the formula, This represents the current iteration number. For the updated frequency, For the current frequency, For frequency deviation, To take the real part of a complex number, This is the introduced damping limiting factor.

[0028] Optionally, the physical fitting residual described in step S104 The calculation model is as follows:

[0029] ,

[0030] in, The L2 norm of a vector. For low-dimensional compressed observation vectors, For windowed measurement matrix, and These are the complex exponential basis vector and complex amplitude, respectively, after the final iteration converges.

[0031] Optionally, the observation noise covariance matrix mentioned in step S105 The dynamic update formula is:

[0032] ,

[0033] In the formula, For the first The adaptive observation noise covariance matrix after frame update. The reference observation noise variance is obtained through system noise floor calibration. This is the sensitivity adjustment coefficient. The physical fitting residuals extracted in step S104 For the natural constant An exponential function with base 0.

[0034] Compared with traditional compressed sensing harmonic detection methods, the detection method of this application has the following significant technical advantages:

[0035] 1. This application overcomes the grid mismatch bottleneck of discrete compressed sensing and eliminates the picket fence effect. By constructing a joint dictionary and first-order Taylor expansion in the continuous frequency domain, and utilizing the gradient guidance mechanism of derivative atoms, this application drives the harmonic and interharmonic frequencies to accurately approximate the real physical frequencies in the continuous real number domain, effectively overcoming the inherent spectral leakage and frequency offset truncation errors of traditional discrete grid methods.

[0036] 2. This invention breaks through the physical lower limit of compression ratio in traditional interpolation algorithms, significantly reducing the communication burden on terminals. Existing correction algorithms based on multi-spectral line interpolation heavily rely on the redundancy of the observation dimension, making it extremely difficult to achieve a compression ratio below 30%. This application combines the excellent sidelobe suppression capability of the Kaiser optimization window with the precise decoupling of Newton's approximation, achieving high-precision reconstruction across the entire frequency band at a low compression ratio of 9% without any interpolation formulas, greatly improving the data compression efficiency of edge sensing terminals.

[0037] 3. The residual-driven adaptive covariance update mechanism significantly enhances measurement robustness in noisy environments. This application uses the physical fitting residual generated by the front-end CS algorithm as the driving source for post-filtering, breaking the static assumption of constant observation noise in traditional filtering. When low compression ratios lead to matrix ill-conditioning or encounter sudden strong electromagnetic interference, the system can actively identify the degradation of single-frame reconstruction quality and automatically adjust the smoothing weights, effectively shielding transient frequency divergence and reducing the steady-state measurement error by 1 to 2 orders of magnitude under noisy conditions. Attached Figure Description

[0038] Figure 1This is a schematic diagram of the overall process of the detection method described in the embodiments of this application.

[0039] Figure 2 This is a schematic diagram of the continuous frequency domain meshless optimization mechanism described in the embodiments of this application.

[0040] Figure 3 The graphs show the frequency offset trajectory and relative error convergence curve of the algorithm described in the embodiments of this application; wherein, 3(a) is the frequency offset trajectory curve of continuous frequency domain joint optimization, and 3(b) is the relative error convergence curve of representative frequency components.

[0041] Figure 4 The figures are comparisons of the relative errors of parameter reconstruction under strong background white noise interference in the embodiments of this application; wherein, 4(a) is the comparison curve of the relative error of steady-state average frequency, 4(b) is the comparison curve of the relative error of steady-state average amplitude, and 4(c) is the comparison curve of the relative error of steady-state average phase.

[0042] Figure 5 This is a system block diagram of the actual hardware testing platform described in the embodiments of this application. Detailed Implementation

[0043] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to illustrate the technical solutions of the present application and do not constitute a limitation on the scope of protection.

[0044] Example 1: High-precision continuous frequency domain meshless extraction of complex harmonic and interharmonic signals

[0045] This embodiment aims to verify the fundamental performance of the method described in this application in eliminating the picket fence effect, resolving grid mismatch, and improving the accuracy of characteristic parameter measurements under ideal operating conditions with no background noise. A complex integrated power test signal containing the fundamental frequency, seven odd-order integer harmonics, and two non-integer interharmonics is constructed, and its continuous-time expression is as follows:

[0046]

[0047] The specific physical characteristic parameters of the test signal are set as shown in Table 1:

[0048] Table 1. Parameter settings for complex harmonic and interharmonic test signals

[0049] fundamental wave 50 1.00 60 3rd harmonic 150 0.30 120 Interharmonic 1 187 0.15 130 5th harmonic 250 0.25 110 Interharmonic 2 317 0.12 140 7th harmonic 350 0.20 120 9th harmonic 450 0.15 110 11th harmonic 550 0.10 150 13th harmonic 650 0.05 60

[0050] Among them, the interharmonic components (187Hz and 317Hz) are not located on the discrete grid in the spectrum, which will cause a severe fence effect.

[0051] The parameters of the signal are extracted using the method described in this application, and the specific steps are as follows:

[0052] S101, Set the system sampling frequency Number of equally spaced sampling points in the time domain (Corresponding sampling time window is) ), to obtain discrete time series Set the shape parameters of the Kaiser optimization window function. Calculate the diagonal matrix of the window function. , and Gaussian random measurement matrix Perform multiplication to generate a windowed measurement matrix. ;use right Perform compressed sampling and dimensionality reduction projection, and set the effective compression ratio to be... Obtain low-dimensional compressed observation vectors ;

[0053] S102, construct an initial sensing matrix based on a preset discrete Fourier basis, and use a sparse adaptive matching pursuit algorithm to process the observation vector. Greedy iterative calculations are performed to extract the discrete frequency points corresponding to the peak values ​​of the spectral lines in the positive frequency domain, which are denoted as the initial frequency point set. Due to the limitation of the discrete grid step size, the non-integer frequency inter-harmonics (187Hz and 317Hz) have a maximum value at the 0th iteration. The truncation mismatch error, this coarse measurement is used as the starting point for continuous domain optimization;

[0054] S103, Construct complex exponential basis vectors for corresponding frequency points in the continuous frequency domain. and its partial derivative vector with respect to frequency At the initial frequency point The basis vectors are linearized by performing a first-order Taylor joint expansion, and the linearized basis atoms are then processed. With derivative atoms Mapped to a compressed observation domain, the data is stitched together to construct an enhanced joint continuous sensing matrix. ;

[0055] S104, using the least squares formula The parameters are decoupled and solved to extract the complex amplitude. and auxiliary variables ; Utilize update rules Calculate the frequency deviation and update it using the damped Gauss-Newton rule. Perform step-size iterations;

[0056] S105, under noise-free ideal operating conditions, after 5-8 iterations of convergence, the parameters can be directly extracted, and the true frequency is the value of the last Newton iteration. The original signal's true amplitude was calculated as follows: The initial phase directly corresponds to the argument of the complex coefficients. .

[0057] Under the same compression ratio, the algorithm of this application is compared with the traditional improved algorithm using discrete windowing interpolation (Comparison Model A: using 4 terms of 3rd order Nuttall window combined with 3 spectral lines interpolation; Comparison Model B: using 6 terms of 5th order cosine window combined with 4 spectral lines interpolation). The statistical comparison of the relative frequency errors obtained by each algorithm is shown in Table 2.

[0058] Table 2. Relative errors of test signal parameters for different algorithms at the same compression ratio.

[0059] fundamental wave <![CDATA[7.9075×10 −7 ]]> <![CDATA[1.0881×10 −4 ]]> <![CDATA[1.3576×10 −15 ]]> 3rd harmonic <![CDATA[6.6455×10 −6 ]]> <![CDATA[1.3677×10 −5 ]]> <![CDATA[5.6843×10 −14 ]]> Interharmonic 1 <![CDATA[2.5977×10 −6 ]]> <![CDATA[2.8384×10 −7 ]]> <![CDATA[1.2159×10 −13 ]]> 5th harmonic <![CDATA[1.4085×10 −7 ]]> <![CDATA[2.1846×10 −5 ]]> <![CDATA[1.5916×10 −13 ]]> Interharmonic 2 <![CDATA[1.2797×10 −6 ]]> <![CDATA[9.4468×10 −5 ]]> <![CDATA[2.1518×10 −13 ]]> 7th harmonic <![CDATA[2.9235×10 −6 ]]> <![CDATA[1.5651×10 −4 ]]> <![CDATA[8.1205×10 −14 ]]> 9th harmonic <![CDATA[4.4577×10 −7 ]]> <![CDATA[1.2211×10 −5 ]]> <![CDATA[1.2632×10 −14 ]]> 11th harmonic <![CDATA[6.7321×10 −7 ]]> <![CDATA[1.0028×10 −5 ]]> <![CDATA[5.1676×10 −13 ]]> 13th harmonic <![CDATA[4.3935×10 −6 ]]> <![CDATA[7.9834×10 −6 ]]> <![CDATA[1.9239×10 −13 ]]>

[0060] As shown in Table 2, traditional algorithms A and B are limited by the discrete grid resolution. Even with spectral interpolation, their frequency reconstruction errors are still distributed within... to The order of magnitude. However, the algorithm in this application, through continuous frequency domain meshless approximation, eliminates the systematic errors caused by the picket fence effect and mesh mismatch, controlling the reconstruction accuracy within... to This represents an order of magnitude improvement in accuracy.

[0061] Example 2: Robustness Test of Adaptive Filtering under Strong Background Noise Conditions

[0062] This embodiment aims to verify the noise robustness of this application under the dual harsh conditions of low compression ratio sampling and strong electromagnetic interference. The effective compression ratio of windowed compression sampling is set to... And superimpose a signal-to-noise ratio of into the signal. Strong Gaussian white noise.

[0063] Due to the compression ratio being reduced to The observation space is extremely compressed, and the joint continuous sensing matrix This leads to severe multicollinearity and a deterioration of the condition number. Under these extreme conditions, without dynamic filtering, even minute low-level noise will be greatly amplified when mapped back to the high-dimensional space (noise amplification factor). This leads to instability in the single-frame algorithm, causing the calculated Newton step size to tend towards infinity, resulting in frequency reconstruction distortion and transient divergence. To address this problem, this application utilizes a dynamic filtering smoothing mechanism:

[0064] (1) Establish a physical state space model: use the single-frame measurement parameters (frequency, amplitude, initial phase) finally output by the CS algorithm in each frame as the input vector of the filter for updating;

[0065] (2) Fitting residual-driven adaptive covariance update: The compressed domain physical fitting residual obtained by back-calculation after the iteration convergence in the current frame in step S104 is used. Incorporate dynamic update rules: .

[0066] When subjected to sudden strong noise interference or ill-conditioned single-frame calculations that degrade the reconstruction quality, the physical fitting residuals... If the anomaly increases significantly, the algorithm actively amplifies the observation noise covariance matrix of the current frame through an exponential mechanism. This allows the filter to actively reduce the trust weight of the current single-frame observation when calculating the filter gain, and instead refer more to historical steady-state predictions, thereby effectively smoothing and shielding the transient frequency divergence induced by low compression ratio.

[0067] Experimental results show that, Low compression ratio and Under harsh operating conditions with both high noise and strong noise, the unfiltered single-frame output is extremely sensitive to random noise. However, after introducing residual-driven adaptive filtering in this invention, the steady-state measurement relative error across the entire frequency band (including fundamental, higher harmonics, and interharmonics) is generally reduced by one to two orders of magnitude in terms of frequency, amplitude, and phase. The frequency measurement error is reduced from... The magnitude steadily decreased to The system successfully smoothed out random jitter caused by noise, demonstrating excellent dynamic locking and noise-resistant tracking performance.

[0068] Example 3: Lightweight Edge Awareness Detection Verification Based on Physical Hardware Platform

[0069] To further verify the engineering feasibility of this application under real physical environments and actual hardware limitations, this embodiment constructs a practical hardware test platform for power quality disturbances. This test platform mainly consists of four parts connected in series: a high-precision power standard source Fluke 6105A, a signal conditioning circuit (including voltage transformer isolation step-down, low-pass filtering, and range conversion circuits), a 16-bit high-precision USB data acquisition card (ADC), and a host computer analysis terminal based on edge computing power.

[0070] In actual testing, the specific engineering implementation steps are as follows:

[0071] The voltage physical signal generated by the Fluke 6105A standard power source contains complex broadband distortion characteristics, intentionally mixed with fundamental frequency fluctuations and severe picket fence effects. and Complex interharmonic components; the physical voltage signal is sent to the signal conditioning circuit for isolation, voltage reduction, and anti-aliasing low-pass filtering, and then sent to the USB acquisition card for 16-bit high-precision analog-to-digital conversion (ADC); after receiving the complete discrete sampling data, the host computer terminal, in order to simulate the lightweight operation state of the distribution network edge sensing terminal under low uplink bandwidth constraints, forces the downsampling projection of the Gaussian random observation matrix, retaining only the... The amount of observation data; the algorithm of this application embedded in the terminal performs joint calculation on this set of dimensionality-reduced data and directly outputs the final measured parameters.

[0072] Under the dual adverse conditions of hardware quantization truncation error and background thermal noise interference from the conditioning circuit, the measurement results of harmonic / interharmonic parameters of the actual hardware platform are shown in Table 3:

[0073] Table 3 Measurement results of harmonic / interharmonic parameters on the actual hardware platform

[0074] fundamental wave 50.1 50.1015 <![CDATA[1.50×10 -3 ]]> 60 60.08 <![CDATA[1.40×10 −3 ]]> 3rd harmonic 150 150.0023 <![CDATA[2.30×10 -3 ]]> 120 120.12 <![CDATA[2.10×10 −3 ]]> Interharmonic 1 187 186.9772 <![CDATA[-2.28×10 -2 ]]> 130 129.85 <![CDATA[-2.62×10 −3 ]]> 5th harmonic 250 250.0035 <![CDATA[3.50×10 -3 ]]> 110 110.18 <![CDATA[3.14×10 −3 ]]> Interharmonic 2 317 317.0141 <![CDATA[1.41×10 -2 ]]> 140 139.78 <![CDATA[-3.84×10 −3 ]]> 7th harmonic 350 349.9952 <![CDATA[-4.80×10 -3 ]]> 120 120.25 <![CDATA[4.36×10 −3 ]]> 9th harmonic 450 450.0056 <![CDATA[5.60×10 -3 ]]> 110 109.72 <![CDATA[-4.89×10 −3 ]]> 11th harmonic 550 550.0064 <![CDATA[6.40×10 -3 ]]> 150 150.32 <![CDATA[5.58×10 −3 ]]> 13th harmonic 650 649.9725 <![CDATA[-2.75×10 -2 ]]> 60 59.65 <![CDATA[-6.11×10 −3 ]]>

[0075] Table 3 shows that, based on objective analysis of the hardware measurement data, this application exhibits extremely stable dynamic tracking performance in the face of fundamental frequency fluctuations, with the fundamental frequency and phase measurement errors controlled within a certain range. and ; and for the non-integer interharmonic components most severely affected by the fence effect ( Hz and (Hz), the maximum absolute error of its measured frequency is only The maximum absolute phase error is only .

[0076] The above-mentioned real physical test results further verify that the "adaptive residual-driven meshless detection architecture" described in this application can still ensure high-precision feature extraction across orders of magnitude even under the superimposed real physical hardware limitations (such as ADC quantization truncation, temperature drift of conditioning circuit, and component noise floor). It has fully feasible engineering applicability and can provide solid technical support for the high-precision, ultra-low bandwidth deployment of lightweight edge sensing terminals in future new power distribution networks.

Claims

1. An adaptive residual-driven meshless power harmonic detection method, characterized in that, Includes the following steps: S101, Time-domain sampling is performed on the signal to be measured in the power system to obtain a sampling sequence. Construct a Gaussian random measurement matrix And combined with shape parameters as The Kaiser optimized window function is diagonalized to generate a windowed measurement matrix that incorporates the characteristics of the window function. ; Using the windowed measurement matrix For the sampling sequence Compressed sampling and dimensionality reduction projection are performed to obtain low-dimensional compressed observation vectors. ; S102, construct an initial sensing matrix based on a preset discrete Fourier basis, and use a sparse adaptive matching pursuit algorithm to process the compressed observation vector. Iterative calculations are performed to extract the discrete frequency points corresponding to the peak values ​​of the spectral lines in the positive frequency domain, which are used as the initial frequency point set for continuous domain optimization. S103, Define the complex exponential basis vectors for the corresponding frequency points in the continuous frequency domain. and the partial derivative vector of the basis vector with respect to frequency ; The complex exponential basis vectors are linearized by a first-order Taylor joint expansion at the initial frequency point set, and the linearized basis atoms and derivative atoms are mapped to the compressed observation domain to construct an enhanced joint continuous sensing matrix. ; S104, the enhanced joint continuous sensing matrix is ​​decoupled and solved using the least squares method to calculate the complex amplitude and frequency deviation. Introducing a damping limiting factor A damped Gaussian-Newton update rule is constructed with the maximum jump bandwidth to iteratively update the frequency step size, driving the frequency points to adaptively approximate within the continuous real domain until convergence is achieved in outputting single-frame harmonic parameters. Simultaneously, based on the joint continuous sensing matrix and sparse complex coefficient vector after the final iteration convergence, the physical fitting residual of the current frame data in the compressed observation domain is calculated. ; S105, establish a physical state-space model of harmonic characteristic parameters, using the single-frame harmonic parameters as filter input; utilize the compressed observation domain physical fitting residual calculated in step S104. The observation noise covariance matrix of the post-filter is dynamically adjusted in real time. The weight of single-frame observations of sudden strong electromagnetic interference is actively reduced, and the filter gain is calculated and the state posterior estimation is performed based on this. In this way, the harmonic parameters of the single frame are subjected to time-domain inertial smoothing and transient divergence shielding, and the final harmonic and interharmonic characteristic parameters are output.

2. The method according to claim 1, characterized in that, The windowed measurement matrix for fusing window function characteristics described in step S101 Represented as ;in, The Gaussian random measurement matrix, To obtain the diagonal matrix of the Kaiser optimization window function after diagonalization, the discrete-time expression of the Kaiser optimization window function is as follows: ,in For the first kind of zeroth order modified Bessel function, The number of sampling points. The shape parameter has a value of .

3. The method according to claim 1, characterized in that, The complex exponential basis vector mentioned in step S103 And the partial derivative vector of the complex exponential basis vector with respect to frequency Specifically, it is expressed as follows: ， ， in, For the frequency points to be solved, The sampling time interval, Represents vector transpose; the enhanced joint continuous sensing matrix From the base atoms mapped to the compressed observation domain With derivative atoms It is constructed by splicing.

4. The method according to claim 1, characterized in that, The specific formula for decoupling and solving using the least squares method described in step S104 is as follows: ， in, Let the complex amplitude be the unknown. It is an auxiliary variable and satisfies , This represents the conjugate transpose of a matrix. For an enhanced joint continuous sensing matrix, This is a low-dimensional compressed observation vector.

5. The method according to claim 1, characterized in that, The damped Gaussian-Newton update rule described in step S104 satisfies the following iterative relationship: ， ， in, This represents the current iteration number. For the updated frequency, For the current frequency, For frequency deviation, To take the real part of a complex number, This is the introduced damping limiting factor.

6. The method according to claim 1, characterized in that, The physical fitting residual described in step S104 The calculation model is as follows: ， in, The L2 norm of a vector. For low-dimensional compressed observation vectors, For windowed measurement matrix, and These are the complex exponential basis vector and complex amplitude, respectively, after the final iteration converges.

7. The method according to claim 1, characterized in that, The observation noise covariance matrix mentioned in step S105 The dynamic update formula is: ， In the formula, For the first The adaptive observation noise covariance matrix after frame update. The reference observation noise variance is obtained through system noise floor calibration. This is the sensitivity adjustment coefficient. The physical fitting residuals extracted in step S104 For the natural constant An exponential function with base 0.

Images