A signal processing method suitable for various signal characteristics

Abstract

This invention discloses a signal processing method applicable to various signal characteristics, relating to the fields of communication and radar signal processing technology. By combining the construction of a multi-dimensional signal feature library with real-time spectrum sensing, this invention achieves rapid analysis and accurate characterization of deep-space radar signal characteristics. The feature library is generated based on unsupervised clustering of historical data, covering various signal templates such as transient impulses, periodic coherence, and steady-state noise. It not only provides rich prior knowledge references but also supports online sliding window incremental updates, ensuring that feature representation always conforms to actual environmental changes. The real-time spectrum sensing employs adaptive window time-frequency analysis, with the window length dynamically adjusted according to the signal's instantaneous bandwidth and local stationarity, effectively balancing the contradiction between frequency resolution and time resolution and overcoming the limitations of traditional fixed-window analysis in non-stationary scenarios.

Description

Technical Field

[0001] This invention relates to the field of communication and radar signal processing technology, and in particular to a signal processing method applicable to a variety of signal characteristics. Background Technology

[0002] Ground-based deep space exploration radars achieve high-precision measurements of deep space targets such as the Moon and near-Earth asteroids by actively emitting electromagnetic waves and receiving reflected signals from targets. As the detection distance expands to hundreds of millions of kilometers, the intensity of the target echo signal decreases sharply to near noise levels. At the same time, the signal is affected by complex time-varying disturbances when passing through the Earth's ionosphere and interstellar plasma. These disturbances include carrier phase abrupt changes caused by ionospheric scintillation, Doppler frequency spread, and signal distortion caused by the interstellar medium, which makes the echo exhibit strong non-stationary and low signal-to-noise ratio characteristics.

[0003] Existing technologies employ adaptive processing methods in some approaches, such as time-varying Wiener filtering to track the statistical characteristics of ionospheric scintillation and suppress time-varying noise by updating filter coefficients in real time. Compressed sensing spectral analysis utilizes the sparsity of the signal in the fractional Fourier domain to improve the detection probability of weak echoes. In addition, some approaches introduce cyclostationary feature detection to distinguish periodic target signals from random noise. While these methods can improve the signal-to-noise ratio under certain conditions, they still have significant limitations: the performance of time-varying Wiener filtering depends on high-precision channel estimation, which is difficult to achieve in real-time channel modeling in deep space environments; compressed sensing methods are sensitive to the assumption of signal sparsity, and the reconstruction failure rate increases significantly when target motion or plasma disturbances complicate the signal structure; and cyclostationary detection struggles to handle scintillation noise with rapidly decaying correlation.

[0004] The fundamental bottleneck of the above-mentioned solutions lies in the mismatch between the processing link and the dynamic characteristics of the signal. The time-frequency characteristics of deep space target echoes and interference change in real time with the detection distance, orbital position, and space environment, but traditional algorithms use a fixed processing framework: either relying on preset parameters, such as the iteration step size of Wiener filtering, or being limited by prior assumptions, such as the sparse basis of compressed sensing. This leads to two typical problems: first, parameter lag, when ionospheric scintillation changes abruptly at the millisecond level, the filter cannot respond in time; second, the lack of model generalization, the same processing link cannot take into account both the broadband signals of near-Earth asteroids and the narrowband reflections of the lunar surface. There is an urgent need for a method that can autonomously sense signal characteristics and dynamically adapt the processing mechanism to cope with the highly heterogeneous and non-stationary signal environment in deep space exploration. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a signal processing method applicable to various signal characteristics to solve the problem in deep space radar detection where weak echoes are caused by ionospheric scintillation and interstellar medium disturbances, resulting in non-stationary signals and low signal-to-noise ratios. Existing adaptive methods are difficult to dynamically match signal characteristics due to fixed processing frameworks and parameter lags.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a signal processing method applicable to various signal characteristics, comprising:

[0009] Step S1: Receive the input signal; perform real-time spectral feature extraction under the condition of not exceeding the preset time delay budget. The real-time spectral features include at least: spectral sparsity, number of dominant spectral lines, intensity of cyclostationary components, spectral kurtosis, and estimated signal-to-noise ratio.

[0010] Step S2: Match the real-time spectral features with the multi-dimensional signal feature library to obtain a set of candidate templates and their matching scores;

[0011] Step S3: Based on the candidate template, calculate the selection score for multiple signal processing links in the processing model library. The score is based on: template matching score, real-time spectral characteristics, current signal-to-noise ratio, time delay budget and computing power constraints.

[0012] Step S4: Determine the target link based on the selection scoring and conflict resolution rules;

[0013] Step S5: Execute the selected link to obtain the output signal;

[0014] Step S6: Calculate the reward feedback index and update the adjustable parameters within the link online accordingly.

[0015] As a preferred embodiment of the signal processing method applicable to multiple signal features described in this invention, the construction of the multidimensional signal feature library includes:

[0016] Unsupervised clustering and silhouette coefficient filtering are performed on historical signal data to generate feature templates;

[0017] Stored categorized by signal type; each template records at least the spectral sparsity range, the range of the number of dominant spectral lines, the intensity range of cyclic stationary components, the spectral kurtosis range, the correlation time range, the typical signal-to-noise ratio range, and a list of recommended links;

[0018] The library supports incremental updates via sliding windows with human review and approval flags during the online phase.

[0019] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the extraction of real-time spectral features employs adaptive window time-frequency analysis;

[0020] The window length is dynamically adjusted based on instantaneous bandwidth and local stationarity metrics, and features are updated using a sliding window recursive method to reduce latency.

[0021] When strong nonstationarity is detected, the window length is automatically shortened and the update frequency is increased.

[0022] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the selection score includes at least two parts: feature matching score and resource constraint score; if the scores of multiple links are similar, conflict resolution is performed in the order of priority period robustness, second priority energy aggregation, and then priority time delay.

[0023] When both cyclic stationary characteristics and multiple dominant spectral lines exist simultaneously, coherent accumulation links are preferred.

[0024] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the signal processing link includes at least:

[0025] A) Transient impact chain: Performs variational mode decomposition and subsequent impulse noise reduction / reconstruction;

[0026] B) Coherent accumulation link: Performs coherent accumulation / coherent compensation and fractional-order domain energy accumulation;

[0027] C) Steady-state noise suppression link: Perform time-varying adaptive filtering / spectral entropy minimization and band-limited reconstruction;

[0028] All links can be fine-tuned by the online parameter adjustment module during operation.

[0029] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the transient impulse link is preferentially selected when the spectral sparsity is high and the number of dominant spectral lines does not exceed a preset upper limit; when the intensity of the cyclic stationary component exceeds a threshold or the number of dominant spectral lines exceeds the upper limit, the coherent accumulation link is preferentially selected; the threshold is determined by dynamic compensation based on the historical statistical median of similar templates and combined with the real-time estimated signal-to-noise ratio, and the compensation amount is adjusted over time according to the window update strategy.

[0030] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the penalty factor of variational mode decomposition, the phase / Doppler compensation parameter in coherent accumulation, and the search range of the rotation angle in fractional domain processing are all finely adjusted by the online parameter adjustment module based on the reward feedback index, and boundaries and strides are set to prevent divergence and overfitting.

[0031] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the reward feedback index includes at least one of the following:

[0032] a) Energy concentration of the output signal;

[0033] b) Estimated signal-to-noise ratio of the output signal;

[0034] c) Cyclic autocorrelation peak value and sidelobe ratio or detection statistic;

[0035] d) Matching bias to known calibration subsequences;

[0036] During the online phase, the single objective or weighted combination of this indicator is used as the optimization objective;

[0037] In step S6, to unify the indicators a) energy concentration, b) estimated signal-to-noise ratio, c) cyclic correlation peak and sidelobe ratio or detection statistic, and d) calibration subsequence matching bias into the same optimization framework, the indicators are first subjected to direction unification and dimensionless mapping to obtain vectors. ; The transpose operator is used; based on this, a combined objective is constructed, and the normalization and online updating of the weights are provided;

[0038] 1) In each sliding window Internally, a weighted sum is used as the optimization objective and is employed to drive the update of parameters within the link:

[0039] ,

[0040] in, Display window The combined objective value, dimensionless. Indicates the first Individual indicators in the window The weights, ranging from [0,1], are dimensionless. Indicates the first Dimensionless returns of each indicator Pick These correspond to the four categories of indicators mentioned above. This is the index of the current window, in units of [number].

[0041] 2) Introduce a temperature-based Softmax mapping to obtain weights on the probabilistic simplex:

[0042] ,

[0043] in, Indicates the first The preference scores for each indicator are initialized from template suggestions and are dimensionless. The parameter is temperature, dimensionless. To indicate the summation index; perform lower bound and renormalization to suppress weight collapse:

[0044] , ,

[0045] in, As the lower limit of the weight, It is a vector consisting entirely of 1s;

[0046] 3) In the online phase, the preference score is adaptively driven by the advantage signal, and steady-state constraints are applied through the learning rate and update cycle:

[0047] ,

[0048] in, This indicates the updated preference score in this window. The learning rate is for weights. It is dimensionless. For immediate gains, The moving average baseline for this indicator, with an update period of [missing information]. By default, it triggers every 1-5 windows; if In recent If the value is continuously below the threshold within a window, then it will revert to the most recent stable value. Link parameters.

[0049] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the online parameter adjustment module adopts a reinforcement learning policy network, whose inputs include real-time spectral features, template matching scores, link operation status and reward feedback indicators, and whose output is the link parameter increment; the network is pre-trained with historical data and periodically updated with policy gradients during the online phase; to meet engineering safety requirements, an upper limit for the update period, an upper limit for the learning rate, a performance rollback mechanism and a delay constraint are set.

[0050] As a preferred embodiment of the signal processing method applicable to multiple signal characteristics described in this invention, the calculation of the spectral characteristics and the setting of the threshold in step S1 are as follows:

[0051] Common settings: Within the sliding window, first obtain the normalized power spectrum vector using Welch / STFT, and then use statistics from similar templates and real-time SNR for threshold adaptation. To maintain consistency, the threshold is set using the following formula:

[0052] ,

[0053] in, Representation of features The threshold for judgment, in units of Consistent, Indicates categories in the template library Historical median Indicates the SNR compensation coefficient. This indicates the estimated signal-to-noise ratio for this window, in dB. Indicate category Historical SNR median, in dB;

[0054] a. Spectral sparsity: Hoyer sparsity measures the concentration of spectral energy.

[0055] ,

[0056] in, The sparsity is expressed as [0, 1]. The normalized power spectrum is dimensionless. Frequency points, in points. They are respectively and Norm;

[0057] Threshold caliber: based on Sparsity is determined by default. , initial empirical values ;

[0058] b. Number of dominant spectral lines: Estimate the noise floor on the logarithmic spectrum using the moving median and set a peak threshold and minimum frequency interval. Count the dominant peaks: If And the distance between adjacent peaks This is recorded as one dominant spectral line;

[0059] Threshold caliber: Upper limit criterion is adopted. ;default , initial empirical values According to the common formula SNR compensation is obtained , here Indicates spectral resolution, in Hz. This represents the minimum physical separation bandwidth, measured in Hz. Noise floor estimate obtained by moving median filtering;

[0060] c. Cyclic stationary component strength: Based on the spectral correlation density (SCD) in the candidate cyclic frequency set. Energy ratio characterization:

[0061] ,

[0062] in, The cyclic steady-state strength is dimensionless. The SCD is estimated by FAM, with units consistent with the power spectrum. This is a conventional power spectrum. Cyclic frequency, in Hz. For the first Frequency points, in Hz;

[0063] Threshold caliber: based on Determine if a cycle is stationary; default. Recommended by the template, initial experience value and according to the common formula Perform SNR compensation;

[0064] d. Spectral kurtosis: For each frequency point, the fourth-order moment ratio is calculated in time, and the full-spectrum statistic is used as the indicator.

[0065] ,by or Warehousing;

[0066] Threshold caliber: based on Non-Gaussian components are significant; empirical initial values. Compensation is made according to a common formula, where, Frequency point Time series amplitude or power spectrum estimation samples, This represents the average time within the window. Dimensionless;

[0067] e. Estimating the signal-to-noise ratio: using the noise floor set With signal set Estimating power ratio by partition and converting to dB, Select a frequency point below the noise floor threshold. Take the remaining frequency points or template annotation bands;

[0068] Threshold caliber: Provides the operating threshold. For link selection and compensation baseline, empirical initial values Template categories are provided Use as a reference.

[0069] The beneficial effects of this invention are as follows: By combining the construction of a multi-dimensional signal feature library with real-time spectrum sensing, this invention achieves rapid analysis and accurate characterization of deep-space radar signal characteristics. The feature library is generated based on unsupervised clustering of historical data, covering various signal templates such as transient impulses, periodic coherence, and steady-state noise. It not only provides rich prior knowledge references but also supports online sliding window incremental updates, ensuring that the feature representation always conforms to changes in the actual environment. The real-time spectrum sensing adopts adaptive window time-frequency analysis. The window length is dynamically adjusted according to the instantaneous bandwidth and local stationarity of the signal, effectively balancing the contradiction between frequency resolution and time resolution, and overcoming the limitations of traditional fixed window analysis in non-stationary scenarios.

[0070] The dynamic link selection mechanism, based on a combination of feature matching scoring and resource constraints, intelligently selects the optimal processing path. When a cyclic stationary component or a highly sparse spectrum is detected, coherent accumulation or transient impact processing links are automatically triggered, avoiding the lag and subjectivity of manual intervention. Parallel evaluation of multiple links and conflict resolution rules further enhance the robustness of the system's decision-making, ensuring stable output even in complex signal environments.

[0071] The online parameter tuning module incorporates a reinforcement learning policy network, using multi-objective weighted reward feedback as the optimization direction to continuously fine-tune key parameters such as the variational mode decomposition penalty factor and the fractional Fourier transform rotation angle. Weights are adaptively updated through temperature-based Softmax mapping and advantage-driven mechanisms, ensuring both the flexibility of the optimization objective and maintaining system stability through lower bound constraints and a backoff mechanism. This closed-loop interaction between parameter optimization and link execution significantly improves the algorithm's ability to track time-varying disturbances and its anti-divergence performance.

[0072] This invention deeply integrates feature extraction, decision logic, and parameter optimization to form a fully adaptive processing chain that connects perception, decision-making, execution, and feedback. It not only solves the core pain point of dynamic mismatch of signal characteristics in deep space radar, but also has the versatility to be transferred to diverse scenarios such as industrial detection and medical diagnosis. It provides a complete, self-consistent, and engineering-feasible technical framework for signal processing in complex environments. Attached Figure Description

[0073] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0074] Figure 1 This is a flowchart illustrating the signal processing method applicable to various signal characteristics in Example 1. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0076] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0077] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0078] This application proposes a signal processing method applicable to multiple signal characteristics, combining... Figure 1 As shown, the method includes:

[0079] Step S1: Receive the input signal; extract real-time spectral features under the condition that the delay does not exceed the preset time budget. The real-time spectral features include at least: spectral sparsity, number of dominant spectral lines, intensity of cyclostationary components, spectral kurtosis and estimated signal-to-noise ratio.

[0080] In this embodiment, the time delay budget refers to the end-to-end processing time from the input signal to obtaining all real-time spectral features of this step, which is given by the real-time performance index of the host system in engineering. The number of dominant spectral lines refers to the number of significant spectral peaks above the noise floor that meet the minimum frequency spacing condition, which is obtained by peak search on the normalized power spectrum. The intensity of the cyclostationary component refers to the energy proportion or significance measure on the candidate cyclostationary frequency set, which comes from the estimation result of the spectral correlation density. The spectral kurtosis is used to identify non-Gaussian components. The estimated signal-to-noise ratio is obtained by converting the noise floor to the energy ratio within the signal band. Specifically, the default time delay budget is 50 ms, which can be adjusted from 20 to 100 ms, depending on the system response and the real-time constraints of the upper-layer application. The default sliding window length is a short window that adapts to the sampling rate (e.g., covering 25-75 ms), and the overlap is 50%-75%, which is determined based on station stability tests and offline playback verification. Furthermore, if this step includes feature calculation formulas, only the engineering interpretation of the symbols will be explained: weight, revenue, and window index correspond to feature weight, dimensionless revenue, and time window number, respectively. Their value ranges are given in the parameter section of the specification and do not change the calculation relationship. Optionally, when resources are scarce or the signal is temporarily mismatched, the process degenerates to extracting only spectral sparsity and estimated signal-to-noise ratio to ensure the lowest possible output delay. If necessary, when input data is missing or power spectrum estimation fails, the result of the previous valid window is used as a temporary measure and marked as low confidence. The mark is automatically cleared after the next window returns to normal.

[0081] The calculation of spectral features and the setting of threshold in step S1 are as follows:

[0082] Common settings: Within the sliding window, first obtain the normalized power spectrum vector using Welch / STFT, and then use statistics from similar templates and real-time SNR for threshold adaptation. To maintain consistency, the threshold is set using the following formula:

[0083] ,

[0084] in, Representation of features The threshold for judgment, in units of Consistent, Indicates categories in the template library The historical median of (transient, cyclically stationary, steady-state noise, etc.) This represents the SNR compensation coefficient, with a default value of 0.05 and an adjustable value of [0.01, 0.2]. This indicates the estimated signal-to-noise ratio for this window, in dB. Indicate category Historical SNR median, in dB;

[0085] a. Spectral sparsity: Hoyer sparsity measures the concentration of spectral energy.

[0086] ,

[0087] in, The sparsity is expressed as [0, 1]. For normalized power spectrum ( ), dimensionless, Frequency points, in points. They are respectively and Norm;

[0088] Threshold caliber: based on Sparsity is determined by default. , initial empirical values Adjustable [0.40, 0.80];

[0089] b. Number of dominant spectral lines: Estimate the noise floor on the logarithmic spectrum using the moving median and set a peak threshold and minimum frequency interval. Count the dominant peaks: If And the distance between adjacent peaks This is recorded as one dominant spectral line;

[0090] Threshold caliber: Upper limit criterion is adopted. ;default Adjustable [1.5, 6.0] initial empirical values Adjustable [1,10]; according to the common formula SNR compensation is obtained , here Indicates spectral resolution, in Hz. This represents the minimum physical separation bandwidth, measured in Hz. Noise floor estimate obtained by moving median filtering;

[0091] c. Cyclic stationary component strength: Based on the spectral correlation density (SCD) in the candidate cyclic frequency set. Energy ratio characterization:

[0092] ,

[0093] in, The cyclic steady-state strength is dimensionless. The SCD is estimated by FAM, with units consistent with the power spectrum. This is a conventional power spectrum. Cyclic frequency, in Hz. For the first Frequency points, in Hz;

[0094] Threshold caliber: based on Determine if a cycle is stationary; default. Recommended by the template, initial experience value Adjustable [0.05, 0.40], and calculated according to the common formula. Perform SNR compensation;

[0095] d. Spectral kurtosis: For each frequency point, calculate the fourth moment ratio over time, and use the full-spectrum statistic (such as the maximum value or 95th percentile) as the indicator.

[0096] ,by or Warehousing;

[0097] Threshold caliber: based on Non-Gaussian components are significant; empirical initial values. The adjustable range [0.10, 1.00] is compensated according to a common formula, where, Frequency point Time series amplitude or power spectrum estimation samples, This represents the average time within the window. Dimensionless;

[0098] e. Estimating the signal-to-noise ratio: using the noise floor set With signal set Estimating power ratio by partition and converting to dB, Select a frequency point below the noise floor threshold. Take the remaining frequency points or template annotation bands;

[0099] Threshold caliber: Provides the operating threshold. For link selection and compensation baseline, empirical initial values Adjustable [-6,6] dB, template categories provided. For reference;

[0100] Specifically, this step combines windowed spectral estimation with template statistics without exceeding the time delay budget constraint to form a stable and adaptive set of discriminants. When the sparsity is high, the energy is concentrated in a few frequency points, which is more suitable for transient or pulse-type processing. The number of dominant spectral lines is used as a direct quantification of structured information, which is helpful to distinguish between low-peak and high-peak spectral shapes. Cyclic stationary intensity is oriented towards periodic coherent characteristics and can directly serve coherent accumulation and compensation links. Spectral kurtosis is used to identify frequency bands that deviate from Gaussian noise and support the extraction of anomalous components. SNR estimation not only participates in link screening but also provides a compensation baseline for other thresholds that varies with the environment.

[0101] The common threshold is anchored to the template median and linearly corrected by real-time SNR, which facilitates cross-scenario migration and online steady-state convergence. The above features are given default values ​​and adjustable ranges according to a unified paradigm. In engineering, they can be gradually calibrated by combining sliding windows and human review marks to ensure that the consistency and operability of the judgment are maintained under low signal-to-noise ratio and non-stationary conditions.

[0102] In this embodiment, power spectrum estimation uses a practical window function and maintains a fixed overlap ratio to reduce leakage and variance. The window type and segment length are determined by offline verification before deployment. Threshold adaptation is based on the template median, linearly adjusted with the real-time signal-to-noise ratio, and limited by the maximum change range. Specifically, the default overlap ratio is 50%-75%, and a general-purpose window with sidelobe suppression characteristics is preferred. When the sampling rate changes or the frequency resolution requirement changes abruptly, the segment length is preferably kept no less than three times the minimum resolvable bandwidth to ensure the stability of peak counting and cyclic feature evaluation. Optionally, when computational resources are insufficient, the configuration degenerates to a shorter segment length and lower frequency resolution, while the minimum frequency distance for peak counting is increased on the template side to maintain criterion consistency.

[0103] Step S2: Match the real-time spectral features with the multi-dimensional signal feature library to obtain a set of candidate templates and their matching scores;

[0104] Specifically, candidate templates refer to entries organized and recorded according to signal type, showing typical feature ranges and recommended links; the matching score refers to the score of the fit between real-time features and the template statistical range, obtained by aggregating normalized deviations or interval landing points. For example, the default is to return the first 3 candidate templates, adjustable from 1 to 5; the default matching threshold is 0.6, adjustable from 0.4 to 0.9, set based on a trade-off between hit rate and false alarm rate on the historical replay set. Optionally, when multiple template scores are close and the difference is less than 0.05, a parallel candidate marking is triggered, and the next step is handled by resource constraints and conflict resolution. In abnormal situations, such as library retrieval timeout or library absence, the process reverts to the most recent successful matching result, using a maximum of 3 window periods.

[0105] Step S3: Based on the candidate template, calculate the selection score for multiple signal processing links in the processing model library. The score is based on: template matching score, real-time spectral characteristics, current signal-to-noise ratio, time delay budget and computing power constraints.

[0106] In this embodiment, the selection score is a comprehensive measure of the feasibility and expected effect of candidate links, which is obtained by weighting template fitting, real-time feature preference, and resource constraints in engineering. Computational constraints refer to resource status such as the number of currently available processing cores and the upper limit of computation time. The latency budget follows the definition from the previous step. Furthermore, the default weight configuration prioritizes template fitting (highest priority), followed by resource constraints, and then real-time feature preference; the specific proportions are given in the parameter section of the specification and can be fine-tuned online. Optionally, when the host system passes a low-latency priority flag, the weight of the latency item in the selection score is temporarily increased by one level. If necessary, if any link is expected to exceed the latency budget, its selection score is directly set to zero to avoid default.

[0107] Step S4: Determine the target link based on the selection scoring and conflict resolution rules;

[0108] Specifically, the conflict resolution rules refer to the priority and decision criteria when multiple link scores are close. In engineering, the order is: cycle robustness first, energy concentration second, and latency third. This order is used for resolution when the score difference is below a preset threshold. Further, the default proximity threshold is 0.03, adjustable from 0.01 to 0.10, based on the stability statistics of the offline validation set. Optionally, if the computing power is insufficient to support the preferred link, it automatically switches to the second-best link and records the resource cause label to ensure continuous output timing. In abnormal scenarios, if all link scores are below the minimum feasible threshold, the output maintains the previous window of links and reduces the update frequency until the scores recover.

[0109] Step S5: Execute the selected link to obtain the output signal;

[0110] In this embodiment, the output signal refers to the data format consistent with the selected link, including but not limited to time-domain reconstruction sequences or frequency-domain energy accumulation results, for direct consumption by subsequent detection or estimation units within the same system. To ensure interface consistency, the output includes a timestamp and link identifier. Furthermore, the default end-to-end latency limit is consistent with the latency budget. If a link execution failure within a single window occurs, a retry mechanism will be triggered (no more than once). If the retry still fails, the valid output of the previous window will be used with an exception flag.

[0111] Step S6: Calculate the reward feedback index and update the adjustable parameters within the link online accordingly;

[0112] The multidimensional signal feature library is organized according to types such as transient impulses, periodic coherence, steady-state or near-steady-state noise, and each template contains the statistical range of features and the mapping relationship of the recommended processing link.

[0113] Specifically, the reward feedback metrics refer to measurable quantities of output quality, including energy concentration, estimated signal-to-noise ratio, cyclic correlation peak and sidelobe ratio or detection statistics, calibration subsequence matching bias, etc. Online updates are triggered periodically, with the update interval and learning rate uniformly managed by a safety boundary. Further, the default update interval is adjustable every 1-5 windows, and the learning rate is set to a small step size range, determined based on both offline playback stability and online volatility. Optionally, when the reward continuously decreases and falls below a threshold, the trigger parameters revert to the most recent stable snapshot and the update cycle is frozen for one iteration.

[0114] In one embodiment, the construction of a multidimensional signal feature library includes:

[0115] Unsupervised clustering and silhouette coefficient filtering are performed on historical signal data to generate feature templates;

[0116] Stored categorized by signal type; each template records at least the spectral sparsity range, the range of the number of dominant spectral lines, the intensity range of cyclic stationary components, the spectral kurtosis range, the correlation time range, the typical signal-to-noise ratio range, and a list of recommended links;

[0117] The library supports incremental updates via a sliding window with human review and verification flags during the online phase;

[0118] Similarly, sliding window incremental updates refer to revising template statistics only based on newly added samples within the most recent period without rebuilding the entire database; the human review flag refers to a status bit where the validity of the template is manually confirmed after automatic updates. Furthermore, the default incremental update cycle is 10-50 windows, with an adjustable range set according to the data arrival rate; the default minimum number of samples must be accumulated before data can be added to the database, which can be set to 100-1000 records depending on the scenario. Optionally, when the human review delay exceeds the preset limit, the template is still allowed to take effect within a low-risk range, but the update magnitude is limited to no more than 10% of the existing range per update to reduce drift risk.

[0119] In one embodiment, the extraction of real-time spectral features employs adaptive window time-frequency analysis;

[0120] The window length is dynamically adjusted based on instantaneous bandwidth and local stationarity metrics, and features are updated using a sliding window recursive method to reduce latency.

[0121] When strong nonstationarity is detected, the window length is automatically shortened and the update frequency is increased.

[0122] In this embodiment, strong non-stationarity refers to a situation where statistics change significantly within a short period of time. In engineering practice, this can be triggered by a joint criterion of increased spectral entropy, decreased correlation time, or weakened cyclic characteristics. Instantaneous bandwidth and local stationarity measures are derived from existing feature calculations in this step, without adding new processing modules. Furthermore, the default window length adjustment step does not exceed 50% of the previous window length, and the update frequency is increased to a maximum of twice the original, set based on the upper limit measured by end-to-end latency and CPU usage. Optionally, when extremely low signal-to-noise ratios cause instability in the criterion, a minimum window length is used as a baseline, and the upper limit of the update frequency is temporarily relaxed by one level.

[0123] In one embodiment, the selection score includes at least two parts: feature matching score and resource constraint score; if multiple links have similar scores, conflict resolution is carried out in the order of priority cycle robustness, second priority energy aggregation, and then priority latency.

[0124] When both cyclic stationary characteristics and multiple dominant spectral lines exist simultaneously, coherent accumulation links are preferred.

[0125] For example, the aperture of multiple dominant spectral lines is given by the corresponding category of the template, usually limited by the upper limit of peak count and the minimum frequency spacing. When the estimated signal-to-noise ratio is in the marginal range, the priority remains unchanged, but a rapid comparison of coherent accumulation links and transient impulse links is allowed within a single window, with the immediate improvement of reward feedback serving as an additional basis for parallel resolution. If necessary, if neither link can meet the time delay budget, a steady-state noise suppression link is used to degrade to ensure output continuity.

[0126] In one embodiment, the signal processing link includes at least:

[0127] A) Transient impact chain: Performs variational mode decomposition and subsequent impulse noise reduction / reconstruction;

[0128] B) Coherent accumulation link: Performs coherent accumulation / coherent compensation and fractional-order domain energy accumulation;

[0129] C) Steady-state noise suppression link: Perform time-varying adaptive filtering / spectral entropy minimization and band-limited reconstruction;

[0130] All links allow for fine-tuning by the online parameter adjustment module during operation;

[0131] Furthermore, fine-tuning refers to small incremental updates within predetermined parameter boundaries, without altering the link structure; these parameter boundaries are given before deployment and calibrated via offline playback. Specifically, the step size of a single update does not exceed 10% of the parameter search range, and consecutive failed updates will trigger a halving of the step size until the minimum step size limit is reached. Optionally, hierarchical fine-tuning can be enabled for highly sensitive parameters such as rotation angle search or phase compensation, starting with coarse-grained adjustments and gradually fine-grained adjustments to reduce oscillations.

[0132] In one embodiment, when the spectral sparsity is high and the number of dominant spectral lines does not exceed a preset upper limit, the transient impact link is preferred; when the intensity of the cyclic stationary component exceeds the threshold or the number of dominant spectral lines exceeds the upper limit, the coherent accumulation link is preferred; the threshold is determined by the historical statistical median of the same type of template and combined with the real-time estimated signal-to-noise ratio for dynamic compensation, and the compensation amount is adjusted over time according to the window update strategy.

[0133] In this embodiment, dynamic compensation refers to the process of linearly or piecewise linearly adjusting the threshold based on the template median as the anchor point and the real-time estimated signal-to-noise ratio. The gain coefficient is calibrated using offline validation data and adjusted in small steps during the online phase. To avoid overcompensation, the threshold change within a single window is limited to the maximum variation. Optionally, when the estimated signal-to-noise ratio confidence is low, threshold updates are paused, and the previous valid threshold is used for one update cycle.

[0134] In one embodiment, the penalty factor of variational mode decomposition, the phase / Doppler compensation parameter in coherent accumulation, and the search range of the rotation angle for fractional domain processing are all finely adjusted by the online parameter adjustment module based on the reward feedback index, and boundaries and strides are set to prevent divergence and overfitting.

[0135] Specifically, divergence monitoring is achieved by observing the monotonicity and variance of the most recent windows of reward feedback; once the divergence criterion is triggered, it immediately reverts to the most recent stable parameter and reduces the learning rate. Furthermore, the default revert memory depth covers the most recent 3-10 windows; when necessary, in resource-constrained or frequent revert scenarios, updates of high-sensitivity parameters are temporarily frozen, retaining only the fine-tuning of low-sensitivity parameters.

[0136] In one embodiment, the reward feedback metric includes at least one of the following:

[0137] a) Energy concentration of the output signal;

[0138] b) Estimated signal-to-noise ratio of the output signal;

[0139] c) Cyclic autocorrelation peak value and sidelobe ratio or detection statistic;

[0140] d) Matching bias to known calibration subsequences;

[0141] During the online phase, the single objective or weighted combination of this indicator is used as the optimization objective;

[0142] In this embodiment, the weights of the weighted combination are in a probabilistic simplex and are subject to a lower bound constraint to prevent collapse; the payoff vector is the result of direction unification and dimensionless mapping; weight updates are performed in an advantage-driven manner and are equipped with three types of stability protection: learning rate, update period, and backoff threshold. Optionally, when the system is in a low signal-to-noise ratio or when the noise floor suddenly increases, the weight temperature is temporarily increased to distract attention and reduce policy jitter.

[0143] In step S6, to unify the indicators a) energy concentration, b) estimated signal-to-noise ratio, c) cyclic correlation peak and sidelobe ratio or detection statistic, and d) calibration subsequence matching bias (taking negative gains) into the same optimization framework, the indicators are first oriented and dimensionless mapped to obtain vectors. ; The transpose operator is used; based on this, a combined objective is constructed, and the normalization and online updating of the weights are provided;

[0144] 1) In each sliding window Internally, a weighted sum is used as the optimization objective and is employed to drive the update of parameters within the link:

[0145] ,

[0146] in, Display window The combined objective value, dimensionless. Indicates the first Individual indicators in the window The weights, ranging from [0,1], are dimensionless. Indicates the first Dimensionless returns of each indicator Pick These correspond to the four categories of indicators mentioned above. This is the index of the current window, in units of [number].

[0147] 2) Introduce a temperature-modified Softmax mapping to obtain weights on the probability simplex, thus avoiding negative weights and sums that are not equal to 1:

[0148] ,

[0149] in, Indicates the first The preference scores for each indicator are initialized from template suggestions and are dimensionless. This is a temperature parameter, default value 0.5, adjustable [0.2, 1.0], dimensionless. To indicate the summation index; perform lower bound and renormalization to suppress weight collapse:

[0150] , ,

[0151] in, This is the lower limit of the weight, defaulting to 0.05, adjustable to [0.0, 0.2]. It is a vector consisting entirely of 1s;

[0152] 3) In the online phase, the preference score is adaptively driven by the advantage signal, and steady-state constraints are applied through the learning rate and update cycle:

[0153] ,

[0154] in, This indicates the updated preference score in this window. This is the weighted learning rate, default 0.05, adjustable [0.01, 0.20], dimensionless. For immediate gains, This is the baseline of the moving average for this indicator (smoothing factor defaults to 0.2, adjustable [0.05, 0.5]), and the update period is... By default, it triggers every 1-5 windows; the number of windows can be adjusted to [1, 10]. In recent If the value is continuously below the threshold within a window, then it will revert to the most recent stable value. With link parameters, Default is 5, adjustable [3,10]. Default value is 0.02, adjustable [0.005, 0.05], all units are dimensionless;

[0155] 4) a) and c) should be as large as possible, directly mapped to [0,1]. b) is truncated to [0,1] after linear stretching. d) should be as small as possible, first taking the negative and then normalizing. Initially... and Read the suggested values ​​given by the template category (transient / cyclic stationary / steady-state noise), and recursively apply the above formula during the online phase;

[0156] 5) The weight space remains in a simplex form; it can be temporarily increased when the noise floor or signal-to-noise ratio is low. By distributing weights, when system latency is critical, [the following can be done]: If no improvement is made, the update is frozen once. If the output marked as valid by the reviewer decreases due to external factors, the rollback mechanism is paused for one round.

[0157] Specifically, this step unifies the four types of feedback indicators into the same benefit vector, and forms a single objective through weighted summation, which facilitates the formation of a clear optimization direction within the link; the weights are not statically configured, but are normalized by temperature and fall into an interpretable probability space, and then adapt to changes in the scenario in an incremental way driven by advantages.

[0158] This approach highlights the contribution of coherent metrics during periods of strong cyclical characteristics and enhances the impact of energy accumulation-related benefits when transient shocks dominate. The lower bound and renormalization of weights prevent excessive bias towards a single metric, while the learning rate, update cycle, and backoff threshold provide stability boundaries, preventing short-term fluctuations from causing strategy jitter. Combined with template initialization, it provides a usable default configuration in the initial deployment phase, gradually aligning with the actual distribution through small adjustments during the online phase. This strategy is decoupled from yet interacts with link parameter updates, providing consistent external driving signals for subsequent adjustable parameters such as phase compensation, penalty factors, and rotation angles.

[0159] In one embodiment, the online parameter adjustment module employs a reinforcement learning policy network. Its inputs include real-time spectral features, template matching scores, link operation status, and reward feedback indicators, and its output is the link parameter increment. The network is pre-trained with historical data and periodically updated with policy gradients during the online phase. To meet engineering safety requirements, an upper limit for the update period, an upper limit for the learning rate, a performance rollback mechanism, and latency constraints are set.

[0160] Furthermore, the pre-trained samples cover three typical template types: transient impact, periodic coherence, and steady-state noise. Policy updates during the online phase are performed on a window-based cycle, with version numbers and timestamps recorded for auditing and reproducibility. Optionally, within a window interval where low returns occur consecutively and the detection threshold is approached, policy updates switch to a conservative mode, allowing only minor adjustments to low-sensitivity parameters to ensure latency constraints are met.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0162] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A signal processing method applicable to multiple signal characteristics, characterized in that, include, Step S1: Receive input signal; Real-time spectral feature extraction is performed under the condition that the delay does not exceed the preset time budget. The real-time spectral features include at least: spectral sparsity, number of dominant spectral lines, intensity of cyclostationary components, spectral kurtosis, and estimated signal-to-noise ratio. Step S2: Match the real-time spectral features with the multi-dimensional signal feature library to obtain a set of candidate templates and their matching scores; Step S3: Based on the candidate template, calculate the selection score for multiple signal processing links in the processing model library. The score is based on: template matching score, real-time spectral characteristics, current signal-to-noise ratio, time delay budget and computing power constraints. Step S4: Determine the target link based on the selection scoring and conflict resolution rules; Step S5: Execute the selected link to obtain the output signal; Step S6: Calculate the reward feedback index and update the adjustable parameters within the link online accordingly.

2. The signal processing method applicable to multiple signal characteristics as described in claim 1, characterized in that, The construction of the multidimensional signal feature library includes: Unsupervised clustering and silhouette coefficient filtering are performed on historical signal data to generate feature templates; Stored categorized by signal type; each template records at least the spectral sparsity range, the range of the number of dominant spectral lines, the intensity range of cyclic stationary components, the spectral kurtosis range, the correlation time range, the typical signal-to-noise ratio range, and a list of recommended links; The library supports incremental updates via sliding windows with human review and approval flags during the online phase.

3. The signal processing method applicable to multiple signal characteristics as described in claim 1, characterized in that, Real-time spectral features are extracted using adaptive window time-frequency analysis; The window length is dynamically adjusted based on instantaneous bandwidth and local stationarity metrics, and features are updated using a sliding window recursive method to reduce latency. When strong nonstationarity is detected, the window length is automatically shortened and the update frequency is increased.

4. The signal processing method applicable to multiple signal characteristics as described in claim 1, characterized in that, The selection score includes at least two parts: feature matching score and resource constraint score; if multiple links have similar scores, conflict resolution is carried out in the order of priority cycle robustness, second priority energy aggregation, and then priority time delay. When both cyclic stationary characteristics and multiple dominant spectral lines exist simultaneously, coherent accumulation links are preferred.

5. The signal processing method applicable to multiple signal characteristics as described in claim 1, characterized in that, The signal processing link includes at least: A) Transient impact chain: Performs variational mode decomposition and subsequent impulse noise reduction / reconstruction; B) Coherent accumulation link: Performs coherent accumulation / coherent compensation and fractional-order domain energy accumulation; C) Steady-state noise suppression link: Perform time-varying adaptive filtering / spectral entropy minimization and band-limited reconstruction; All links can be fine-tuned by the online parameter adjustment module during operation.

6. The signal processing method applicable to multiple signal characteristics as described in claim 5, characterized in that, When the spectral sparsity is high and the number of dominant spectral lines does not exceed the preset upper limit, the transient impact link is preferred; when the intensity of the cyclic stationary component exceeds the threshold or the number of dominant spectral lines exceeds the upper limit, the coherent accumulation link is preferred. The threshold is determined by dynamic compensation based on the historical statistical median of similar templates and real-time estimated signal-to-noise ratio. The compensation amount is adjusted over time according to a window update strategy.

7. The signal processing method applicable to multiple signal characteristics as described in claim 5, characterized in that, The penalty factor of variational mode decomposition, the phase / Doppler compensation parameter in coherent accumulation, and the search range of rotation angle in fractional domain processing are all finely adjusted by the online parameter adjustment module based on the reward feedback index, and boundaries and strides are set to prevent divergence and overfitting.

8. The signal processing method applicable to multiple signal characteristics as described in claim 1, characterized in that, The reward feedback metrics include at least one of the following: a) Energy concentration of the output signal; b) Estimated signal-to-noise ratio of the output signal; c) Cyclic autocorrelation peak value and sidelobe ratio or detection statistic; d) Matching bias to known calibration subsequences; During the online phase, the single objective or weighted combination of this indicator is used as the optimization objective; In step S6, to unify the indicators a) energy concentration, b) estimated signal-to-noise ratio, c) cyclic correlation peak and sidelobe ratio or detection statistic, and d) calibration subsequence matching bias into the same optimization framework, the indicators are first subjected to direction unification and dimensionless mapping to obtain vectors. ; The transpose operator is used; based on this, a combined objective is constructed, and the normalization and online updating of the weights are provided; 1) In each sliding window Internally, a weighted sum is used as the optimization objective and is employed to drive the update of parameters within the link: ， in, Display window The combined objective value, dimensionless. Indicates the first Individual indicators in the window The weights, ranging from [0,1], are dimensionless. Indicates the first Dimensionless returns of each indicator Pick These correspond to the four categories of indicators mentioned above. This is the index of the current window, in units of [number]. 2) Introduce a temperature-based Softmax mapping to obtain weights on the probabilistic simplex: ， in, Indicates the first The preference scores for each indicator are initialized from template suggestions and are dimensionless. The parameter is temperature, dimensionless. To indicate the summation index; perform lower bound and renormalization to suppress weight collapse: ， ， in, As the lower limit of the weight, It is a vector consisting entirely of 1s; 3) In the online phase, the preference score is adaptively driven by the advantage signal, and steady-state constraints are applied through the learning rate and update cycle: ， in, This indicates the updated preference score in this window. The learning rate is for weights; it is dimensionless. For immediate gains, The moving average baseline for this indicator, with an update period of [missing information]. By default, it triggers every 1-5 windows; if In recent If the value is continuously below the threshold within a window, then it will revert to the most recent stable value. Link parameters.

9. The signal processing method applicable to multiple signal characteristics as described in claim 5, characterized in that, The online parameter adjustment module employs a reinforcement learning policy network. Its inputs include real-time spectral features, template matching scores, link operation status, and reward feedback indicators, while the output is the link parameter increment. The network is pre-trained with historical data and periodically updated using policy gradients during the online phase. To ensure engineering safety, an upper limit for the update cycle, an upper limit for the learning rate, a performance rollback mechanism, and latency constraints are set.

10. The signal processing method applicable to multiple signal characteristics as described in claim 1, characterized in that, The calculation of the spectral features and the setting of the threshold in step S1 are as follows: Common settings: Within the sliding window, first obtain the normalized power spectrum vector using Welch / STFT, and then use statistics from similar templates and real-time SNR for threshold adaptation. To maintain consistency, the threshold is set using the following formula: ， in, Representation of features The threshold for judgment, in units of Consistent, Indicates categories in the template library Historical median Indicates the SNR compensation coefficient. This indicates the estimated signal-to-noise ratio for this window, in dB. Indicate category Historical SNR median, in dB; a. Spectral sparsity: Hoyer sparsity measures the concentration of spectral energy. ， in, The sparsity is expressed as [0, 1]. The normalized power spectrum is dimensionless. Frequency points, in points. They are respectively and Norm; Threshold caliber: based on Sparsity is determined by default. , initial empirical values ; b. Number of dominant spectral lines: Estimate the noise floor on the logarithmic spectrum using the moving median and set a peak threshold and minimum frequency interval. Count the dominant peaks: If And the distance between adjacent peaks This is recorded as one dominant spectral line; Threshold caliber: Upper limit criterion is adopted. ;default , initial empirical values According to the common formula SNR compensation is obtained , here Indicates spectral resolution, in Hz. This represents the minimum physical separation bandwidth, measured in Hz. Noise floor estimate obtained by moving median filtering; c. Cyclic stationary component strength: Based on the spectral correlation density (SCD) in the candidate cyclic frequency set. Energy ratio characterization: ， in, The cyclic steady-state strength is dimensionless. The SCD is estimated by FAM, with units consistent with the power spectrum. This is a conventional power spectrum. Cyclic frequency, in Hz. For the first Frequency points, in Hz; Threshold caliber: based on Determine if a cycle is stationary; default. Recommended by the template, initial experience value and according to the common formula Perform SNR compensation; d. Spectral kurtosis: For each frequency point, the fourth-order moment ratio is calculated in time, and the full-spectrum statistic is used as the indicator. ,by or Warehousing; Threshold caliber: based on Non-Gaussian components are significant; empirical initial values. Compensation is made according to a common formula, where, Frequency point Time series amplitude or power spectrum estimation samples, This represents the average time within the window. Dimensionless; e. Estimating the signal-to-noise ratio: using the noise floor set With signal set Estimating power ratio by partition and converting to dB, Select a frequency point below the noise floor threshold. Take the remaining frequency points or template annotation bands; Threshold caliber: Provides the operating threshold. For link selection and compensation baseline, empirical initial values Template categories are provided Use as a reference.

Images