Feature real-time extraction method based on ultrahigh frequency original signal

Abstract

The invention discloses a feature real-time extraction method based on an ultrahigh-frequency original signal, which relates to the field of signal feature extraction, and comprises the following steps: setting a dynamic threshold to trigger acquisition logic according to transient pulse time domain distribution features of the ultrahigh-frequency signal; capturing an original signal segment in real time based on acquisition logic, and synchronously completing baseline calibration and amplitude normalization preprocessing; pre-processed signal segments are acquired, a space-time correlation filtering model is constructed, and adaptive noise filtering is performed based on energy distribution correlation of signals at adjacent acquisition moments so as to retain high-frequency details of transient characteristics; acquisition logic is triggered through a dynamic threshold value, the threshold value is adaptively adjusted according to pulse density, effective signal segments are accurately captured, baseline calibration and amplitude normalization are synchronously completed, signal preprocessing stability is improved, and invalid data interference is reduced.

Description

Technical Field

[0001] This invention relates to the field of signal feature extraction technology, specifically a real-time feature extraction method based on ultra-high frequency raw signals. Background Technology

[0002] UHF raw signals are widely used in communication, radar, and electrical discharge applications. These signals are unprocessed, retaining their original waveforms and spectral characteristics, and serve as fundamental data for signal analysis and equipment testing.

[0003] Patent application No. 202010883082.4 discloses a method for detecting partial discharge in power distribution network switchgear using an ultra-high frequency (UHF) sensor. The method includes: an UHF intelligent sensing module receiving a partial discharge UHF signal emitted by power equipment; filtering and amplifying the UHF signal to improve the signal-to-noise ratio; performing logarithmic detection on the UHF signal to extract the envelope signal and reduce its frequency; processing the signal in an MCU after logarithmic detection and peak hold, encapsulating the partial discharge data for each power frequency cycle into a data frame and uploading the data; and a partial discharge pulse signal extraction module extracting the characteristic parameters of the partial discharge pulse signal in real time. This application aims to address the problems of "current issues in the detection and diagnosis of live-line equipment in power distribution rooms, such as heavy workload for maintenance personnel, low cost-effectiveness of detection methods, and low accuracy in detecting old and faulty equipment. Furthermore, the original signal collected by the UHF sensor is of low quality, contains a large amount of noise, and has a low signal-to-noise ratio, making it unsuitable for direct analysis and diagnosis."

[0004] However, for feature extraction techniques of UHF raw signals, the accuracy or timeliness of feature extraction in existing technologies need to be improved.

[0005] To address this, we propose a real-time feature extraction method based on UHF raw signals. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a real-time feature extraction method based on ultra-high frequency raw signals, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions;

[0008] This invention discloses a real-time feature extraction method based on ultra-high frequency raw signals, comprising:

[0009] Based on the transient pulse time-domain distribution characteristics of UHF signals, a dynamic threshold is set to trigger the acquisition logic. The original signal segments are captured in real time based on this logic, and baseline calibration and amplitude normalization preprocessing are performed simultaneously. The preprocessed signal segments are acquired, and a spatiotemporal correlation filtering model is constructed. Adaptive noise filtering is performed based on the energy distribution correlation of the signal at adjacent acquisition times to preserve high-frequency details of the transient features. The decomposition scale is adaptively determined for the denoised signal based on the spectral peak distribution, and multi-band feature components are separated and extracted. The temporal abrupt change slope and frequency domain energy spectrum entropy of each frequency band component are extracted and integrated into a multi-dimensional feature set, forming an initial feature vector for quantifying transient feature attributes. Based on the distribution density of the initial feature vector, a real-time feature contribution evaluation logic is set, and the weights of each parameter are dynamically and iteratively adjusted based on the evaluation logic. High-confidence core feature sequences are selected based on the parameter weights and a feature stability verification model. A dynamic update interface for the feature sequences is constructed simultaneously, and the application interface receives new feature data and updates the output core feature sequences.

[0010] Furthermore, the dynamic threshold-triggered acquisition logic includes:

[0011] Calculate the dynamic trigger threshold based on the time-domain amplitude statistical characteristics of transient pulses of UHF signals:

[0012] ;

[0013] In the formula: For dynamic trigger thresholds; This is the adaptive adjustment coefficient; The standard deviation of the background noise during the data collection period; This represents the mean of the background noise.

[0014] When the instantaneous amplitude of the signal exceeds The system acquires signal segments triggered by time and synchronously records the time-domain coordinates of the trigger moment.

[0015] in, ∈[2.5, 4], and its value follows the following: when ρ≥5 pulses / second, k increases linearly with ρ and is controlled within (3.0, 4.0]; when ρ<5 pulses / second, k decreases linearly with ρ and is controlled within [2.5~3.0]. ρ represents the signal pulse density, that is, the number of pulses exceeding the initial threshold per unit time. The initial threshold is... .

[0016] Furthermore, the baseline calibration and amplitude normalization preprocessing includes:

[0017] The baseline signal is obtained by estimating the baseline of the acquired original signal segments using a piecewise linear fitting method, and the baseline correction signal is obtained by subtracting the baseline signal from the original signal.

[0018] Normalization is performed based on the maximum amplitude of the baseline correction signal to obtain the normalized signal:

[0019] ;

[0020] In the formula: For normalized signals; This is the baseline correction signal; To correct the maximum amplitude of the signal;

[0021] The amplitude range of the normalized signal is constrained to... .

[0022] Furthermore, when constructing the spatiotemporal correlation filtering model, the signal energy correlation factor between adjacent acquisition times is denoted as... Adaptive noise filtering is performed using correlation factors;

[0023] Among them, correlation factors The value follows:

[0024] ;

[0025] In the formula: for Moment signal energy; This indicates taking the maximum value within the parentheses; This is the time decay coefficient; Indicates adjacent data collection times;

[0026] when High-frequency detail components are preserved when the value is ≥0.6; otherwise, noise components are attenuated.

[0027] Among them, the time decay coefficient ∈[0.01,0.1], the value is larger when the signal pulse interval is less than 5ms, and smaller when the signal pulse interval is greater than 20ms.

[0028] Furthermore, the decomposition scale is adaptively determined through the following steps:

[0029] Step 1: Calculate the decomposition scale based on the peak distribution of the denoised signal's spectrum:

[0030] ;

[0031] In the formula: For decomposition scale; These are the maximum and minimum frequencies of the spectral peak value; The average frequency interval between adjacent spectral peaks; This is the scale correction factor. ∈ (0.8, 1.2); Indicates rounding up;

[0032] Step 2: Based on the decomposition scale Separate the signal into Each frequency band characteristic component;

[0033] Among them, when the average frequency interval between adjacent peaks is small, The larger the value, the greater the average frequency interval between adjacent peaks. The smaller the value.

[0034] Furthermore, when extracting the temporal abrupt change slope of each frequency band component, the temporal abrupt change slope of each frequency band component is calculated through a sliding window, and the width of the sliding window is set to 2N+1, where N is a positive integer;

[0035] ;

[0036] In the formula: Let be the slope of the time-domain abrupt change of the j-th frequency band component; Let be the amplitude of the j-th frequency band component at the n-th point within the sliding window;

[0037] Among them, the slope threshold is manually set by the user. When the slope is not less than the slope threshold, the j-th frequency band component is marked as a valid mutation feature.

[0038] Furthermore, the real-time evaluation logic for the feature contribution is as follows:

[0039] Calculate the contribution of each feature based on the distribution density of the initial feature vector:

[0040] ;

[0041] In the formula: The contribution of the m-th feature; Let m be the local density of the m-th feature; Let be the distance between the m-th feature and the high-density features; The dimension of the initial feature vector;

[0042] Based on the above Dynamically iteratively adjust the weights of each parameter:

[0043] Set initial weight values ;

[0044] Weight update amount in the k-th iteration ;

[0045] Updated weights ;

[0046] in, Represents the iteration step size coefficient. ∈(0.1,0.3), when When the fluctuation range is large in consecutive iterations The smaller the value, the more... When the fluctuation range is small, The larger the value, the better.

[0047] Furthermore, the feature stability verification model is as follows:

[0048] Calculate the stability index of the core feature sequence ;

[0049] when If the value is ≥0.85, the feature sequence is considered stable; otherwise, feature re-selection is triggered.

[0050] In the formula: The length of the core feature sequence; Let be the amplitude of the p-th core feature at time t.

[0051] Furthermore, the dynamic update interface, when constructed, conforms to:

[0052] The feature update trigger condition is set as follows: the cosine similarity between the new feature data and the current core feature sequence is less than a preset threshold, and the preset threshold is set to [0,7,0.85].

[0053] The cosine similarity is:

[0054] ;

[0055] In the formula: For feature dimensions; These are the q-th component of the new feature data and the q-th component of the current core feature sequence, respectively.

[0056] When an update is triggered, an updated core feature sequence is generated based on a weighted fusion of the new feature data and historical feature data:

[0057] ;

[0058] In the formula: The updated core feature sequence; This represents the total number of historical feature data entries. These are the weighting coefficients; Let q be the q-th component of the k-th historical feature data;

[0059] in, , Indicates the time of data collection for new features. This represents the collection time of the k-th historical feature data, and the sum of the weight coefficients of all historical feature data is 1.

[0060] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0061] This invention provides a real-time feature extraction method based on UHF raw signals. During the execution of this method, the acquisition logic is triggered by a dynamic threshold, the threshold is adaptively adjusted according to the pulse density, the effective signal segments are accurately captured, and baseline calibration and amplitude normalization are completed simultaneously, thereby improving the stability of signal preprocessing and reducing interference from invalid data.

[0062] Based on the spatiotemporal correlation filtering model, the adjacent energy correlation is used to adaptively denoise, effectively preserving transient high-frequency details and enhancing the signal-to-noise ratio. The decomposition scale is adaptively determined according to the spectral peak distribution to achieve accurate separation of multi-band features.

[0063] By dynamically adjusting parameter weights in conjunction with contribution assessment, filtering high-confidence core features through stability verification, and simultaneously building a dynamic update interface to integrate new data, the feature sequence is continuously optimized to improve the real-time performance, accuracy, and reliability of transient feature extraction. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0065] Figure 1 This is a flowchart illustrating a method for real-time feature extraction from UHF raw signals. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] The present invention will be further described below with reference to embodiments.

[0068] Example:

[0069] This embodiment presents a real-time feature extraction method based on ultra-high frequency raw signals, such as... Figure 1 As shown, it includes:

[0070] Based on the transient pulse time-domain distribution characteristics of UHF signals, a dynamic threshold is set to trigger the acquisition logic. Based on the acquisition logic, the original signal segments are captured in real time and baseline calibration and amplitude normalization preprocessing are completed simultaneously.

[0071] The dynamic threshold-triggered acquisition logic includes:

[0072] Calculate the dynamic trigger threshold based on the time-domain amplitude statistical characteristics of transient pulses of UHF signals:

[0073] ;

[0074] In the formula: For dynamic trigger thresholds; This is the adaptive adjustment coefficient; The standard deviation of the background noise during the data collection period; This represents the mean of the background noise.

[0075] The above formula combines the time-domain amplitude statistical characteristics of transient pulses of UHF signals, and takes the standard deviation and mean of background noise as the basis. The threshold is dynamically adjusted by adaptive adjustment coefficient k. The coefficient k will change flexibly according to the signal pulse density ρ per unit time. When the pulse is dense, k is increased to avoid false sampling, and k is decreased to prevent missed sampling when the pulse is sparse, so that the signal segment acquisition can accurately respond to different pulse intensity scenarios.

[0076] When the instantaneous amplitude of the signal exceeds The system acquires signal segments triggered by time and synchronously records the time-domain coordinates of the trigger moment.

[0077] in, ∈[2.5, 4], and its value follows the following: when ρ≥5 pulses / second, k increases linearly with ρ and is controlled within (3.0, 4.0]; when ρ<5 pulses / second, k decreases linearly with ρ and is controlled within [2.5~3.0]. ρ represents the signal pulse density, that is, the number of pulses exceeding the initial threshold per unit time. The initial threshold is... ;

[0078] The baseline calibration and amplitude normalization preprocessing steps include:

[0079] The baseline signal is obtained by estimating the baseline of the acquired original signal segments using a piecewise linear fitting method, and the baseline correction signal is obtained by subtracting the baseline signal from the original signal.

[0080] Normalization is performed based on the maximum amplitude of the baseline correction signal to obtain the normalized signal:

[0081] ;

[0082] In the formula: For normalized signals; This is the baseline correction signal; To correct the maximum amplitude of the signal;

[0083] The amplitude range of the normalized signal is constrained to... ;

[0084] The preprocessed signal segments are acquired, a spatiotemporal correlation filtering model is constructed, and adaptive noise filtering is performed based on the correlation of energy distribution of the signal at adjacent acquisition times to preserve high-frequency details of transient features.

[0085] When constructing the spatiotemporal correlation filtering model, the signal energy correlation factor between adjacent acquisition times is denoted as... Adaptive noise filtering is performed using correlation factors;

[0086] Among them, correlation factors The value follows:

[0087] ;

[0088] In the formula: for Moment signal energy; This indicates taking the maximum value within the parentheses; This is the time decay coefficient; Indicates adjacent data collection times;

[0089] The above formula calculates the ratio of signal energy at adjacent acquisition times and introduces a time attenuation coefficient C to reflect the energy correlation. When the correlation factor is not less than 0.6, high-frequency details are preserved, and noise is attenuated. The value of C is dynamically adjusted with the signal pulse interval. When the interval is small, a larger value is taken to enhance the correlation judgment, and when the interval is large, a smaller value is taken to weaken the interference, thereby achieving adaptive noise filtering.

[0090] when High-frequency detail components are preserved when the value is ≥0.6; otherwise, noise components are attenuated.

[0091] Among them, the time decay coefficient ∈[0.01,0.1], the value is larger when the signal pulse interval is less than 5ms, and smaller when the signal pulse interval is greater than 20ms;

[0092] The decomposition scale is adaptively determined based on the spectral peak distribution of the denoised signal to separate and extract multi-band feature components;

[0093] The decomposition scale is adaptively determined through the following steps:

[0094] Step 1: Calculate the decomposition scale based on the peak distribution of the denoised signal's spectrum:

[0095] ;

[0096] In the formula: For decomposition scale; These are the maximum and minimum frequencies of the spectral peak value; The average frequency interval between adjacent spectral peaks; This is the scale correction factor. ∈ (0.8, 1.2); Indicates rounding up;

[0097] The above formula is based on the maximum and minimum frequency range of the peak values ​​in the denoised signal spectrum, the average interval between adjacent peak values, and the scaling correction coefficient. Determine the number of frequency bands to be decomposed and the correction factor. The value will vary with the peak interval. When the interval is small, a larger value is used to improve the scale adaptability, and when the interval is large, a smaller value is used to avoid over-decomposition, so that the separation of multi-band characteristic components is more in line with the characteristics of the signal spectrum distribution.

[0098] Step 2: Based on the decomposition scale Separate the signal into Each frequency band characteristic component;

[0099] Among them, when the average frequency interval between adjacent peaks is small, The larger the value, the greater the average frequency interval between adjacent peaks. The smaller the value;

[0100] Extract the temporal abrupt change slope and frequency domain energy spectrum entropy value of each frequency band component, integrate the temporal abrupt change slope and frequency domain energy spectrum entropy value into a multi-dimensional feature set, and form an initial feature vector for quantifying transient feature attributes;

[0101] When extracting the temporal abrupt change slope of each frequency band component, the temporal abrupt change slope of each frequency band component is calculated through a sliding window, and the width of the sliding window is set to 2N+1, where N is a positive integer;

[0102] ;

[0103] In the formula: Let be the slope of the time-domain abrupt change of the j-th frequency band component; Let be the amplitude of the j-th frequency band component at the n-th point within the sliding window;

[0104] The above formula calculates the slope by the ratio of the amplitude difference at both ends of the sliding window to the window width. The window width is set to 2N+1 to cover sufficient signal details. When the slope is not less than the user-defined threshold, it is marked as an effective abrupt change feature, which intuitively captures the degree of abrupt change of each frequency band component in the time domain and accurately identifies valuable transient changes.

[0105] Among them, the slope threshold is manually set by the user. When the slope is not less than the slope threshold, the j-th frequency band component is marked as an effective abrupt change feature;

[0106] Based on the distribution density of the initial feature vector, a real-time evaluation logic for feature contribution is set, and the weights of each parameter are dynamically and iteratively adjusted based on the evaluation logic.

[0107] The real-time evaluation logic for feature contribution is as follows:

[0108] Calculate the contribution of each feature based on the distribution density of the initial feature vector:

[0109] ;

[0110] In the formula: The contribution of the m-th feature; Let m be the local density of the m-th feature; Let be the distance between the m-th feature and the high-density features; The dimension of the initial feature vector;

[0111] based on Dynamically iteratively adjust the weights of each parameter:

[0112] Set initial weight values ;

[0113] Weight update amount in the k-th iteration ;

[0114] Updated weights ;

[0115] in, Represents the iteration step size coefficient. ∈(0.1,0.3), when When the fluctuation range is large in consecutive iterations The smaller the value, the more... When the fluctuation range is small, The larger the value;

[0116] The above formula combines the local density of features and the distance to high-density features to calculate the contribution. Based on this, the parameter weights are dynamically adjusted iteratively. During the iteration process, the step size coefficient... The weighting is adjusted flexibly according to the fluctuation range of the contribution. When the fluctuation is large, a smaller value is used for stable updates, and when the fluctuation is small, a larger value is used to accelerate convergence, so that the weighting can adapt to the changes in the importance of features in real time.

[0117] High-confidence core feature sequences are selected based on parameter weights combined with feature stability verification models. A dynamic update interface for feature sequences is constructed simultaneously. The application interface receives new feature data and updates the output core feature sequences.

[0118] The feature stability verification model is as follows:

[0119] Calculate the stability index of the core feature sequence ;

[0120] when If the value is ≥0.85, the feature sequence is considered stable; otherwise, feature re-selection is triggered.

[0121] In the formula: The length of the core feature sequence; Let be the amplitude of the p-th core feature at time t;

[0122] The above formula quantifies feature stability by calculating the ratio of the sum of amplitude differences between adjacent time steps of the core feature sequence to the sequence length. A feature is considered stable when the index is not lower than 0.85; otherwise, a re-screening is triggered. This simple and intuitive difference calculation reflects the fluctuation of features over time, thereby ensuring the reliability of the core features.

[0123] When a dynamically updated interface is built, it follows the following rules:

[0124] The feature update trigger condition is set as follows: the cosine similarity between the new feature data and the current core feature sequence is less than a preset threshold, and the preset threshold is set to [0,7,0.85].

[0125] The cosine similarity is:

[0126] ;

[0127] In the formula: For feature dimensions; These are the q-th component of the new feature data and the q-th component of the current core feature sequence, respectively.

[0128] When an update is triggered, an updated core feature sequence is generated based on a weighted fusion of the new feature data and historical feature data:

[0129] ;

[0130] In the formula: The updated core feature sequence; This represents the total number of historical feature data entries. These are the weighting coefficients; Let q be the q-th component of the k-th historical feature data;

[0131] in, , Indicates the time of data collection for new features. This indicates the collection time of the k-th historical feature data, and the sum of the weight coefficients of all historical feature data is 1;

[0132] The above optimization settings measure the difference between the new feature data and the current core feature sequence by calculating the product of each dimension component and the ratio of the product of the modulus. When the similarity is lower than the preset threshold, an updated sequence is generated based on the weighted fusion of the new features and historical features. The weight of historical features decays over time, incorporating new information while retaining historical accumulation, ultimately achieving dynamic optimization of the feature sequence.

[0133] In this embodiment, the method described above precisely triggers signal acquisition through dynamic thresholds, simultaneously completes baseline calibration and normalization to reduce interference, effectively filters out noise and retains high-frequency details through spatiotemporal correlation filtering, achieves accurate feature separation through adaptive frequency band decomposition, constructs a feature set by combining temporal domain abrupt change slope and frequency domain entropy value, dynamically iterates weights and verifies stability to ensure the reliability of core features, and dynamically updates the interface to continuously optimize the feature sequence, effectively improving the real-time performance, accuracy and stability of UHF signal feature extraction, while enhancing the reliability of signal analysis in application scenarios.

[0134] Based on the method in the above embodiments, an application example of this method is provided:

[0135] In the online monitoring of partial discharge in a 110kV GIS equipment of a substation, a real-time feature extraction method based on UHF raw signals was used to achieve accurate identification of discharge signals. The specific application process is as follows:

[0136] The monitoring system first acquires the UHF signals generated during equipment operation in real time. By analyzing historical monitoring data, the mean background noise in this scenario is determined to be 0.2V, with a standard deviation of 0.1V. Based on the time-domain distribution characteristics of the transient pulses of the UHF signal, the system initiates dynamic threshold-triggered acquisition logic:

[0137] When the number of pulses exceeding the initial threshold per unit time is 6 / second, the adaptive adjustment coefficient k increases linearly with ρ to 3.5; at this time, the dynamic trigger threshold is calculated to be 1.05V. When the instantaneous amplitude of the signal exceeds 1.05V, the system immediately triggers signal segment acquisition and synchronously records the time-domain coordinates of the trigger moment.

[0138] After acquisition, the system performs baseline calibration and amplitude normalization preprocessing on the original signal segments. The baseline signal is estimated by a piecewise linear fitting method, and the baseline correction signal is obtained by subtracting the baseline signal from the original signal. Then, normalization processing is performed based on the maximum amplitude of the correction signal to constrain the signal amplitude within the range of [-1, 1], resulting in a standardized signal segment.

[0139] To effectively filter out noise, the system constructs a spatiotemporal correlation filtering model. When processing signals from adjacent acquisition times, the energy correlation factor between the two is calculated: if the signal energy at the previous time is 0.6mW and at the current time is 0.5mW, the maximum value is 0.6mW. Since the signal pulse interval is 3ms and the time attenuation coefficient C is 0.08, the correlation factor, rounded to three decimal places, is 0.067. Because this value is less than 0.6, the system attenuates this noise component.

[0140] When the signal energies of another set of adjacent time points are 0.7mW and 0.65mW respectively, with a pulse interval of 4ms, the correlation factor is calculated to retain three decimal places as 0.074, which is still less than 0.6. The noise is further attenuated, and the high-frequency detail signal containing transient features is finally preserved.

[0141] After denoising, the system adaptively determines the decomposition scale based on the spectral peak distribution. Spectral analysis shows that the maximum frequency of the signal's spectral peaks is 1500MHz, the minimum frequency is 500MHz, and the average frequency interval between adjacent spectral peaks is 200MHz. Due to the large interval, the scale correction coefficient β is set to 0.9. The decomposition scale is calculated by rounding up to 5, thus separating the signal into 5 frequency bands of characteristic components.

[0142] For each frequency band component, the system calculates the time-domain abrupt change slope using a sliding window. For example, the amplitudes of the third frequency band component within the window are -0.3, -0.1, 0.2, 0.4, and 0.6 respectively, resulting in a calculated time-domain abrupt change slope of 0.25.

[0143] The user presets a slope threshold of 0.2. Since 0.25 ≥ 0.2, the frequency band component is marked as a valid abrupt change feature. At the same time, the frequency domain energy spectrum entropy value of each frequency band component is extracted, where the entropy value of the valid frequency band is 0.68. The temporal abrupt change slope and the frequency domain energy spectrum entropy value are integrated to form a 5-dimensional initial feature vector.

[0144] Based on the initial feature vector distribution density, the system initiates real-time feature contribution evaluation logic. Taking the second feature as an example, its local density is 0.8, its distance from high-density features is 0.4, and its initial feature vector dimension is 5, resulting in a calculated contribution of 0.4. The system sets the initial weight value to 0.2, and the iteration step size coefficient λ is set to 0.2 due to the small fluctuation in contribution. The weight update amount in the first iteration is 0.08, and the updated weight is 0.28. Through multiple iterations, the weights of each parameter are dynamically adjusted.

[0145] Combining a feature stability verification model, the system filters core feature sequences: the core feature sequence has a length of 3, and the calculated stability index is 0.9. Since 0.9 ≥ 0.85, the feature sequence is determined to be stable, and three high-confidence core features are identified. Simultaneously, a dynamic update interface is constructed, with a preset cosine similarity threshold of 0.8 between new feature data and the core feature sequence. When the cosine similarity calculated between newly collected feature data and the current core feature sequence is 0.75 (less than 0.8), an update is triggered: three historical feature data points are merged with the new data, where the weights of the historical data are distributed according to time decay as 0.3, 0.2, and 0.2, and the weight of the new data is 0.3 (the total weight sum is 1), generating an updated core feature sequence, thus achieving dynamic optimization of the monitoring model.

[0146] In summary, the method described in the above embodiments, during execution, triggers the acquisition logic with a dynamic threshold, adaptively adjusts the threshold based on pulse density, accurately captures effective signal segments, and simultaneously completes baseline calibration and amplitude normalization, improving signal preprocessing stability and reducing interference from invalid data. Simultaneously, based on a spatiotemporal correlation filtering model, it utilizes the correlation between adjacent energy sources for adaptive denoising, effectively preserving transient high-frequency details and enhancing the signal-to-noise ratio. It adaptively determines the decomposition scale based on the spectral peak distribution, achieving accurate separation of multi-band features. Furthermore, it dynamically adjusts parameter weights in conjunction with contribution evaluation, filters high-confidence core features through stability verification, and simultaneously constructs a dynamically updated interface to fuse new data, continuously optimizing feature sequences and improving the real-time performance, accuracy, and reliability of transient feature extraction.

[0147] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time feature extraction based on ultra-high frequency raw signals, characterized in that, include: Based on the transient pulse time-domain distribution characteristics of UHF signals, a dynamic threshold is set to trigger the acquisition logic. Based on the acquisition logic, the original signal segments are captured in real time and baseline calibration and amplitude normalization preprocessing are completed simultaneously. The preprocessed signal segments are acquired, a spatiotemporal correlation filtering model is constructed, and adaptive noise filtering is performed based on the correlation of energy distribution of the signal at adjacent acquisition times to preserve high-frequency details of transient features. The decomposition scale is adaptively determined based on the spectral peak distribution of the denoised signal to separate and extract multi-band feature components; Extract the temporal abrupt change slope and frequency domain energy spectrum entropy value of each frequency band component, integrate the temporal abrupt change slope and frequency domain energy spectrum entropy value into a multi-dimensional feature set, and form an initial feature vector for quantifying transient feature attributes; Based on the distribution density of the initial feature vector, a real-time evaluation logic for feature contribution is set, and the weights of each parameter are dynamically and iteratively adjusted based on the evaluation logic. High-confidence core feature sequences are selected based on parameter weights combined with a feature stability verification model. A dynamic update interface for the feature sequences is constructed simultaneously. The application interface receives new feature data and updates the output core feature sequences.

2. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, The dynamic threshold-triggered acquisition logic includes: Calculate the dynamic trigger threshold based on the time-domain amplitude statistical characteristics of transient pulses of UHF signals: ； In the formula: For dynamic trigger thresholds; This is the adaptive adjustment coefficient; The standard deviation of the background noise during the data collection period; This represents the mean of the background noise. When the instantaneous amplitude of the signal exceeds The system acquires signal segments triggered by time and synchronously records the time-domain coordinates of the trigger moment. in, ∈[2.5, 4], and its value follows the following: when ρ≥5 pulses / second, k increases linearly with ρ and is controlled within (3.0, 4.0]; when ρ<5 pulses / second, k decreases linearly with ρ and is controlled within [2.5~3.0]. ρ represents the signal pulse density, that is, the number of pulses exceeding the initial threshold per unit time. The initial threshold is... .

3. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, The baseline calibration and amplitude normalization preprocessing process includes: The baseline signal is obtained by estimating the baseline of the acquired original signal segments using a piecewise linear fitting method, and the baseline correction signal is obtained by subtracting the baseline signal from the original signal. Normalization is performed based on the maximum amplitude of the baseline correction signal to obtain the normalized signal: ； In the formula: For normalized signals; This is the baseline correction signal; To correct the maximum amplitude of the signal; The amplitude range of the normalized signal is constrained to [-1, 1].

4. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, When constructing the spatiotemporal correlation filtering model, the signal energy correlation factor between adjacent acquisition times is denoted as... Adaptive noise filtering is performed using correlation factors; Among them, correlation factors The value follows: ； In the formula: for Moment signal energy; This indicates taking the maximum value within the parentheses; This is the time decay coefficient; Indicates adjacent data collection times; when High-frequency detail components are preserved when the value is ≥0.6; otherwise, noise components are attenuated. Among them, the time decay coefficient ∈[0.01,0.1], the value is larger when the signal pulse interval is less than 5ms, and smaller when the signal pulse interval is greater than 20ms.

5. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, The decomposition scale is adaptively determined through the following steps: Step 1: Calculate the decomposition scale based on the peak distribution of the denoised signal's spectrum: ； In the formula: For decomposition scale; These are the maximum and minimum frequencies of the spectral peak value; The average frequency interval between adjacent spectral peaks; This is the scale correction factor. ∈ (0.8, 1.2); Indicates rounding up; Step 2: Based on the decomposition scale Separate the signal into Each frequency band characteristic component; Among them, when the average frequency interval between adjacent peaks is small, The larger the value, the greater the average frequency interval between adjacent peaks. The smaller the value.

6. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, When extracting the temporal abrupt change slope of each frequency band component, the temporal abrupt change slope of each frequency band component is calculated through a sliding window, and the width of the sliding window is set to 2N+1, where N is a positive integer. ； In the formula: Let be the slope of the time-domain abrupt change of the j-th frequency band component; Let be the amplitude of the j-th frequency band component at the n-th point within the sliding window; Among them, the slope threshold is manually set by the user. When the slope is not less than the slope threshold, the j-th frequency band component is marked as a valid mutation feature.

7. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, The real-time evaluation logic for feature contribution is as follows: Calculate the contribution of each feature based on the distribution density of the initial feature vector: ； In the formula: The contribution of the m-th feature; Let m be the local density of the m-th feature; Let be the distance between the m-th feature and the high-density features; The dimension of the initial feature vector; Based on the above Dynamically iteratively adjust the weights of each parameter: Set initial weight values ; Weight update amount in the k-th iteration ; Updated weights ; in, Represents the iteration step size coefficient. ∈(0.1,0.3), when When the fluctuation range is large in consecutive iterations The smaller the value, the more... When the fluctuation range is small, The larger the value, the better.

8. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, The feature stability verification model is as follows: Calculate the stability index of the core feature sequence ; when If the value is ≥0.85, the feature sequence is considered stable; otherwise, feature re-selection is triggered. In the formula: The length of the core feature sequence; Let be the amplitude of the p-th core feature at time t.

9. The method for real-time feature extraction based on UHF raw signals according to claim 1, characterized in that, The dynamic update interface, when constructed, conforms to: The feature update trigger condition is set as follows: the cosine similarity between the new feature data and the current core feature sequence is less than a preset threshold, and the preset threshold is set to [0,7,0.85]. The cosine similarity is: ； In the formula: For feature dimensions; These are the q-th component of the new feature data and the q-th component of the current core feature sequence, respectively. When an update is triggered, an updated core feature sequence is generated based on a weighted fusion of the new feature data and historical feature data: ； In the formula: The updated core feature sequence; This represents the total number of historical feature data entries. These are the weighting coefficients; Let q be the q-th component of the k-th historical feature data; in, , Indicates the time of data collection for the new feature. This represents the collection time of the k-th historical feature data, and the sum of the weight coefficients of all historical feature data is 1.

Images