High-frequency discharge signal identification method

Abstract

The invention relates to the technical field of electrical variable measurement, in particular to a high-frequency discharge signal identification method, which comprises the following steps of: based on a voltage signal or a current signal measured in real time, continuously acquiring to form a data stream, carrying out continuous segmentation processing according to a set time window length, arranging segmented signal time slices according to a time sequence, and obtaining a high-frequency discharge signal; and obtaining a segmented high-frequency electric signal sequence. A continuous data flow is formed by collecting voltage or current signals in real time, segmentation is performed by adopting a fixed time window to generate segmented high-frequency electric signal sequences, and signal time sequence integrity and granularity analysis consistency are ensured. After the signal time slice representing the background noise is screened, the dynamic noise statistical threshold is constructed based on the background noise amplitude statistical magnitude of the voltage and current signals, the method can adapt to environmental noise fluctuation, and the misjudgment rate of non-discharge interference signals is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of measuring electrical variables, and in particular to a high-frequency discharge signal recognition method. Background Art

[0002] The purpose of the high-frequency discharge signal recognition method is to identify abnormal discharge signals caused by partial discharge, arc or insulation degradation in power equipment or electronic components by analyzing the time domain and frequency domain characteristics of high-frequency electrical signals, thereby realizing early equipment fault warning, insulation status assessment and maintenance decision optimization.

[0003] In the prior art, the verification of the periodic characteristics of the signal is mostly limited to frequency domain spectrum analysis, and lacks the verification of the time domain autocorrelation function combined with the power frequency period characteristics, making it difficult to effectively identify pseudo-discharge signals synchronized with the power frequency period, such as periodic interference caused by switching operations or load fluctuations. In addition, the existing methods lack dynamic optimization of the timing continuity and window length of signal segmentation, and the short-term high-frequency discharge signal is easily segmented into different segments due to the fixed window length, resulting in fragmented feature extraction and inconsistent discrimination results. Therefore, improvements are needed. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings in the prior art and to propose a high-frequency discharge signal recognition method.

[0005] In order to achieve the above object, the present invention adopts the following technical solution, a high-frequency discharge signal recognition method, comprising the following steps: Based on the real-time measured voltage signal or current signal, the data stream is continuously collected and segmented according to the set time window length, and the segmented signal time segments are arranged in time sequence to obtain a segmented high-frequency electrical signal sequence; Based on the segmented high-frequency electrical signal sequence, a signal time segment representing background noise is selected, background noise amplitude statistics of voltage signals and current signals in the signal time segment are calculated, and a dynamic noise statistical threshold is established based on the background noise amplitude statistics; Based on the segmented high-frequency electrical signal sequence, the signal amplitude characteristics of each signal time segment in the sequence are extracted, all the extracted signal amplitude characteristics are collected to obtain an amplitude feature set of each segment, and based on the amplitude feature set of each segment, each feature is compared with a dynamic noise statistical threshold one by one to generate a candidate discharge pulse set; Based on the candidate discharge pulse set and the associated signal time segments in the corresponding segmented high-frequency electrical signal sequence, an autocorrelation function of the statistical characteristics of the voltage signal or current signal in the associated signal time segment that changes with time is established. Based on the autocorrelation function, whether there is a periodic characteristic fluctuation in the function value is evaluated by comparing it with the power frequency cycle, and a high-frequency discharge identification result is obtained.

[0006] Preferably, the steps of acquiring the segmented high-frequency electrical signal sequence are: Based on the voltage signal or current signal measured in real time, the original voltage instantaneous value or current instantaneous value is continuously collected at a fixed sampling frequency, and the collected voltage or current values ​​are stored in a circular buffer in the order of timestamps to generate an original data stream marked with timestamps; According to a preset time window length parameter, the original data stream is cut into signal time segments of continuous equal time spans in a non-overlapping manner along the time axis, each signal time segment contains a fixed number of voltage or current values, and a discrete signal time segment set is generated; A start timestamp and an end timestamp are marked for each signal time segment in the signal time segment set, and all signal time segments are sorted in ascending order of timestamps to eliminate timing misalignment between adjacent segments and form a segmented high-frequency electrical signal sequence.

[0007] Preferably, the step of obtaining the background noise amplitude statistics is: Based on the segmented high-frequency electrical signal sequence, the absolute average value of the voltage signal amplitude and the absolute average value of the current signal amplitude are calculated for each signal time segment, and the signal time segments in which the absolute average value of the voltage is less than or equal to the preset amplitude threshold and the absolute average value of the current is less than or equal to the preset amplitude threshold are screened out to generate a background noise signal time segment set; Extracting the instantaneous voltage values ​​and the instantaneous current values ​​of all the segments in the background noise signal time segment set, and arranging them into a voltage noise sample matrix and a current noise sample matrix in the order of the segments respectively; The background noise amplitude statistics are calculated according to the voltage noise sample matrix and the current noise sample matrix.

[0008] Preferably, the step of obtaining the dynamic noise statistical threshold is: Based on the background noise amplitude statistics, extract the noise amplitude sequence of all background noise signal time segments, calculate the absolute deviation of each noise amplitude from the sequence median, and generate a set of noise amplitude volatility indicators; Counting the distribution characteristics of the noise amplitude volatility index set, taking the 99% quantile as the upper limit threshold of the noise fluctuation, and generating the noise amplitude fluctuation constraint boundary; A dynamic noise statistical threshold is established according to the noise amplitude fluctuation constraint boundary and the background noise amplitude statistic.

[0009] Preferably, the steps of obtaining the amplitude feature set of each segment are: Based on the segmented high-frequency electrical signal sequence, the voltage signal sampling points and the current signal sampling points in each signal time segment are traversed, and a sliding window traversal method is adopted to extract the maximum absolute value of the voltage signal in each signal time segment as the instantaneous peak value, calculate the root mean square value of all sampling points of the voltage signal as the effective value, count the difference between the highest point and the lowest point of the voltage signal as the peak-to-peak value, calculate the arithmetic mean of the absolute value of the voltage signal sampling points as the absolute mean value, and at the same time count the number of times the current signal crosses the zero point from positive to negative in each time segment as the zero-crossing rate, so as to generate an amplitude parameter set for each signal time segment; The voltage instantaneous peak value, voltage effective value, voltage peak-to-peak value, and voltage absolute average value in the amplitude parameter set of each signal time segment are combined into a voltage amplitude feature vector according to voltage category, the current zero-crossing rate is combined into a current feature vector separately, and the voltage feature vector and the current feature vector are spliced ​​into a multidimensional amplitude feature vector according to the time segment sequence, so as to generate a multidimensional amplitude feature vector set containing the joint features of voltage and current; According to the original timestamp order of the segmented high-frequency electrical signal sequence, all feature vectors in the multidimensional amplitude feature vector set are time-sequentially aligned by timestamp, and the start timestamp and end timestamp labels of the corresponding signal time segment are added to each feature vector to generate amplitude feature sets for each segment.

[0010] Preferably, the step of acquiring the candidate discharge pulse set is: Based on the amplitude feature set of each segment, the voltage peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value and current zero-crossing rate feature items in each signal time segment are traversed, the values ​​of the feature items are extracted one by one, and they are matched one by one with the preset dynamic noise statistical threshold, and a difference value set of each feature item and the dynamic noise statistical threshold is generated; All feature item difference values ​​of each signal time segment in the difference value set are traversed and judged. If the voltage peak difference value is greater than zero, the voltage effective value difference value is greater than zero, the voltage peak-to-peak difference value is greater than zero, the voltage absolute average difference value is greater than zero, or the current zero-crossing rate difference value is greater than zero, then the signal time segment is marked as a feature over-limit state, and a feature item over-limit marking set is generated; According to the feature item over-limit mark set, all signal time segments marked as feature over-limit states are screened, and the time segment data in the corresponding original segmented high-frequency electrical signal sequence are extracted and integrated into a candidate discharge pulse set in timestamp order.

[0011] Preferably, the step of obtaining the autocorrelation function is: Based on the candidate discharge pulse set, traverse each signal time segment, extract discrete sampling values ​​of voltage signals or current signals within a corresponding time stamp range in the associated segmented high-frequency electrical signal sequence, and generate original time domain signal data of the signal time segment; Perform de-averaging on the original time domain signal data of each signal time segment, calculate the arithmetic mean of the signal sampling values, subtract the arithmetic mean from each sampling value, and generate a zero-centered time domain signal sequence; Based on the zero-centered time domain signal sequence, an autocorrelation function of the statistical characteristics of the voltage or current signal is established.

[0012] Preferably, the steps of obtaining the high-frequency discharge identification result are: Based on the autocorrelation function, traverse the autocorrelation function curve corresponding to each candidate discharge pulse, extract the time delay coordinates of all local maximum points on the curve, and generate an autocorrelation function peak time delay set; According to the preset periodic fluctuation threshold of the power frequency cycle assessment function value, the delay intervals of adjacent maximum value points in the autocorrelation function peak delay set are calculated, the deviation rate between the interval sequence and the power frequency cycle is counted, and the delay interval segments with a deviation rate less than 5% are screened to generate a periodic characteristic fluctuation identification set; The periodic characteristic fluctuation identification set is traversed, and if a single candidate discharge pulse has three or more consecutive time delay intervals that meet the power frequency cycle deviation requirements, the pulse is determined to have significant periodic characteristics, and a high-frequency discharge identification result is generated.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention forms a continuous data stream by real-time acquisition of voltage or current signals, and uses a fixed time window for segmentation to generate a segmented high-frequency electrical signal sequence, thereby ensuring the consistency of signal timing integrity and analysis granularity. After screening the signal time segments that characterize the background noise, a dynamic noise statistical threshold is constructed based on the background noise amplitude statistics of the voltage and current signals, which can adapt to environmental noise fluctuations and reduce the misjudgment rate of non-discharge interference signals. The multi-dimensional amplitude characteristics of the voltage instantaneous peak value, effective value, peak-to-peak value, absolute average value and current zero-crossing rate of each signal time segment are extracted to form a multi-dimensional feature set and compared with the dynamic threshold one by one, thereby enhancing the comprehensiveness and discrimination accuracy of feature screening and avoiding the limitations of single feature analysis. For the candidate discharge pulse set, the voltage or current signal autocorrelation function of the associated signal time segment is established, and the periodic characteristic fluctuations are verified in combination with the power frequency period assessment function to distinguish between pseudo-discharge signals caused by the inherent power frequency characteristics of the equipment and real high-frequency discharge events. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] See also Figure 1 The present invention provides a technical solution, a high-frequency discharge signal recognition method, comprising the following steps: Based on the real-time measured voltage signal or current signal, the data stream is continuously collected and segmented according to the set time window length, and the segmented signal time segments are arranged in time sequence to obtain a segmented high-frequency electrical signal sequence; Based on the segmented high-frequency electrical signal sequence, the signal time segment representing the background noise is selected, the background noise amplitude statistics of the voltage signal and the current signal in the signal time segment are calculated, and the dynamic noise statistical threshold is established based on the background noise amplitude statistics; Based on the segmented high-frequency electrical signal sequence, the signal amplitude characteristics of each signal time segment in the sequence are extracted, and all the extracted signal amplitude characteristics are collected to obtain the amplitude feature set of each segment. Based on the amplitude feature set of each segment, each feature is compared with the dynamic noise statistical threshold one by one to generate a set of candidate discharge pulses; Based on the candidate discharge pulse set and the corresponding associated signal time segments in the segmented high-frequency electrical signal sequence, an autocorrelation function of the statistical characteristics of the voltage signal or current signal in the associated signal time segment that changes with time is established. Based on the autocorrelation function, whether there is a periodic characteristic fluctuation in the function value is evaluated by comparing it with the power frequency cycle, and the high-frequency discharge recognition result is obtained.

[0017] The steps for acquiring the segmented high-frequency electrical signal sequence are as follows: Based on the voltage signal or current signal measured in real time, the original voltage instantaneous value or current instantaneous value is continuously collected at a fixed sampling frequency, and the collected voltage or current values ​​are stored in a circular buffer in the order of timestamps to generate an original data stream marked with timestamps; According to the preset time window length parameter, the original data stream is cut into signal time segments of continuous equal time spans in a non-overlapping manner along the time axis, each signal time segment contains a fixed number of voltage or current values, and a discrete signal time segment set is generated; The starting timestamp and the ending timestamp are marked for each signal time segment in the signal time segment set, and all signal time segments are sorted in ascending order of timestamps to eliminate the timing misalignment between adjacent segments and form a segmented high-frequency electrical signal sequence.

[0018] Specifically, based on the real-time measured voltage signal or current signal, it is first necessary to determine the fixed sampling frequency used for data acquisition. The setting of this frequency needs to be based on the spectral characteristics of the high-frequency discharge signal to be identified and the frequency range of the background noise. For example, if the expected high-frequency discharge signal has a main energy distribution in the range of hundreds of kilohertz to several megahertz (MHz), according to the Nyquist sampling theorem, the sampling frequency should be at least twice the highest expected frequency. Considering the filter roll-off characteristics in practical applications and the need to retain sufficient signal details, a higher sampling frequency is usually selected. A specific fixed sampling frequency value is set, such as 20 megahertz (MHz). An analog-to-digital converter (ADC) with a corresponding sampling rate and sufficient bit width (such as 14 bits or 16 bits) is selected to perform synchronous or quasi-synchronous continuous sampling of the voltage and current channels. A high-precision timestamp is immediately attached to the instantaneous voltage and current values ​​obtained at each sampling moment. The timestamp can be derived from a hardware clock or a high-precision system clock synchronized with the ADC. The timestamp accuracy must reach the nanosecond level. Then, the voltage value and current value with the timestamp are paired (for example, ) In the order of the time of collection, they are pushed into a circular buffer running in first-in-first-out (FIFO) mode for temporary storage. The size of the buffer needs to be set in advance, and its capacity should be sufficient to accommodate at least several time windows of data to ensure continuous data flow. For example, if the time window for subsequent processing is 1 millisecond, the buffer can store at least tens of milliseconds of data. When the buffer is full, the new data will automatically overwrite the earliest stored data, thereby continuously generating a raw data stream with precise timestamps.

[0019] According to the preset time window length parameter, the setting of this parameter needs to balance the typical duration of the discharge signal with the time scale of the statistical characteristics of the background noise. For example, if the duration of a single high-frequency discharge pulse is usually between tens of microseconds and hundreds of microseconds, and the background noise statistics are relatively stable on the millisecond time scale, the time window length can be set to a moderate value, such as 0.5 milliseconds (ms). Using the generated raw data stream with timestamps, continuous data points are read in chronological order from the circular buffer or directly from the real-time data stream, and strict non-overlapping cutting is performed on the time axis with the set time window length (0.5 milliseconds) as the unit. The specific operation is to determine the starting point of the first time segment, and then calculate the number of sample points that each segment should contain based on the sampling frequency (for example, 20MHz) and the time window length (0.5ms). The calculation method is: number of sample points = sampling frequency time window length = Sample points are collected, and every 10,000 consecutive pairs of voltage and current data points are combined into a signal time segment. After the first segment ends, the starting point of the second segment follows the end point of the first segment without any time overlap or gap. The original data stream is processed continuously in this way, and the entire data stream is divided into a series of continuous and equal-length signal time segments. Each segment contains a fixed number (10,000) of voltage instantaneous values ​​and corresponding current instantaneous values. Finally, these segments are collected to generate a discrete signal time segment set.

[0020] Each signal time segment in the generated discrete signal time segment set is processed, and the timestamp of the first data point contained in each segment is extracted as the start timestamp of the segment, and the timestamp of the last data point is extracted as the end timestamp of the segment. The two timestamp information are clearly marked on the corresponding signal time segment to form a segment structure with start and end time marks. For example, the first segment is marked as {start: ,termination: ,data: },in and They are the precise timestamps of the first and last samples of the segment respectively. Subsequently, all the signal time segments with marked timestamps in the set are arranged in ascending order according to their starting timestamps, that is, from the earliest segment to the latest segment. Through this timestamp-based sorting operation, it can be ensured that even if temporary order confusion may be introduced during the data processing process (such as parallel processing or buffer scheduling), its strict physical time series relationship can be restored in the end, thereby effectively eliminating any timing misalignment that may occur between adjacent segments, ensuring the continuity and logic of the signal in the time dimension, and finally forming a segmented high-frequency electrical signal sequence that is precisely arranged in time order.

[0021] The steps to obtain the background noise amplitude statistics are: Based on the segmented high-frequency electrical signal sequence, the absolute average value of the voltage signal amplitude and the absolute average value of the current signal amplitude are calculated for each signal time segment, and the signal time segments in which the absolute average value of the voltage is less than or equal to the preset amplitude threshold and the absolute average value of the current is less than or equal to the preset amplitude threshold are screened out to generate a background noise signal time segment set; Extracting the instantaneous voltage values ​​and the instantaneous current values ​​of all the segments in the background noise signal time segment set, and arranging them into a voltage noise sample matrix and a current noise sample matrix according to the sequence of the segments; According to the voltage noise sample matrix and the current noise sample matrix, the background noise amplitude statistics are calculated. The calculation formula is: ; in, is the background noise amplitude statistic, is the total number of time segments of the background noise signal, is the total number of samples in each signal time segment, For the In the noise segment Voltage sample values, For the In the noise segment Current sample value.

[0022] Specifically, based on the segmented high-frequency electrical signal sequence obtained in the previous step, it is necessary to traverse each signal time segment in the sequence, and process the voltage signal sampling point sequence and the current signal sampling point sequence contained in each segment, specifically to calculate the absolute average value of the voltage signal amplitude, that is, first take the absolute value of all voltage sampling values ​​in the segment, and then calculate the arithmetic mean of these absolute values ​​to obtain the absolute average voltage value of the segment. Similarly, take the absolute value of all current sampling values ​​in the segment, calculate the arithmetic mean of these absolute values, and obtain the absolute average current value of the segment. After the calculation is completed, the obtained absolute average voltage value and absolute average current value are compared with the preset amplitude thresholds respectively. These two amplitude thresholds, namely the voltage amplitude threshold and the current amplitude threshold, are the key boundaries for distinguishing background noise from potential signals. Their setting needs to be based on the historical data analysis of the device when it is running in a non-discharge state. For example, a period of time when the acquisition device is operating normally but it is determined that no high-frequency discharge occurs The signal data within a period of time (for example, continuous operation for 1 hour) is segmented in the same way and the absolute average voltage and the absolute average current of all segments are calculated. The distribution of these values ​​is statistically analyzed, and the 95th percentile of the distribution or the statistical maximum value multiplied by a coefficient slightly greater than 1 (for example, 1.2 times) is selected as the threshold. This is to ensure that the threshold is slightly higher than the maximum noise fluctuation under normal circumstances to effectively filter out background noise. For example, by analyzing historical data, the voltage amplitude threshold is set to 0.015 volts (V) and the current amplitude threshold is set to 0.005 amperes (A). Then, for each signal time segment currently being processed, it is determined whether the calculated absolute average voltage is less than or equal to 0.015V, and whether the absolute average current is less than or equal to 0.005A. Only when these two conditions are met at the same time, the signal time segment is determined to belong to background noise. All signal time segments determined to be background noise are collected to generate a background noise signal time segment set.

[0023] Using the background noise signal time segment set generated by screening, it is necessary to extract the original sample data required for subsequent noise characteristic analysis. The specific operation is to traverse each signal time segment in this set that is determined to be background noise. For each segment, all the original voltage instantaneous value samples contained in it are completely extracted according to the time sequence in the segment. At the same time, all the original current instantaneous value samples corresponding to the segment are extracted in the same time sequence. For example, there are a total of fragments, and previously set each fragment to contain Sample points (e.g. ), then for the Noise fragments ( From 1 to ), the voltage sequence will be extracted and current sequence , all The voltage sequences extracted from the fragments are organized and arranged according to the original order of the fragments (usually time order) to form a dimension of The voltage noise sample matrix of The line is The voltage sample sequence of the noise fragments, similarly, all The current sequences extracted from the fragments are also arranged according to the original order of the fragments, forming a dimension of The current noise sample matrix of the matrix The line is The current sample sequence of the noise fragments is finally generated into a structured voltage noise sample matrix and a current noise sample matrix.

[0024] formula: The benefit of the formula is that it provides a comprehensive and robust background noise level estimate by integrating the voltage and current information of multiple background noise time segments. , it first calculates the average absolute amplitude of the voltage and current signals within each identified noise segment, then adds the two average values ​​to obtain the comprehensive amplitude metric of the segment, and finally averages this comprehensive metric for all noise segments. This method considers the noise contribution of both voltage and current dimensions at the same time, and eliminates the influence of positive and negative fluctuations by taking the absolute value, so that the result can better reflect the average level of noise energy. Compared with only using instantaneous peaks or a single signal type, It can represent the basic noise floor more stably, provide a more reliable basis for the subsequent setting of dynamic thresholds, and help improve the distinction between discharge signals and noise. The structural design of the formula (average of inner and outer layers) clearly reflects the statistical averaging process from a single fragment to the overall noise level.

[0025] The steps to obtain the parameters are: Represents the total number of background noise signal time segments. Its value is obtained during the screening process, that is, it counts how many signal time segments in the entire segmented high-frequency electrical signal sequence analyzed simultaneously meet the two conditions that the absolute average voltage value is less than or equal to the preset voltage amplitude threshold (for example, 0.015V) and the absolute average current value is less than or equal to the preset current amplitude threshold (for example, 0.005A). The counting result is For example, after analyzing a total of 5000 signal time segments, the above screening conditions determine that there are a total of The fragments belong to background noise.

[0026] The steps to obtain the parameters are: Represents the total number of sample points contained in each signal time segment. This value is determined by the fixed sampling frequency set during the data acquisition phase. and the time window length set in the data segmentation stage Jointly decided, the calculation formula is , for example, if the sampling frequency Set to 25 MHz, the time window length If set to 0.4 milliseconds (ms), the total number of samples per segment Sample points.

[0027] The steps to obtain the parameters are: Indicated in The first time segment of the signal identified as background noise The voltage sampling instantaneous value, where index The range is from 1 to (For example ),index The range is from 1 to (For example ),For example That is the 500th voltage sampling value in the 10th noise segment.

[0028] The steps to obtain the parameters are: Indicated in The first time segment of the signal identified as background noise The instantaneous value of current sampling, its index and The meaning and scope of Exactly the same ( From 1 to 850, From 1 to 10000), these specific current values ​​are derived from the generated current noise sample matrix, and the That is, directly read the current noise sample matrix Row, No. The element value of the column, for example That is the 500th current sampling value in the 10th noise segment.

[0029] Calculation process: Calculate the background noise amplitude statistics It is necessary to substitute the parameter values ​​according to the voltage noise sample matrix and the current noise sample matrix and The calculation process firstly takes each noise segment as (from 1 to 850), calculate the absolute average value of its internal voltage and current respectively: The absolute average voltage is: , fragment The absolute average current is: , and then add the two together to get the fragment The comprehensive amplitude of: Finally, the combined amplitude of all 850 noise segments is summed and divided by the total number of segments. , and get the final : ; After calculating 850 noise fragments, the comprehensive amplitude of each fragment is obtained. The sum of . Then the background noise amplitude statistic is The calculation is as follows: ; ; The results show that: the calculated background noise amplitude statistics The value of is 0.006 (in volts + amperes), which quantifies the baseline level of the combined average amplitude of the voltage and current signals in the current monitoring environment when there is no high-frequency discharge signal. It represents the average background noise intensity of the system. The value is a core basic parameter for establishing the dynamic noise statistical threshold. A value such as 0.006 means that subsequent thresholds will be adjusted around this baseline to more accurately distinguish potential discharge signal pulses above this average noise level.

[0030] The steps for obtaining the dynamic noise statistical threshold are: Based on the background noise amplitude statistics, the noise amplitude sequence of all background noise signal time segments is extracted, the absolute deviation of each noise amplitude from the sequence median is calculated, and a set of noise amplitude volatility indicators is generated; The distribution characteristics of the noise amplitude volatility index set are statistically analyzed, and the 99% quantile is taken as the upper threshold of the noise fluctuation to generate the noise amplitude fluctuation constraint boundary; According to the noise amplitude fluctuation constraint boundary and the background noise amplitude statistics, the dynamic noise statistical threshold is established. The calculation formula is: ; in, is the dynamic noise statistical threshold, is the background noise amplitude statistic, is the 99% quantile of the noise amplitude volatility index, is the total number of time segments of the background noise signal.

[0031] Specifically, based on the background noise amplitude statistics calculated in the previous stage , and using the identified (e.g. ) fragment of the background noise signal time segment set, first of all, we need to The noise amplitude sequence is constructed from the background noise segments. The “noise amplitude” here refers to the comprehensive amplitude measurement of each background noise segment itself, that is, For each segment Computational , this indivual Values ​​(for example, ) are arranged in order to form a noise amplitude sequence Next, calculate the noise amplitude sequence the median of , which is calculated by converting the sequence All The values ​​are sorted from smallest to largest, because is an even number, so the median is the two values ​​in the middle after sorting (the Position and For example, if the 425th value after sorting is 0.0059 and the 426th value is 0.0060, then the median , then, traverse the noise amplitude sequence Each value in , calculate its difference with the median The absolute deviation , for all This calculation is performed for each value, and we get The absolute deviation value is The absolute deviation values ​​calculated are combined to generate a set of noise amplitude volatility indicators .

[0032] Based on the generated noise amplitude volatility indicator set , it is necessary to perform statistical analysis on the data distribution characteristics of this set in order to determine a boundary value that can represent the upper limit of noise fluctuation. The specific operation is to calculate the 99% quantile of the set (denoted as ), the calculation process first converts all Deviation value Sort from low to high to get an ordered sequence , and then determine the position corresponding to the 99th percentile, which can be calculated by calculating the index Since the index is not an integer, linear interpolation or rounding is used to determine the quantile value. A common method is to take the value corresponding to the integer ranking closest to the index and not less than the index, that is, to take the value corresponding to the quantile after sorting. The value of the bit is taken as the 99% quantile. For example, in the sorted deviation value sequence, the 842nd value is found to be 0.0015, then it is determined ,this The value means that in a statistical sense, the deviation between the comprehensive amplitude of about 99% of the background noise fragments and their sequence median will not exceed this value. The 99% quantile calculated is (For example, 0.0015) is defined as the upper threshold of noise fluctuation and generates the noise amplitude fluctuation constraint boundary.

[0033] formula: The benefit of this formula is that the dynamic noise statistical threshold constructed by this formula Combined with the average level of background noise and its fluctuation characteristics (given by and sample size embodied), realizing the adaptive adjustment of the threshold, The term ensures that the threshold changes as the overall noise baseline changes. The additional term is based on the severity of the noise fluctuation ( ) and statistical sample size ( ) to set a reasonable tolerance boundary, reflects the extreme case of noise fluctuations (99% of fluctuations do not exceed this value), while the denominator Items utilize sample size When used to evaluate the number of noise fragments When the value increases, the estimation of noise fluctuation is more stable, and the allowed additional boundary can be appropriately reduced, otherwise it can be enlarged. This design makes the threshold robust to changes in the noise environment and the amount of data, and can more accurately identify signals that are significantly higher than the background noise as potential discharge pulses, reduce misjudgments caused by noise fluctuations, and improve recognition accuracy.

[0034] The steps to obtain the parameters are: is the background noise amplitude statistic, which quantifies the average integrated amplitude of the signal segments identified as background noise during the current monitoring period. Using the result calculated in the previous step, set .

[0035] The steps to obtain the parameters are: is the 99% quantile of the noise amplitude volatility index set, which represents the upper limit level of the absolute deviation distribution between the comprehensive amplitude of the background noise segment and its sequence median (covering 99% of the deviation values). In this example, set .

[0036] The steps to obtain the parameters are: Represents the background noise characteristics used to calculate (including and ) is the total number of background noise signal time segments. This value is obtained by counting the background noise segments in the first stage. It reflects the sample size based on which the noise analysis is based. This value has been determined in the previous steps. In this example, .

[0037] Calculation process: Substitute the obtained parameter values ​​for calculation. Known parameters: , , . First calculate the logarithmic term in the denominator: ; have to: ; Next, calculate the square root of the denominator: ; Then calculate the fractional term (volatility adjustment term): ; Finally, the average background noise amplitude statistic is added with the fluctuation adjustment term: ; ; ; Considering the accuracy requirements in practical applications, the result can be approximated as .

[0038] The results show that: the calculated dynamic noise statistical threshold About 0.0066, this value is based on the current average background noise level And superimposed a noise volatility and sample size The dynamically calculated adjustment (about 0.0006) represents a dynamic boundary for distinguishing signal from noise. In the subsequent steps, if the comprehensive amplitude feature extracted from a certain signal time segment exceeds , then the segment will be preliminarily judged as a candidate pulse that may contain high-frequency discharge signals, otherwise it will be considered as part of the background noise.

[0039] The steps for obtaining the amplitude feature set of each segment are: Based on the segmented high-frequency electrical signal sequence, the voltage signal sampling points and the current signal sampling points in each signal time segment are traversed, and the sliding window traversal method is adopted to extract the maximum absolute value of the voltage signal in each signal time segment as the instantaneous peak value, calculate the root mean square value of all sampling points of the voltage signal as the effective value, count the difference between the highest point and the lowest point of the voltage signal as the peak-to-peak value, calculate the arithmetic mean of the absolute value of the voltage signal sampling points as the absolute mean value, and at the same time count the number of times the current signal crosses the zero point from positive to negative in each time segment as the zero-crossing rate, so as to generate an amplitude parameter set for each signal time segment; The voltage instantaneous peak value, voltage effective value, voltage peak-to-peak value, and voltage absolute average value in the amplitude parameter set of each signal time segment are combined into a voltage amplitude feature vector according to voltage category, the current zero-crossing rate is combined into a current feature vector separately, and the voltage feature vector and the current feature vector are spliced ​​into a multidimensional amplitude feature vector according to the time segment sequence, so as to generate a multidimensional amplitude feature vector set containing the joint features of voltage and current; According to the original timestamp order of the segmented high-frequency electrical signal sequence, all feature vectors in the multidimensional amplitude feature vector set are aligned in time sequence according to the timestamps, and the start timestamp and end timestamp labels of the corresponding signal time segment are added to each feature vector to generate the amplitude feature set of each segment.

[0040] Specifically, based on the previously generated segmented high-frequency electrical signal sequence, it is necessary to perform independent feature extraction processing on each signal time segment in the sequence, and for the voltage signal sampling point sequence contained in each segment (for example, Voltage samples ) and the corresponding current signal sampling point sequence (including Current samples ), perform the following calculations: First, extract the instantaneous peak value of the voltage signal by traversing all voltage samples in the segment , find the sample value with the largest absolute value, that is, calculate ; Secondly, calculate the effective value (root mean square value) of the voltage signal, that is, first calculate the sum of the squares of all voltage sample values, and then divide it by the total number of samples Get the average square value, and finally take the square root of the average square value; again, count the peak-to-peak value of the voltage signal, that is, find the maximum value among all voltage samples in the segment and minimum value , calculate the difference between the two ; Next, calculate the absolute average value of the voltage signal, that is, calculate the arithmetic mean of the absolute values ​​of all voltage samples, that is, At the same time, the zero-crossing rate of the current signal is counted by checking each pair of adjacent current sampling points in the segment , count the number of s that satisfy the condition of crossing from positive value to non-positive value (i.e. and ) occurs, and the five calculated values ​​(instantaneous peak value of voltage, effective value of voltage, peak-to-peak value of voltage, absolute average value of voltage, and current zero-crossing rate) are combined together to generate the amplitude parameter set corresponding to the current signal time segment.

[0041] For the amplitude parameter set generated for each signal time segment (including four voltage-related parameters, namely, voltage instantaneous peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value, and current zero-crossing rate, which is a current-related parameter), it is necessary to organize these parameters into a structured feature vector. The specific operation is that for each signal time segment, the four voltage-related amplitude parameters are arranged in a fixed order, such as [voltage instantaneous peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value], to form a four-dimensional voltage amplitude feature vector. At the same time, the current zero-crossing rate parameter is independently constructed into a one-dimensional current feature vector, namely, [current zero-crossing rate]. Then, the segment is The voltage amplitude feature vector and current feature vector of the segment are spliced ​​and combined into a multidimensional amplitude feature vector in a predetermined order (for example, voltage feature first, current feature second). In this example, a five-dimensional feature vector is obtained after splicing: [voltage instantaneous peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value, current zero-crossing rate]. The same merging and splicing operations are performed on all signal time segments to ensure that each segment corresponds to a multidimensional amplitude feature vector with the same dimension (five dimensions) and consistent internal element order. Finally, the multidimensional amplitude feature vectors generated by all segments are collected to generate a set of multidimensional amplitude feature vectors containing the joint features of voltage and current.

[0042] Based on the generated multidimensional amplitude feature vector set and referring to the timestamp information of each signal time segment in the original segmented high-frequency electrical signal sequence, these feature vectors need to be time-sequenced and annotated. First, for each five-dimensional feature vector in the multidimensional amplitude feature vector set, find the original signal time segment to which it corresponds, and read its starting timestamp from the original segment. and end timestamp (These two timestamps have been determined and marked when the signal is segmented). The two timestamp information is appended to the corresponding five-dimensional feature vector to form a data structure containing timestamps and feature values, such as Then, all the timestamp-annotated feature vector structures in the collection are sorted according to their starting timestamps The features are arranged in ascending order, that is, from the feature vector corresponding to the earliest segment in time to the feature vector corresponding to the latest segment in time. Through this sorting process, it is ensured that the order of the feature vector set strictly reflects the time evolution order of the original signal, eliminating the disorder caused by the processing process, and finally forming a feature set of amplitude values ​​of each segment arranged in time sequence, and each feature vector has the start and end time marks of its corresponding time segment.

[0043] The steps for obtaining the candidate discharge pulse set are: Based on the amplitude feature set of each segment, the voltage peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value and current zero-crossing rate feature items in each signal time segment are traversed, the values ​​of the feature items are extracted one by one, and they are matched one by one with the preset dynamic noise statistical threshold, and a difference value set of each feature item and the dynamic noise statistical threshold is generated; All feature item difference values ​​of each signal time segment in the difference value set are traversed and judged. If the voltage peak difference value is greater than zero, the voltage effective value difference value is greater than zero, the voltage peak-to-peak difference value is greater than zero, the voltage absolute average difference value is greater than zero, or the current zero-crossing rate difference value is greater than zero, then the signal time segment is marked as a feature over-limit state, and a feature item over-limit marking set is generated; According to the feature item over-limit mark set, all signal time segments marked as feature over-limit states are screened, and the corresponding time segment data in the original segmented high-frequency electrical signal sequence are extracted and integrated into a candidate discharge pulse set in timestamp order.

[0044] Specifically, based on the amplitude feature set of each segment generated in the previous stage (including the start timestamp, end timestamp and five-dimensional feature vector [instantaneous voltage peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value, current zero-crossing rate] of each signal time segment), and using the dynamic noise statistical threshold calculated previously (For example ), it is necessary to evaluate the characteristics of each signal time segment. The specific operation is to traverse each entry in the amplitude feature set of each segment, and extract the values ​​of its five feature items for the current entry: instantaneous peak voltage , voltage RMS , peak-to-peak voltage , the absolute average voltage , and the current zero-crossing rate , each of the five characteristic values ​​is compared with the same dynamic noise statistical threshold Compare and calculate the difference between the two, that is, for each feature item ( From 1 to 5), calculate the difference For example, if the instantaneous peak voltage of a segment is , then the difference value , if the voltage RMS , then the difference value , if the peak-to-peak voltage , then the difference value , if the absolute average voltage , then the difference value , if the current zero-crossing rate , then the difference value , the five difference values ​​calculated for the fragment As a set, it is stored and associated with the timestamp of the segment, and this process is repeated for all signal time segments to generate a set of difference values ​​between each feature item and the dynamic noise statistical threshold.

[0045] The difference value set between each feature item and the dynamic noise statistical threshold is used (where each signal time segment corresponds to a set of five difference values). ), it is necessary to determine the state of each segment. The specific method is to traverse the difference value group corresponding to each signal time segment, check the five difference values ​​in the group, and determine whether there is at least one difference value greater than zero, that is, perform logical judgment: Check Is it established, or Is it established, or Is it established, or Is it established, or Is it true? As long as any one or more of the five conditions are met, it means that the value of at least one characteristic item in the signal time segment exceeds the dynamic noise statistical threshold. In this case, the signal time segment is marked as "feature overlimit state". If the five difference values ​​corresponding to a segment to If all are not greater than zero (i.e., all are less than or equal to zero), the segment does not meet the over-limit condition and is marked as "not over-limit state" or left unmarked. Such a state mark is assigned to each signal time segment in the system (for example, the Boolean value True is used to indicate over-limit and False is used to indicate not over-limit), and these marks are associated with their corresponding time segments to eventually form a set of feature item over-limit marks.

[0046] According to the generated feature item exceeding limit mark set, which indicates which signal time segments are marked as "feature exceeding limit state", it is necessary to screen out these potential discharge pulse segments from the original data. The specific operation is to screen out the identifiers of all signal time segments marked as "feature exceeding limit state" (such as their timestamp range) according to the marking results in the feature item exceeding limit mark set, and then use these identifiers to return to the segmented high-frequency electrical signal sequence that originally stored the complete data, accurately locate these marked original signal time segments, and for each located segment, extract all the original time domain data it contains, that is, the complete voltage sampling point sequence And the corresponding current sampling point sequence , as well as the start timestamp and end timestamp of the fragment, collect all the screened signal time fragments containing the original sampling data, and keep their original time sequence during the collection process, that is, arrange them from small to large according to the start timestamp. The set of original signal data fragments arranged in time sequence obtained by integration is the candidate discharge pulse set.

[0047] The steps to obtain the autocorrelation function are: Based on the candidate discharge pulse set, traverse each signal time segment, extract the discrete sampling values ​​of the voltage signal or current signal within the corresponding time stamp range in the associated segmented high-frequency electrical signal sequence, and generate the original time domain signal data of the signal time segment; Perform de-averaging on the original time domain signal data of each signal time segment, calculate the arithmetic mean of the signal sampling values, subtract the arithmetic mean from each sampling value, and generate a zero-centered time domain signal sequence; Based on the zero-centered time domain signal sequence, the autocorrelation function of the statistical characteristics of the voltage or current signal is established, and the calculation formula is: ; in, Delay The corresponding autocorrelation function value is, For voltage or current signal Zero-centered sample values, For the same signal Zero-centered sample values, is the total number of samples in the signal time segment, is the time delay, and its value range is .

[0048] Specifically, based on the candidate discharge pulse set obtained by screening in the previous stage, the set contains several original signal time segments (including voltage, current sampling sequences and timestamps) that are preliminarily determined to contain discharge signals. It is necessary to prepare time domain data for subsequent autocorrelation analysis for each candidate signal time segment. The specific operation is to traverse each segment in the candidate discharge pulse set, and decide whether to use the voltage signal sequence or current signal sequence of the segment for subsequent analysis based on the preset analysis strategy or signal characteristics. Usually, a signal type with a higher signal-to-noise ratio or that can better reflect the characteristics of the power frequency cycle is selected. For example, the voltage signal is selected for analysis here, and for the currently traversed candidate segment, the complete voltage discrete sampling value sequence contained in it is extracted. ,in The total number of samples for each time segment determined previously (e.g. ), the extracted voltage (or current) sampling value sequence is the original time domain signal data of the signal time segment for subsequent processing.

[0049] The original time domain signal data (such as voltage series) of each candidate signal time segment is obtained. , ) After that, it needs to be de-averaged to eliminate the influence of DC bias on the subsequent autocorrelation function calculation. The specific steps are as follows: , first calculate all The arithmetic mean of the sample values ​​is the sum of all sample values ​​divided by the total number of samples. , get the mean , then, each original sample value in the sequence Subtract the calculated mean from this , generating a new sequence ,in , this newly generated sequence That is, it is a zero-centered time domain signal sequence, whose mean is theoretically zero. This de-meaning operation is performed on the original time domain signal data extracted from each segment in the candidate discharge pulse set to generate a corresponding zero-centered time domain signal sequence set.

[0050] formula: , the formula is useful in that it is used to calculate the autocorrelation function of a signal , which can effectively reveal the periodic components inside the signal, which is crucial for identifying the discharge characteristics related to the power frequency cycle. The autocorrelation function calculates the signal and itself at different time delays. The product average under quantifies the similarity of the signal at different time points. If the signal contains periodic components, the autocorrelation function will also show corresponding periodic fluctuations, and its peak value appears at a delay equal to an integer multiple of the period. At , use the generated zero-centered sequence As input, the interference of DC component can be removed, making the periodic characteristics more prominent, and the normalization factor This ensures that the average correlation is calculated, which helps to stabilize the peak size and position in the subsequent analysis. The periodic characteristics of the curve can determine whether the candidate pulse is related to the power frequency.

[0051] The steps to obtain the parameters are: Represents the total number of samples of the zero-centered time domain signal sequence used to calculate the autocorrelation function, that is, the total number of samples in each signal time segment. This value is determined in the data acquisition and segmentation steps and is determined by the sampling frequency and time window length Decide( ), in this example, we use the previous settings. Sample points.

[0052] The steps to obtain the parameters are: represents the first sample values, where index The range is from 1 to , which is the original time domain signal data minus its arithmetic mean. Each is a specific value, and the unit is the same as the original signal.

[0053] The steps to obtain the parameters are: Represents the same zero-centered time domain signal sequence, relative to the Sample points delayed The sampling value of the sample point, that is, the elements, among which The index range is from arrive , these values ​​are also read directly from the generated zero-centered sequence.

[0054] The steps to obtain the parameters are: is the time delay (time lag), which represents the time difference between the signal and itself when calculating the correlation. The unit is the number of sampling points, and its value range is , autocorrelation function Need to care about everyone The value is calculated to obtain a The function curve usually needs to be calculated from A series of values ​​starting from the beginning until a sufficiently large delay (such as ),this Should be able to cover the expected cycle length, in this formula, the definition is to .

[0055] Calculation process: For a candidate discharge pulse segment corresponding to the zero-centered sequence (in ), whose autocorrelation function needs to be calculated At different delays (from 0 to ) under the condition of . The calculation process is as follows: for each Value (from 0 to ): Calculate the product term: For all allowed (from 1 to ),calculate . Sum: Adding up all these product terms gives . Normalization: Divide the sum by the number of items involved in the summation ,get .

[0056] After executing this formula calculation on the zero-centered time domain signal sequence of each candidate discharge pulse segment, the autocorrelation function sequence of the segment signal is obtained. This sequence describes the self-similarity of the signal under different time delays and is the key basis for analyzing the internal periodic structure of the signal. The sequence itself is not the final recognition result, but the basic data for the next stage of analysis. Whether there is a significant peak fluctuation with a specific period (such as the power frequency period) in the sequence can ultimately determine whether the candidate pulse is a high-frequency discharge related to the power frequency.

[0057] The steps for obtaining high-frequency discharge recognition results are as follows: Based on the autocorrelation function, the autocorrelation function curve corresponding to each candidate discharge pulse is traversed, the time delay coordinates of all local maximum points on the curve are extracted, and the autocorrelation function peak time delay set is generated; According to the preset periodic fluctuation threshold of the power frequency cycle assessment function value, the delay intervals of adjacent maximum value points in the autocorrelation function peak delay set are calculated, the deviation rate between the interval sequence and the power frequency cycle is counted, and the delay interval segments with a deviation rate less than 5% are screened to generate a periodic characteristic fluctuation identification set; Traversing the periodic characteristic fluctuation identification set, if a single candidate discharge pulse contains three or more consecutive time delay intervals that meet the power frequency cycle deviation requirements, the pulse is judged to have significant periodic characteristics and a high-frequency discharge identification result is generated.

[0058] Specifically, based on the autocorrelation function sequence calculated for each candidate discharge pulse in the previous stage , it is necessary to analyze the function to extract its periodic information. The specific operation is to traverse the autocorrelation function sequence corresponding to each candidate discharge pulse (in is the delay, from 0 to ,For example ), which is considered as a delay The curve is used to identify and extract the delays corresponding to all local maximum points on the curve. , not considered when searching The global maximum at The local maximum point can be found by comparing each point The values ​​of its neighboring points are used to achieve this, for example, to determine a point is a local maximum if it is also greater than its previous point and is greater than or equal to the next point (for From 1 to ), for boundary points (e.g. ) requires special processing or is included according to the analysis requirements. The delay values ​​corresponding to all the local maximum points found are collected and arranged in order from small to large. For example, the sequence , this ordered list of time delay values ​​is the set of peak time delays of the autocorrelation function of the candidate discharge pulse.

[0059] Using the autocorrelation function peak delay set generated for each candidate discharge pulse (in To find the number of peak values, it is necessary to examine the regularity of the intervals between these peak delays based on the power system frequency cycle (for example, 50 Hz) and the preset cycle fluctuation threshold (5%). First, determine the number of sampling points corresponding to the power frequency cycle. , which is equal to the power frequency cycle length (for example, 50Hz corresponds to 0.02 seconds) multiplied by the sampling frequency used when collecting data (For example, if it is set to 1MHz before, sample points; Note: This value must be greater than the segment length M = 10000, indicating that a single segment cannot completely cover a power frequency cycle. In applications, a longer time window or detection of higher harmonics is required. For example, if a cycle is The characteristics of the sample points, such as high-order harmonics or the inherent oscillation period of a certain device), then calculate the delay interval between all adjacent maximum points in the autocorrelation function peak delay set, that is, calculate , and get the interval sequence , then, for each calculated delay interval , calculate its difference with the target period The relative deviation rate is calculated as follows: First find the interval With target cycle The absolute value of the difference, then divide this absolute difference by the target period , multiply the result by 100% to get the deviation rate percentage. Finally, compare each calculated deviation rate with the preset period fluctuation threshold of 5% to filter out all delay intervals with a deviation rate less than 5%. , and mark these interval segments that meet the conditions as meeting the periodic characteristics, forming a periodic characteristic fluctuation identification set (for example, a list that records all interval values ​​or their serial numbers that meet the conditions).

[0060] Based on the periodic characteristic fluctuation identification set generated for each candidate discharge pulse (marking which adjacent peak delay intervals meet the requirement of less than 5% deviation rate), it is necessary to make a final statistical judgment on these identifications to determine whether the candidate pulse has significant periodic characteristics. The specific operation is to traverse the periodic characteristic fluctuation identification set corresponding to a single candidate discharge pulse, and check whether there are three or more consecutive delay interval segments that are marked as meeting the power frequency cycle deviation requirement (that is, the deviation rate is less than 5%). A counter is used to track the number of interval segments that meet the conditions continuously. When an interval segment that meets the conditions is encountered, the counter is incremented by one. If an interval segment that meets the conditions is encountered, the counter is incremented by one. If there is an interval segment that does not meet the conditions, the counter will be cleared. During the traversal process, once the counter value reaches 3, it can be determined that the candidate discharge pulse has a significant periodic feature, and the recognition result of this pulse is marked as "high-frequency discharge (periodic)", and there is no need to continue to check the remaining interval segments of the pulse. If the counter has never reached 3 after traversing all the time delay interval segments of the pulse, it is determined that the pulse does not have a significant periodic feature, and its recognition result is marked as "non-periodic pulse" or maintained as a "candidate pulse" for further analysis. Repeat this determination process for all candidate discharge pulses to generate the final high-frequency discharge recognition result.

[0061] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A high-frequency discharge signal recognition method, characterized in that: The following steps are involved: Based on the real-time measured voltage signal or current signal, the data stream is continuously collected and segmented according to the set time window length, and the segmented signal time segments are arranged in time sequence to obtain a segmented high-frequency electrical signal sequence; Based on the segmented high-frequency electrical signal sequence, a signal time segment representing background noise is selected, background noise amplitude statistics of voltage signals and current signals in the signal time segment are calculated, and a dynamic noise statistical threshold is established based on the background noise amplitude statistics; Based on the segmented high-frequency electrical signal sequence, the signal amplitude characteristics of each signal time segment in the sequence are extracted, all the extracted signal amplitude characteristics are collected to obtain an amplitude feature set of each segment, and based on the amplitude feature set of each segment, each feature is compared with a dynamic noise statistical threshold one by one to generate a candidate discharge pulse set; Based on the candidate discharge pulse set and the associated signal time segments in the corresponding segmented high-frequency electrical signal sequence, an autocorrelation function of the statistical characteristics of the voltage signal or current signal in the associated signal time segment that changes with time is established. Based on the autocorrelation function, whether there is a periodic characteristic fluctuation in the function value is evaluated by comparing it with the power frequency cycle, and a high-frequency discharge identification result is obtained.

2. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps of acquiring the segmented high-frequency electrical signal sequence are as follows: Based on the voltage signal or current signal measured in real time, the original voltage instantaneous value or current instantaneous value is continuously collected at a fixed sampling frequency, and the collected voltage or current values ​​are stored in a circular buffer in the order of timestamps to generate an original data stream marked with timestamps; According to a preset time window length parameter, the original data stream is cut into signal time segments of continuous equal time spans in a non-overlapping manner along the time axis, each signal time segment contains a fixed number of voltage or current values, and a discrete signal time segment set is generated; A start timestamp and an end timestamp are marked for each signal time segment in the signal time segment set, and all signal time segments are sorted in ascending order of timestamps to eliminate timing misalignment between adjacent segments and form a segmented high-frequency electrical signal sequence.

3. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps for obtaining the background noise amplitude statistics are as follows: Based on the segmented high-frequency electrical signal sequence, the absolute average value of the voltage signal amplitude and the absolute average value of the current signal amplitude are calculated for each signal time segment, and the signal time segments in which the absolute average value of the voltage is less than or equal to the preset amplitude threshold and the absolute average value of the current is less than or equal to the preset amplitude threshold are screened out to generate a background noise signal time segment set; Extracting the instantaneous voltage values ​​and the instantaneous current values ​​of all the segments in the background noise signal time segment set, and arranging them into a voltage noise sample matrix and a current noise sample matrix in the order of the segments respectively; The background noise amplitude statistics are calculated according to the voltage noise sample matrix and the current noise sample matrix.

4. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps for obtaining the dynamic noise statistical threshold are: Based on the background noise amplitude statistics, extract the noise amplitude sequence of all background noise signal time segments, calculate the absolute deviation of each noise amplitude from the sequence median, and generate a set of noise amplitude volatility indicators; Counting the distribution characteristics of the noise amplitude volatility index set, taking the 99% quantile as the upper limit threshold of the noise fluctuation, and generating the noise amplitude fluctuation constraint boundary; A dynamic noise statistical threshold is established according to the noise amplitude fluctuation constraint boundary and the background noise amplitude statistic.

5. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps for obtaining the amplitude feature set of each segment are: Based on the segmented high-frequency electrical signal sequence, the voltage signal sampling points and the current signal sampling points in each signal time segment are traversed, and a sliding window traversal method is adopted to extract the maximum absolute value of the voltage signal in each signal time segment as the instantaneous peak value, calculate the root mean square value of all sampling points of the voltage signal as the effective value, count the difference between the highest point and the lowest point of the voltage signal as the peak-to-peak value, calculate the arithmetic mean of the absolute value of the voltage signal sampling points as the absolute mean value, and at the same time count the number of times the current signal crosses the zero point from positive to negative in each time segment as the zero-crossing rate, so as to generate an amplitude parameter set for each signal time segment; The voltage instantaneous peak value, voltage effective value, voltage peak-to-peak value, and voltage absolute average value in the amplitude parameter set of each signal time segment are combined into a voltage amplitude feature vector according to voltage category, the current zero-crossing rate is combined into a current feature vector separately, and the voltage feature vector and the current feature vector are spliced ​​into a multidimensional amplitude feature vector according to the time segment sequence, so as to generate a multidimensional amplitude feature vector set containing the joint features of voltage and current; According to the original timestamp order of the segmented high-frequency electrical signal sequence, all feature vectors in the multidimensional amplitude feature vector set are time-sequentially aligned by timestamp, and the start timestamp and end timestamp labels of the corresponding signal time segment are added to each feature vector to generate amplitude feature sets for each segment.

6. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps of acquiring the candidate discharge pulse set are: Based on the amplitude feature set of each segment, the voltage peak value, voltage effective value, voltage peak-to-peak value, voltage absolute average value and current zero-crossing rate feature items in each signal time segment are traversed, the values ​​of the feature items are extracted one by one, and they are matched one by one with the preset dynamic noise statistical threshold, and a difference value set of each feature item and the dynamic noise statistical threshold is generated; All feature item difference values ​​of each signal time segment in the difference value set are traversed and judged. If the voltage peak difference value is greater than zero, the voltage effective value difference value is greater than zero, the voltage peak-to-peak difference value is greater than zero, the voltage absolute average difference value is greater than zero, or the current zero-crossing rate difference value is greater than zero, then the signal time segment is marked as a feature over-limit state, and a feature item over-limit marking set is generated; According to the feature item over-limit mark set, all signal time segments marked as feature over-limit states are screened, and the time segment data in the corresponding original segmented high-frequency electrical signal sequence are extracted and integrated into a candidate discharge pulse set in timestamp order.

7. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps of obtaining the autocorrelation function are: Based on the candidate discharge pulse set, traverse each signal time segment, extract discrete sampling values ​​of voltage signals or current signals within a corresponding time stamp range in the associated segmented high-frequency electrical signal sequence, and generate original time domain signal data of the signal time segment; Perform de-averaging on the original time domain signal data of each signal time segment, calculate the arithmetic mean of the signal sampling values, subtract the arithmetic mean from each sampling value, and generate a zero-centered time domain signal sequence; Based on the zero-centered time domain signal sequence, an autocorrelation function of the statistical characteristics of the voltage or current signal is established.

8. The high-frequency discharge signal recognition method according to claim 1, characterized in that: The steps for obtaining the high-frequency discharge recognition result are as follows: Based on the autocorrelation function, traverse the autocorrelation function curve corresponding to each candidate discharge pulse, extract the time delay coordinates of all local maximum points on the curve, and generate an autocorrelation function peak time delay set; According to the preset periodic fluctuation threshold of the power frequency cycle assessment function value, the delay intervals of adjacent maximum value points in the autocorrelation function peak delay set are calculated, the deviation rate between the interval sequence and the power frequency cycle is counted, and the delay interval segments with a deviation rate less than 5% are screened to generate a periodic characteristic fluctuation identification set; The periodic characteristic fluctuation identification set is traversed, and if a single candidate discharge pulse has three or more consecutive time delay intervals that meet the power frequency cycle deviation requirements, the pulse is determined to have significant periodic characteristics, and a high-frequency discharge identification result is generated.

Images