GIS partial discharge large pulse real-time screening method based on quantile and double threshold fusion strategy

Abstract

The application discloses a GIS partial discharge large pulse real-time screening method based on a quantile and double threshold fusion strategy, and comprises the following steps: acquiring a partial discharge pulse amplitude to be screened; writing the partial discharge pulse amplitude to be screened into a sampling window and covering old elements, and outputting a first large pulse amplitude; determining a target quantile based on the sampling window and comparing, and outputting a second large pulse amplitude; and determining a replacement probability based on the second large pulse amplitude and double thresholds and performing replacement selection. The application takes the quantile as a qualified line, adjusts the replacement strength by the double thresholds, stably retains the large pulse under constant level overhead, reduces the storage and computing power burden on site, and is beneficial to subsequent PRPD feature calculation and insulation diagnosis.

Description

Technical Field

[0001] This invention relates to the field of partial discharge signal recognition technology, and in particular to a real-time screening method for large partial discharge pulses in GIS based on a quantile and dual threshold fusion strategy. Background Technology

[0002] With the development of online monitoring technology for gas-insulated switchgear (GIS), the monitoring system is now able to continuously collect partial discharge pulses around the clock and at high frequency, thereby fully recording the dynamic data of partial discharge during equipment operation.

[0003] In related technologies, field monitoring devices typically save all acquired raw partial discharge pulses and then calculate characteristic parameters such as partial discharge phase-resolved spectra in the backend to identify defect types. However, long-term continuous monitoring generates massive amounts of dense raw pulse data. On the one hand, the storage capacity and computing resources of field monitoring devices are limited, making it difficult to save all raw pulses for extended periods. This can easily lead to data overflow, data loss, and device lag, affecting the stable operation of the monitoring system. On the other hand, phase-resolved spectra calculation and insulation condition diagnosis mainly rely on large pulses within fixed time intervals. Redundant, invalid, and interfering pulses not only increase the computational load but also reduce feature extraction efficiency and diagnostic accuracy.

[0004] To compress the amount of stored data, existing methods often use fixed amplitude thresholds or a single dynamic threshold to select pulses. Fixed thresholds are difficult to adapt to changes in operating conditions; when using only dual thresholds for screening, convergence is faster, but it is easily affected by sudden fluctuations in the data stream, resulting in an unstable proportion of large pulses in the queue; when using only the quantile of the sliding sampling window as a threshold, it can approximately track the global distribution of large pulses, but it can only distinguish whether the acceptable threshold has been reached, and it is difficult to further distinguish pulses with higher amplitudes among the acceptable large pulses, resulting in limited improvement in the overall amplitude level of the queue. Therefore, existing screening methods are unable to keep up with changes in the data stream and stably retain large pulses for subsequent diagnosis under the constraints of single-judgment and constant-level computational overhead. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy. This method addresses the problem of how to screen continuously acquired partial discharge pulses in real time under the conditions of limited storage and computing power in on-site monitoring devices, stably retain large pulses, and reduce the interference of invalid pulses on subsequent diagnosis.

[0006] Technical solution

[0007] A real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy, the method comprising:

[0008] Obtain the amplitude of the partial discharge pulse to be screened;

[0009] The partial discharge pulse amplitude is written into the sampling window, overwriting the old elements in the sampling window, and the output is the first large pulse amplitude; the first large pulse amplitude is the partial discharge pulse amplitude after being written into the sampling window.

[0010] Based on the sampling window, the target quantile is determined, and the first largest pulse amplitude is compared with the target quantile. The output is the second largest pulse amplitude; the second largest pulse amplitude is the first largest pulse amplitude, which is greater than the target quantile.

[0011] The replacement probability is determined based on the second largest pulse amplitude and the dual thresholds, and a replacement selection is performed on the second largest pulse amplitude based on the replacement probability; the dual thresholds include a low threshold and a high threshold.

[0012] Beneficial effects

[0013] This invention determines the target quantile based on a sampling window, compares the amplitude of the first largest pulse with the target quantile to obtain the amplitude of the second largest pulse, and then determines the replacement probability based on the interval relationship between the second largest pulse amplitude and the low and high thresholds. Based on the replacement probability, it decides whether to write the second largest pulse amplitude into the queue. Amplitudes that do not reach the target quantile are discarded, while amplitudes that reach the target quantile enter a fixed-length queue according to different replacement probabilities. This allows the monitoring device to obtain large pulses for phase-resolved spectrum calculations while only storing the sampling window and the queue, and suppresses queue amplitude fluctuations caused by data stream fluctuations. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0015] Figure 2 This is a schematic diagram of the application environment and monitoring system structure of the method of the present invention;

[0016] Figure 3 This is a schematic diagram of the sampling window cyclic coverage and the first large pulse amplitude output of the present invention;

[0017] Figure 4 This is a schematic diagram of the target quantile determination and second largest pulse amplitude output process of the present invention;

[0018] Figure 5 This is a schematic diagram of the dual-threshold exponential moving average update of the present invention;

[0019] Figure 6 This is a schematic diagram of the replacement probability determination and replacement selection process of the present invention;

[0020] Figure 7 This is a schematic diagram of the complete real-time processing loop of the present invention;

[0021] Figure 8 This is a structural block diagram of the screening device of the present invention;

[0022] Figure 9 This is an analysis chart of the convergence curve of the large number proportion and the tracking curve of the threshold (quantile) in this invention. Detailed Implementation

[0023] Definitions:

[0024] Partial discharge pulse amplitude refers to the amplitude of a single partial discharge pulse acquired during online monitoring of gas-insulated switchgear. In real-time filtering, each newly arriving pulse corresponds to a partial discharge pulse amplitude to be filtered. For ease of description, the currently arriving partial discharge pulse amplitude can be denoted as x in the example below.

[0025] A queue is a fixed-capacity storage area used to store large pulses after filtering. It has a preset queue capacity, denoted as N.

[0026] Exponential Moving Average (EMA) is a weighted moving average method where more recent data is assigned greater weight, while the weight of earlier data gradually decreases exponentially. For low thresholds, the EMA is updated using the lower threshold's update factor; for high thresholds, it is updated using the higher threshold's update factor.

[0027] Replacement probability refers to the probability value of deciding whether to replace an element at a certain position in the queue with the second largest pulse amplitude after obtaining the second largest pulse amplitude.

[0028] Replacement selection refers to generating a random number r with a value between 0 and 1, comparing r with the replacement probability to determine whether to write the second largest pulse amplitude into the queue: if r is not greater than the replacement probability, then a replacement position is randomly determined within the queue index range and updated to the second largest pulse amplitude; if r is greater than the replacement probability, then the queue remains unchanged.

[0029] The technical solution of the present invention will be further described below with reference to the embodiments and accompanying drawings.

[0030] Example 1

[0031] like Figure 1 and Figure 2As shown in the embodiments of this application, the real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy can be applied to an online monitoring system for partial discharge in gas-insulated switchgear. The monitoring system includes a monitoring terminal installed on-site and an optional monitoring server. The monitoring terminal continuously acquires partial discharge pulse signals through sensors to obtain the amplitude of the partial discharge pulses to be screened. The screening method can be executed by the monitoring terminal or by a combination of the monitoring terminal and the monitoring server. The data storage system can be integrated into the monitoring terminal or located on a server or in the cloud, used to store the large pulses screened in the queue and the partial discharge phase-resolved maps generated accordingly.

[0032] In a further embodiment of this invention, the monitoring terminal can be a partial discharge online monitoring device, an edge computing device, or an embedded device with data acquisition and processing capabilities deployed at a substation or switchyard. The monitoring server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0033] like Figure 1 As shown, a real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy includes the following steps:

[0034] Step 1: Obtain the amplitude of the partial discharge pulse to be screened;

[0035] The partial discharge pulse amplitudes to be screened are derived from the raw partial discharge pulse data collected during the online monitoring of gas-insulated switchgear. After extracting the amplitude of the continuously acquired pulse signals, the monitoring terminal uses each newly arrived amplitude as a partial discharge pulse amplitude to be screened and processes it in real time.

[0036] In a further implementation of this embodiment, the monitoring terminal receives pulse amplitude x sequentially according to the acquisition time sequence. Upon reaching each x, a complete window update, quantile comparison, and replacement selection are performed without waiting for the batch of data to be collected before processing, thus meeting the real-time requirements of online monitoring.

[0037] Step 2: Write the partial discharge pulse amplitude into the sampling window, overwriting the old elements in the sampling window, and output the first largest pulse amplitude. The first largest pulse amplitude is the current pulse amplitude obtained after writing the partial discharge pulse amplitude to be filtered into the sampling window and overwriting the old elements in the window. The first largest pulse amplitude is the same as the partial discharge pulse amplitude value written into the sampling window this time, only the window has been updated.

[0038] The sampling window is a fixed-capacity buffer used to store the amplitude values ​​of the most recent partial discharge pulses and to calculate the target quantile. The sampling window has a preset window size, denoted as M. M is a small positive integer, such as 20 to 100. The sampling window only retains the most recently written M amplitude values; newly written amplitude values ​​overwrite old elements within the window, thus forming a sliding window. After writing is complete, the current partial discharge pulse amplitude is the first largest pulse amplitude.

[0039] The old element is the historical amplitude that will be written over in the next sampling window. In an implementation using cyclic overwriting, the old element is the amplitude that entered the sampling window earliest in the writing order and is located at the current overwriting position.

[0040] Further implementations of this embodiment, such as... Figure 3 As shown, the sampling window is maintained using a cyclic coverage method. Let the number of processed pulses be i, the preset window size be M, and the coverage position loc satisfy loc = i modulo M. By updating the old element at position loc in the sampling window with the amplitude of the first largest pulse, the window can be slid in constant time, ensuring that the sampling window always stores the latest M amplitudes of the first largest pulse.

[0041] Step 3: Determine the target quantile based on the sampling window, compare the first largest pulse amplitude with the target quantile, and output the second largest pulse amplitude. The second largest pulse amplitude is the first largest pulse amplitude that is greater than the target quantile; that is, the second largest pulse amplitude is the pulse amplitude that has passed the target quantile screening and entered the replacement selection stage. If the first largest pulse amplitude is not greater than the target quantile, a second largest pulse amplitude is not formed, and the first largest pulse amplitude is discarded.

[0042] The target quantile is a quantile threshold determined based on the amplitude of the first largest pulse currently stored in the sampling window, serving as the criterion for determining whether a large pulse meets the acceptable threshold. In a specific application, the target quantile is the 80th quantile of the sampling window, denoted as q80, which is the quantile value satisfying P(X≤q80)=0.8. A quantile is a point in a continuous distribution function corresponding to probability p: if 0<p<1, the quantile Za of a random variable X or its probability distribution is a real number satisfying P(X≤Za)=a. The amplitude of the first largest pulse is compared with the target quantile: if the amplitude of the first largest pulse is greater than the target quantile, it is determined as the second largest pulse amplitude and enters the subsequent replacement selection; if the amplitude of the first largest pulse is not greater than the target quantile, it is discarded, the queue is not updated, and the system waits for the next partial discharge pulse amplitude.

[0043] Further implementations of this embodiment, such as... Figure 4As shown, the amplitude of the first large pulse in the sampling window is sorted to obtain the sorting result. Then, the target quantile is determined from the sorting result according to the preset quantile ratio. The preset quantile ratio is used to characterize the qualification criterion for large pulses. In this embodiment, the preset quantile ratio is 80%, and the corresponding target quantile is the 80th quantile q80 of the sampling window. Since the sorting and quantile calculation are only performed on a small window with a capacity of M, when M is on the order of 20 to 100, the time cost of a single processing can still be controlled within the constant range.

[0044] Step 4: Determine the replacement probability based on the second largest pulse amplitude and the dual thresholds. Perform replacement selection on the second largest pulse amplitude based on the replacement probability. The dual thresholds include a low threshold (L) and a high threshold (H). Both the low and high thresholds are dynamically updated with the data stream using an exponential moving average, and the updated high threshold is greater than the low threshold. The low threshold provides a lower limit to the interval when the replacement probability is determined, while the high threshold distinguishes whether to increase the base replacement probability.

[0045] In this approach, given the second-largest pulse amplitude, the target quantile is no longer the sole criterion for whether to write an element into the queue. Instead, the replacement probability is determined by combining the synchronously updated low and high thresholds, and then the second-largest pulse amplitude is used to replace elements in the queue based on this probability. Thus, the target quantile serves as the pass / fail selection criterion, while the dual thresholds adjust the replacement intensity, with both acting sequentially.

[0046] In this embodiment, the target quantile is determined by a sampling window, which can approximate the distribution of large pulses in the current data stream with a smaller window capacity. This ensures that the amplitude of the second largest pulse entering the replacement selection stage meets the large pulse qualification criteria without having to save all historical pulses. By determining the replacement probability and performing replacement selection using dual thresholds, the probability of writing to the queue can be adjusted according to the intervals of the low and high thresholds for pulses that have exceeded the target quantile. This prevents all pulses exceeding the qualification line from entering the queue, thereby suppressing queue fluctuations and increasing the retention rate of higher amplitude pulses in the queue. The screening process is completed one by one as the pulses arrive, and the computational overhead remains constant, making it suitable for continuous acquisition scenarios in online monitoring of gas-insulated switchgear.

[0047] Before obtaining the amplitude of the partial discharge pulse to be screened, the process also includes: setting characteristic parameters and initializing screening information. The characteristic parameters include queue, sampling window and dual thresholds. The queue is set with a preset queue capacity, and the sampling window is set with a preset window capacity, which is smaller than the preset queue capacity.

[0048] In a further implementation of this embodiment, initialization is performed only once, specifically including: setting a preset queue capacity N, a preset window capacity M, a low threshold update coefficient α, a high threshold update coefficient β, and a basic replacement probability p. Wherein, M is less than N; the recommended value range for α is 0.01 to 0.05; β is less than α; and the recommended value range for p is 0.6 to 0.8. In a specific example, N is 1000, M is 100, α is 0.02, β is 0.005, p is 0.6, and the first replacement probability p1 is 0.8.

[0049] In a further implementation of this embodiment, the queue is initialized with a length of N and the initial values ​​of each position in the queue are set to 0; the sampling window is initialized with a length of M and the initial values ​​of each position in the sampling window are set to 0; and the dual thresholds are initialized by setting both the low threshold L and the high threshold H to 0, which are then dynamically updated with the data flow.

[0050] In a further implementation of this embodiment, a replacement selection rule is set: only when the second largest pulse amplitude is obtained will the decision to write it into the queue be made based on the replacement probability; if the second largest pulse amplitude is not obtained, the replacement selection will not be performed.

[0051] In this embodiment, the sampling window capacity is set to be less than the queue capacity, ensuring that quantile calculations are always performed within the small window, while the queue is used to store the filtered large pulses. Initializing both the low and high thresholds to 0 allows for the rapid establishment of a dual-threshold structure during the initial data stream phase, and maintains the high threshold greater than the low threshold in subsequent updates.

[0052] like Figure 3 As shown, writing the partial discharge pulse amplitude into the sampling window includes: determining the coverage position of the partial discharge pulse amplitude in the sampling window based on the number of processed pulses and the preset window capacity; and updating the old element to the first largest pulse amplitude.

[0053] A further implementation of this embodiment determines the coverage position of the partial discharge pulse amplitude within the sampling window based on the number of processed pulses and the preset window capacity, including: taking the remainder of the number of processed pulses divided by the preset window capacity and outputting the coverage position.

[0054] In a further implementation of this embodiment, the current value is denoted as the i-th arriving partial discharge pulse amplitude, with i incrementing from 0 or 1 in arrival order. The coverage position loc = i % M, where % represents the modulo operation. The old element at index loc in the sampling window is updated to the first largest pulse amplitude, thus completing one window write. When i is less than M, the element initialized to zero is covered; when i is not less than M, the first largest pulse amplitude written M pulses ago is covered, which is the oldest element written in the window and located at this coverage position.

[0055] In a further implementation of this embodiment, the sampling window can be implemented using a one-dimensional array. Each write operation addresses only one location and overwrites it, without moving the remaining elements within the window, resulting in a time complexity of O(1). The window always stores the amplitude values ​​of the most recent M largest pulses, thereby maintaining the ability to track the current data stream distribution.

[0056] In this embodiment, the coverage position is determined by taking the remainder of the number of processed pulses with respect to the preset window capacity. This enables a sliding window to be implemented in a cyclical coverage manner, avoiding the need to maintain complex queue head and tail pointers or frequently move data, and ensuring that window updates meet the constraints of real-time filtering on the single processing latency.

[0057] like Figure 4 As shown, determining the target quantile based on the sampling window includes: sorting the amplitude of the first large pulse and outputting the sorting results; determining the target quantile from the sorting results based on a preset quantile ratio; the preset quantile ratio is 80%.

[0058] It should be noted that the first largest pulse amplitude being sorted here refers to all the first largest pulse amplitudes currently stored in the sampling window, not just the first largest pulse amplitude newly written this time.

[0059] In a further implementation of this embodiment, the M amplitude values ​​of the first large pulse in the sampling window are arranged in ascending order to obtain a sorting result. When the preset quantile ratio is 80%, the target quantile is the amplitude value at the 80% position in the sorting result. In this embodiment, the target quantile q80 is determined with a = 0.8, and q80 is used as the large pulse qualification line.

[0060] In a further implementation of this embodiment, since M is relatively small, the amplitudes within the window can be completely sorted before the quantile positions are read; alternatively, a quantile approximation algorithm suitable for small samples can be used. Regardless of the specific sorting implementation used, the quantile calculation object is limited to the sampling window, and not to all historical pulses or the entire queue.

[0061] In this embodiment, using the 80th percentile of the sampling window as the target quantile allows the pass / fail criterion to closely match the amplitude distribution of the most recent data stream, rather than using a fixed threshold. Compared to a queue that only stores pass / fail pulses, the sampling window has a smaller capacity and updates faster, making it suitable as an approximate means of tracking the global pulse distribution in real time.

[0062] Further implementations of this embodiment, such as... Figure 4 As shown, the amplitude of the first largest pulse is compared with the target quantile, and the amplitude of the second largest pulse is output. This includes: if the amplitude of the first largest pulse is greater than the target quantile, the amplitude of the first largest pulse is determined as the amplitude of the second largest pulse; if the amplitude of the first largest pulse is not greater than the target quantile, the amplitude of the first largest pulse is discarded.

[0063] In a further implementation of this embodiment, a comparison is performed on the current first largest pulse amplitude x′: if x′>q80, then the second largest pulse amplitude=x′, and the dual threshold update and replacement selection is entered; if x′≤q80, then x′ is discarded, the low threshold and high threshold are not updated, no random number is generated, the queue is not rewritten, and the system directly waits for the next partial discharge pulse amplitude.

[0064] In a further implementation of this embodiment, the comparison is a basic screening in the fusion strategy to determine whether a passing threshold has been reached. The technical solution emphasizes comparing the newly arrived amplitude with the target quantile only once to meet the time constraint of a single processing step. The dual threshold does not replace this comparison; instead, it adjusts the replacement probability after the comparison passes.

[0065] In this embodiment, the first largest pulse amplitude that does not exceed the target quantile is not entered into the queue. This removes amplitudes that clearly do not meet the current data stream's large pulse distribution from the source, reducing the storage and subsequent phase-resolved spectrum calculations occupied by invalid pulses. The first largest pulse amplitude that exceeds the target quantile is treated as the second largest pulse amplitude and enters for secondary adjustment, preventing all pulses on the qualified line from being indiscriminately written into the queue.

[0066] Further implementations of this embodiment, such as... Figure 5 As shown, before determining the replacement probability based on the second largest pulse amplitude and the dual thresholds, the process includes: setting a low threshold update coefficient and updating the low threshold by an exponential moving average based on the second largest pulse amplitude; setting a high threshold update coefficient and updating the high threshold by an exponential moving average based on the second largest pulse amplitude; wherein the high threshold update coefficient is less than the low threshold update coefficient, and the high threshold is greater than the low threshold after the update.

[0067] The exponential moving average is a weighted moving average where more recent data has a higher weight, while the weight of earlier data decreases exponentially. A low threshold is used to follow changes in the data stream, providing a lower bound for determining the interval of replacement probability; a high threshold follows the data stream more slowly and is used to determine whether to increase the base replacement probability to raise the overall amplitude level of the queue.

[0068] In a further implementation of this embodiment, let the low threshold before the update be Lt-1, the high threshold be Ht-1, and the current second largest pulse amplitude be x″. Then, an implementable update method is as follows:

[0069] Lt=α·x″+(1-α)·Lt-1;

[0070] Ht=β·x″+(1-β)·Ht-1;

[0071] And keep Ht > Lt after the update. If Ht is not greater than Lt after the update, the high threshold can be adjusted to be greater than the low threshold. For example, the high threshold can be set to the sum of the low threshold and the preset interval, or the coefficient can be updated according to the high threshold and iterated until Ht > Lt is satisfied.

[0072] In a further implementation of this embodiment, α is set to 0.02 and β to 0.005. Since β < α, the lower threshold is more sensitive to changes in the data stream and can keep up with the shift in amplitude level more quickly; the higher threshold rises more slowly and serves as a benchmark for judging higher amplitude pulses, avoiding excessive raising of the threshold due to individual maxima and also avoiding drastic fluctuations in the threshold due to short-term volatility.

[0073] It should be noted that in this embodiment, the updates to the low and high thresholds occur after the second largest pulse amplitude has been obtained, that is, after the first largest pulse amplitude has exceeded the target quantile. Amplitudes that fail to pass the target quantile comparison are not included in this dual threshold update, so as to prevent amplitudes that do not meet the passing threshold from pulling the low and high thresholds to a lower level.

[0074] In this embodiment, the low and high thresholds are maintained with different update coefficients. This allows for the establishment of a set of dual thresholds that slowly change with the data flow after the target quantile has been determined as the pass line. The high threshold is always greater than the low threshold, ensuring that subsequent replacement probability judgments have a clear interval division, rather than treating the dual thresholds as two independent quantities that are unrelated to each other.

[0075] Further implementations of this embodiment, such as... Figure 6 As shown, the replacement probability is determined based on the second largest pulse amplitude and two thresholds, including: if the second largest pulse amplitude is greater than the high threshold, the basic replacement probability is increased, the first replacement probability is output and determined as the replacement probability; if the second largest pulse amplitude is greater than the low threshold but not greater than the high threshold, the basic replacement probability is determined as the replacement probability.

[0076] In a further implementation of this embodiment, the base replacement probability p takes a preset value between 0.6 and 0.8, and the first replacement probability p1 is greater than p, for example, p = 0.6, p1 = 0.8, or p = 0.7, p1 = 0.8.

[0077] In a further implementation of this embodiment, when the amplitude of the second largest pulse x″ > H, it indicates that the amplitude has exceeded the high threshold after already exceeding the target quantile. The replacement probability is set to p1 to increase the possibility of writing it into the queue and to prioritize replacing the relatively lower amplitude in the queue.

[0078] In a further implementation of this embodiment, when L<x″≤H, it indicates that the amplitude has passed the target quantile screening but has not yet exceeded the high threshold. The replacement probability is set to p, and it participates in the replacement selection according to the basic probability, so as to avoid such amplitudes rewriting the higher amplitude pulses already stored in the queue at too high a frequency.

[0079] In a further implementation of this embodiment, since the amplitude of the second largest pulse entering this step is already greater than the target quantile, this step no longer repeats the comparison with the target quantile. Instead, it selects p or p1 based solely on the interval relationship between the amplitude of the second largest pulse and the high and low thresholds. The role of the low threshold in this step is to provide the lower limit of the interval corresponding to the basic replacement probability, which, together with the high threshold, constitutes a dual-threshold interval, rather than adding another comparison with the pass line equivalent to the target quantile.

[0080] In this embodiment, the replacement probability varies with the range of the second largest pulse amplitude, rather than using the same write probability for all amplitudes exceeding the target quantile. Therefore, without changing the basic screening structure of "comparing with the target quantile once first," the chance of higher amplitude pulses entering the queue can be increased, while frequent replacements of the queue by medium amplitude pulses can be suppressed.

[0081] Further implementations of this embodiment, such as... Figure 6 As shown, the replacement selection of the second largest pulse amplitude is performed based on the replacement probability, including: generating a random number, the random number value ranges from 0 to 1; if the random number is not greater than the replacement probability, the replacement position is determined from the queue and updated to the second largest pulse amplitude; if the random number is greater than the replacement probability, the queue remains unchanged.

[0082] A further implementation of this embodiment, determining the replacement position from the queue, includes: randomly determining the replacement position within the index range of the queue.

[0083] In a further implementation of this embodiment, after determining the replacement probability P, a random number r is generated that follows a uniform distribution from 0 to 1. r can be understood as a sample from the replacement space: when P = 0.7, the range of r falling between 0 and 0.7 is greater than the range falling between 0.7 and 1. Therefore, replacement is performed when r ≤ P, and not when r > P. That is, the higher the replacement probability, the greater the chance that the second largest pulse amplitude will actually be written into the queue.

[0084] In a further implementation of this embodiment, when r is not greater than P, a replacement position is randomly determined within the queue index range of 0 to N-1, and the element at that position is updated to the second largest pulse amplitude. When r is greater than P, no position in the queue is rewritten, the queue remains unchanged, and the system waits for the next partial discharge pulse amplitude.

[0085] In a further implementation of this embodiment, the random determination of the replacement position can be achieved using uniform random rounding, ensuring that each position in the queue has an equal chance of being selected and avoiding local stagnation caused by always covering the head or tail of the queue. The replacement selection is triggered only after the second largest pulse amplitude is obtained; if the first largest pulse amplitude is not greater than the target quantile, no random number is generated.

[0086] In this embodiment, the decision to write is made by comparing a random number with the replacement probability, thus avoiding replacing the queue with all amplitudes exceeding the target quantile, thereby reducing queue oscillation. By randomly determining the replacement position within the index range, newly entering higher amplitude pulses can replace existing elements at any position in the queue. This is beneficial because after a long period of operation, most of the pulses retained in the queue are large pulses that should be retained.

[0087] like Figure 7 As shown in the above embodiments, the monitoring terminal performs the following loop for each newly arrived partial discharge pulse amplitude x until the end of the current monitoring period.

[0088] Step A: Receive x as the amplitude of the partial discharge pulse to be screened.

[0089] Step B: Calculate the coverage position loc = i % M, update the old element at position loc in the sampling window to x, and obtain the amplitude of the first large pulse. i is the number of pulses processed.

[0090] Step C: Sort the amplitude of the first largest pulse in the sampling window and determine the target quantile q80 according to the 80% preset quantile ratio.

[0091] Step D: Compare the amplitude of the first largest pulse with q80. If the amplitude of the first largest pulse is not greater than q80, discard the amplitude of the first largest pulse and proceed to step G; if the amplitude of the first largest pulse is greater than q80, determine the amplitude of the first largest pulse as the amplitude of the second largest pulse and proceed to step E.

[0092] Step E: Update the low threshold L using an exponential moving average based on the low threshold update coefficient α, and update the high threshold H using an exponential moving average based on the high threshold update coefficient β, while keeping H > L after the update.

[0093] Step F: If the amplitude of the second largest pulse is greater than H, the replacement probability is set to the first replacement probability p1; if the amplitude of the second largest pulse is greater than L but not greater than H, the replacement probability is set to the base replacement probability p. Generate a random number r between 0 and 1. If r is not greater than the replacement probability, randomly determine the replacement position from 0 to N-1 and update that position with the amplitude of the second largest pulse; if r is greater than the replacement probability, keep the queue unchanged.

[0094] Step G: Wait for the amplitude of the next partial discharge pulse, increment i, and repeat steps A to G.

[0095] In the above loop, window covering, remainder taking, one quantile comparison, double threshold update, one random number comparison, and one random position writing can all be completed when a single pulse arrives. The overall time complexity is on the order of O(1), which can enable continuous operation under the condition that the storage capacity and computing resources of the field monitoring device are limited.

[0096] In a further implementation of this embodiment, the pulses retained in the queue are identified as large pulses and stored, and a partial discharge phase-resolved spectrum is generated based on the stored large pulses. Since the queue capacity is fixed at N, the field device only needs to maintain a set of large pulses of length N and a sampling window of length M, without having to store a massive number of original pulses for a long time. This reduces the risk of data overflow, data loss, and device lag, and also reduces the interference of redundant pulses on feature extraction and insulation condition diagnosis.

[0097] Example 2

[0098] A further implementation of this embodiment, to illustrate the role of each part in the fusion strategy, may also provide a screening process using only exponential moving average dual thresholds. This process is not a necessary implementation of this application and is only used for comparative illustration.

[0099] In a further implementation of this embodiment, the queue capacity N, update coefficient α, low threshold L = 0, high threshold H = 0, and a base replacement probability p are initialized. For each newly arrived amplitude x, it is compared only once with the low threshold: if x > L, x is replaced at a random position in the queue according to the replacement probability; if x ≤ L, x is discarded directly. Both the low and high thresholds are updated with the data stream using an exponential moving average. The high threshold is used to adjust the replacement probability when x > L: the replacement probability is increased when x > H, and the base replacement probability is maintained when L < x ≤ H.

[0100] The comparison process converges relatively quickly, but it is highly dependent on exponential moving average updates. When sudden fluctuations occur in the data stream, the low threshold can be affected by short-term extreme values, and the proportion of large pulses in the queue is prone to fluctuations. Therefore, the fusion strategy does not use the low threshold as the passing line, but rather the target quantile of the sampling window, and moves the dual thresholds to the replacement probability adjustment stage.

[0101] Example 3

[0102] A further implementation of this embodiment may also provide a screening process using only a quantile approximation sliding window, which is also only for comparative illustration.

[0103] In a further implementation of this embodiment, the queue capacity N and sampling window capacity M are initialized, and the sampling window is initially filled with zeros. For each newly arrived amplitude x, the old elements in the sampling window are first overwritten by i%M, and then the 80th percentile q80 of the window is calculated. If x > q80, a position in the queue is randomly replaced; otherwise, it is discarded. Each new amplitude overwrites the oldest position in the window, so that the window maintains an approximation of the distribution of the most recent M amplitudes.

[0104] This comparison process can closely match the global large pulse distribution, and its accuracy in determining compliance is higher than that of the process using only dual thresholds. However, it can only distinguish whether the compliance threshold has been reached, and cannot adjust the intensity of the pulses written into the queue within the already compliance amplitude range, thus limiting the overall improvement in the amplitude level of the queue. The fusion strategy, while retaining the compliance threshold, adds range adjustment of the replacement probability using dual thresholds.

[0105] Example 4

[0106] like Figure 9 As shown, in a further implementation of this embodiment, to visually compare the three processing procedures described above, 100,000 randomly generated values ​​are used for verification. The queue capacity is fixed at 1000. The parameters are set as follows: low threshold update coefficient α = 0.02, high threshold update coefficient β = 0.005, sampling window capacity M = 100, base replacement probability p = 0.6, and first replacement probability p1 = 0.8. The changes in the proportion of large pulses with the amount of processed data, and the tracking curves of the low threshold or target quantile with the amount of processed data are statistically analyzed.

[0107] The large pulse proportion refers to the percentage of large pulses in the queue that should be retained. A higher tracking curve value indicates a higher threshold for entry into the queue. The tracking target for the dual-threshold control process is its low threshold; the tracking target for both the quantile control and fusion processes is the 80th percentile of the window; the fusion process also provides its low-threshold tracking curve to observe changes in the secondary adjustment baseline.

[0108] From the perspective of large pulse proportion convergence, all three processes can quickly reach a relatively stable range. The process using only dual thresholds converges the fastest, but the accuracy of the large pulse proportion is lower and it is more affected by data stream fluctuations; in the example results, the large pulse proportion is 81.90%. The process using only quantile sliding windows is the second fastest, better matching the global large pulse distribution; in the example results, the large pulse proportion is 90.40%, but it can only filter pulses that meet the acceptable threshold. The fusion process converges the slowest, but the large pulse proportion fluctuates less and has high stability; in the example results, the large pulse proportion is 95.50%.

[0109] From the perspective of threshold tracking, as the amount of processed data increases, the low thresholds or target quantiles of each process gradually rise and tend to stabilize, ultimately adapting to the stable distribution of the data, demonstrating dynamic adaptability. The fusion process exhibits minimal fluctuations after the curve stabilizes, indicating that the joint adjustment of the pass line and dual thresholds is less prone to misjudgment due to local fluctuations.

[0110] The above comparison shows that, under the premise that the time complexity is constant, using the target quantile of the sampling window as the pass line, using the dual thresholds for adjusting the replacement probability interval, and then using random numbers to complete the replacement selection can simultaneously take into account the pass determination and the improvement of the queue amplitude. This ensures that the pulses retained in the queue after long-term operation meet the storage target of "at least most of them should be large pulses", thereby supporting the subsequent partial discharge phase resolution spectrum calculation and insulation status diagnosis when field device resources are limited.

[0111] Example 5

[0112] like Figure 8 As shown in the figure, a further embodiment of this invention provides a real-time screening device for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy. This device corresponds one-to-one with the method embodiment described above and includes: an amplitude acquisition module for acquiring the amplitude of the partial discharge pulse to be screened; a window update module for writing the partial discharge pulse amplitude into a sampling window, overwriting the old elements in the sampling window, and outputting the first large pulse amplitude; a quantile comparison module for determining the target quantile based on the sampling window, comparing the first large pulse amplitude with the target quantile, and outputting the second large pulse amplitude; and a replacement module for determining the replacement probability based on the second large pulse amplitude and the dual thresholds, and performing a replacement selection on the second large pulse amplitude based on the replacement probability. Specific limitations of the device can be found in the limitations of the method described above, and will not be repeated here. Each module in the above device can be implemented entirely or partially through software, hardware, or a combination thereof.

[0113] Example 6

[0114] A further embodiment of this invention provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. This computer device can be an embedded controller in a field monitoring terminal or a server. In an exemplary embodiment, a computer-readable storage medium and a computer program product are also provided, wherein the computer program stored thereon, when executed by a processor, implements the steps in the above-described method embodiments.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. The patent protection scope of this application shall be determined by the appended claims.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these modifications or substitutions do 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 real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy, characterized in that, The method includes: Obtain the amplitude of the partial discharge pulse to be screened; The partial discharge pulse amplitude is written into the sampling window, overwriting the old elements in the sampling window, and the output is the first large pulse amplitude; the first large pulse amplitude is the partial discharge pulse amplitude after being written into the sampling window. The target quantile is determined based on the sampling window. The first largest pulse amplitude is compared with the target quantile, and the second largest pulse amplitude is output. The second largest pulse amplitude is greater than the first largest pulse amplitude of the target quantile. The replacement probability is determined based on the second largest pulse amplitude and the dual thresholds, and a replacement selection is performed on the second largest pulse amplitude based on the replacement probability; the dual thresholds include a low threshold and a high threshold.

2. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 1, characterized in that, Before obtaining the amplitude of the partial discharge pulse to be screened, the method further includes: Set feature parameters and initialize filtering information, the feature parameters including queue, sampling window and dual threshold; The queue is configured with a preset queue capacity, and the sampling window is configured with a preset window capacity, wherein the preset window capacity is smaller than the preset queue capacity.

3. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 2, characterized in that, The step of writing the partial discharge pulse amplitude into the sampling window includes: Based on the number of processed pulses and the preset window capacity, the coverage position of the partial discharge pulse amplitude in the sampling window is determined; The old element is updated to the amplitude of the first large pulse.

4. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 3, characterized in that, Determining the coverage position of the partial discharge pulse amplitude within the sampling window based on the number of processed pulses and the preset window capacity includes: The coverage position is output by taking the remainder of the number of processed pulses divided by the preset window capacity.

5. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 1, characterized in that, The step of determining the target quantile based on the sampling window includes: sorting the amplitude of the first large pulse and outputting the sorting result; determining the target quantile from the sorting result based on a preset quantile ratio; the preset quantile ratio is 80%.

6. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 1, characterized in that, The step of comparing the first largest pulse amplitude with the target quantile and outputting the second largest pulse amplitude includes: If the amplitude of the first large pulse is greater than the target quantile, the amplitude of the first large pulse is determined as the amplitude of the second large pulse. If the amplitude of the first large pulse is not greater than the target quantile, the amplitude of the first large pulse is discarded.

7. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 1, characterized in that, Before determining the replacement probability based on the second large pulse amplitude and the dual thresholds, the method further includes: Set a low threshold update coefficient, and update the low threshold by an exponential moving average based on the second large pulse amplitude; Set a high threshold update coefficient, and update the high threshold by an exponential moving average based on the second large pulse amplitude; Wherein, the high threshold update coefficient is less than the low threshold update coefficient, and the high threshold is greater than the low threshold after the update.

8. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 1, characterized in that, The determination of the replacement probability based on the second largest pulse amplitude and the dual thresholds includes: If the amplitude of the second large pulse is greater than the high threshold, the base replacement probability is increased, the first replacement probability is output and determined as the replacement probability; If the amplitude of the second large pulse is greater than the low threshold and not greater than the high threshold, the basic replacement probability is determined as the replacement probability.

9. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 2, characterized in that, The process of performing replacement selection on the second large pulse amplitude based on the replacement probability includes: Generate random numbers, wherein the random numbers range from 0 to 1; The random number is not greater than the replacement probability. The replacement position is determined from the queue and updated to the second largest pulse amplitude. If the random number is greater than the replacement probability, the queue remains unchanged.

10. The real-time screening method for large partial discharge pulses in GIS based on a quantile and dual-threshold fusion strategy according to claim 9, characterized in that, Determining the replacement position from the queue includes: The replacement position is randomly determined within the index range of the queue.

Images