Substation partial discharge detection method and system

Abstract

The invention relates to the technical field of substation operation and maintenance, particularly provides a substation partial discharge detection method and system, and aims to improve the precision of partial discharge detection and the intelligent level of equipment insulation state assessment. The partial discharge signal data set comprises a plurality of monitoring signal sequences composed of discharge pulse waveforms, then performing feature extraction on the partial discharge signal data set, generating a time domain and frequency domain feature set of each monitoring signal sequence to comprehensively reflect the intrinsic characteristics of the discharge signals, and then utilizing a preset discharge mode analysis model to analyze the partial discharge signals. And performing discharge mode analysis on the time domain and frequency domain feature set, accurately identifying discharge types, and finally, generating an equipment insulation state evaluation strategy according to a discharge type classification result, and transmitting the equipment insulation state evaluation strategy to a substation monitoring system in real time to trigger early warning operation, thereby effectively ensuring safe and stable operation of a power system.

Description

Technical Field

[0001] The present invention relates to the technical field of substation operation and maintenance, and more particularly, to a method and system for detecting partial discharge in a substation. Background Art

[0002] In the power system, as the core node for power transmission and distribution, the insulation status of substation equipment is directly related to the safe and stable operation of the power grid. Partial discharge, as an important indicator of equipment insulation deterioration, its accurate detection and timely warning are of crucial significance for preventing equipment failures and ensuring power grid safety. However, the existing substation partial discharge detection technologies still have many deficiencies and are difficult to meet the growing power grid safety requirements.

[0003] Traditional partial discharge detection methods often rely on single signal acquisition and analysis means, such as only evaluating the equipment insulation status by monitoring limited parameters such as the amplitude and frequency of discharge pulses. This method not only has difficulty in comprehensively capturing the complex characteristics of discharge signals, but is also easily affected by factors such as electromagnetic interference and noise, resulting in a significant reduction in the accuracy and reliability of detection results. In addition, there is a lack of efficient and intelligent discharge pattern recognition mechanisms in the existing technologies, mostly relying on manual experience or simple threshold judgment, making it difficult to accurately classify discharge types and thus unable to provide strong support for the evaluation of equipment insulation status.

[0004] More critically, the existing partial discharge detection technologies often lack deep integration with substation monitoring systems, and cannot achieve real-time feedback and warning of detection results. Even if partial discharge signals can be detected, the information is often not passed to operation and maintenance personnel in a timely manner or corresponding protection measures are not automatically triggered, thus missing the best opportunity to prevent failures. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, the present invention provides a method for detecting partial discharge in a substation, the method comprising: Obtaining a set of partial discharge signal data of target substation equipment, the set of partial discharge signal data including a plurality of monitoring signal sequences, each monitoring signal sequence consisting of at least one discharge pulse waveform; Performing signal feature extraction processing on the set of partial discharge signal data to obtain a time-domain feature set and a frequency-domain feature set for each monitoring signal sequence; Based on a preset discharge pattern analysis model, performing discharge pattern analysis on the time-domain feature set and the frequency-domain feature set to generate a discharge type classification result for the monitoring signal sequence; Generating an equipment insulation status evaluation strategy according to the discharge type classification result, and transmitting the equipment insulation status evaluation strategy to the substation monitoring system to trigger a warning operation.

[0006] In another aspect, the present invention further provides a substation partial discharge detection system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the present invention significantly improves the accuracy and comprehensiveness of substation partial discharge detection by integrating multi-source partial discharge signal data and fusing time-domain and frequency-domain feature analysis. Specifically, by obtaining a signal data set containing rich discharge pulse waveforms, and on this basis extracting time-domain and frequency-domain feature sets, the internal laws and feature differences of the discharge signals are effectively mined. Further, with the help of a preset discharge mode analysis model, the discharge types can be intelligently analyzed and classified, breaking through the limitations of traditional methods that only rely on single features or simple threshold judgments, and realizing a deep understanding and accurate discrimination of the discharge phenomenon. Finally, based on the discharge type classification results, an equipment insulation state evaluation strategy is dynamically generated and transmitted to the substation monitoring system in real time to trigger early warning operations, which not only improves the accuracy and timeliness of equipment insulation state evaluation, but also enhances the safety and reliability of substation operation, thus contributing to the intelligent operation and maintenance and fault prevention of the power system. Description of the Drawings

[0008] Figure 1 is a schematic flowchart of the execution process of the substation partial discharge detection method provided by the embodiment of the present invention.

[0009] Figure 2 is a schematic diagram of exemplary hardware and software components of the substation partial discharge detection system provided by the embodiment of the present invention. Detailed Embodiments

[0010] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flowchart of the substation partial discharge detection method provided by an embodiment of the present invention. The substation partial discharge detection method will be introduced in detail below.

[0011] Step S110: Obtain a partial discharge signal data set of the target substation equipment. The partial discharge signal data set includes a plurality of monitoring signal sequences, and each monitoring signal sequence is composed of at least one discharge pulse waveform.

[0012] In this embodiment, a large substation is used as an example for illustration. There are many electrical equipment in the substation, which is a key node for regional power supply. A variety of partial discharge monitoring devices are deployed for different equipment in the station, and these partial discharge monitoring devices continuously collect relevant data.

[0013] For example, in a certain high-voltage switchgear area of a substation, there is a partial discharge monitoring system. The sensors in it collect electromagnetic signals in the surrounding space to obtain partial discharge signals. One of the monitoring signal sequences has a discharge pulse waveform with a specific shape. Its amplitude may show a trend of first rising rapidly and then decreasing slowly during a certain period, accompanied by some small fluctuations. Through continuous collection at each monitoring point, a partial discharge signal data set containing multiple monitoring signal sequences is finally formed.

[0014] Step S120: Perform signal feature extraction processing on the partial discharge signal data set to obtain the time-domain feature set and frequency-domain feature set of each monitoring signal sequence.

[0015] Taking a certain monitoring signal sequence obtained from the high-voltage switchgear area as an example, detailed signal feature extraction work is carried out on it.

[0016] Step S121: Perform time-domain segmentation processing on the discharge pulse waveform in the monitoring signal sequence to obtain multiple pulse segmentation data blocks.

[0017] For the discharge pulse waveform of this monitoring signal sequence, time-domain segmentation is carried out according to the characteristics of the discharge pulse waveform. Since the discharge pulse waveform has a certain periodicity and amplitude change law, it can be divided according to a fixed time interval or according to significant change points of the amplitude. For example, set a time interval of Δt. Starting from the starting time t0 of the discharge pulse waveform, each segment is divided every Δt time. In this way, within the first Δt time period, pulse segmentation data block 1 is formed, which contains all the sampling point information of the discharge pulse waveform within the time period from t0 to t0 + Δt. In the second Δt time period, pulse segmentation data block 2 is generated, and so on. If there is an obvious mutation in the amplitude of the discharge pulse waveform at a certain moment, such as the amplitude suddenly rising or falling sharply from a relatively stable value, this mutation moment can be used as the segmentation basis. For example, taking this mutation moment as the boundary, two pulse segmentation data blocks before and after are divided, so that each data block can more accurately reflect the characteristics of the discharge pulse waveform in different stages.

[0018] Step S122: Call the preset time-domain feature extraction module to perform peak detection processing on each pulse segmentation data block to generate the time-domain peak distribution feature.

[0019] For each pulse segmented data block, call the preset time-domain feature extraction module. Taking the pulse segmented data block 1 as an example, this time-domain feature extraction module traverses the sampling point sequence therein one by one. Suppose the sampling point sequence is S1, S2, S3... Sn, and the time interval between adjacent sampling points is Δt_sample. The module compares the amplitude difference between adjacent sampling points with the preset differential threshold D. For example, for sampling points S1 and S2, calculate their amplitude difference as ΔS = |S2 - S1|. If ΔS is greater than the differential threshold D, it is considered that there may be a pulse starting point or other key feature points between S1 and S2. By comparing sampling points one by one, determine the pulse starting point position and the pulse ending point position. Suppose after comparison, the sampling point corresponding to the pulse starting point position is Si, and the sampling point corresponding to the pulse ending point position is Sj. Generate a pulse width sequence according to the time interval between the pulse starting point position and the pulse ending point position (i.e., (j - i)*Δt_sample). At the same time, extract the amplitude extreme points of all pulse waveforms in this pulse segmented data block. Suppose the amplitude extreme points in this data block are M1, M2, M3... Mk. Generate a pulse amplitude density feature based on the distribution density of the amplitude extreme points. For example, divide the amplitude range into several small intervals, count the number of amplitude extreme points in each small interval to determine the amplitude density. For example, divide the amplitude range from the minimum value to the maximum value into m small intervals. There are x amplitude extreme points in the nth small interval and y amplitude extreme points in the (n + 1)th small interval, etc. Describe the distribution of amplitude extreme points in different amplitude intervals in this way, and then obtain the pulse amplitude density feature. Finally, combine the pulse width sequence and the pulse amplitude density feature to form a time-domain peak distribution feature, which comprehensively reflects the peak-related characteristics of the pulse in the time domain.

[0020] Step S123: Perform statistical quantity calculation processing on the time-domain peak distribution feature to obtain a time-domain statistical feature set, where the time-domain statistical feature set includes a peak mean value, a pulse interval variance, and a rise time skewness.

[0021] For the generated time-domain peak distribution feature, perform detailed statistical quantity calculation to obtain a time-domain statistical feature set.

[0022] First, calculate the peak mean value. Suppose the amplitude extreme points in the time-domain peak distribution feature are M1, M2, M3... Mk. The calculation method of the peak mean value is to add the amplitudes of these amplitude extreme points and then divide by the number k of amplitude extreme points, that is, the peak mean value = (M1 + M2 + M3 +... + Mk) / k.

[0023] Next, calculate the variance of the pulse intervals. Specifically, first determine the time intervals between adjacent pulses. Assume that in the time-domain peak distribution characteristics, the starting points and ending points of a series of pulses are determined, and then the pulse intervals are obtained. For example, if the end time of pulse 1 is t11 and the start time of pulse 2 is t21, then the pulse interval is t21 - t11. By analogy, a series of pulse interval values I1, I2, I3... In are obtained. Calculate the average value of these pulse interval values as I_avg = (I1 + I2 + I3 +... + In) / n. Then calculate the variance of the pulse intervals, that is, the sum of the squares of the differences between each pulse interval value and the average value, divided by the number of pulse interval values. The variance of the pulse intervals = [(I1 - I_avg)² + (I2 - I_avg)² + (I3 - I_avg)² +... + (In - I_avg)²] / n.

[0024] Finally, calculate the skewness of the rise time. Specifically, first determine the rise time of each pulse waveform. Assume that the time from the starting point to the peak point of the pulse waveform is the rise time, and the rise time values R1, R2, R3... Rm are obtained respectively. Calculate the average value of these rise time values as R_avg = (R1 + R2 + R3 +... + Rm) / m. Then calculate the skewness of the rise time. The formula is skewness of the rise time = [(R1 - R_avg)³ + (R2 - R_avg)³ + (R3 - R_avg)³ +... + (Rm - R_avg)³] / [m * (standard deviation of the rise time values)³]. Through these calculations, a set of time-domain statistical characteristics including the peak mean, variance of the pulse intervals, and skewness of the rise time is obtained, which depicts the distribution characteristics of the discharge pulses in the time domain from different perspectives.

[0025] Step S124: Perform frequency-domain conversion processing on the discharge pulse waveforms in the monitoring signal sequence to generate frequency-domain energy distribution characteristics.

[0026] Continuing with the example of the monitoring signal sequence obtained from the high-voltage switchgear area, perform frequency-domain conversion processing on its discharge pulse waveforms. Use a set frequency-domain conversion method, such as Fourier transform, to convert the time-domain discharge pulse waveforms to the frequency domain. Assume that the time-domain discharge pulse waveform can be represented as the function f(t), and it is converted to the frequency-domain function F(f) through Fourier transform, where f represents the frequency. In the frequency domain, the magnitude of the corresponding amplitude of different frequency components reflects the energy distribution. For example, at a certain frequency f1, the amplitude of the frequency-domain function F(f1) is relatively large, indicating that the discharge pulse waveform has higher energy near this frequency; while at another frequency f2, the amplitude is smaller, meaning that the energy at this frequency is relatively low. Through this conversion, frequency-domain energy distribution characteristics are generated, which depict the energy distribution of the discharge pulse waveforms at different frequencies.

[0027] Step S125: Perform frequency band division processing on the frequency domain energy distribution characteristics to obtain multiple frequency band energy proportion characteristics, and associate and store the time domain statistical feature set with the frequency band energy proportion characteristics.

[0028] For example, assume that the frequency spectrum range is from f_min to f_max. According to the preset frequency band width Δf, it is divided into frequency band interval 1 (from f_min to f_min + Δf), frequency band interval 2 (from f_min + Δf to f_min + 2Δf) …… frequency band interval n (from f_min+(n - 1)Δf to f_max). The width of each frequency band interval can be dynamically adjusted according to the historical discharge signal spectrum characteristics. For instance, through the analysis of the spectrum characteristics of a large number of historical discharge signals, it is found that the probability of discharge characteristics appearing in a certain frequency band is relatively high and has a greater impact on the operation of the equipment. Then, the width of this frequency band interval is appropriately reduced to more precisely analyze the energy distribution in this frequency band.

[0029] Then calculate the energy integral value within each frequency band interval. Assume that within frequency band interval i, the frequency domain function is F_i(f), then the energy integral value E_i within this frequency band interval is E_i = ∫[f_i1, f_i2]|F_i(f)|²df, where f_i1 and f_i2 are the starting frequency and ending frequency of frequency band interval i respectively. Normalize the energy integral values of all frequency band intervals to obtain the energy proportion coefficient of each frequency band interval. For example, for frequency band interval 1, its energy proportion coefficient is E1 / (E1 + E2 + E3 + …… + En), where En is the sum of the energy integral values of all frequency band intervals. Sort the frequency band intervals according to the size of the energy proportion coefficient, and select the first preset number of frequency band intervals in the sorting result as the key frequency band characteristics. Assume that the preset number is p, and select the first p frequency band intervals with the largest energy proportion coefficients as the key frequency band characteristics. And bind the energy proportion coefficients of the key frequency band characteristics with the corresponding frequency band identifiers. Finally, associate and store the time domain statistical feature set with the frequency band energy proportion characteristics. Using the unique identifier of the signal sequence in the database as the index, store the corresponding time domain statistical feature set and frequency band energy proportion characteristics in the same record, which is convenient for quickly obtaining and associating relevant feature information when analyzing the discharge mode and the insulation state of the equipment later.

[0030] Step S130: Based on a preset discharge mode analysis model, perform discharge mode analysis on the time domain feature set and the frequency domain feature set to generate a discharge type classification result for the monitoring signal sequence.

[0031] Taking the monitoring signal sequence obtained from the high - voltage switchgear area as an example, use the preset discharge mode analysis model to perform discharge mode analysis on it.

[0032] Step S131: Map the time-domain statistical feature set in the time-domain feature set to a first feature vector, map the frequency-band energy ratio feature in the frequency-domain feature set to a second feature vector, and perform a normalization transformation on the first feature vector and the second feature vector.

[0033] Suppose the time-domain statistical feature set includes the peak mean M_avg, the pulse interval variance V, and the rise time skewness S, and map them to the first feature vector X = [M_avg, V, S]. For the frequency-band energy ratio feature in the frequency-domain feature set, suppose the energy ratio coefficient of frequency-band interval 1 is E1, the energy ratio coefficient of frequency-band interval 2 is E2... the energy ratio coefficient of frequency-band interval q is Eq, and map them to the second feature vector Y = [E1, E2, E3... Eq].

[0034] Then perform a normalization transformation on the first feature vector X and the second feature vector Y. The normalization transformation aims to make different features have the same scale for subsequent processing. For example, for each element in the first feature vector X, a normalization method is adopted. Suppose for the element x_i, the normalized element x_i' = (x_i - μ) / σ, where μ is the mean of this feature dimension and σ is the standard deviation. For the peak mean dimension, the mean μ_M and the standard deviation σ_M of this dimension are obtained through statistical analysis of a large amount of historical data, and then the peak mean is normalized. Similarly, the pulse interval variance and the rise time skewness are normalized. For the second feature vector Y, another normalization method is adopted. Suppose for the element y_i, the normalized element y_i' = (y_i - y_min) / (y_max - y_min), where y_min and y_max are the minimum and maximum values of this feature dimension respectively. By analyzing all the frequency-band energy ratio coefficients in the frequency-domain feature set, y_min and y_max are determined, so as to perform a normalization transformation on each element in the second feature vector Y, and obtain the normalized first feature vector X' and the second feature vector Y'.

[0035] Step S132: Call the feature fusion layer in the discharge mode analysis model to splice the normalized first feature vector and the second feature vector to generate a fused feature vector.

[0036] In the discharge mode analysis model, the feature fusion layer is called. The first feature vector X' and the second feature vector Y' after normalization transformation are input into the feature fusion layer. Assume the first feature vector X' = [x1', x2', x3'], and the second feature vector Y' = [y1', y2', y3'... yq']. The feature fusion layer concatenates these two vectors to generate a fused feature vector Z = [x1', x2', x3', y1', y2', y3'... yq']. Through this concatenation, the features in the time domain and frequency domain are fused together to form a fused feature vector containing richer information, providing more comprehensive feature data for subsequent discharge mode classification.

[0037] Step S133: Perform a vector dimension matching verification on the fused feature vector to confirm that its dimension is consistent with the number of input nodes of the classification decision layer of the discharge mode analysis model.

[0038] Perform a dimension matching verification on the generated fused feature vector Z. Assume the number of input nodes of the classification decision layer of the discharge mode analysis model is N. Count the dimension of the fused feature vector Z, assume it is n. Compare n with N. If n is equal to N, it means the dimension matches and subsequent processing can continue; if n is not equal to N, the fused feature vector needs to be adjusted. When n is less than N, some set values (such as zero elements) can be added to the fused feature vector to make its dimension reach N; when n is greater than N, some features may need to be screened or merged to ensure that the dimension is consistent with the number of input nodes of the classification decision layer. For example, if n is 2 less than N, add two zero elements at the end of the fused feature vector; if n is 3 more than N, by analyzing the correlation of each feature, merge some features with higher correlation to reduce the dimension to N.

[0039] Step S134: Perform a weight matrix multiplication process on the fused feature vector through the fully connected layer in the classification decision layer to generate an initial mode score vector.

[0040] In the classification decision layer, the fused feature vector Z is processed through the fully connected layer. Assume the weight matrix of the fully connected layer is W, and its dimension is related to the dimension of the fused feature vector Z and the number of discharge mode categories. The number of rows of the weight matrix W is equal to the number of discharge mode categories, and the number of columns is equal to the dimension of the fused feature vector Z. Represent the fused feature vector Z as a column vector and perform a matrix multiplication operation with the weight matrix W. The initial mode score vector S = W * Z. Through such a matrix multiplication operation, each feature in the fused feature vector Z is weighted and combined according to the weight matrix W to obtain an initial mode score vector containing scores for different discharge modes. This initial mode score vector reflects the likelihood scores of each discharge mode based on the current feature combination.

[0041] Step S135: Perform a non - linear activation function mapping on the initial pattern score vector to obtain the non - linear score values for each discharge pattern category.

[0042] Perform a non - linear activation function mapping on the initial pattern score vector S. Use a set non - linear activation function, such as the ReLU function or the Sigmoid function. Assume the initial pattern score vector S = [s1, s2, s3... sm]. After being processed by the non - linear activation function f, the non - linear score values N for each discharge pattern category are obtained as N = [f(s1), f(s2), f(s3)... f(sm)]. The non - linear activation function can introduce non - linear factors, enabling the model to learn more complex pattern relationships, thus better classifying the discharge patterns. For example, if the ReLU function is used, when si > 0, f(si)=si; when si ≤ 0, f(si)=0. Through such a transformation, the values in the initial pattern score vector are adjusted to highlight the score differences of certain discharge patterns.

[0043] Step S136: Input the non - linear score values into a probability normalization layer, calculate the probability values for each discharge pattern category and perform probability ranking to generate a discharge pattern prediction probability distribution sorted in descending order of probability values.

[0044] Input the non - linear score values N into the probability normalization layer. The probability normalization layer uses a set method, such as the Softmax function, to convert the non - linear score values into probability values. Assume the non - linear score values N = [n1, n2, n3... nm]. After being processed by the Softmax function, the probability values P for each discharge pattern category are P = [p1, p2, p3... pm], where pi = exp(ni) / ∑[j = 1 to m]exp(nj). Then these probability values are sorted in descending order of probability values to generate a discharge pattern prediction probability distribution sorted in descending order of probability values, which shows the order of the likelihood of each discharge pattern occurring under the current monitored signal sequence.

[0045] Step S137: Determine the discharge type classification result corresponding to the monitored signal sequence according to the discharge pattern prediction probability distribution, where the discharge type classification result includes at least one of surface discharge, internal air gap discharge, and floating potential discharge.

[0046] Based on the predicted probability distribution of the discharge pattern, determine the classification result of the discharge type corresponding to the monitoring signal sequence. For example, assume that in the predicted probability distribution of the discharge pattern, the probability value of surface discharge is p_surface, the probability value of internal air gap discharge is p_gap, and the probability value of floating potential discharge is p_suspension, and p_surface > p_gap > p_suspension. Then, determine that the discharge type corresponding to this monitoring signal sequence is surface discharge. By comparing the magnitudes of the probability values of different discharge patterns, select the discharge pattern with the largest probability value as the classification result.

[0047] Step S140: Generate an evaluation strategy for the insulation status of the device according to the classification result of the discharge type, and transmit the evaluation strategy for the insulation status of the device to the substation monitoring system to trigger a warning operation.

[0048] After determining the classification result of the discharge type of the monitoring signal sequence, taking surface discharge as an example, generate an evaluation strategy for the insulation status of the device.

[0049] Step S141: Match the preset insulation degradation level mapping table according to the classification result of the discharge type, and determine the current insulation degradation level of the target substation device.

[0050] In this embodiment, a preset insulation degradation level mapping table is provided, which details the corresponding relationship between different discharge types and insulation degradation levels. For example, for surface discharge, the mapping table stipulates that its corresponding insulation degradation level is a certain specific level, assumed to be level L. By matching the determined discharge type, surface discharge, with this mapping table, the current insulation degradation level of the target substation device is determined to be level L. This level reflects the current degradation degree of the insulation performance of the device and provides a basis for further evaluating the device status.

[0051] Step S142: Obtain the historical insulation detection records of the target substation device, and extract the degradation trend parameters from the historical insulation detection records.

[0052] In this embodiment, the historical insulation detection records of the target substation device are stored in a dedicated database, which details the relevant information of each insulation detection of the device, including the detection time, detection method, and various insulation performance indicators obtained.

[0053] Taking the device whose previously determined discharge type is surface discharge and insulation degradation level is level L as an example, retrieve the historical insulation detection records related to this device from the database. These historical insulation detection records cover multiple detection data from the time the device was put into use to the present or a past period.

[0054] For each detection record, it includes a series of insulation performance indicators, such as insulation resistance, dielectric loss factor, etc. These indicators reflect the insulation performance status of the equipment from different aspects. Assume that the insulation resistance indicator is represented by R, and the dielectric loss factor indicator is represented by D. In the detection records at different time points, R and D will have corresponding values. For example, in the detection at time t1, the insulation resistance is R1 and the dielectric loss factor is D1; in the detection at time t2, the insulation resistance is R2 and the dielectric loss factor is D2, and so on.

[0055] To extract the degradation trend parameters, first analyze the variation of the insulation performance indicators over time. For the insulation resistance R, calculate the change amount between two adjacent detections, that is, ΔR = R2 - R1. Similarly, for the dielectric loss factor D, calculate ΔD = D2 - D1. Then, combined with the detection time interval, for example, the time interval is Δt = t2 - t1, calculate the change rate of the insulation resistance as VR = ΔR / Δt, and the change rate of the dielectric loss factor as VD = ΔD / Δt. These change rates are part of the degradation trend parameters.

[0056] At the same time, other factors related to insulation degradation can also be considered. For example, analyze the fluctuation of the insulation resistance and dielectric loss factor over a period of time. Assume that within a relatively long time period T, there are multiple detection time points t1, t2, t3... tn, and the corresponding insulation resistance values are R1, R2, R3... Rn, and the dielectric loss factor values are D1, D2, D3... Dn. Calculate the fluctuation amplitude of the insulation resistance. For example, it is represented by calculating the average value of the absolute values of the differences between adjacent detection values, that is, the fluctuation amplitude AR = (|R2 - R1| + |R3 - R2| +... + |Rn - R(n - 1)|) / (n - 1). Similarly, calculate the fluctuation amplitude AD of the dielectric loss factor. These fluctuation amplitudes are also components of the degradation trend parameters.

[0057] Through comprehensive analysis of various insulation performance indicators and their changes in the historical insulation detection records, multiple degradation trend parameters are extracted. These degradation trend parameters comprehensively reflect the change trend of the equipment insulation performance over time.

[0058] Step S143: Generate a predicted value of the remaining life of the equipment by combining the current insulation degradation level and the degradation trend parameters, and determine the maintenance priority label according to the predicted value of the remaining life of the equipment.

[0059] After obtaining the current insulation degradation level as level L and a series of degradation trend parameters (such as the change rate VR of the insulation resistance, the change rate VD of the dielectric loss factor, the fluctuation amplitude AR of the insulation resistance, the fluctuation amplitude AD of the dielectric loss factor, etc.), start generating the predicted value of the remaining life of the equipment.

[0060] First, analyze the insulation performance thresholds corresponding to the current insulation degradation level. Assume that the insulation resistance threshold corresponding to level L is RL_threshold, and the dielectric loss factor threshold is DL_threshold. At the same time, obtain the currently detected insulation resistance value as R_current, and the dielectric loss factor value as D_current.

[0061] Calculate the remaining performance margin in terms of insulation resistance, that is, MR = RL_threshold - R_current. Similarly, calculate the remaining performance margin in terms of dielectric loss factor MD = DL_threshold - D_current.

[0062] Consider the change rate in the degradation trend parameters. Taking the insulation resistance change rate VR as an example, since it reflects the change amount of insulation resistance per unit time, it is used as the attenuation amount of insulation resistance per unit time. For the dielectric loss factor change rate VD, it is also used as the attenuation amount of the dielectric loss factor per unit time.

[0063] To generate the predicted value of the remaining life of the equipment, it is necessary to comprehensively consider the remaining performance margins and attenuation amounts of different insulation performance indicators. One way is to calculate the predicted values of the remaining life based on insulation resistance and dielectric loss factor respectively, and then conduct a comprehensive calculation.

[0064] The predicted value of the remaining life based on insulation resistance TR = MR / VR, and the predicted value of the remaining life based on dielectric loss factor TD = MD / VD.

[0065] However, in actual situations, the influence degrees of different insulation performance indicators on the remaining life of the equipment may be different. Therefore, corresponding weights need to be assigned to insulation resistance and dielectric loss factor. Assume that the weight of insulation resistance is WR, the weight of dielectric loss factor is WD, and WR + WD = 1.

[0066] The comprehensively predicted value of the remaining life of the equipment T = WR * TR + WD * TD.

[0067] After obtaining the predicted value T of the remaining life of the equipment, determine the maintenance priority label according to the pre-set rules. For example, set different intervals of the predicted value of the remaining life corresponding to different maintenance priorities. Assume that when T is greater than a certain value T1, the maintenance priority label is "lower"; when T is less than or equal to T1 and greater than another value T2, the maintenance priority label is "medium"; when T is less than or equal to T2, the maintenance priority label is "higher". In this way, accurately determine the maintenance priority label according to the predicted value of the remaining life of the equipment, providing guidance for formulating a reasonable maintenance plan.

[0068] Step S144: Generate an equipment insulation status evaluation strategy including a maintenance time window and suggestions for maintenance measures based on the maintenance priority label.

[0069] After determining the maintenance priority label, an equipment insulation status evaluation strategy is generated based on this.

[0070] If the maintenance priority label is "high", it indicates that the insulation condition of the equipment is relatively severe and maintenance is required as soon as possible. At this time, according to relevant industry standards, the equipment operation status, and the experience of previous similar cases, the maintenance time window is determined. Suppose through analysis, for such a situation, the maintenance time window should be set within a relatively short time period starting from the current time, for example, set as the time interval [t_current, t_current + Δt1], where t_current is the current time and Δt1 is a relatively short time interval determined according to specific circumstances.

[0071] For the case where surface discharge leads to an insulation degradation level of L and a high maintenance priority, specific maintenance measure suggestions are generated. First, for the outer surface of the equipment, a comprehensive cleaning work is required to remove possible attached dust, dirt, and other impurities, because these impurities may affect the insulation performance of the equipment. When cleaning, special insulation cleaners and cleaning tools can be used and operated according to the set operation process.

[0072] Secondly, check whether there are any damages, peeling, etc. on the insulation coating of the equipment outer surface. If any damage to the insulation coating is found, it needs to be repaired or recoated in a timely manner. The repair or recoating work should select appropriate insulation coatings and follow the corresponding process requirements to ensure that the thickness, uniformity, etc. of the coating meet the standards.

[0073] In addition, conduct a fastening inspection on the connection parts of the equipment. Due to factors such as vibration during the operation of the equipment, the connection parts may become loose, which may lead to uneven local electric field distribution and thus cause partial discharge. Use professional fastening tools to tighten the connection parts according to the specified torque value.

[0074] If the maintenance priority label is "medium", the maintenance time window can be set as a relatively long time interval [t_current, t_current + Δt2], where Δt2 is greater than Δt1. In terms of maintenance measure suggestions, in addition to the above-mentioned cleaning, checking the insulation coating, and tightening the connection parts, the equipment also needs to be more deeply detected, such as local discharge location detection, to further determine the specific location and degree of the discharge and provide a basis for more accurate maintenance in the future.

[0075] If the maintenance priority label is "low", the maintenance time window is set to a longer time interval [t_current, t_current + Δt3], where Δt3 is greater than Δt2. At this time, the maintenance measures recommended mainly focus on regular inspections, closely monitoring the operating status of the equipment, recording the changes in various insulation performance indicators, and at the same time monitoring and regulating the environmental parameters around the equipment to ensure that the equipment operates under suitable environmental conditions.

[0076] By formulating different maintenance time windows and detailed maintenance measure recommendations according to the maintenance priority label, a complete equipment insulation status evaluation strategy is formed, providing comprehensive and scientific guidance for the maintenance management of substation equipment.

[0077] Step S145: Generate a warning level identifier according to the maintenance priority label in the equipment insulation status evaluation strategy.

[0078] Assume that the maintenance priority label is "high", and according to the preset corresponding relationship, a matching warning level identifier is generated. For example, it is stipulated that the "high" maintenance priority label corresponds to the warning level identifier "severe". If the maintenance priority label is "medium", the corresponding warning level identifier is "relatively severe"; if the maintenance priority label is "low", the corresponding warning level identifier is "general". Through this clear corresponding relationship, the maintenance priority label is converted into a warning level identifier that is convenient for identification and processing.

[0079] Step S146: Package the warning level identifier and the equipment insulation status evaluation strategy into a standardized warning message.

[0080] Integrate and package the generated warning level identifier with the contents of the equipment insulation status evaluation strategy (including the maintenance time window, maintenance measure recommendations, etc.). For example, according to the set message format, place the warning level identifier in the set field position of the message, such as the first field in the message header; place the maintenance time window information in a certain middle field of the message, and the maintenance measure recommendations in the subsequent fields of the message, etc. Through this standardized packaging method, a standardized warning message is formed to ensure the unity of the message format and facilitate transmission and processing between different systems.

[0081] Step S147: Send the standardized warning message to the central control unit of the substation monitoring system through a preset communication protocol, triggering the central control unit to perform the corresponding sound and light alarm and log recording operations according to the warning level identifier.

[0082] Using a preset communication protocol, such as a communication protocol commonly used in a certain power system, the standardized warning message is sent out. This communication protocol stipulates rules such as data transmission methods, rates, and checksums. According to the protocol requirements, the standardized warning message is sent to the central control unit of the substation monitoring system through a network line or a wireless communication link.

[0083] When the central control unit receives the standardized warning message, it first analyzes the warning level identifier in the message. If the warning level identifier is "severe", the central control unit immediately triggers a strong audible and visual alarm. For example, it starts a high-decibel alarm to emit a sharp alarm sound, and at the same time flashes a prominent red warning light to attract the high attention of the staff. And, the central control unit will record this warning event in detail in the system log, including the warning time, warning level, corresponding equipment information, equipment insulation status assessment strategy, etc. If the warning level identifier is "relatively severe", a relatively weaker audible and visual alarm is triggered, such as an alarm sound of moderate volume and the flashing of an orange warning light, and relevant information is recorded in the log at the same time. If the warning level identifier is "general", a relatively gentle audible and visual alarm is triggered, such as a low-volume reminder sound and the flashing of a yellow warning light, and a detailed warning log is also recorded. In this way, according to different warning level identifiers, the corresponding audible and visual alarms and log recording operations are accurately triggered to promptly inform the staff of the insulation status of the equipment so as to take corresponding measures.

[0084] Furthermore, in a possible implementation manner, the embodiment of the present invention may further include a training step for the discharge mode analysis model, including: Step S210: Obtain a training data set, where the training data set includes a plurality of sample monitoring signal sequences with discharge type labels, and each sample monitoring signal sequence is correspondingly extracted with a sample time-domain statistical feature set and a sample frequency band energy ratio feature.

[0085] When training the discharge mode analysis model, first obtain the training data set. Taking this substation as an example, a plurality of sample monitoring signal sequences are screened out from the accumulated data in the past and the real-time monitoring data. These sample monitoring signal sequences come from the partial discharge monitoring of different equipment and different operating states in the substation.

[0086] Each sample monitoring signal sequence has a clear discharge type label, such as "surface discharge", "internal air gap discharge", "floating potential discharge", etc. At the same time, the same signal feature extraction process as before is performed on each sample monitoring signal sequence to obtain a sample time-domain statistical feature set and a sample frequency band energy ratio feature.

[0087] Taking a sample monitoring signal sequence as an example, the discharge pulse waveform is segmented in the time domain. Assuming segmentation is performed at time intervals of Δt_sample, multiple pulse segmentation data blocks are obtained. Peak detection processing is carried out on each pulse segmentation data block. By traversing the sampling point sequence, the amplitude differences between adjacent sampling points are compared with the differential threshold D_sample to determine the positions of the pulse start point and end point. Furthermore, a pulse width sequence is generated and the amplitude extreme points are extracted, and the time-domain peak distribution characteristics are obtained by combination. Statistical calculations are performed on the time-domain peak distribution characteristics to obtain a set of sample time-domain statistical characteristics, including the sample peak mean, sample pulse interval variance, sample rise time skewness, etc.

[0088] Frequency domain conversion processing is carried out on the discharge pulse waveform of the sample monitoring signal sequence to obtain frequency domain energy distribution characteristics. According to the preset frequency band division rule, the frequency spectrum range from f_min_sample to f_max_sample is divided into multiple frequency band intervals, and the width of each frequency band interval is Δf_sample. The energy integral value within each frequency band interval is calculated, and after normalization, the sample frequency band energy proportion characteristics are obtained. In this way, for each sample monitoring signal sequence, the corresponding set of sample time-domain statistical characteristics and sample frequency band energy proportion characteristics are obtained to construct a training dataset.

[0089] Step S211: Perform mean-variance normalization processing on the set of sample time-domain statistical characteristics in the training dataset to generate a sample normalized time-domain feature vector, where the mean parameter and variance parameter of the mean-variance normalization processing are respectively from the statistical results of historical data in the corresponding feature dimension.

[0090] For the set of sample time-domain statistical characteristics in the training dataset, mean-variance normalization processing is performed. Assume that the set of sample time-domain statistical characteristics includes the sample peak mean M_avg_sample, the sample pulse interval variance V_sample, and the sample rise time skewness S_sample.

[0091] The mean and variance of the corresponding feature dimension are statistically calculated from historical data. For example, for the sample peak mean dimension, through the analysis of a large number of historical sample peak mean data, the mean of this dimension is μ_M_sample and the variance is σ_M_sample. For the sample pulse interval variance dimension, the statistically obtained mean is μ_V_sample and the variance is σ_V_sample. For the sample rise time skewness dimension, the statistically obtained mean is μ_S_sample and the variance is σ_S_sample.

[0092] Normalize the sample peak mean value, and the formula is M_avg_sample' = (M_avg_sample - μ_M_sample) / σ_M_sample. Similarly, normalize the sample pulse interval variance, V_sample' = (V_sample - μ_V_sample) / σ_V_sample, and normalize the sample rise time skewness, S_sample' = (S_sample - μ_S_sample) / σ_S_sample. Combine the normalized sample peak mean value, sample pulse interval variance, and sample rise time skewness into a sample normalized time-domain feature vector X_sample = [M_avg_sample', V_sample', S_sample']. Through this mean-variance normalization process, the time-domain statistical features of different samples have the same scale, which is convenient for subsequent model training.

[0093] Step S212: Perform range scaling on the sample frequency band energy ratio features in the training dataset to generate a sample normalized frequency-domain feature vector, where the scaling coefficient of the range scaling is dynamically calculated according to the maximum historical value and the minimum historical value of each frequency band energy ratio feature.

[0094] For the sample frequency band energy ratio features in the training dataset, perform range scaling. Assume that the sample frequency band energy ratio features include the energy ratio coefficients of multiple frequency band intervals, such as E1_sample, E2_sample... En_sample.

[0095] For the energy ratio coefficient of each frequency band interval, dynamically calculate the scaling coefficient according to its maximum historical value and minimum historical value. For example, for the energy ratio coefficient E1_sample of frequency band interval 1, find the maximum value of the energy ratio coefficient of this frequency band interval in the historical data as E1_max_sample, and the minimum value as E1_min_sample. Calculate the scaling coefficient k1_sample = 1 / (E1_max_sample - E1_min_sample). Perform range scaling on E1_sample, E1_sample' = k1_sample * (E1_sample - E1_min_sample).

[0096] Similarly, for the energy ratio coefficient E2_sample of frequency band interval 2, calculate its scaling coefficient k2_sample = 1 / (E2_max_sample - E2_min_sample), and after processing, get E2_sample' = k2_sample * (E2_sample - E2_min_sample), and so on, process the energy ratio coefficients of all frequency band intervals.

[0097] Combine the energy ratio coefficients of each processed frequency band interval into a sample standardized frequency domain feature vector Y_sample = [E1_sample', E2_sample', ……, En_sample']. Through range scaling processing, map the sample frequency band energy ratio features to a set range to make them easier to process and compare in model training.

[0098] Step S213: Construct an initial discharge pattern analysis model. The initial discharge pattern analysis model includes the feature fusion layer and the classification decision layer. The number of input nodes of the feature fusion layer is set to the sum of the dimensions of the sample standardized time domain feature vector and the sample standardized frequency domain feature vector. The number of output nodes of the classification decision layer is the same as the number of preset discharge pattern categories.

[0099] Construct an initial discharge pattern analysis model, which mainly consists of a feature fusion layer and a classification decision layer.

[0100] First, determine the number of input nodes of the feature fusion layer. Since the dimension of the sample standardized time domain feature vector X_sample is assumed to be m, and the dimension of the sample standardized frequency domain feature vector Y_sample is assumed to be n, then the number of input nodes of the feature fusion layer is set to m + n.

[0101] Initialize the weight matrix of the feature fusion layer. The column dimensions of the weight matrix match the dimensions of the sample standardized time domain feature vector and the sample standardized frequency domain feature vector respectively, that is, the first m columns of the weight matrix correspond to the dimension of the sample standardized time domain feature vector, and the last n columns correspond to the dimension of the sample standardized frequency domain feature vector. The row dimension of the weight matrix is set to the preset fusion feature dimension, assumed to be p.

[0102] In the classification decision layer, configure a fully connected layer and a probability normalization layer. The weight dimension of the fully connected layer is determined according to the fusion feature dimension p and the number of discharge pattern categories q. Assume that the weight matrix of the fully connected layer is W_fc, with q rows and p columns. The input dimension of the probability normalization layer is the same as the output dimension of the fully connected layer, that is, q.

[0103] At the same time, perform dimension verification on the output of the feature fusion layer. When it is detected that the dimensions of the sample standardized time domain feature vector and the sample standardized frequency domain feature vector change, for example, due to data processing or new data feature extraction methods, the dimension of X_sample becomes m' and the dimension of Y_sample becomes n', then trigger the dimension adaptive adjustment operation of the weight matrix. Adjust the column dimension of the weight matrix so that its first m' columns correspond to the new dimension of the sample standardized time domain feature vector, and the last n' columns correspond to the new dimension of the sample standardized frequency domain feature vector, ensuring that the splicing operation of the feature fusion layer satisfies the vector dimension consistency constraint.

[0104] Verify the dimensional consistency of each feature dimension during the weight update process of the fully connected layer in the classification decision layer. During the training process, when it is detected that the input feature dimension units are inconsistent, for example, due to data entry errors or feature calculation errors, the dimension of a certain feature does not match that of other features, the training process is interrupted and a feature re-normalization operation is triggered. Re-check and process the relevant features until the dimension units of all input features pass the uniformity check, ensuring the accuracy and stability of model training.

[0105] Step S214: Input the sample standardized time-domain feature vector and the sample standardized frequency-domain feature vector into the initial discharge mode analysis model batch by batch, calculate the cross-entropy loss value between the discharge mode prediction probability distribution output by the initial discharge mode analysis model and the discharge type label, and perform backpropagation gradient update on the model parameters of the initial discharge mode analysis model according to the cross-entropy loss value.

[0106] The sample standardized time-domain feature vector and the sample standardized frequency-domain feature vector are input into the initial discharge mode analysis model in sequence according to a certain sample batch division method. Assume that each sample batch contains b samples. For each batch, the sample standardized time-domain feature vectors X_sample_batch and the sample standardized frequency-domain feature vectors Y_sample_batch of these b samples are respectively combined into matrix forms for easy model processing.

[0107] In this embodiment, first, through the feature fusion layer, X_sample_batch and Y_sample_batch are concatenated. The feature fusion layer processes the two vectors according to the previously set weight matrix to generate a fused feature vector matrix Z_batch. In this process, each row of the weight matrix performs a multiplication operation with the elements of the corresponding sample standardized time-domain feature vector and the sample standardized frequency-domain feature vector, and the results are added to obtain each element of the fused feature vector matrix Z_batch.

[0108] Next, the fused feature vector matrix Z_batch enters the fully connected layer of the classification decision layer. The weight matrix W_fc of the fully connected layer performs a matrix multiplication operation with Z_batch to obtain an initial mode score vector matrix S_batch. The specific calculation process is that each row of W_fc performs a multiplication operation with each column of Z_batch and accumulates to obtain each element of S_batch.

[0109] Then, perform a non-linear activation function mapping on the initial pattern score vector matrix S_batch, for example, using the Softmax function. The Softmax function calculates each element in S_batch to generate the discharge pattern prediction probability distribution matrix P_batch. Assuming the element in S_batch is s_ij (i represents the sample serial number in the sample batch, and j represents the discharge pattern category serial number), after being processed by the Softmax function, the element p_ij in P_batch = exp(s_ij) / ∑[k = 1 to q]exp(s_ik), where q is the number of discharge pattern categories.

[0110] At this time, compare the discharge pattern prediction probability distribution matrix P_batch with the discharge type label matrix L_batch corresponding to this batch of samples. The discharge type label matrix L_batch is a matrix with the same dimension as P_batch, where the element at the position corresponding to the actual discharge type of the sample is 1, and the rest are 0.

[0111] Calculate the cross-entropy loss value to measure the difference between the prediction result and the actual label. For each sample i, the cross-entropy loss value CE_i = -∑[j = 1 to q]L_ij * log(p_ij). The cross-entropy loss value of the entire batch CE_batch = ∑[i = 1 to b]CE_i.

[0112] According to the cross-entropy loss value CE_batch, perform backpropagation gradient update on the model parameters of the initial discharge pattern analysis model. The backpropagation process starts from the classification decision layer. First, calculate the error gradient between the output of the probability normalization layer (i.e., P_batch) and the label (L_batch). Then, backpropagate this error gradient to the fully connected layer to calculate the gradient of the fully connected layer weight matrix W_fc. Next, the error gradient continues to backpropagate to the feature fusion layer to calculate the gradient of the feature fusion layer weight matrix.

[0113] Taking the fully connected layer weight matrix W_fc as an example, assuming the element of W_fc is w_kl, the gradient calculation process involves the partial derivative of the error with respect to w_kl. Through the chain rule, the gradient of w_kl is gradually derived from the error of the output layer. For example, first calculate the partial derivative of the error with respect to the output of the fully connected layer (i.e., S_batch), and then multiply it by the partial derivative of S_batch with respect to w_kl to obtain the gradient of w_kl.

[0114] After obtaining the gradients of the weight matrices of each layer, the model parameters are updated according to the gradient descent algorithm. Assuming the learning rate is η, the update formula for the weight matrix W_fc of the fully connected layer is W_fc' = W_fc - η * ∇W_fc, where ∇W_fc is the gradient of W_fc. Similarly, the weight matrix of the feature fusion layer is updated to adjust the model parameters in the direction of reducing the cross-entropy loss value, gradually optimizing the performance of the model.

[0115] Step S215: During the gradient update process, a dynamic learning rate adjustment strategy is adopted to decay and control the learning rate parameter of the optimizer. The dynamic learning rate adjustment strategy is adaptively adjusted based on the product factor of the training iteration number and the descent rate of the cross-entropy loss value.

[0116] During the gradient update process of the model parameters, to better optimize the model performance, a dynamic learning rate adjustment strategy is adopted. First, define the training iteration number as t, and the cross-entropy loss value calculated in each iteration is CE_t.

[0117] Calculate the descent rate of the cross-entropy loss value. Assume that at the current iteration t and the previous iteration t - 1, the cross-entropy loss values are CE_t and CE_(t - 1) respectively. Then the descent rate of the cross-entropy loss value DR = (CE_(t - 1) - CE_t) / CE_(t - 1).

[0118] Dynamically adjust the learning rate according to the product factor of the training iteration number t and the descent rate DR of the cross-entropy loss value. Let the product factor be F = t * DR.

[0119] Assume the initial learning rate is η_0, and adjust the learning rate according to the product factor F. For example, a simple adjustment method can be adopted, and the new learning rate η_t = η_0 / (1 + F). As the training progresses, if the cross-entropy loss value drops rapidly, that is, DR is large, and at the same time the training iteration number t increases, the product factor F will increase, resulting in a gradual decrease in the learning rate η_t. In this way, in the initial stage of training, the learning rate is large, and the update amplitude of the model parameters is large, enabling rapid exploration of the solution space; while in the later stage of training, the learning rate gradually decreases, and the update of the model parameters is more refined, avoiding missing the optimal solution or falling into a local optimum. Through this dynamic learning rate adjustment strategy, the model can converge more stably and efficiently during training, improving the generalization ability and prediction accuracy of the model.

[0120] Furthermore, for example, in a possible implementation manner, after determining the discharge type classification result corresponding to the monitoring signal sequence according to the predicted probability distribution of the discharge mode, it further includes: Step S310: Generate an incremental training sample set for the discharge pattern analysis model based on the discharge type classification result and the historical execution record of the equipment insulation state evaluation strategy. The incremental training sample set includes the time-domain statistical feature set corresponding to the current monitoring signal sequence, the frequency band energy ratio feature, and the corrected discharge type label.

[0121] After determining the discharge type classification result corresponding to the monitoring signal sequence according to the discharge pattern prediction probability distribution, start generating the incremental training sample set for the discharge pattern analysis model in combination with the historical execution record of the equipment insulation state evaluation strategy.

[0122] Taking a certain determined discharge type classification result as an example, assume that the monitoring signal sequence is determined to be a certain discharge type through analysis, such as "surface discharge". At the same time, review the historical execution record of the equipment insulation state evaluation strategy to view information such as insulation detection, maintenance measures taken for this discharge type and related equipment, and the feedback of the subsequent equipment operation status.

[0123] Extract the time-domain statistical feature set and the frequency band energy ratio feature from the current monitoring signal sequence. The extraction process of the time-domain statistical feature set is the same as before. Perform time-domain segmentation on the discharge pulse waveform, and segment it into multiple pulse segmentation data blocks according to the time interval Δt. For each pulse segmentation data block, by traversing the sampling point sequence, compare the amplitude difference between adjacent sampling points with the differential threshold D to determine the starting and ending positions of the pulse, generate the pulse width sequence and extract the amplitude extreme points, and then obtain the time-domain peak distribution feature. Calculate the statistical quantities of the time-domain peak distribution feature to obtain the time-domain statistical feature set including the peak mean, pulse interval variance, and rise time skewness.

[0124] Perform frequency-domain conversion processing on the discharge pulse waveform to obtain the frequency-domain energy distribution feature. According to the preset frequency band division rule, divide the frequency spectrum range from f_min to f_max into multiple frequency band intervals, and the width of each frequency band interval is Δf. Calculate the energy integral value within each frequency band interval, and obtain the frequency band energy ratio feature after normalization processing.

[0125] Correct the discharge type label according to the historical execution record. For example, if the historical execution record shows that in some specific cases, the situation originally considered to be "surface discharge" is found to be more accurately the "internal air gap discharge" type after further detection and analysis, then correct the discharge type label of the current monitoring signal sequence to "internal air gap discharge".

[0126] Combine the time-domain statistical feature set corresponding to the current monitoring signal sequence, the frequency band energy ratio feature, and the corrected discharge type label into an incremental training sample. By performing the same processing on multiple such monitoring signal sequences, an incremental training sample set for the discharge pattern analysis model is generated. This incremental training sample set contains new actual monitoring data and label information corrected according to historical experience, providing new training data for further optimizing the discharge pattern analysis model.

[0127] Step S311: Perform sliding window mean filtering on the time-domain statistical feature set in the incremental training sample set to generate a denoised time-domain statistical feature vector, and perform dynamic normalization on the frequency band energy ratio feature based on the historical energy distribution curve to generate a normalized frequency-domain feature vector.

[0128] For the time-domain statistical feature set in the incremental training sample set, sliding window mean filtering is used to remove possible noise interference. Assume that the time-domain statistical feature set includes the peak mean M_avg, the pulse interval variance V, and the rise time skewness S.

[0129] Taking the peak mean as an example, define a sliding window size of window_size. Arrange the peak mean data in a set order, such as in the order of monitoring time. Starting from the first data point, take window_size data points and calculate their average. Assume these window_size data points are M1, M2, ……, M_window_size, then the average of the first window is M_avg_1 = (M1 + M2 + …… + M_window_size) / window_size. Then, slide the window one data point backward, take M2, M3, ……, M_window_size + 1 these window_size data points, and calculate the average M_avg_2 = (M2 + M3 + …… + M_window_size + 1) / window_size, and so on. In this way, perform sliding window mean filtering on the peak mean data to obtain the denoised peak mean sequence.

[0130] Apply the same method to the pulse interval variance and rise time skewness data to obtain the denoised pulse interval variance sequence and rise time skewness sequence respectively. Combine the denoised peak mean, pulse interval variance, and rise time skewness into a denoised time-domain statistical feature vector X_denoised.

[0131] For the frequency band energy ratio feature, perform dynamic normalization based on the historical energy distribution curve. First, obtain the global maximum energy value E_max and the minimum energy value E_min recorded for each frequency band energy ratio feature during the historical training phase.

[0132] For each energy proportion coefficient \(E_i\) of each frequency band interval in the current frequency band energy proportion feature, map it to the interval \([0, 1]\) to generate a normalized energy proportion coefficient \(E_i'\). The mapping formula is: \(E_i'=(E_i - E_{min}) / (E_{max}-E_{min})\).

[0133] Perform a distribution uniformity test on the normalized energy proportion coefficients. Assume a set outlier threshold \(threshold\). When it is detected that the normalized energy proportion coefficient \(E_i'\) of any frequency band interval exceeds \(threshold\), replace it with the moving average of the historical energy proportion coefficients of the same frequency band. For example, for frequency band interval \(j\), if \(E_j'>threshold\), obtain the energy proportion coefficients \(E_{j1}, E_{j2}, \cdots, E_{jn}\) of this frequency band interval in the past several trainings from historical data, calculate the moving average \(E_{j\_avg}=(E_{j1}+E_{j2}+\cdots+E_{jn}) / n\), and replace \(E_j'\) with \(E_{j\_avg}\). Through such processing, ensure that the dimension of all frequency band energy proportion features is unified as a dimensionless proportional coefficient, and the distribution is more reasonable, generating a normalized frequency domain feature vector \(Y_{normalized}\).

[0134] Step S312: Invoke a pre-set model structure adapter to perform dimension verification on the number of input layer nodes of the discharge mode analysis model. When it is detected that the sum of the dimensions of the denoised time-domain statistical feature vector and the normalized frequency domain feature vector does not match the current number of input layer nodes, trigger an input layer node expansion operation, and add an equal number of input nodes to the dimension difference and initialize them with zero weights.

[0135] Invoke a pre-set model structure adapter to perform dimension verification on the number of input layer nodes of the discharge mode analysis model. Assume that the dimension of the denoised time-domain statistical feature vector \(X_{denoised}\) is \(m_{denoised}\), the dimension of the normalized frequency domain feature vector \(Y_{normalized}\) is \(n_{normalized}\), and the current number of input layer nodes of the discharge mode analysis model is \(N_{input}\).

[0136] Calculate the sum of the dimensions of the denoised time-domain statistical feature vector and the normalized frequency domain feature vector as \(m_{denoised}+n_{normalized}\). Compare \(m_{denoised}+n_{normalized}\) with \(N_{input}\).

[0137] If \(m_{denoised}+n_{normalized}\) is equal to \(N_{input}\), it means that the current number of input layer nodes can meet the input requirements of the new feature vector, and no adjustment is required.

[0138] If m_denoised + n_normalized is less than N_input, no additional operation is required either, because the model can handle inputs with smaller dimensions.

[0139] When m_denoised + n_normalized is greater than N_input, it indicates that the number of nodes in the current input layer is insufficient, and the input layer node expansion operation needs to be triggered. Calculate the dimension difference ΔN = (m_denoised + n_normalized) - N_input.

[0140] Add an equal number of input nodes to the input layer as the dimension difference ΔN. The weights of these newly added nodes are set to zero initially, so that during the subsequent training process, as new data is input, the model can learn and adjust the weights of these nodes according to the actual situation. In this way, ensure that the input layer of the discharge pattern analysis model can adapt to the new feature vector dimension and guarantee that the model can properly process the data in the incremental training sample set.

[0141] Step S313: Input the denoised time-domain statistical feature vector and the normalized frequency-domain feature vector into the updated discharge pattern analysis model in chronological order of timestamps for incremental training, calculate the loss gradient between the predicted probability distribution output by the model and the corrected discharge type label, and use the momentum optimization algorithm to iteratively update the model weight matrix.

[0142] Input the denoised time-domain statistical feature vector X_denoised and the normalized frequency-domain feature vector Y_normalized into the updated discharge pattern analysis model in chronological order of timestamps for incremental training.

[0143] The model first processes the input feature vectors. Similar to the previous training process, X_denoised and Y_normalized are concatenated through the feature fusion layer to generate the fused feature vector Z_incremental. The feature fusion layer performs multiplication and accumulation operations on the elements of X_denoised and Y_normalized according to the current weight matrix to obtain each element of Z_incremental.

[0144] The fused feature vector Z_incremental enters the fully connected layer of the classification decision layer. The weight matrix W_fc of the fully connected layer performs matrix multiplication with Z_incremental to obtain the initial pattern score vector S_incremental. The specific calculation method is to perform multiplication and accumulation operations on each row of W_fc and each column of Z_incremental to determine the value of each element of S_incremental.

[0145] Next, perform a non - linear activation function mapping on the initial pattern score vector \(S_{incremental}\). For example, use the Softmax function again to generate the predicted probability distribution \(P_{incremental}\) of the model output. The Softmax function calculates for each element in \(S_{incremental}\). Assume the element in \(S_{incremental}\) is \(s_{ij}\) (where \(i\) represents the sample number and \(j\) represents the discharge pattern category number), then the element \(p_{ij}\) in \(P_{incremental}\) is \(p_{ij}=\frac{\exp(s_{ij})}{\sum_{k = 1}^{q}\exp(s_{ik})}\), where \(q\) is the number of discharge pattern categories.

[0146] Compare the predicted probability distribution \(P_{incremental}\) output by the model with the corrected discharge type label. The corrected discharge type label is also represented in matrix form, denoted as \(L_{incremental}\), where the element at the position corresponding to the actual corrected discharge type of the sample is 1, and the rest are 0.

[0147] Calculate the loss gradient between the predicted probability distribution \(P_{incremental}\) and the corrected discharge type label \(L_{incremental}\). Taking the cross - entropy loss function as an example, for each sample \(i\), the cross - entropy loss value \(CE_i =-\sum_{j = 1}^{q}L_{ij}\times\log(p_{ij})\). By taking the derivative of the cross - entropy loss value with respect to the elements of the weight matrices of each layer of the model, the loss gradient is obtained. For example, for the element \(w_{kl}\) of the fully - connected layer weight matrix \(W_{fc}\), its gradient is calculated through the chain rule. First, calculate the partial derivative of the cross - entropy loss value with respect to the output of the fully - connected layer (i.e., \(S_{incremental}\)), and then multiply it by the partial derivative of \(S_{incremental}\) with respect to \(w_{kl}\) to obtain the gradient of \(w_{kl}\).

[0148] Use the momentum optimization algorithm to iteratively update the model weight matrices. The momentum optimization algorithm introduces a momentum factor \(\gamma\). Assume the current iteration number is \(t\), the update amount of the weight matrix \(W_{fc}\) in the previous iteration is \(\Delta W_{fc}(t - 1)\), and the currently calculated gradient of the weight matrix \(W_{fc}\) is \(\nabla W_{fc}^t\). Then the update amount of the weight matrix \(W_{fc}\) this time is \(\Delta W_{fc}^t=\gamma\times\Delta W_{fc}(t - 1)-\eta\times\nabla W_{fc}^t\), where \(\eta\) is the learning rate. The weight matrix \(W_{fc}\) is updated to \(W_{fc}' = W_{fc}+\Delta W_{fc}^t\). Similarly, the weight matrix of the feature fusion layer is also updated according to the momentum optimization algorithm. In this way, using the momentum optimization algorithm can accelerate the convergence speed of the model, avoid falling into local optimal solutions during training, further optimize the performance of the discharge pattern analysis model, and enable it to better adapt to new incremental training samples.

[0149] Step S314: After each iterative update, perform a dimensional alignment verification on the concatenated output vector of the feature fusion layer. When it is detected that the dimensional changes of the denoised time-domain statistical feature vector and the normalized frequency-domain feature vector result in the dimension of the concatenated vector exceeding the input node capacity of the classification decision layer, automatically insert padding nodes and apply a weight decay penalty term to suppress overfitting.

[0150] After each iterative update of the model weight matrix using the momentum optimization algorithm, perform a dimensional alignment verification on the concatenated output vector of the feature fusion layer.

[0151] Suppose the dimension of the denoised time-domain statistical feature vector X_denoised becomes m_denoised' after a certain iterative update, and the dimension of the normalized frequency-domain feature vector Y_normalized becomes n_normalized'. The dimension of the vector generated after concatenation through the feature fusion layer is m_denoised'+n_normalized'.

[0152] The input node capacity of the classification decision layer is N_classification. Compare m_denoised'+n_normalized' with N_classification.

[0153] If m_denoised'+n_normalized' is less than or equal to N_classification, it indicates that the dimension of the concatenated vector meets the input requirements of the classification decision layer, and no additional operations are required.

[0154] When m_denoised'+n_normalized' is greater than N_classification, it means that the dimension of the concatenated vector exceeds the input node capacity of the classification decision layer. At this time, automatically insert padding nodes to match the dimension of the concatenated vector with the input node capacity of the classification decision layer. The number of inserted padding nodes is ΔN_fill=(m_denoised'+n_normalized')-N_classification.

[0155] To suppress the overfitting phenomenon, apply a weight decay penalty term to these padding nodes. The role of the weight decay penalty term is to penalize the nodes with larger weights during the model training process, making the model tend to learn smaller weights, thereby preventing the model from overfitting to the training data. Specifically, for the weights corresponding to the inserted padding nodes, in each iterative update, multiply their weight values by a decay factor β less than 1 (e.g., β = 0.99), so that the weights gradually decrease.

[0156] Meanwhile, during the subsequent training process, continuously monitor the dimensional changes of the denoised time-domain statistical feature vector and the normalized frequency-domain feature vector. Because in practical applications, with the continuous addition of new data or the adjustment of data processing methods, the dimensions of these two feature vectors may change again.

[0157] Suppose that in a subsequent training iteration, due to changes in data acquisition or processing methods, the dimension of the denoised time-domain statistical feature vector X_denoised changes from m_denoised' to m_denoised'', and the dimension of the normalized frequency-domain feature vector Y_normalized changes from n_normalized' to n_normalized''. Recalculate the dimension of the concatenated vector m_denoised'' + n_normalized'', and compare it with the input node capacity N_classification of the classification decision layer.

[0158] If m_denoised'' + n_normalized'' is still greater than N_classification, and the difference is different from before, for example, the difference becomes ΔN_fill', then the number of padding nodes needs to be readjusted. If ΔN_fill' is greater than ΔN_fill, then ΔN_fill' - ΔN_fill more padding nodes need to be inserted, and the weight decay penalty term is also applied to the weights of these newly inserted nodes, that is, multiply their weights by β during each iteration update. If ΔN_fill' is less than ΔN_fill, then ΔN_fill - ΔN_fill' padding nodes need to be deleted, and at the same time, the weights of the remaining padding nodes continue to be updated according to the rules of the weight decay penalty term.

[0159] If m_denoised'' + n_normalized'' is less than N_classification, it indicates that the current concatenated vector dimension is smaller than the input node capacity of the classification decision layer. Although the model can run normally at this time, in order to make full use of the computing resources of the classification decision layer, it may be necessary to further adjust the model structure. A possible adjustment method is to select some relatively unimportant feature dimensions for merging or removal based on the importance analysis of the current feature vector, so as to improve the computing efficiency and performance of the model. For example, by analyzing the correlation between features, if it is found that the correlation between two dimensions in the denoised time-domain statistical feature vector is extremely high, such as the correlation between the peak mean and another statistic related to the amplitude reaches a certain threshold (assumed to be 0.9), then these two dimensions can be considered merged into a new dimension. The merging method can be to perform a weighted sum of the feature values of these two dimensions, and the weights can be determined according to their contribution degrees to the discharge pattern classification in the historical data. Suppose through the analysis of historical data, it is found that the peak mean contributes more to the classification, with a weight of 0.6, and the weight of the other related statistic is 0.4, then the value of the new dimension is 0.6 * peak mean + 0.4 * the value of the other related statistic.

[0160] After such processing, recalculate the dimension of the denoised time-domain statistical feature vector, assuming it becomes m_denoised''', and then add it to the dimension n_normalized'' of the normalized frequency-domain feature vector to obtain the new concatenated vector dimension m_denoised''' + n_normalized''. Compare it with the input node capacity N_classification of the classification decision layer again to ensure dimension matching. If it still does not match, continue to adjust according to the above rules until the concatenated vector dimension matches the input node capacity of the classification decision layer, ensuring that the model can run stably and efficiently, and effectively suppressing the overfitting phenomenon, thereby improving the accuracy and generalization ability of the model for classifying different discharge patterns.

[0161] In addition, during the entire incremental training process, it is also necessary to closely monitor other performance indicators of the model, such as accuracy, recall, etc. For accuracy, it is defined as the ratio of the number of correctly classified samples to the total number of samples. Suppose after a certain incremental training, the discharge pattern classification of a batch of test samples is carried out. There are a total of N_test test samples, and the number of correctly classified samples is N_correct, then the accuracy Accuracy = N_correct / N_test. By continuously monitoring the accuracy, observe the performance changes of the model during the incremental training process. If the accuracy fluctuates or stagnates during the training process, it may be necessary to further adjust the training parameters, such as the learning rate, momentum factor, etc., or check whether there are problems in the data processing process.

[0162] Recall rate is also an important performance metric, which reflects the model's ability to identify a certain type of discharge pattern. Taking "surface discharge" as an example, assume that the number of samples that are actually "surface discharge" in the test samples is N_surface_true, and the number of "surface discharge" samples correctly identified by the model is N_surface_correct. Then the recall rate of the "surface discharge" pattern Recall_surface = N_surface_correct / N_surface_true. Similarly, for other discharge patterns, such as "internal air gap discharge" and "floating potential discharge", their recall rates can also be calculated respectively. By comprehensively analyzing performance metrics such as accuracy and recall rate, comprehensively evaluate the performance of the model during the incremental training process, promptly discover and solve possible problems, and ensure that the model can accurately and reliably classify the partial discharge patterns of substation equipment.

[0163] For example, after generating the predicted remaining life of the equipment by combining the current insulation deterioration level and the deterioration trend parameter, it further includes: Step S410, generating a set of monitoring parameter adjustment coefficients according to the predicted remaining life of the equipment and a preset dynamic adjustment rule for monitoring parameters. The set of monitoring parameter adjustment coefficients includes a sampling frequency adjustment factor and a detection threshold scaling ratio.

[0164] , after obtaining the predicted remaining life of the equipment, generate a set of monitoring parameter adjustment coefficients according to the preset dynamic adjustment rule for monitoring parameters. These rules are formulated based on a large amount of historical data and long-term research on the operating conditions of the equipment. For example, when the predicted remaining life of the equipment is within a certain range, corresponding different adjustment strategies are adopted. Assume that the predicted remaining life of the equipment is T_predicted, and there are multiple thresholds T1, T2 (T1 < T2). If T_predicted <= T1, it indicates that the insulation condition of the equipment is relatively severe and more intensive and precise monitoring is required. At this time, according to the rule, the sampling frequency should be significantly increased and the detection threshold should be made more stringent to capture more subtle changes in partial discharge signals. Through the correlation analysis of different stages of the predicted remaining life of the equipment and the monitoring effect in historical data, the sampling frequency adjustment factor is determined to be α1 and the detection threshold scaling ratio is β1. If T1 < T_predicted <= T2, the insulation condition of the equipment is relatively good but still needs attention. Correspondingly, the sampling frequency adjustment factor is α2 and the detection threshold scaling ratio is β2, where α2 is less than α1 and β2 is greater than β1. In this way, according to the predicted remaining life of the equipment, generate a set of monitoring parameter adjustment coefficients including the sampling frequency adjustment factor and the detection threshold scaling ratio to adapt to different insulation states of the equipment and improve the effectiveness of monitoring.

[0165] Step S420: Obtain the original sampling frequency and the original pulse amplitude detection threshold of the partial discharge signal acquisition module currently operating for the target substation equipment.

[0166] After generating the monitoring parameter adjustment coefficient set, it is necessary to obtain the relevant parameters of the partial discharge signal acquisition module currently operating for the target substation equipment. This acquisition module is responsible for collecting partial discharge signals. Its original sampling frequency is set to f_original, and the original pulse amplitude detection threshold is set to V_original. These parameters determine the frequency range of the acquired signals and the detection sensitivity to pulse amplitudes. For example, f_original determines the number of signal samples acquired per unit time. A higher sampling frequency can capture the details of the discharge pulse waveform more precisely, but it will also increase the data volume and processing complexity; V_original is used to determine whether the acquired signal is a valid discharge pulse. Only when the signal amplitude exceeds V_original will it be recognized as a possible discharge pulse for subsequent processing.

[0167] Step S430: Multiply the sampling frequency adjustment factor by the original sampling frequency to generate a dynamic sampling frequency value, and verify the dimension unit of the dynamic sampling frequency value to confirm that its unit is the preset hertz unit.

[0168] Multiply the sampling frequency adjustment factor α in the monitoring parameter adjustment coefficient set by the original sampling frequency f_original to obtain the dynamic sampling frequency value f_dynamic = α * f_original. When performing this operation, ensure the consistency of their dimensions. Since the sampling frequency adjustment factor α is a dimensionless proportional coefficient and the unit of the original sampling frequency f_original is hertz (Hz), the unit of the product result f_dynamic should also be hertz. Verify the dimension unit of f_dynamic to confirm that its unit conforms to the preset hertz unit. For example, if α = 1.5 and f_original = 1000 Hz, then f_dynamic = 1.5 * 1000 Hz = 1500 Hz. Through such operations and verifications, a dynamic sampling frequency value that meets the requirements is obtained.

[0169] Step S440: Multiply the detection threshold scaling ratio by the original pulse amplitude detection threshold to generate a dynamic detection threshold, and perform a range boundary check on the dynamic detection threshold to ensure that it is within the preset amplitude safety range.

[0170] Multiply the detection threshold scaling ratio β in the monitoring parameter adjustment coefficient set by the original pulse amplitude detection threshold V_original to obtain the dynamic detection threshold V_dynamic = β * V_original. Also, pay attention to the dimensional consistency. The detection threshold scaling ratio β is dimensionless, and the original pulse amplitude detection threshold V_original has an amplitude unit (assumed to be volts V), so the unit of V_dynamic is also volts. After obtaining V_dynamic, perform a range boundary check. Preset an amplitude safety interval [V_min, V_max], and check whether V_dynamic satisfies V_min <= V_dynamic <= V_max. For example, if β = 0.8, V_original = 5V, V_min = 3V, and V_max = 6V, then V_dynamic = 0.8 * 5V = 4V, which satisfies 3V <= 4V <= 6V and is within the safety interval. If V_dynamic exceeds this interval, it needs to be adjusted according to the exceeding situation. If V_dynamic < V_min, it may be necessary to appropriately increase the value of β and recalculate; if V_dynamic > V_max, it may be necessary to decrease the value of β and recalculate until V_dynamic is within the preset amplitude safety interval, ensuring that the detection threshold can effectively detect partial discharge pulses without misjudgment or missed judgment due to too high or too low a threshold.

[0171] Step S450, update the configuration parameters of the partial discharge signal acquisition module according to the dynamic sampling frequency value and the dynamic detection threshold, and trigger a re-acquisition operation of the partial discharge signal data set. Among them, the updated configuration parameters monitor the dimensional matching between the time-domain segmentation length and the frequency-domain resolution of each discharge pulse waveform in real time during the acquisition process. When it is detected that the number of sampling points corresponding to the time-domain segmentation length does not satisfy the preset Fourier transform point number constraint, automatically adjust the dynamic sampling frequency value to the nearest integer multiple frequency point value to force dimensional consistency.

[0172] After obtaining the dynamic sampling frequency value \(f_{dynamic}\) and the dynamic detection threshold \(V_{dynamic}\) that meet the requirements, use these two parameters to update the configuration parameters of the partial discharge signal acquisition module. After the update is completed, trigger the re-acquisition operation of the partial discharge signal data set. During the acquisition process, monitor the dimensional matching between the time-domain segment length and the frequency-domain resolution of each discharge pulse waveform in real time. Assume that the time-domain segment length is \(L_{time}\), the corresponding number of sampling points is \(N_{sample}\), and the frequency-domain resolution parameter is \(\Delta f_{freq}\). When performing frequency-domain analysis such as Fourier transform, there is a preset constraint relationship for the number of Fourier transform points. For example, in some cases, it is required that there is a set proportional relationship between \(N_{sample}\) and \(1 / \Delta f_{freq}\) (assumed to be \(k\), i.e., \(N_{sample}=k / \Delta f_{freq}\)). When it is detected that \(N_{sample}\) and \(k / \Delta f_{freq}\) do not satisfy this relationship, it means that the number of sampling points corresponding to the time-domain segment length and the frequency-domain resolution parameter do not satisfy the preset Fourier transform point constraint. At this time, in order to force the dimensional consistency, automatically adjust the dynamic sampling frequency value \(f_{dynamic}\) to the nearest integer multiple frequency point value. For example, if the current \(f_{dynamic}\) makes \(N_{sample}\) not meet the requirements, calculate the nearest integer multiple frequency point value \(f_{dynamic\_new}\) that can make \(N_{sample}\) satisfy the \(k / \Delta f_{freq}\) relationship, and update \(f_{dynamic}\) to \(f_{dynamic\_new}\), so as to ensure that during the acquisition process, the dimensional matching of time-domain and frequency-domain related parameters is achieved, enabling accurate subsequent signal analysis.

[0173] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a substation partial discharge detection system 100 that can implement the inventive concept provided by some embodiments of the present invention. For example, the processor 120 can be used on the substation partial discharge detection system 100 and is used to execute the functions in the present invention.

[0174] The substation partial discharge detection system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the substation partial discharge detection method of the present invention. Although only one server is shown in the present invention, for convenience, the functions described in the present invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0175] For example, the substation partial discharge detection system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the substation partial discharge detection system 100 may further include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention can be implemented according to these program instructions. The substation partial discharge detection system 100 further includes an I / O interface 150 between the computer and other input / output devices.

[0176] For the sake of convenience of description, only one processor is described in the substation partial discharge detection system 100. However, it should be noted that the substation partial discharge detection system 100 in the present invention may further include multiple processors. Therefore, the steps performed by one processor described in the present invention may also be jointly performed or separately performed by multiple processors. For example, if the processor of the substation partial discharge detection system 100 performs step A and step B, it should be understood that step A and step B may also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0177] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned substation partial discharge detection method is implemented.

[0178] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for detecting partial discharge in a substation, characterized in that, The method includes: Obtaining a partial discharge signal data set of target substation equipment, where the partial discharge signal data set includes a plurality of monitoring signal sequences, and each monitoring signal sequence is composed of at least one discharge pulse waveform; Performing signal feature extraction processing on the partial discharge signal data set to obtain a time-domain feature set and a frequency-domain feature set for each monitoring signal sequence; Based on a preset discharge pattern analysis model, performing discharge pattern analysis on the time-domain feature set and the frequency-domain feature set to generate a discharge type classification result for the monitoring signal sequence; Generating an equipment insulation state evaluation strategy according to the discharge type classification result, and transmitting the equipment insulation state evaluation strategy to the substation monitoring system to trigger an early warning operation.

2. The method for detecting partial discharge in a substation according to claim 1, characterized in that, The performing signal feature extraction processing on the partial discharge signal data set to obtain a time-domain feature set and a frequency-domain feature set for each monitoring signal sequence includes: Performing time-domain segmentation processing on the discharge pulse waveforms in the monitoring signal sequence to obtain a plurality of pulse segmentation data blocks; Invoking a preset time-domain feature extraction module to perform peak detection processing on each pulse segmentation data block to generate a time-domain peak distribution feature; Performing statistic calculation processing on the time-domain peak distribution feature to obtain a time-domain statistic feature set, where the time-domain statistic feature set includes a peak mean value, a pulse interval variance, and a rise time skewness; Performing frequency-domain conversion processing on the discharge pulse waveforms in the monitoring signal sequence to generate a frequency-domain energy distribution feature; Performing frequency band division processing on the frequency-domain energy distribution feature to obtain a plurality of frequency band energy ratio features, and associatively storing the time-domain statistic feature set and the frequency band energy ratio features.

3. The substation partial discharge detection method according to claim 2, characterized in that The invoking a preset time-domain feature extraction module to perform peak detection processing on each pulse segmentation data block to generate a time-domain peak distribution feature includes: Traversing the sampling point sequence in each pulse segmentation data block, comparing the amplitude difference between adjacent sampling points with a preset differential threshold to determine the pulse start point position and the pulse end point position; Generating a pulse width sequence according to the time interval between the pulse start point position and the pulse end point position, and extracting the amplitude extreme points of all pulse waveforms in the pulse segmentation data block; Generating a pulse amplitude density feature based on the distribution density of the amplitude extreme points, and combining the pulse width sequence and the pulse amplitude density feature to obtain a time-domain peak distribution feature.

4. The method for detecting partial discharge in a substation according to claim 2, characterized in that, The performing frequency band division processing on the frequency-domain energy distribution feature to obtain a plurality of frequency band energy ratio features includes: Dividing the spectrum range corresponding to the frequency-domain energy distribution feature into equal-width continuous frequency band intervals according to preset typical partial discharge frequency bands, and dynamically adjusting the width of each frequency band interval according to the historical discharge signal spectrum characteristics; Calculating the energy integral value in each frequency band interval, and normalizing the energy integral values of all frequency band intervals to obtain the energy ratio coefficient of each frequency band interval. Sort the frequency band intervals according to the magnitudes of the energy proportion coefficients, select the first preset number of frequency band intervals in the sorting result as the key frequency band features, and bind the energy proportion coefficients of the key frequency band features to the corresponding frequency band identifiers.

5. The partial discharge detection method for a substation according to claim 1, wherein Based on the preset discharge mode analysis model, perform discharge mode analysis on the time domain feature set and the frequency domain feature set to generate the discharge type classification result of the monitoring signal sequence, including: Map the time domain statistical feature set in the time domain feature set to a first feature vector, map the frequency band energy proportion feature in the frequency domain feature set to a second feature vector, and perform standardization transformation on the first feature vector and the second feature vector; Call the feature fusion layer in the discharge mode analysis model to splice the standardized first feature vector and the second feature vector to generate a fused feature vector; Perform verification on the vector dimension matching after standardizing the fused feature vector to confirm that its dimension is consistent with the number of input nodes of the classification decision layer of the discharge mode analysis model; Perform weight matrix multiplication processing on the fused feature vector through the fully connected layer in the classification decision layer to generate an initial mode score vector; Perform non-linear activation function mapping on the initial mode score vector to obtain the non-linear score values of each discharge mode category; Input the non-linear score values into the probability normalization layer, calculate the probability values of each discharge mode category and perform probability sorting to generate a discharge mode prediction probability distribution arranged in descending order of probability values; Determine the discharge type classification result corresponding to the monitoring signal sequence according to the discharge mode prediction probability distribution, and the discharge type classification result includes at least one of surface discharge, internal air gap discharge, and floating potential discharge.

6. The method for detecting partial discharge in a substation according to claim 5, wherein The training method of the discharge mode analysis model includes: Obtain a training data set, where the training data set includes multiple sample monitoring signal sequences with discharge type labels, and each sample monitoring signal sequence is correspondingly extracted with a sample time domain statistical feature set and a sample frequency band energy proportion feature; Perform mean-variance standardization processing on the sample time domain statistical feature set in the training data set to generate a sample standardized time domain feature vector, where the mean parameter and variance parameter of the mean-variance standardization processing are respectively from the historical data statistical results of the corresponding feature dimensions; Perform range scaling processing on the sample frequency band energy proportion features in the training data set to generate sample standardized frequency domain feature vectors, where the scaling coefficient of the range scaling processing is dynamically calculated according to the maximum historical value and the minimum historical value of each frequency band energy proportion feature; Construct an initial discharge mode analysis model, where the initial discharge mode analysis model includes the feature fusion layer and the classification decision layer, the number of input nodes of the feature fusion layer is set to the sum of the dimensions of the sample standardized time domain feature vector and the sample standardized frequency domain feature vector, and the number of output nodes of the classification decision layer is consistent with the preset number of discharge mode categories; Input the standardized time-domain feature vectors and standardized frequency-domain feature vectors of the samples into the initial discharge mode analysis model batch by batch, calculate the cross-entropy loss value between the predicted probability distribution of the discharge mode output by the initial discharge mode analysis model and the discharge type label, and perform backpropagation gradient update on the model parameters of the initial discharge mode analysis model according to the cross-entropy loss value; During the gradient update process, adopt a dynamic learning rate adjustment strategy to control the decay of the learning rate parameter of the optimizer, and the dynamic learning rate adjustment strategy is adaptively adjusted based on the product factor of the training iteration number and the decline rate of the cross-entropy loss value; Among them, the construction of the initial discharge mode analysis model includes: Initialize the weight matrix of the feature fusion layer so that the column dimension of the weight matrix matches the dimensions of the standardized time-domain feature vector and the standardized frequency-domain feature vector of the sample respectively, and the row dimension is set to a preset fusion feature dimension; Configure a fully connected layer and a probability normalization layer in the classification decision layer. The weight dimension of the fully connected layer is determined according to the fusion feature dimension and the number of discharge mode categories, and the input dimension of the probability normalization layer is the same as the output dimension of the fully connected layer; Perform dimension verification on the output of the feature fusion layer. When it is detected that the dimensions of the standardized time-domain feature vector and the standardized frequency-domain feature vector of the sample change, trigger the dimension adaptive adjustment operation of the weight matrix to ensure that the splicing operation of the feature fusion layer meets the vector dimension consistency constraint; Verify the dimensional consistency of each feature dimension during the weight update process of the fully connected layer in the classification decision layer. When it is detected that the input feature dimension units are inconsistent, interrupt the training process and trigger the feature re-standardization operation until the dimension units of all input features pass the uniformity verification.

7. The method for detecting partial discharge in a substation according to claim 1, characterized in that, The generation of the equipment insulation status evaluation strategy according to the discharge type classification result includes: Match the preset insulation deterioration level mapping table according to the discharge type classification result to determine the current insulation deterioration level of the target substation equipment; Obtain the historical insulation detection records of the target substation equipment and extract the deterioration trend parameters from the historical insulation detection records; Combine the current insulation deterioration level and the deterioration trend parameters to generate a predicted value of the remaining life of the equipment, and determine the maintenance priority label according to the predicted value of the remaining life of the equipment; Generate an equipment insulation status evaluation strategy including a maintenance time window and suggestions for maintenance measures based on the maintenance priority label.

8. The method for detecting partial discharge in a substation according to claim 7, characterized in that, The combination of the current insulation deterioration level and the deterioration trend parameters to generate a predicted value of the remaining life of the equipment includes: Extract the numerical value and corresponding dimension unit of the historical deterioration rate stored in the deterioration trend parameters. The historical deterioration rate is the average change amount of the insulation performance index within the historical time period; Perform a time dimension verification on the dimension unit of the historical deterioration rate to confirm that its unit is the ratio of the preset insulation performance unit to the time unit; When the dimension unit of the historical degradation rate does not include the time dimension, convert the historical degradation rate based on the total duration of the historical time period to generate a standardized degradation rate including time units; Use the absolute value of the standardized degradation rate as the insulation performance attenuation amount per unit time, and calibrate the confidence interval of the insulation performance attenuation amount based on the historical data fluctuation range; Determine the remaining performance margin based on the difference between the performance threshold corresponding to the current insulation degradation level and the current performance value; Calculate the ratio of the remaining performance margin to the insulation performance attenuation amount per unit time to obtain the predicted value of the remaining life of the device, and perform confidence calibration processing on the predicted value of the remaining life of the device.

9. The method for detecting partial discharge in a substation according to claim 1, wherein, The transmitting the device insulation state evaluation strategy to the substation monitoring system to trigger an early warning operation includes: Generating an early warning level identifier according to the maintenance priority label in the device insulation state evaluation strategy; Encapsulating the early warning level identifier and the device insulation state evaluation strategy into a standardized early warning message; Sending the standardized early warning message to the central control unit of the substation monitoring system through a preset communication protocol to trigger the central control unit to execute the corresponding audible and visual alarm and log recording operations according to the early warning level identifier.

10. A partial discharge detection system for a substation, characterized in that, It includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the substation partial discharge detection method according to any one of claims 1-9 above.

Images