A method and system for detecting partial discharge in a substation

Abstract

The present invention relates to the technical field of substation operation and maintenance, and specifically provides a substation partial discharge detection method and system, which aim to improve the accuracy of partial discharge detection and the intelligence level of equipment insulation status assessment. First, a partial discharge signal data set of target substation equipment is obtained, which includes multiple monitoring signal sequences composed of discharge pulse waveforms. Then, feature extraction is performed on the partial discharge signal data set to generate time domain and frequency domain feature sets for each monitoring signal sequence to fully reflect the intrinsic characteristics of the discharge signal. Then, a preset discharge pattern analysis model is used to perform discharge pattern analysis on the time domain and frequency domain feature sets to accurately identify the discharge type. Finally, an equipment insulation status assessment strategy is generated based on the discharge type classification result, and transmitted to the substation monitoring system in real time to trigger an early warning operation, thereby effectively ensuring the safe and stable operation of the power system.

Description

Technical Field

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

[0002] In power systems, substations are core nodes for energy transmission and distribution. The insulation condition of their equipment is directly related to the safe and stable operation of the power grid. Partial discharge (PD), a key indicator of insulation degradation, makes accurate detection and timely warning crucial for preventing equipment failures and ensuring grid security. However, existing substation PD detection technology still has numerous shortcomings, making it difficult to meet the growing demand for grid security.

[0003] Traditional partial discharge detection methods often rely on a single signal acquisition and analysis method, such as assessing the insulation condition of equipment by monitoring only limited parameters such as the amplitude and frequency of the discharge pulse. This approach not only struggles to fully capture the complex characteristics of the discharge signal but is also susceptible to factors such as electromagnetic interference and noise, significantly compromising the accuracy and reliability of the detection results. Furthermore, existing technologies lack efficient and intelligent discharge pattern recognition mechanisms, relying primarily on manual experience or simple threshold judgments. This makes it difficult to accurately classify discharge types, and consequently, fails to provide strong support for evaluating the insulation condition of equipment.

[0004] Crucially, existing partial discharge detection technologies often lack deep integration with substation monitoring systems, preventing real-time feedback and early warning of detection results. Even when partial discharge signals are detected, they often fail to promptly transmit this information to maintenance personnel or automatically trigger appropriate protective measures, missing the optimal 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:

[0006] Acquire a partial discharge signal data set of target substation equipment, wherein the partial discharge signal data set includes a plurality of monitoring signal sequences, each monitoring signal sequence consisting of at least one discharge pulse waveform;

[0007] 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 of each monitoring signal sequence;

[0008] Based on a preset discharge pattern analysis model, performing a discharge pattern analysis on the time domain feature set and the frequency domain feature set to generate a discharge type classification result of the monitoring signal sequence;

[0009] An equipment insulation status assessment strategy is generated according to the discharge type classification result, and the equipment insulation status assessment strategy is transmitted to a substation monitoring system to trigger an early warning operation.

[0010] On the other hand, the present invention also provides a substation partial discharge detection system, including a processor and a machine-readable storage medium, wherein 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.

[0011] 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 extracting time domain and frequency domain feature sets on this basis, the inherent laws and characteristic differences of the discharge signal are effectively excavated. Furthermore, with the help of a preset discharge pattern analysis model, the discharge type can be intelligently analyzed and classified, breaking through the limitations of traditional methods that rely only on a single feature or simple threshold judgment, and achieving a deep understanding and accurate judgment of the discharge phenomenon. Finally, based on the discharge type classification results, the equipment insulation status assessment strategy is dynamically generated and transmitted to the substation monitoring system in real time to trigger the early warning operation, which not only improves the accuracy and timeliness of the equipment insulation status assessment, but also enhances the safety and reliability of substation operation, thereby contributing to the intelligent operation and maintenance and fault prevention of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The figure is a schematic diagram of the execution flow of the substation partial discharge detection method provided by an embodiment of the present invention.

[0013] Figure 2 FIG. 4 is a schematic diagram of exemplary hardware and software components of a substation partial discharge detection system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 FIG1 is a flow chart of a method for detecting partial discharge in a substation provided by an embodiment of the present invention. The method for detecting partial discharge in a substation is described in detail below.

[0015] Step S110: obtaining a partial discharge signal data set of target substation equipment, wherein 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.

[0016] This example uses a large substation as an example scenario. This substation houses numerous electrical devices and serves as a key node for regional power supply. Multiple partial discharge monitoring devices are deployed within the substation for different devices, continuously collecting relevant data.

[0017] For example, a partial discharge monitoring system is installed in a high-voltage switchgear area of ​​a substation. Sensors collect electromagnetic signals from the surrounding space to acquire partial discharge signals. The discharge pulse waveform of one monitoring signal sequence exhibits a specific shape. The amplitude may exhibit a rapid rise followed by a slow decline within a certain period, accompanied by minor fluctuations. Continuous data collection at each monitoring point ultimately generates a partial discharge signal data set containing multiple monitoring signal sequences.

[0018] Step S120: 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 of each monitoring signal sequence.

[0019] Taking a monitoring signal sequence obtained from the high-voltage switchgear area as an example, detailed signal feature extraction is carried out.

[0020] Step S121: performing time-domain segmentation processing on the discharge pulse waveform in the monitoring signal sequence to obtain a plurality of pulse segmentation data blocks.

[0021] For the discharge pulse waveform of the monitoring signal sequence, time domain segmentation is performed based on the characteristics of the discharge pulse waveform. Since the discharge pulse waveform has a certain periodicity and amplitude variation pattern, it can be divided according to a fixed time interval or according to the point of significant amplitude change. For example, a time interval is set as Δt, starting from the starting time t0 of the discharge pulse waveform, and divided into a segment 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 from t0 to t0+Δt. Within the second Δt time period, pulse segmentation data block 2 is generated, and so on. If the amplitude of the discharge pulse waveform shows a significant mutation at a certain moment, such as the amplitude suddenly rises or falls sharply from a relatively stable value, the mutation moment can be used as the basis for segmentation. For example, the two pulse segmentation data blocks before and after the mutation moment are divided, so that each data block can more accurately reflect the characteristics of the discharge pulse waveform at different stages.

[0022] Step S122: calling a preset time domain feature extraction module to perform peak detection processing on each pulse segment data block to generate a time domain peak distribution feature.

[0023] For each pulse segmented data block, a pre-defined time-domain feature extraction module is invoked. Taking pulse segmented data block 1 as an example, the module iterates through the sequence of sampling points. Assume the sampling point sequence is S1, S2, S3, ..., Sn, with the time interval between adjacent sampling points being Δt_sample. The module compares the amplitude difference between adjacent sampling points with a preset differential threshold D. For example, for sampling points S1 and S2, the amplitude difference is calculated as ΔS = |S2 - S1|. If ΔS is greater than the differential threshold D, it is assumed that the pulse onset or other key feature point may exist between S1 and S2. By comparing the sampling points one by one, the pulse onset and pulse end points are determined. Assume that the sampling point corresponding to the pulse onset is Si, and the sampling point corresponding to the pulse end point is Sj. A pulse width sequence is generated based on the time interval between the pulse onset and pulse end points (i.e., (ji) * Δt_sample). Simultaneously, the amplitude extremes of all pulse waveforms in the pulse segmented data block are extracted. Assume that in this data block, the amplitude extreme value points are M1, M2, M3...Mk. Generate the pulse amplitude density feature based on the distribution density of the amplitude extreme value points. For example, divide the amplitude range into several small intervals, count the number of amplitude extreme value points in each small interval, and 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 value points in the nth small interval, and y amplitude extreme value points in the n+1th small interval, etc. In this way, the distribution of amplitude extreme value points in different amplitude intervals is described, and the pulse amplitude density feature is obtained. Finally, the pulse width sequence is combined with the pulse amplitude density feature to form the time domain peak distribution feature, which fully reflects the peak correlation characteristics of the pulse in the time domain.

[0024] Step S123: performing statistical calculation processing on the time domain peak distribution characteristics to obtain a time domain statistical feature set, wherein the time domain statistical feature set includes a peak mean, a pulse interval variance, and a rise time skewness.

[0025] For the generated time-domain peak distribution features, detailed statistical calculations are performed to obtain a time-domain statistical feature set.

[0026] First, calculate the peak mean. Assuming the amplitude extremes in the time-domain peak distribution are M1, M2, M3, ..., and Mk, the peak mean is calculated by adding the amplitudes of these extremes and then dividing by the number of extremes, k. That is, peak mean = (M1 + M2 + M3 + ... + Mk) / k.

[0027] Next, calculate the pulse interval variance. Specifically, first determine the time interval between adjacent pulses. Assume that the start and end points of a series of pulses are determined in the time-domain peak distribution characteristics, and thus the pulse interval is obtained. For example, if pulse 1 ends at t11 ​​and pulse 2 begins at t21, then the pulse interval is t21 - t11. Similarly, a series of pulse interval values, I1, I2, I3, ..., In, are obtained. The average of these pulse interval values ​​is calculated as I_avg = (I1 + I2 + I3 + ... + In) / n. Next, calculate the pulse interval variance: the sum of the squares of the differences between each pulse interval value and the average, divided by the number of pulse interval values. Pulse interval variance = [(I1 - I_avg)² + (I2 - I_avg)² + (I3 - I_avg)² + ... + (In - I_avg)²] / n.

[0028] Finally, the rise time skewness is calculated. Specifically, the rise time of each pulse waveform is first determined. Assuming the time from the starting point to the peak of the pulse waveform is the rise time, the rise time values ​​R1, R2, R3, …, and Rm are obtained. The average of these rise time values ​​is calculated as R_avg = (R1 + R2 + R3 + … + Rm) / m. The rise time skewness is then calculated using the formula: rise time skewness = [(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]. These calculations yield a set of time-domain statistical features, including the peak mean, pulse interval variance, and rise time skewness. These features characterize the time-domain distribution characteristics of the discharge pulse from different perspectives.

[0029] Step S124: performing frequency domain conversion processing on the discharge pulse waveform in the monitoring signal sequence to generate frequency domain energy distribution characteristics.

[0030] Continuing with the example of a monitoring signal sequence acquired from the high-voltage switchgear area, its discharge pulse waveform is subjected to frequency domain conversion. Using a specified frequency domain conversion method, such as the Fourier transform, the time-domain discharge pulse waveform is converted to the frequency domain. Assuming that the time-domain discharge pulse waveform can be represented as a function f(t), the Fourier transform is used to convert it into a frequency-domain function F(f), where f represents frequency. In the frequency domain, the amplitudes corresponding to different frequency components reflect the energy distribution. For example, at a certain frequency f1, the amplitude of the frequency-domain function F(f1) is large, indicating that the discharge pulse waveform near this frequency has high energy; at another frequency f2, the amplitude is small, indicating that the energy at this frequency is relatively low. This conversion generates a frequency-domain energy distribution feature that depicts the energy distribution of the discharge pulse waveform at different frequencies.

[0031] Step S125: performing frequency band division processing on the frequency domain energy distribution feature to obtain multiple frequency band energy proportion features, and associating the time domain statistical feature set with the frequency band energy proportion features for storage.

[0032] For example, assuming the spectrum ranges from f_min to f_max, it can be divided into frequency band 1 (f_min to f_min+Δf), frequency band 2 (f_min+Δf to f_min+2Δf), and finally frequency band n (f_min+(n-1)Δf to f_max) according to a preset frequency band width Δf. The width of each frequency band can be dynamically adjusted based on the spectral characteristics of historical discharge signals. For example, if analysis of the spectral characteristics of a large number of historical discharge signals reveals a high probability of discharge characteristics occurring within a certain frequency band, significantly impacting device operation, the width of that frequency band can be appropriately reduced to more accurately analyze the energy distribution within that frequency band.

[0033] Then, calculate the energy integral within each frequency band. Assuming that the frequency domain function within frequency band i is F_i(f), the energy integral within this frequency band is E_i = ∫[f_i1, f_i2]|F_i(f)|²df, where f_i1 and f_i2 are the starting and ending frequencies of frequency band i, respectively. Normalize the energy integrals of all frequency bands to obtain the energy contribution coefficient for each frequency band. For example, for frequency band 1, its energy contribution coefficient is E1 / (E1+E2+E3+…+En), where En is the sum of the energy integrals of all frequency bands. Sort the frequency bands by their energy contribution coefficients, and select the first preset number of frequency bands from the sorted results as key frequency band features. Assuming the preset number is p, select the first p frequency bands with the largest energy contribution coefficients as key frequency band features. The energy contribution coefficients of the key frequency band features are then bound to the corresponding frequency band identifiers. Finally, the time domain statistical feature set and the frequency band energy ratio feature are associated and stored, and the unique identifier of the monitoring signal sequence in the database is used as the index. The corresponding time domain statistical feature set and frequency band energy ratio feature are stored in the same record, which facilitates the subsequent rapid acquisition and association of relevant feature information when analyzing the discharge mode and equipment insulation status.

[0034] Step S130: Based on a preset discharge pattern analysis model, a discharge pattern analysis is performed on the time domain feature set and the frequency domain feature set to generate a discharge type classification result of the monitoring signal sequence.

[0035] Taking the monitoring signal sequence obtained from the high-voltage switchgear area as an example, the discharge pattern analysis is performed using the preset discharge pattern analysis model.

[0036] Step S131: Mapping the time domain statistical feature set in the time domain feature set into a first feature vector, and mapping the frequency band energy proportion feature in the frequency domain feature set into a second feature vector, and performing standardization conversion on the first feature vector and the second feature vector.

[0037] 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, and map them to the first eigenvector X = [M_avg, V, S]. For the frequency domain feature set, assume that the energy share coefficient of frequency band interval 1 is E1, the energy share coefficient of frequency band interval 2 is E2, and the energy share coefficient of frequency band interval q is Eq, and map them to the second eigenvector Y = [E1, E2, E3, ..., Eq].

[0038] The first eigenvector X and the second eigenvector Y are then normalized. Normalization aims to bring different features to the same scale, facilitating subsequent processing. For example, for each element in the first eigenvector X, a normalization method is employed: for element x_i, the normalized element x_i' = (x_i - μ) / σ, where μ is the mean of that feature dimension and σ is the standard deviation. For the peak-mean dimension, the mean μ_M and 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 rise time skewness are normalized. For the second eigenvector Y, another normalization method is employed: for 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 that feature dimension, respectively. By analyzing the energy proportion coefficients of all frequency bands in the frequency domain feature set, y_min and y_max are determined, and then each element in the second eigenvector Y is normalized to obtain the normalized first eigenvector X' and second eigenvector Y'.

[0039] Step S132: calling the feature fusion layer in the discharge pattern analysis model to perform splicing processing on the first feature vector and the second feature vector after the normalization transformation to generate a fused feature vector.

[0040] In the discharge pattern analysis model, the feature fusion layer is called. The first eigenvector X' and the second eigenvector Y', which have undergone normalization, are input into the feature fusion layer. Assume that the first eigenvector X' = [x1', x2', x3'], and the second eigenvector 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 time domain and frequency domain features are fused together to form a fused feature vector containing richer information, providing more comprehensive feature data for subsequent discharge pattern classification.

[0041] Step S133: performing vector dimension matching verification on the standardized fusion feature vector to confirm that its dimension is consistent with the number of input nodes of the classification decision layer of the discharge pattern analysis model.

[0042] Verify the dimensionality matching of the generated fused feature vector Z. Assume that the number of input nodes of the classification decision layer of the discharge pattern analysis model is N. Count the dimension of the fused feature vector Z, assuming it is n. Compare n with N. If n is equal to N, it means that the dimensions match 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 filtered 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 to the end of the fused feature vector; if n is 3 more than N, analyze the correlation of each feature and merge some features with higher correlation to reduce the dimension to N.

[0043] Step S134: performing weight matrix multiplication processing on the fused feature vector through the fully connected layer in the classification decision layer to generate an initial pattern score vector.

[0044] In the classification decision layer, the fused feature vector Z is processed by the fully connected layer. Assume that the weight matrix of the fully connected layer is W, whose dimension is related to the dimension of the fused feature vector Z and the number of discharge pattern categories. The number of rows in the weight matrix W is equal to the number of discharge pattern categories, and the number of columns is equal to the dimension of the fused feature vector Z. The fused feature vector Z is represented as a column vector and matrix multiplication is performed with the weight matrix W. The initial pattern score vector S = W * Z. Through this matrix multiplication, the various features in the fused feature vector Z are weighted and combined according to the weight matrix W, resulting in an initial pattern score vector containing different discharge pattern scores. This initial pattern score vector reflects the probability score of each discharge pattern based on the current feature combination.

[0045] Step S135: performing nonlinear activation function mapping on the initial pattern score vector to obtain nonlinear score values ​​of each discharge pattern category.

[0046] The initial pattern score vector S is mapped using a nonlinear activation function. A set nonlinear activation function is used, such as the ReLU function or the Sigmoid function. Assuming that the initial pattern score vector S = [s1, s2, s3…sm], after being processed by the nonlinear activation function f, the nonlinear score value N = [f(s1), f(s2), f(s3)…f(sm)] of each discharge pattern category is obtained. The nonlinear activation function can introduce nonlinear factors, enabling the model to learn more complex pattern relationships, thereby better classifying discharge patterns. For example, if the ReLU function is used, when si is greater than 0, f(si) = si; when si is less than or equal to 0, f(si) = 0. Through this transformation, the values ​​in the initial pattern score vector are adjusted to highlight the score differences of certain discharge patterns.

[0047] Step S136: inputting the nonlinear score value into the probability normalization layer, calculating the probability value of each discharge mode category and performing probability sorting, and generating a discharge mode prediction probability distribution arranged in descending order of probability value.

[0048] The nonlinear score value N is input into the probability normalization layer. The probability normalization layer uses a set method, such as the Softmax function, to convert the nonlinear score value into a probability value. Assuming the nonlinear score value N = [n1, n2, n3…nm], after processing the Softmax function, the probability value of each discharge pattern category P = [p1, p2, p3…pm], where pi = exp(ni) / ∑[j=1tom]exp(nj). These probability values ​​are then sorted and arranged in descending order of probability value to generate a discharge pattern prediction probability distribution arranged in descending order of probability value. This shows the order of likelihood of each discharge pattern occurring in the current monitoring signal sequence.

[0049] Step S137: determining a discharge type classification result corresponding to the monitoring signal sequence according to the discharge mode prediction probability distribution, wherein the discharge type classification result includes at least one of surface discharge, internal air gap discharge, and floating potential discharge.

[0050] Based on the generated discharge pattern prediction probability distribution, the discharge type classification result corresponding to the monitoring signal sequence is determined. For example, if the discharge pattern prediction probability distribution has a probability value of p_surface for surface discharge, a probability value of p_gap for internal air gap discharge, and a probability value of p_suspension for suspension potential discharge, and p_surface > p_gap > p_suspension, then the discharge type corresponding to the monitoring signal sequence is determined to be surface discharge. By comparing the probability values ​​of different discharge patterns, the discharge pattern with the highest probability value is selected as the classification result.

[0051] Step S140: generating an equipment insulation status assessment strategy according to the discharge type classification result, and transmitting the equipment insulation status assessment strategy to a substation monitoring system to trigger an early warning operation.

[0052] After determining the discharge type classification results of the monitoring signal sequence, the equipment insulation status assessment strategy is generated by taking surface discharge as an example.

[0053] Step S141: matching a preset insulation degradation level mapping table according to the discharge type classification result to determine the current insulation degradation level of the target substation equipment.

[0054] In this embodiment, a pre-defined insulation degradation level mapping table details the correspondence between different discharge types and insulation degradation levels. For example, for surface discharge, the mapping table specifies a specific insulation degradation level, assuming it is Level L. By matching the determined discharge type, surface discharge, with this mapping table, the current insulation degradation level of the target substation equipment is determined to be Level L. This level reflects the current degree of insulation degradation of the equipment and provides a basis for subsequent assessment of the equipment's condition.

[0055] Step S142: Obtain historical insulation detection records of the target substation equipment, and extract degradation trend parameters from the historical insulation detection records.

[0056] In this embodiment, the historical insulation test records of the target substation equipment are stored in a dedicated database, which records in detail the relevant information of each insulation test on the equipment, including the test time, test method, and various insulation performance indicators obtained.

[0057] For example, for a device previously identified as having surface discharge and insulation degradation level L, historical insulation test records related to the device are retrieved from the database. These historical insulation test records cover multiple test data from the time the device was put into operation to the present, or over a period of time.

[0058] Each test record includes a series of insulation performance indicators, such as insulation resistance and dielectric loss factor. These indicators reflect the insulation performance of the equipment from different aspects. Assume that the insulation resistance indicator is represented by R and the dielectric loss factor is represented by D. In the test records at different time points, R and D will have corresponding values. For example, in the test at time t1, the insulation resistance is R1 and the dielectric loss factor is D1; ​​in the test at time t2, the insulation resistance is R2 and the dielectric loss factor is D2, and so on.

[0059] To extract degradation trend parameters, we first analyze how insulation performance indicators change over time. For insulation resistance R, we calculate the change between two consecutive tests: ΔR = R2 - R1. Similarly, for dielectric loss factor D, we calculate ΔD = D2 - D1. Then, based on the test interval (for example, Δt = t2 - t1), we calculate the rate of change of insulation resistance: VR = ΔR / Δt, and the rate of change of dielectric loss factor: VD = ΔD / Δt. These rates of change form part of the degradation trend parameters.

[0060] Other factors related to insulation degradation can also be considered. For example, analyze the fluctuations of insulation resistance and dielectric loss factor over a period of time. Assume that within a long time period T, there are multiple test time points t1, t2, t3, ..., tn, with corresponding insulation resistance values ​​R1, R2, R3, ..., Rn, and dielectric loss factor values ​​D1, D2, D3, ..., Dn. Calculate the fluctuation amplitude of insulation resistance, for example by taking the average of the absolute values ​​of the differences between adjacent test values: fluctuation amplitude AR = (|R2-R1| + |R3-R2| + ... + |Rn-R(n-1)|) / (n-1). Similarly, calculate the fluctuation amplitude AD of dielectric loss factor. These fluctuation amplitudes are also considered components of the degradation trend parameter.

[0061] Through comprehensive analysis of various insulation performance indicators and their changes in historical insulation test records, multiple degradation trend parameters are extracted. These degradation trend parameters comprehensively reflect the changing trend of equipment insulation performance over time.

[0062] Step S143: generating a predicted value of the remaining life of the equipment in combination with the current insulation degradation level and the degradation trend parameter, and determining a maintenance priority label according to the predicted value of the remaining life of the equipment.

[0063] After obtaining the current insulation degradation level as level L and a series of degradation trend parameters (such as the insulation resistance change rate VR, the dielectric loss factor change rate VD, the insulation resistance fluctuation amplitude AR, the dielectric loss factor fluctuation amplitude AD, etc.), the remaining life prediction value of the equipment is generated.

[0064] 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. Also, obtain the currently detected insulation resistance value, R_current, and the dielectric loss factor value, D_current.

[0065] 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.

[0066] Considering the rate of change in the degradation trend parameters, for example, the insulation resistance rate of change (VR) reflects the change in insulation resistance per unit time and is therefore considered the attenuation of insulation resistance per unit time. Similarly, the dielectric loss factor rate of change (VD) is considered the attenuation of dielectric loss factor per unit time.

[0067] To generate a prediction for the remaining life of a device, it is necessary to comprehensively consider the remaining performance margin and attenuation of different insulation performance indicators. One approach is to calculate the remaining life prediction values ​​based on insulation resistance and dielectric loss factor separately and then combine them.

[0068] The remaining life prediction value based on insulation resistance is TR=MR / VR, and the remaining life prediction value based on dielectric loss factor is TD=MD / VD.

[0069] However, in practice, different insulation performance indicators may have different impacts on the remaining life of the equipment, so it is necessary to assign corresponding weights to the 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.

[0070] The comprehensive predicted value of the remaining life of the equipment is T=WR*TR+WD*TD.

[0071] After obtaining the predicted remaining life value T for the equipment, the maintenance priority label is determined according to pre-defined rules. For example, different ranges of the predicted remaining life value can correspond to different maintenance priorities. For example, when T is greater than a certain value T1, the maintenance priority label is "low"; when T is less than or equal to T1 and greater than another value T2, the maintenance priority label is "medium"; and when T is less than or equal to T2, the maintenance priority label is "high". In this way, the maintenance priority label is accurately determined based on the predicted remaining life value of the equipment, providing guidance for formulating a reasonable maintenance plan.

[0072] Step S144: generating an equipment insulation status assessment strategy including a maintenance time window and maintenance measure suggestions based on the maintenance priority tag.

[0073] After the maintenance priority labels are determined, an equipment insulation status assessment strategy is generated based on them.

[0074] If the maintenance priority label is "High," this indicates that the equipment's insulation condition is critical and requires urgent maintenance. In this case, a maintenance window is determined based on relevant industry standards, equipment operating conditions, and experience from previous similar cases. Assume that analysis indicates that for this situation, the maintenance window should be set within a short period of time from the current time, for example, the time interval [t_current, t_current + Δt1], where t_current is the current time and Δt1 is a short time interval determined based on the specific situation.

[0075] For situations where insulation degradation due to surface discharge reaches Level L and has a high maintenance priority, specific repair measures are generated. First, the equipment's exterior surface requires a thorough cleaning to remove any dust, dirt, and other impurities that may affect the equipment's insulation performance. This cleaning should be done using specialized insulation cleaners and tools, following a defined procedure.

[0076] Next, inspect the insulation coating on the equipment's exterior for damage or peeling. If the insulation coating is damaged, repair or recoat it promptly. For repair or recoating, select the appropriate insulation coating and follow the appropriate process requirements to ensure that the coating's thickness and uniformity meet standards.

[0077] Additionally, check the tightness of the equipment's connections. Due to vibration and other factors during equipment operation, connections may become loose, which can lead to uneven local electric field distribution and, in turn, partial discharge. Use specialized tightening tools and tighten the connections to the specified torque.

[0078] If the maintenance priority is "Medium," the maintenance window can be set to a relatively long time interval [t_current, t_current + Δt2], where Δt2 is greater than Δt1. Regarding recommended maintenance measures, in addition to the aforementioned cleaning, inspection of insulation coatings, and tightening of connections, more in-depth equipment testing is required, such as partial discharge location testing, to further determine the specific location and extent of the discharge, providing a basis for more precise subsequent maintenance.

[0079] If the maintenance priority is "low," the maintenance window is set to a longer time interval [t_current, t_current + Δt3], where Δt3 is greater than Δt2. Recommended maintenance measures in this case primarily include regular inspections, closely monitoring the equipment's operating status, recording changes in various insulation performance indicators, and monitoring and regulating the environmental parameters surrounding the equipment to ensure that the equipment operates under appropriate environmental conditions.

[0080] By formulating different maintenance time windows and detailed maintenance measures based on maintenance priority tags, a complete equipment insulation status assessment strategy is formed, providing comprehensive and scientific guidance for the maintenance and management of substation equipment.

[0081] Step S145: Generate an early warning level identifier according to the maintenance priority tag in the equipment insulation status assessment strategy.

[0082] Assuming a maintenance priority tag of "High," a matching warning level identifier is generated based on a pre-defined correspondence. For example, a maintenance priority tag of "High" corresponds to a warning level identifier of "Critical." If the maintenance priority tag is "Medium," the corresponding warning level identifier is "Critical." If the maintenance priority tag is "Low," the corresponding warning level identifier is "Normal." This clear correspondence transforms maintenance priority tags into warning level identifiers that are easy to identify and process.

[0083] Step S146: Encapsulate the warning level identifier and the equipment insulation status assessment strategy into a standardized warning message.

[0084] The generated warning level identifier is integrated and encapsulated with the various components of the equipment insulation status assessment strategy (including maintenance time windows and recommended repair measures). For example, according to the established message format, the warning level identifier is placed in a designated field of the message, such as the first field in the message header; the maintenance time window information is placed in a middle field of the message, and the recommended repair measures are placed in subsequent fields of the message. This standardized encapsulation method creates a standardized warning message, ensuring a uniform message format and facilitating transmission and processing across different systems.

[0085] Step S147: sending the standardized early warning message to the central control unit of the substation monitoring system through a preset communication protocol, triggering the central control unit to execute the sound and light alarm and log recording operations corresponding to the early warning level identifier.

[0086] Standardized warning messages are sent using a pre-defined communication protocol, such as one commonly used in power systems. This protocol specifies the data transmission method, rate, and verification rules. Following the protocol requirements, the standardized warning messages are sent to the central control unit of the substation monitoring system via a network or wireless communication link.

[0087] When the central control unit receives a standardized early warning message, it first analyzes the warning level indicator in the message. If the warning level is "severe," the central control unit immediately triggers a strong audible and visual alarm. For example, a high-decibel siren sounds sharply and a bright red warning light flashes simultaneously to draw personnel's attention. The central control unit also records the warning event in detail in the system log, including the warning time, warning level, corresponding equipment information, and the equipment insulation status assessment strategy. If the warning level is "severe," a relatively weaker audible and visual alarm is triggered, such as a moderately loud alarm and a flashing orange warning light, and the relevant information is recorded in the log. If the warning level is "moderate," a milder audible and visual alarm is triggered, such as a low-volume beep and a flashing yellow warning light, and a detailed warning log is also recorded. In this way, the corresponding audible and visual alarms and logging operations are accurately triggered according to the different warning levels, promptly notifying personnel of the equipment's insulation status so that appropriate measures can be taken.

[0088] Furthermore, in a possible implementation manner, the embodiment of the present invention may further include a step of training the discharge pattern analysis model, including:

[0089] Step S210: obtaining a training data set, wherein the training data set includes a plurality of sample monitoring signal sequences with discharge type labels, and each sample monitoring signal sequence corresponds to an extracted sample time domain statistical feature set and a sample frequency band energy proportion feature.

[0090] When training the discharge pattern analysis model, the first step is to obtain a training dataset. Taking this substation as an example, multiple sample monitoring signal sequences are selected from historically accumulated data and real-time monitoring data. These sample monitoring signal sequences come from partial discharge monitoring of different equipment and operating conditions within the substation.

[0091] Each sample monitoring signal sequence is clearly labeled with the discharge type, such as "surface discharge," "internal air gap discharge," or "floating potential discharge." Simultaneously, each sample monitoring signal sequence undergoes the same signal feature extraction process as before, yielding a set of sample time-domain statistical features and a sample frequency band energy ratio feature.

[0092] Taking a sample monitoring signal sequence as an example, the discharge pulse waveform is segmented in the time domain, assuming that it is segmented according to the time interval Δt_sample, to obtain multiple pulse segment data blocks. Peak detection is performed on each pulse segment data block. By traversing the sampling point sequence, the amplitude difference between adjacent sampling points is compared with the differential threshold D_sample to determine the pulse start and end points. This is then used to generate a pulse width sequence and extract the amplitude extreme points, which are combined to obtain the time domain peak distribution characteristics. Statistical calculations are performed on the time domain peak distribution characteristics to obtain a set of sample time domain statistical features, including the sample peak mean, sample pulse interval variance, and sample rise time skewness.

[0093] The discharge pulse waveform of the sample monitoring signal sequence is converted into the frequency domain to obtain the frequency domain energy distribution characteristics. According to the preset frequency band division rules, the 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 processing, the sample frequency band energy ratio characteristics are obtained. In this way, the corresponding sample time domain statistical feature set and sample frequency band energy ratio characteristics are obtained for each sample monitoring signal sequence to construct a training data set.

[0094] Step S211: performing mean-variance normalization processing on the sample time-domain statistical feature set in the training data set to generate a sample normalized time-domain feature vector, wherein the mean parameter and variance parameter of the mean-variance normalization processing are respectively derived from the historical data statistical results of the corresponding feature dimensions.

[0095] For the sample time domain statistical feature set in the training data set, mean-variance normalization is performed. Assume that the sample time domain statistical feature set includes the sample peak mean M_avg_sample, the sample pulse interval variance V_sample, and the sample rise time skewness S_sample.

[0096] The mean and variance of the corresponding feature dimension are calculated from historical data. For example, for the sample peak mean dimension, by analyzing a large amount 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 mean is μ_V_sample and the variance is σ_V_sample. For the sample rise time skewness dimension, the mean is μ_S_sample and the variance is σ_S_sample.

[0097] The sample peak mean is normalized using the formula M_avg_sample' = (M_avg_sample - μ_M_sample) / σ_M_sample. Similarly, the sample pulse interval variance is normalized using V_sample' = (V_sample - μ_V_sample) / σ_V_sample, and the sample rise time skewness is normalized using S_sample' = (S_sample - μ_S_sample) / σ_S_sample. The normalized sample peak mean, sample pulse interval variance, and sample rise time skewness are combined into the sample normalized time domain feature vector X_sample = [M_avg_sample', V_sample', S_sample']. This mean-variance normalization ensures that the time domain statistical features of different samples are scaled uniformly, facilitating subsequent model training.

[0098] Step S212: Perform range scaling processing on the sample frequency band energy proportion features in the training data set to generate a sample standardized frequency domain feature vector, wherein the scaling coefficient of the range scaling processing is dynamically calculated based on the maximum historical value and the minimum historical value of the energy proportion feature of each frequency band.

[0099] For the sample frequency band energy ratio feature in the training data set, perform range scaling. Assume that the sample frequency band energy ratio feature contains energy ratio coefficients of multiple frequency band intervals, such as E1_sample, E2_sample...En_sample.

[0100] For each frequency band's energy contribution coefficient, a scaling factor is dynamically calculated based on its historical maximum and minimum values. For example, for frequency band 1's energy contribution coefficient E1_sample, the maximum value of the energy contribution coefficient for this frequency band is E1_max_sample, and the minimum value is E1_min_sample, as found in historical data. The scaling factor k1_sample is calculated as 1 / (E1_max_sample - E1_min_sample). E1_sample is scaled for its range, resulting in E1_sample' = k1_sample * (E1_sample - E1_min_sample).

[0101] Similarly, for the energy proportion coefficient E2_sample of frequency band interval 2, its scaling factor k2_sample=1 / (E2_max_sample-E2_min_sample) is calculated. After processing, E2_sample'=k2_sample*(E2_sample-E2_min_sample) is obtained. And so on, the energy proportion coefficients of all frequency band intervals are processed.

[0102] The processed energy proportion coefficients of each frequency band are combined into the sample standardized frequency domain feature vector Y_sample=[E1_sample', E2_sample', ..., En_sample']. Through range scaling processing, the sample frequency band energy proportion feature is mapped to a set range, making it easier to process and compare during model training.

[0103] Step S213: Construct an initial discharge pattern analysis model, which 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 normalized time domain feature vector and the sample normalized frequency domain feature vector, and the number of output nodes of the classification decision layer is consistent with the preset number of discharge pattern categories.

[0104] An initial discharge pattern analysis model is constructed, which mainly consists of a feature fusion layer and a classification decision layer.

[0105] First, determine the number of input nodes of the feature fusion layer. Since the dimension of the sample normalized time domain feature vector X_sample is assumed to be m, and the dimension of the sample normalized frequency domain feature vector Y_sample is assumed to be n, the number of input nodes of the feature fusion layer is set to m+n.

[0106] Initialize the weight matrix of the feature fusion layer. The column dimensions of the weight matrix match the dimensions of the sample normalized time-domain feature vector and the sample normalized frequency-domain feature vector, respectively. That is, the first m columns of the weight matrix correspond to the dimensions of the sample normalized time-domain feature vector, and the last n columns correspond to the dimensions of the sample normalized frequency-domain feature vector. The row dimension of the weight matrix is ​​set to the preset fusion feature dimension, assuming it is p.

[0107] 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 by 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.

[0108] At the same time, the output of the feature fusion layer is dimensionally checked. When it is detected that the dimensions of the sample-normalized time-domain feature vector and the sample-normalized frequency-domain feature vector have changed, for example, due to data processing or a new data feature extraction method, the dimension of the weight matrix is ​​adaptively adjusted. The column dimensions of the weight matrix are adjusted so that the first m' columns correspond to the new sample-normalized time-domain feature vector dimensions, and the last n' columns correspond to the new sample-normalized frequency-domain feature vector dimensions, ensuring that the splicing operation of the feature fusion layer meets the vector dimension consistency constraint.

[0109] Verify the dimensional consistency of each feature dimension during the weight update process of the fully connected layer of the classification decision layer. During training, if inconsistent input feature dimensional units are detected, such as due to data entry errors or feature calculation errors, resulting in a mismatch between the dimensional units of a feature and other features, the training process is interrupted and feature renormalization is triggered. The relevant features are rechecked and processed until the dimensional units of all input features pass the consistency check to ensure the accuracy and stability of model training.

[0110] Step S214: Input the sample normalized time domain feature vector and the sample normalized frequency domain feature vector into the initial discharge pattern analysis model according to sample batches, calculate the cross entropy loss value between the discharge pattern prediction probability distribution output by the initial discharge pattern analysis model and the discharge type label, and perform back propagation gradient update on the model parameters of the initial discharge pattern analysis model according to the cross entropy loss value.

[0111] The sample-normalized time-domain feature vectors and sample-normalized frequency-domain feature vectors are sequentially input into the initial discharge pattern analysis model according to a specific sample batch partitioning method. Assuming each sample batch contains b samples, for each batch, the sample-normalized time-domain feature vectors X_sample_batch and sample-normalized frequency-domain feature vectors Y_sample_batch of these b samples are combined into matrices for easier model processing.

[0112] In this embodiment, X_sample_batch and Y_sample_batch are first concatenated through the feature fusion layer. 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 is ​​multiplied by the corresponding element of the sample-normalized time-domain feature vector and the sample-normalized frequency-domain feature vector, and the results are added to obtain each element of the fused feature vector matrix Z_batch.

[0113] Next, the fused feature vector matrix Z_batch enters the fully connected layer of the classification decision layer. The fully connected layer's weight matrix W_fc is multiplied by Z_batch to produce the initial pattern score vector matrix S_batch. The specific calculation process is to multiply each row of W_fc by each column of Z_batch and accumulate them to obtain each element of S_batch.

[0114] Next, the initial pattern score vector matrix S_batch is mapped using a nonlinear activation function, such as the Softmax function. The Softmax function calculates each element in S_batch to generate the discharge pattern prediction probability distribution matrix P_batch. Assuming that the elements in S_batch are s_ij (i represents the sample number in the sample batch, and j represents the discharge pattern category number), after the Softmax function is applied, the element p_ij in P_batch = exp(s_ij) / ∑[k=1toq]exp(s_ik), where q is the number of discharge pattern categories.

[0115] At this point, the discharge pattern prediction probability distribution matrix P_batch is compared with the discharge type label matrix L_batch corresponding to the batch of samples. The discharge type label matrix L_batch is a matrix with the same dimensions as P_batch, where the elements at the positions corresponding to the actual discharge type of the sample are 1, and the remaining positions are 0.

[0116] The difference between the predicted result and the actual label is measured by calculating the cross-entropy loss. For each sample i, the cross-entropy loss value CE_i = -∑[j=1toq]L_ij*log(p_ij). The cross-entropy loss value of the entire batch is CE_batch = ∑[i=1tob]CE_i.

[0117] Based on the cross-entropy loss value CE_batch, the model parameters of the initial discharge pattern analysis model are back-propagated and gradient updated. The back-propagation process begins at the classification decision layer, first calculating the error gradient between the probability normalization layer output (i.e., P_batch) and the label (L_batch). This error gradient is then back-propagated to the fully connected layer to calculate the gradient of the fully connected layer weight matrix W_fc. The error gradient is then back-propagated to the feature fusion layer to calculate the gradient of the feature fusion layer weight matrix.

[0118] Taking the fully connected layer weight matrix W_fc as an example, assuming that the element of W_fc is w_kl, its gradient calculation process involves the partial derivative of the error with respect to w_kl. Using 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 fully connected layer output (i.e., S_batch), then multiply it by the partial derivative of S_batch with respect to w_kl to obtain the gradient of w_kl.

[0119] After obtaining the gradients of the weight matrices for each layer, the model parameters are updated using the gradient descent algorithm. Assuming a learning rate of η, the update formula for the fully connected layer weight matrix W_fc 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, adjusting the model parameters to minimize the cross-entropy loss and gradually optimizing the model's performance.

[0120] Step S215: During the gradient update process, a dynamic learning rate adjustment strategy is used to perform attenuation control on the learning rate parameter of the optimizer, wherein the dynamic learning rate adjustment strategy is adaptively adjusted based on a multiplication factor of the number of training iterations and the rate of decrease of the cross entropy loss value.

[0121] In the gradient update process of model parameters, in order to better optimize model performance, a dynamic learning rate adjustment strategy is adopted. First, the number of training iterations is defined as t, and the cross entropy loss value calculated at each iteration is CE_t.

[0122] Calculate the rate of decrease of the cross entropy loss. Assuming that the cross entropy loss values ​​at the current iteration t and the previous iteration t-1 are CE_t and CE_(t-1) respectively, the rate of decrease of the cross entropy loss value DR = (CE_(t-1)-CE_t) / CE_(t-1).

[0123] The learning rate is dynamically adjusted based on the product of the number of training iterations t and the rate of decrease of the cross entropy loss DR. Let the product factor be F=t*DR.

[0124] Assuming the initial learning rate is η_0, the learning rate is adjusted based on the multiplication factor F. For example, a simple adjustment method can be used: the new learning rate η_t = η_0 / (1 + F). As training progresses, if the cross-entropy loss value decreases rapidly, that is, the DR is large, and the number of training iterations t increases, the multiplication factor F will increase, causing the learning rate η_t to gradually decrease. In this way, in the early stages of training, the learning rate is high, the model parameter updates are large, and the solution space can be quickly explored; in the later stages of training, the learning rate gradually decreases, and the model parameter updates are 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 model's generalization ability and prediction accuracy.

[0125] Furthermore, for example, in a possible implementation, after determining the discharge type classification result corresponding to the monitoring signal sequence according to the discharge mode prediction probability distribution, the method further includes:

[0126] Step S310: Based on the discharge type classification result and the historical execution record of the equipment insulation status assessment strategy, an incremental training sample set of the discharge mode analysis model is generated, and the incremental training sample set includes a time domain statistical feature set corresponding to the current monitoring signal sequence, a frequency band energy proportion feature and a corrected discharge type label.

[0127] After the discharge type classification result corresponding to the monitoring signal sequence is determined according to the discharge pattern prediction probability distribution, the incremental training sample set of the discharge pattern analysis model is generated in combination with the historical execution records of the equipment insulation status assessment strategy.

[0128] For example, consider a discharge classification result identified in a specific event. For example, assume that the monitored signal sequence is analyzed and identified as a specific discharge type, such as "surface discharge." Simultaneously, review the historical execution records of the equipment insulation condition assessment strategy to examine information such as insulation testing and maintenance measures implemented for this discharge type and the associated equipment, as well as subsequent feedback on the equipment's operating status.

[0129] The time-domain statistical feature set and frequency band energy ratio features are extracted from the current monitoring signal sequence. The extraction process of the time-domain statistical feature set is the same as before. The discharge pulse waveform is segmented in the time domain and segmented according to the time interval Δt to obtain multiple pulse segmentation data blocks. For each pulse segmentation data block, the sampling point sequence is traversed and the amplitude difference between adjacent sampling points is compared with the differential threshold D to determine the pulse starting and ending points. The pulse width sequence is generated and the amplitude extreme points are extracted to obtain the time-domain peak distribution characteristics. The time-domain peak distribution characteristics are statistically calculated to obtain a time-domain statistical feature set including the peak mean, pulse interval variance, and rise time skewness.

[0130] The discharge pulse waveform is transformed into the frequency domain to obtain the frequency domain energy distribution characteristics. According to the preset frequency band division rules, the spectrum range from f_min to f_max is divided into multiple frequency bands, each with a width of Δf. The energy integral value within each frequency band is calculated and normalized to obtain the frequency band energy proportion characteristics.

[0131] The discharge type label is modified based on historical execution records. For example, if historical execution records show that in certain specific cases, the situation originally thought to be "surface discharge" is found to be more accurately described as "internal air gap discharge" after further testing and analysis, the discharge type label of the current monitoring signal sequence is modified to "internal air gap discharge".

[0132] The time-domain statistical feature set, frequency band energy ratio features, and corrected discharge type labels corresponding to the current monitoring signal sequence are combined 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 includes new actual monitoring data and label information corrected based on historical experience, providing new training data for further optimizing the discharge pattern analysis model.

[0133] 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 processing on the frequency band energy proportion feature based on the historical energy distribution curve to generate a normalized frequency domain feature vector.

[0134] For the time domain statistical feature set in the incremental training sample set, a sliding window mean filter is used to remove possible noise interference. Assume that the time domain statistical feature set includes the peak mean M_avg, pulse interval variance V, and rise time skewness S.

[0135] Taking the peak mean as an example, define a sliding window size of window_size. Arrange the peak mean data in a predetermined order, for example, by monitoring time. Starting from the first data point, take window_size data points and calculate their average. Assuming these window_size data points are M1, M2, ..., M_window_size, the average of the first window is M_avg_1 = (M1 + M2 + ... + M_window_size) / window_size. Then, slide the window backward one data point, taking window_size data points M2, M3, ..., M_window_size + 1, and calculate the average value M_avg_2 = (M2 + M3 + ... + M_window_size + 1) / window_size, and so on. In this way, the peak mean data is subjected to sliding window mean filtering to obtain a denoised peak mean sequence.

[0136] The same method is applied to the pulse interval variance and rise time skewness data to obtain the denoised pulse interval variance series and rise time skewness series respectively. The denoised peak mean, pulse interval variance and rise time skewness are combined into the denoised time domain statistical feature vector X_denoised.

[0137] For the frequency band energy ratio feature, dynamic normalization processing is performed based on the historical energy distribution curve. First, the global maximum energy value E_max and minimum energy value E_min recorded in the historical training phase of each frequency band energy ratio feature are obtained.

[0138] For each frequency band energy share coefficient E_i in the current frequency band energy share feature, map it to the [0, 1] interval to generate the normalized energy share coefficient E_i'. The mapping formula is: E_i' = (E_i - E_min) / (E_max - E_min).

[0139] Perform a distribution uniformity test on the normalized energy share coefficient. Assume that an outlier threshold threshold is set. When it is detected that the normalized energy share coefficient E_i' of any frequency band interval exceeds the threshold, it is replaced by the moving average of the historical energy share coefficients of the same frequency band. For example, for frequency band interval j, if E_j'>threshold, obtain the energy share coefficients E_j1, E_j2, ..., E_jn of the frequency band interval in the past several trainings from the historical data, calculate the moving average E_j_avg=(E_j1+E_j2+...+E_jn) / n, and replace E_j' with E_j_avg. Through such processing, it is ensured that the dimensions of the energy share features of all frequency bands are unified into dimensionless proportional coefficients, and the distribution is more reasonable, generating the normalized frequency domain feature vector Y_normalized.

[0140] Step S312: Call the preset model structure adapter to perform dimensionality check on the number of input layer nodes of the discharge mode analysis model. When it is detected that the sum of the dimension of the denoised time domain statistical feature vector and the dimension of the normalized frequency domain feature vector does not match the current number of input layer nodes, the input layer node expansion operation is triggered, and new input nodes are added with a number equal to the dimensional difference and initialized to zero weight.

[0141] Call the pre-configured model structure adapter to perform dimension verification on the number of input layer nodes of the discharge pattern 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 number of input layer nodes of the current discharge pattern analysis model is N_input.

[0142] 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.

[0143] If m_denoised + n_normalized is equal to N_input, it means that the current number of input layer nodes can meet the new feature vector input requirements and no adjustment is required.

[0144] If m_denoised+n_normalized is less than N_input, no additional operation is required because the model can handle inputs with smaller dimensions.

[0145] When m_denoised + n_normalized is greater than N_input, it indicates that the number of input layer nodes is insufficient and the input layer node expansion operation needs to be triggered. Calculate the dimension difference ΔN = (m_denoised + n_normalized) - N_input.

[0146] Add a number of input nodes to the input layer equal to the dimension difference ΔN. Initially, the weights of these new nodes are set to zero. This allows the model to learn and adjust these node weights based on the actual situation as new data is input during subsequent training. This ensures that the input layer of the discharge pattern analysis model can adapt to the new feature vector dimensions and that the model can properly process the data in the incremental training sample set.

[0147] Step S313: Input the denoised time domain statistical eigenvector and the normalized frequency domain eigenvector into the updated discharge pattern analysis model in timestamp order for incremental training, calculate the loss gradient between the predicted probability distribution output by the model and the corrected discharge type label, and iteratively update the model weight matrix using the momentum optimization algorithm.

[0148] The denoised time domain statistical feature vector X_denoised and the normalized frequency domain feature vector Y_normalized are input into the updated discharge pattern analysis model in sequence according to the timestamp order for incremental training.

[0149] The model first processes the input feature vector. Similar to the previous training process, the feature fusion layer concatenates X_denoised and Y_normalized to generate the fused feature vector Z_incremental. The feature fusion layer multiplies and accumulates the elements of X_denoised and Y_normalized based on the current weight matrix to obtain the elements of Z_incremental.

[0150] The fused feature vector Z_incremental is fed into the fully connected layer of the classification decision layer. The fully connected layer's weight matrix W_fc is multiplied by Z_incremental to produce the initial pattern score vector S_incremental. This calculation involves multiplying each row of W_fc by each column of Z_incremental and adding the sums to determine the value of each element in S_incremental.

[0151] Next, a nonlinear activation function mapping is performed on the initial pattern score vector S_incremental, such as the Softmax function, to generate the predicted probability distribution P_incremental output by the model. The Softmax function is calculated for each element in S_incremental. Assuming that the elements in S_incremental are s_ij (i represents the sample number and j represents the discharge pattern category number), then the elements in P_incremental p_ij = exp(s_ij) / ∑[k=1toq]exp(s_ik), where q is the number of discharge pattern categories.

[0152] 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 corresponding to the actual corrected discharge type of the sample is 1, and the remaining elements are 0.

[0153] 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 = -∑[j=1toq]L_ij*log(p_ij). The loss gradient is obtained by taking the derivative of the cross-entropy loss value with respect to the elements of the weight matrix of each layer of the model. For example, for the element w_kl of the fully connected layer weight matrix W_fc, its gradient is calculated using the chain rule. First, calculate the partial derivative of the cross-entropy loss value with respect to the fully connected layer output (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.

[0154] The momentum optimization algorithm is used to iteratively update the model weight matrix. The momentum optimization algorithm introduces a momentum factor, γ. Assuming the current iteration number is t, the update to the weight matrix W_fc in the previous iteration is ΔW_fc_(t-1), and the gradient of the currently calculated weight matrix W_fc is ∇W_fc_t. Therefore, the update to the weight matrix W_fc this time, ΔW_fc_t, is calculated as γ*ΔW_fc_(t-1)-η*∇W_fc_t, where η is the learning rate. The weight matrix W_fc is updated to W_fc'=W_fc+ΔW_fc_t. Similarly, the weight matrix of the feature fusion layer is also updated using the momentum optimization algorithm. In this way, the momentum optimization algorithm is used to accelerate the model's convergence while avoiding falling into local optimal solutions during training. This further optimizes the performance of the discharge pattern analysis model, enabling it to better adapt to new incremental training samples.

[0155] Step S314: After each iterative update, the spliced ​​output vector of the feature fusion layer is dimensionally aligned and verified. When it is detected that the dimensional change of the denoised time domain statistical feature vector and the normalized frequency domain feature vector causes the spliced ​​vector dimension to exceed the input node capacity of the classification decision layer, a padding node is automatically inserted and a weight decay penalty term is applied to suppress overfitting.

[0156] After each iterative update of the model weight matrix using the momentum optimization algorithm, the dimension alignment of the spliced ​​output vector of the feature fusion layer is verified.

[0157] Assume that 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 splicing through the feature fusion layer is m_denoised'+n_normalized'.

[0158] The input node capacity of the classification decision layer is N_classification. Compare m_denoised'+n_normalized' with N_classification.

[0159] If m_denoised'+n_normalized' is less than or equal to N_classification, it means that the concatenated vector dimension meets the input requirements of the classification decision layer and no additional operation is required.

[0160] When m_denoised'+n_normalized' is greater than N_classification, it means that the concatenated vector dimension exceeds the input node capacity of the classification decision layer. In this case, filler nodes are automatically inserted to make the concatenated vector dimension match the input node capacity of the classification decision layer. The number of filler nodes inserted is ΔN_fill = (m_denoised'+n_normalized') - N_classification.

[0161] To prevent overfitting, a weight decay penalty is applied to these filler nodes during model training. This penalty penalizes nodes with larger weights, causing the model to learn smaller weights, thereby preventing overfitting to the training data. Specifically, the weights corresponding to the inserted filler nodes are multiplied by a decay factor β less than 1 (e.g., β = 0.99) during each iteration, causing the weights to gradually decrease.

[0162] At the same time, during subsequent training, the dimensionality changes of the denoised time-domain statistical eigenvector and the normalized frequency-domain eigenvector are continuously monitored. This is because in practical applications, the dimensions of these two eigenvectors may change again as new data is continuously added or the data processing method is adjusted.

[0163] Suppose that in a subsequent training iteration, the dimension of the denoised time-domain statistical feature vector X_denoised changes from m_denoised' to m_denoised'' due to a change in the data acquisition or processing method, and the dimension of the normalized frequency-domain feature vector Y_normalized changes from n_normalized' to n_normalized''. Recalculate the concatenated vector's dimension m_denoised'' + n_normalized'' and compare it with the input node capacity N_classification of the classification decision layer.

[0164] 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 filler nodes needs to be readjusted. If ΔN_fill' is greater than ΔN_fill, then ΔN_fill'-ΔN_fill filler nodes need to be inserted, and the weights of these newly inserted nodes are also subject to the weight decay penalty term, that is, their weights are multiplied by β at each iterative update. If ΔN_fill' is less than ΔN_fill, then ΔN_fill-ΔN_fill' filler nodes need to be deleted, and the weights of the remaining filler nodes continue to be updated according to the rules of the weight decay penalty term.

[0165] If m_denoised + n_normalized is less than N_classification, the current concatenated vector dimension is smaller than the input node capacity of the classification decision layer. While the model can function normally, further adjustments to the model structure may be necessary to fully utilize the computational resources of the classification decision layer. One possible approach is to analyze the importance of the current feature vector and select or remove relatively unimportant feature dimensions to improve the model's computational efficiency and performance. For example, by analyzing the correlation between features, if the correlation between two dimensions in the denoised time-domain statistical feature vector is extremely high, such as the correlation between the peak-mean value and another amplitude-related statistic reaching a certain threshold (assuming 0.9), these two dimensions can be merged into a new dimension. This merging can be accomplished by taking a weighted sum of the feature values ​​of these two dimensions, with the weights determined based on their contributions to discharge pattern classification in historical data. For example, if historical data analysis reveals that the peak-mean value contributes significantly to classification, its weight is set to 0.6, and the weight of another related statistic is set to 0.4. The value of the new dimension is 0.6 * peak-mean + 0.4 * the value of the other related statistic.

[0166] After this processing, recalculate the dimension of the denoised time-domain statistical feature vector, assuming it becomes m_denoised''', and then add it to the dimension of the normalized frequency-domain feature vector, n_normalized'', to obtain the new concatenated vector dimension, m_denoised'''+n_normalized''. Compare this again with the input node capacity, N_classification, of the classification decision layer to ensure a dimensional match. If there is still a mismatch, continue adjusting according to the above rules until the concatenated vector dimension matches the input node capacity of the classification decision layer. This ensures the model can run stably and efficiently, effectively suppresses overfitting, and thus improves the accuracy and generalization of the model's classification of different discharge patterns.

[0167] In addition, during the entire incremental training process, it is also necessary to pay close attention to other performance indicators of the model, such as accuracy and recall. For accuracy, it is defined as the ratio of the number of correctly classified samples to the total number of samples. Assume that after a certain incremental training, a batch of test samples are classified into discharge patterns. There are a total of N_test test samples, of which the number of correctly classified samples is N_correct. Then the accuracy is 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.

[0168] The recall rate is also an important performance indicator, which reflects the model's recognition ability for a certain type of discharge pattern. Taking "surface discharge" as an example, assume that the number of samples actually being "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 indicators 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.

[0169] For example, after generating the predicted remaining life value of the equipment by combining the current insulation deterioration level and the deterioration trend parameter, it further includes:

[0170] Step S410, generate a set of monitoring parameter adjustment coefficients according to the predicted remaining life value of the equipment and the preset dynamic adjustment rules of the monitoring parameters. The set of monitoring parameter adjustment coefficients includes a sampling frequency adjustment factor and a detection threshold scaling ratio.

[0171] , after obtaining the predicted remaining life value of the equipment, generate a set of monitoring parameter adjustment coefficients according to the preset dynamic adjustment rules of the 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 value of the equipment is within a certain range, corresponding different adjustment strategies are available. Assume the predicted remaining life value 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 rules, the sampling frequency should be significantly increased and the detection threshold should be made more stringent to capture more subtle changes in the partial discharge signal. Through the correlation analysis of the different remaining life stages of the equipment and the monitoring effects in historical data, determine that the sampling frequency adjustment factor is 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, generate a set of monitoring parameter adjustment coefficients including the sampling frequency adjustment factor and the detection threshold scaling ratio according to the predicted remaining life value of the equipment to adapt to different insulation states of the equipment and improve the effectiveness of monitoring.

[0172] Step S420: obtaining the original sampling frequency and original pulse amplitude detection threshold of the partial discharge signal acquisition module currently operating in the target substation equipment.

[0173] After generating the set of monitoring parameter adjustment coefficients, it is necessary to obtain the relevant parameters of the partial discharge signal acquisition module currently operating in the target substation equipment. This acquisition module is responsible for collecting partial discharge signals, with its original sampling frequency set to f_original and the original pulse amplitude detection threshold set to V_original. These parameters determine the frequency range of the acquired signal and the detection sensitivity to the pulse amplitude. For example, f_original determines the number of signal samples collected per unit time. A higher sampling frequency can more accurately capture the details of the discharge pulse waveform, but it also increases the amount of data and processing complexity. V_original is used to determine whether the collected signal is a valid discharge pulse. Only when the signal amplitude exceeds V_original is it identified as a possible discharge pulse for subsequent processing.

[0174] Step S430 : performing a product operation on the sampling frequency adjustment factor and the original sampling frequency to generate a dynamic sampling frequency value, and performing dimension unit verification on the dynamic sampling frequency value to confirm that the unit is a preset Hertz unit.

[0175] Multiply the sampling frequency adjustment factor α from 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. During this calculation, ensure the dimensionality of the two values ​​is consistent. Because the sampling frequency adjustment factor α is a dimensionless proportionality factor and the original sampling frequency f_original is in Hertz (Hz), the product f_dynamic should also be in Hertz. Verify the dimensionality of f_dynamic to ensure that it 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 this calculation and verification, the dynamic sampling frequency value that meets the requirements is obtained.

[0176] In step S440 , the detection threshold scaling ratio is multiplied by the original pulse amplitude detection threshold to generate a dynamic detection threshold, and a range boundary check is performed on the dynamic detection threshold to ensure that it is within a preset amplitude safety range.

[0177] 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 thresholds.

[0178] 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 segment 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 segment length does not satisfy the preset Fourier transform point constraint with the frequency-domain resolution parameter, automatically adjust the dynamic sampling frequency value to the nearest integer multiple frequency point value to force dimensional consistency.

[0179] After obtaining the required dynamic sampling frequency f_dynamic and dynamic detection threshold V_dynamic, these two parameters are used to update the configuration parameters of the partial discharge signal acquisition module. Once the update is complete, the re-acquisition of the partial discharge signal data set is triggered. During the acquisition process, the dimensional consistency between the time domain segment length and the frequency domain resolution of each discharge pulse waveform is monitored in real time. Assuming the time domain segment length is L_time, the corresponding number of sampling points is N_sample, and the frequency domain resolution parameter is Δf_freq. Frequency domain analysis, such as Fourier transform, is subject to preset constraints on the number of Fourier transform points. For example, in some cases, N_sample and 1 / Δf_freq must meet a predetermined proportional relationship (assuming k, i.e., N_sample = k / Δf_freq). If N_sample and k / Δf_freq do not meet this relationship, it indicates that the number of sampling points corresponding to the time domain segment length and the frequency domain resolution parameter do not meet the preset Fourier transform point constraints. In this case, to enforce dimensional consistency, the dynamic sampling frequency f_dynamic is automatically adjusted to the nearest integer multiple of the frequency point. For example, if the current f_dynamic makes N_sample not meet the requirements, the nearest integer multiple frequency point value f_dynamic_new that can make N_sample meet the k / Δf_freq relationship is found through calculation, and f_dynamic is updated to f_dynamic_new, thereby ensuring that the dimensions of the time domain and frequency domain related parameters match during the acquisition process, so that subsequent signal analysis can be carried out accurately.

[0180] Figure 2 The following diagram shows exemplary hardware and software components of a substation partial discharge detection system 100 that can implement the present invention according to some embodiments of the present invention. For example, the processor 120 can be used in the substation partial discharge detection system 100 to perform the functions of the present invention.

[0181] 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 this invention, for convenience, the functions described in this invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0182] 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 various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. For example, the substation partial discharge detection system 100 may also 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 may be implemented based on these program instructions. The substation partial discharge detection system 100 also includes an I / O interface 150 between the computer and other input and output devices.

[0183] For ease of explanation, 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 of the present invention may also include multiple processors. Therefore, the steps described herein as being performed by a single processor may also be performed jointly or individually by multiple processors. For example, if a processor of the substation partial discharge detection system 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by a single processor. For example, the first processor may perform step A, the second processor may perform step B, or the first and second processors may perform steps A and B jointly.

[0184] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned substation partial discharge detection method is implemented.

[0185] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for detecting partial discharge in a substation, characterized in that: The method comprises: Acquire a partial discharge signal data set of target substation equipment, wherein the partial discharge signal data set includes 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 partial discharge signal data set to obtain a time domain feature set and a frequency domain feature set of each monitoring signal sequence; Based on a preset discharge pattern analysis model, performing a discharge pattern analysis on the time domain feature set and the frequency domain feature set to generate a discharge type classification result of the monitoring signal sequence; generating an equipment insulation status assessment strategy based on the discharge type classification result, and transmitting the equipment insulation status assessment strategy to a substation monitoring system to trigger an early warning operation; The performing of discharge pattern analysis on the time domain feature set and the frequency domain feature set based on a preset discharge pattern analysis model to generate a discharge type classification result of the monitoring signal sequence includes: Mapping the time domain statistical feature set in the time domain feature set into a first feature vector, mapping the frequency band energy proportion feature in the frequency domain feature set into a second feature vector, and performing a standardization transformation on the first feature vector and the second feature vector; Calling a feature fusion layer in the discharge pattern analysis model to perform splicing processing on the first feature vector and the second feature vector after the normalization transformation to generate a fused feature vector; Performing a vector dimension matching verification on the standardized fusion feature vector to confirm that its dimension is consistent with the number of input nodes of the classification decision layer of the discharge pattern analysis model; Performing weight matrix multiplication processing on the fused feature vector through the fully connected layer in the classification decision layer to generate an initial pattern score vector; Performing nonlinear activation function mapping on the initial pattern score vector to obtain nonlinear score values ​​for each discharge pattern category; Inputting the nonlinear score value into the probability normalization layer, calculating the probability value of each discharge mode category and performing probability sorting, and generating a discharge mode prediction probability distribution arranged in descending order of probability value; Determining a discharge type classification result corresponding to the monitoring signal sequence according to the discharge mode prediction probability distribution, the discharge type classification result including at least one of surface discharge, internal air gap discharge, and floating potential discharge; After determining the discharge type classification result corresponding to the monitoring signal sequence according to the discharge mode prediction probability distribution, the method further includes: Based on the discharge type classification result and the historical execution record of the equipment insulation status assessment strategy, an incremental training sample set of the discharge mode analysis model is generated, wherein the incremental training sample set includes a time domain statistical feature set corresponding to the current monitoring signal sequence, a frequency band energy proportion feature, and a corrected discharge type label; Performing a sliding window mean filtering process on the time domain statistical feature set in the incremental training sample set to generate a denoised time domain statistical feature vector, and performing a dynamic normalization process on the frequency band energy proportion feature based on a historical energy distribution curve to generate a normalized frequency domain feature vector; Calling a preset model structure adapter to perform dimensionality check on the number of input layer nodes of the discharge pattern analysis model; when it is detected that the sum of the dimension of the denoised time domain statistical feature vector and the dimension of the normalized frequency domain feature vector does not match the number of current input layer nodes, triggering an input layer node expansion operation, adding input nodes equal to the dimensional difference and initializing them to zero weight; Inputting the denoised time-domain statistical feature vector and the normalized frequency-domain feature vector into the updated discharge pattern analysis model in timestamp order for incremental training, calculating the loss gradient between the predicted probability distribution output by the model and the corrected discharge type label, and iteratively updating the model weight matrix using a momentum optimization algorithm; After each iterative update, the dimension alignment of the spliced ​​output vector of the feature fusion layer is verified. When it is detected that the dimension change of the denoised time domain statistical feature vector and the normalized frequency domain feature vector causes the spliced ​​vector dimension to exceed the input node capacity of the classification decision layer, padding nodes are automatically inserted and a weight decay penalty term is applied to suppress overfitting.

2. The substation partial discharge detection method 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 of each monitoring signal sequence includes: Performing time-domain segmentation processing on the discharge pulse waveform in the monitoring signal sequence to obtain a plurality of pulse segmentation data blocks; Calling the preset time domain feature extraction module to perform peak detection processing on each pulse segment data block to generate time domain peak distribution features; Performing statistical calculation processing on the time-domain peak distribution characteristics to obtain a time-domain statistical feature set, wherein the time-domain statistical feature set includes a peak mean, a pulse interval variance, and a rise time skewness; Performing frequency domain conversion processing on the discharge pulse waveform in the monitoring signal sequence to generate frequency domain energy distribution characteristics; The frequency domain energy distribution feature is subjected to frequency band division processing to obtain a plurality of frequency band energy proportion features, and the time domain statistical feature set is associated with the frequency band energy proportion features and stored.

3. The method for detecting partial discharge in a substation according to claim 2, wherein: The calling of a preset time domain feature extraction module to perform peak detection processing on each pulse segment data block to generate a time domain peak distribution feature includes: Traverse the sampling point sequence in each pulse segment data block, compare the amplitude difference between adjacent sampling points with the preset difference threshold, and determine the pulse starting point position and pulse ending point position; generating a pulse width sequence according to the time interval between the pulse starting point position and the pulse ending point position, and extracting the amplitude extreme value points of all pulse waveforms in the pulse segmented data block; A pulse amplitude density feature is generated based on the distribution density of the amplitude extreme value points, and the pulse width sequence is combined with 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, wherein: The frequency domain energy distribution feature is subjected to frequency band division processing to obtain multiple frequency band energy proportion features, including: Dividing the frequency spectrum corresponding to the frequency domain energy distribution characteristics into continuous frequency bands of equal width according to a preset typical frequency band of partial discharge, and dynamically adjusting the width of each frequency band according to the frequency spectrum characteristics of historical discharge signals; Calculate the energy integral value within each frequency band interval, and normalize the energy integral values ​​of all frequency band intervals to obtain the energy proportion coefficient of each frequency band interval; The frequency band intervals are sorted according to the size of the energy proportion coefficient, a preset number of frequency band intervals in the sorting result are selected as key frequency band features, and the energy proportion coefficient of the key frequency band feature is bound to the corresponding frequency band identifier.

5. The method for detecting partial discharge in a substation according to claim 1, wherein: The training method of the discharge pattern analysis model includes: Obtaining a training data set, the training data set including a plurality of sample monitoring signal sequences with discharge type labels, each sample monitoring signal sequence corresponding to a sample time domain statistical feature set and a sample frequency band energy ratio feature extracted; Performing mean-variance normalization on the sample time-domain statistical feature set in the training data set to generate a sample-normalized time-domain feature vector, wherein the mean parameter and variance parameter of the mean-variance normalization are respectively derived from the historical data statistical results of the corresponding feature dimension; Performing range scaling processing on the sample frequency band energy proportion features in the training data set to generate sample normalized frequency domain feature vectors, wherein the scaling coefficient of the range scaling processing is dynamically calculated based on the maximum and minimum historical values ​​of the energy proportion features of each frequency band; Constructing an initial discharge pattern analysis model, the initial discharge pattern analysis model including the feature fusion layer and the classification decision layer, the number of input nodes of the feature fusion layer being set to the sum of the dimensions of the sample normalized time domain feature vector and the sample normalized frequency domain feature vector, and the number of output nodes of the classification decision layer being consistent with the number of preset discharge pattern categories; Inputting the sample normalized time domain feature vector and the sample normalized frequency domain feature vector into the initial discharge pattern analysis model according to sample batches, calculating the cross entropy loss value between the discharge pattern prediction probability distribution output by the initial discharge pattern analysis model and the discharge type label, and performing back propagation gradient update on the model parameters of the initial discharge pattern analysis model according to the cross entropy loss value; During the gradient update process, a dynamic learning rate adjustment strategy is used to attenuate the learning rate parameter of the optimizer, wherein the dynamic learning rate adjustment strategy is adaptively adjusted based on a multiplication factor of the number of training iterations and the rate of decrease of the cross entropy loss value; The step of constructing an initial discharge mode analysis model includes: Initializing the weight matrix of the feature fusion layer so that the column dimensions of the weight matrix match the dimensions of the sample normalized time domain feature vector and the sample normalized frequency domain feature vector, respectively, and the row dimensions are set to the preset fusion feature dimensions; A fully connected layer and a probability normalization layer are configured in the classification decision layer, wherein the weight dimension of the fully connected layer is determined according to the fusion feature dimension and the number of discharge pattern categories, and the input dimension of the probability normalization layer is consistent with the output dimension of the fully connected layer; Performing dimensionality check on the output of the feature fusion layer. When a dimensional change of the sample normalized time domain feature vector and the sample normalized frequency domain feature vector is detected, triggering a dimensional adaptive adjustment operation of the weight matrix to ensure that the splicing operation of the feature fusion layer satisfies the vector dimensional consistency constraint. Verify the dimensional consistency of each feature dimension during the weight update process of the fully connected layer of the classification decision layer. When inconsistent input feature dimensional units are detected, interrupt the training process and trigger feature re-normalization operation until the dimensional units of all input features pass the uniformity check.

6. The method for detecting partial discharge in a substation according to claim 1, wherein: Generating a device insulation status assessment strategy according to the discharge type classification result includes: Matching a preset insulation degradation level mapping table according to the discharge type classification result to determine the current insulation degradation level of the target substation equipment; Obtaining historical insulation test records of the target substation equipment, and extracting degradation trend parameters from the historical insulation test records; generating a predicted value of the remaining life of the equipment in combination with the current insulation degradation level and the degradation trend parameter, and determining a maintenance priority label according to the predicted value of the remaining life of the equipment; An equipment insulation status assessment strategy including a maintenance time window and maintenance measure recommendations is generated based on the maintenance priority label.

7. The method for detecting partial discharge in a substation according to claim 6, wherein: The generating of a predicted value of the remaining life of the equipment by combining the current insulation degradation level with the degradation trend parameter includes: Extracting the numerical value and corresponding dimensional unit of the historical degradation rate stored in the degradation trend parameter, wherein the historical degradation rate is the average change of the insulation performance index within a historical time period; Verify the dimensional unit of the historical degradation rate in terms of time dimension to confirm that the unit is a ratio of a preset insulation performance unit to a time unit; When the dimension unit of the historical degradation rate does not include a time dimension, converting the historical degradation rate based on the total duration of the historical time period to generate a normalized degradation rate including a time unit; The absolute value of the normalized degradation rate is used as the insulation performance degradation amount per unit time, and the insulation performance degradation amount is calibrated with a confidence interval based on the fluctuation range of historical data; determining a remaining performance margin based on a difference between a performance threshold corresponding to the current insulation degradation level and a current performance value; The ratio of the remaining performance margin to the insulation performance attenuation per unit time is calculated to obtain the predicted value of the remaining life of the equipment, and a confidence calibration process is performed on the predicted value of the remaining life of the equipment.

8. The substation partial discharge detection method according to claim 1, characterized in that: The transmitting the equipment insulation status assessment strategy to the substation monitoring system to trigger an early warning operation includes: Generate an early warning level identifier according to the maintenance priority tag in the equipment insulation status assessment strategy; Encapsulating the warning level identifier and the equipment insulation status assessment strategy into a standardized warning message; The standardized early warning message is sent to the central control unit of the substation monitoring system through a preset communication protocol, triggering the central control unit to execute the sound and light alarm and log recording operations corresponding to the early warning level identifier.

9. A partial discharge detection system for a substation, characterized in that: The invention comprises a processor and a memory, wherein 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 to 8.

Images