A method and device for diagnosing GIS insulation defects by using variable-frequency narrow-band phase correlation

Abstract

The application discloses a GIS insulation defect variational frequency narrowband phase correlation diagnosis method and device, relates to the GIS insulation detection technical field, and obtains the narrowband signal stream of the target frequency band to be detected equipment, is divided into a plurality of continuous narrowband sub-frequency bands and is subjected to noise reduction processing to obtain an effective narrowband sub-frequency band; a frequency-phase feature matrix is constructed according to the distribution of the effective narrowband sub-frequency band on the phase of the power system power frequency cycle to identify an effective discharge signal; the frequency interval parameters of each effective narrowband sub-frequency band, the phase distribution characteristic vector of the effective discharge signal and the amplitude change rate are fused, a GIS insulation defect diagnosis model established based on an improved random forest algorithm is used to realize insulation defect type diagnosis and positioning of the equipment to be detected. The single feature is easy to overlap and cannot distinguish the same frequency and coherent interference, the defect feature recognition precision is low under the problem of multiple types of gas chamber structures, and the accuracy and applicability of the insulation defect detection are improved.

Description

Technical Field

[0001] This application relates to the field of GIS insulation testing technology, specifically a variable frequency narrowband phase correlation diagnostic method and device for GIS insulation defects. Background Technology

[0002] The insulation condition of gas-insulated metal-enclosed switchgear (GIS) directly affects the safe and stable operation of power systems. Partial discharge, as a key early sign of insulation degradation, requires accurate detection as a core means of fault early warning and condition-based maintenance. However, current GIS partial discharge detection technology still faces significant challenges in complex electromagnetic environments. Firstly, GIS consists of various functional chambers (such as busbar chambers, circuit breaker chambers, disconnector chambers, and instrument transformer chambers), each with significantly different structures, resulting in varying internal electromagnetic wave propagation characteristics (such as resonant frequency and attenuation characteristics). For example, the busbar chamber has a long waveguide structure, making it more suitable for low-frequency signal monitoring; while the circuit breaker chamber has a complex structure and suffers from severe attenuation of high-frequency components. Existing detection systems generally use fixed-frequency sensing and acquisition modules, lacking adaptability to different chamber structures. This makes it difficult to achieve universal deployment for various types of GIS equipment, often requiring customized detection solutions for individual components, leading to high costs, low efficiency, and limited applicability.

[0003] On the other hand, in actual operating environments, partial discharge signals undergo multiple reflections and mode conversions within the GIS cavity, resulting in severe waveform distortion. Simultaneously, external electromagnetic interference (such as switching operations and communication signals) often shares the same frequency or even the same power frequency phase as the actual discharge signal, creating "co-frequency and co-phase" interference. Existing broadband detection methods suffer from low signal-to-noise ratios and lack an effective correlation mechanism between frequency characteristics and power frequency phase information. This makes it impossible to distinguish between signals originating from different sources but appearing in the same phase interval (e.g., a 50kHz actual discharge and a 500kHz interference signal may appear in phase), easily leading to misjudgments.

[0004] In summary, under complex electromagnetic environments, existing GIS partial discharge detection technologies are unable to effectively suppress co-frequency and co-phase interference and lack the ability to detect overlapping features of multiple types of air chamber structures. This results in low accuracy and insufficient reliability in insulation defect identification. There is an urgent need for an intelligent diagnostic method that combines adaptive frequency band matching and multi-dimensional feature fusion discrimination. Summary of the Invention

[0005] This application addresses the problem that existing GIS partial discharge detection technologies struggle to effectively suppress in-phase and co-frequency interference in complex electromagnetic environments and lack adaptive overlapping feature detection for various types of air chamber structures, resulting in low accuracy in insulation defect detection. It proposes a variable-frequency narrowband phase correlation diagnostic method and device for GIS insulation defects. By dividing the narrowband signal stream into multiple sub-bands in the frequency domain to avoid full-band noise interference, a feature matrix is ​​constructed based on phase and discharge correlation to distinguish and identify effective discharge signals from a temporal dimension. This integrates multi-dimensional features for defect detection, overcoming the problem that single features easily overlap and cannot distinguish in-phase and co-frequency interference, leading to low accuracy in defect feature identification under various types of air chamber structures. This improves the accuracy and applicability of insulation defect detection.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows:

[0007] In a first aspect, embodiments of this application provide a variable-frequency narrowband phase correlation diagnosis method for GIS insulation defects, the method comprising:

[0008] Based on the target frequency band, acquire the characteristic signals of the device under test and obtain the narrowband signal stream;

[0009] The narrowband signal stream is divided into multiple consecutive narrowband sub-bands and noise reduction is performed on them to obtain effective narrowband sub-bands;

[0010] A frequency-phase feature matrix is ​​constructed based on the distribution of effective narrowband sub-bands on the power frequency cycle phase of the power system. Effective discharge signals are then identified based on the frequency-phase feature matrix combined with an interference feature library.

[0011] By integrating the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate, a GIS insulation defect diagnosis model based on an improved random forest algorithm is used to diagnose and locate the insulation defect type of the equipment under test.

[0012] In this scheme, by limiting signal acquisition to the target frequency band, broadband noise pollution that is avoided during full-band acquisition can be prevented, providing a high signal-to-noise ratio input for subsequent refined processing. By dividing the frequency band into sub-bands and reducing noise, the spectral components that may be excited by different defects are further separated within the target frequency band, preventing weak discharge signals from being masked by interference, improving the detection signal-to-noise ratio and feature resolution. This allows for the construction of an accurate frequency-phase feature matrix using the physical coupling relationship between discharge and power frequency voltage, enabling the differentiation of effective discharge signals with fixed phase patterns and irregular interference in the effective narrowband sub-bands from a time-series perspective, thereby ensuring feature accuracy and reducing the false positive rate of pulse signals. Finally, by fusing three-dimensional features and using an improved random forest algorithm for defect diagnosis, the limitations of single feature detection can be overcome, achieving accurate diagnosis of defect type and location, and improving detection accuracy.

[0013] Optionally, the step of acquiring the characteristic signal of the device under test based on the target frequency band and acquiring the narrowband signal stream includes:

[0014] Construct a multi-dimensional feature mapping database containing one-to-one correspondences between GIS component locations, defect types, sensitive frequency bands, and typical phases;

[0015] The signal spectrum distribution of the device under test is identified based on a multidimensional feature mapping database to determine the target frequency band range.

[0016] The signal acquisition frequency division parameters are dynamically adjusted based on the target frequency band range. The signal acquisition frequency division parameters include at least the sampling rate and the filtering bandwidth.

[0017] Based on the adjusted frequency division parameters, continuous analog signal acquisition is performed on the target frequency band interval, and the signal is divided into discrete narrowband signal streams by a digital filter.

[0018] Optionally, the step of dividing the narrowband signal stream into multiple consecutive narrowband sub-bands and performing noise reduction processing on them to obtain effective narrowband sub-bands includes:

[0019] The narrowband signal stream is divided into multiple consecutive narrowband sub-bands in the frequency domain, and each narrowband sub-band is numbered.

[0020] A noise reduction threshold is determined based on the statistical characteristics of the narrowband signal in each narrowband sub-band. Wavelet threshold filtering is then applied to the narrowband sub-band signal based on the noise reduction threshold to determine candidate narrowband sub-bands.

[0021] Based on pulse signal constraints, the candidate narrowband sub-frequency bands are checked for leaks. If a pulse signal exists in the candidate narrowband sub-frequency band and the pulse signal attributes meet the pulse standard, then the corresponding candidate narrowband sub-frequency band is taken as a valid narrowband sub-frequency band.

[0022] Optionally, before dividing the narrowband sub-bands, the narrowband signal is divided and judged based on the signal density and signal time-frequency continuity of the narrowband signal stream. When the signal density is greater than a preset threshold and the signal is continuous, the narrowband signal stream is divided into narrowband sub-bands.

[0023] Optionally, the step of constructing a frequency-phase feature matrix based on the distribution of effective narrowband sub-bands on the power system's power frequency cycle phase, and identifying effective discharge signals based on the frequency-phase feature matrix combined with an interference feature library, includes:

[0024] The effective narrowband sub-bands are marked with effective sub-band identifiers, and the phase intervals are divided and numbered according to the phase angle of the power system power frequency cycle;

[0025] Narrowband signal data is extracted based on the effective narrowband sub-frequency bands, the power frequency period corresponding to each narrowband signal is identified, and its corresponding phase interval is numbered and marked.

[0026] Statistically analyze the pulse amplitude of all narrowband signals within the target time period under each effective sub-band-phase combination, effectively verify the narrowband signals based on the pulse amplitude, and obtain the average pulse amplitude of the effective narrowband signals;

[0027] The average pulse amplitude of the effective narrowband signal is used as a matrix element, and a frequency-phase feature matrix is ​​constructed with the effective narrowband sub-band number as the horizontal axis and the phase interval number as the vertical axis.

[0028] The effective discharge signal is identified based on the frequency-phase feature matrix combined with the interference feature library.

[0029] Optionally, the step of identifying valid discharge signals based on the frequency-phase feature matrix combined with the interference feature library includes:

[0030] The pulse signal amplitude ratio of the effective sub-band is obtained based on the real-time frequency-phase feature matrix, and its overlap with the standard phase amplitude ratio is calculated.

[0031] The overlap is compared with a set overlap threshold. If the overlap is less than a first overlap threshold, it is determined to be an interference signal; if the overlap is greater than or equal to the first overlap threshold and less than a second overlap threshold, it is determined to be a suspected discharge signal; if the overlap is greater than or equal to the second overlap threshold, it is determined to be a valid discharge signal; the first overlap threshold is less than the second overlap threshold.

[0032] The effective discharge signal or the suspected discharge signal is compared with an interference feature library containing frequency-phase-duration information to perform sudden interference verification on the effective discharge signal and obtain the final effective discharge signal.

[0033] Optionally, the process of fusing the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate, and using a GIS insulation defect diagnosis model based on an improved random forest algorithm to diagnose and locate the insulation defect type of the device under test, includes:

[0034] Based on the phase interval number of each effective discharge signal, the distribution ratio of effective discharge signals in each phase interval is statistically analyzed to obtain the phase distribution feature vector;

[0035] The amplitude change rate per unit time is calculated based on the peak value of the effective discharge signal continuously monitored within the effective narrowband sub-band.

[0036] Extract the frequency range parameters of each effective narrowband sub-band, and integrate them with the phase distribution feature vector and the amplitude change rate to form a structured three-dimensional feature parameter set;

[0037] A standard defect feature library was obtained, and a random forest algorithm with feature importance weighted splitting and differentiating weights for different defect types as optimization objectives was used to establish a GIS insulation defect diagnosis model.

[0038] Based on the GIS insulation defect diagnosis model, the three-dimensional feature parameter set of the equipment to be tested is analyzed, and the insulation defect type and corresponding component location of the equipment to be tested are output.

[0039] Optionally, the step of obtaining the standard defect feature library, and establishing a GIS insulation defect diagnosis model using a random forest algorithm with feature importance-weighted splitting and differentiated weights for different defect types as optimization objectives, includes:

[0040] Based on various GIS defect samples calibrated in the laboratory, the frequency range, phase distribution feature vector, and amplitude change rate of each type of defect were extracted to establish a standard defect feature library;

[0041] Calculate the classification contribution of the three-dimensional features in the standard defect feature library to the defect type, and determine the split weights of the frequency range, phase distribution feature vector and amplitude change rate based on the classification contribution.

[0042] The sample splitting weight for each type of defect sample is determined by linear weighting based on the number and severity of different defect samples;

[0043] Weighted embedding is performed on individual GIS defect samples based on the sample splitting weights, and the three-dimensional features of individual GIS defect samples are weighted based on the splitting weights. The weighted dataset is then divided into training and testing sets.

[0044] The single decision tree-based learner is trained based on the training set, and the basic GIS insulation defect diagnosis model generated by the integrated decision tree-based learner is obtained. The basic GIS insulation defect diagnosis model is then validated and optimized based on the test set to obtain the target GIS insulation defect diagnosis model.

[0045] Optionally, the method further includes effectively verifying the suspected discharge signal, specifically including:

[0046] Based on the suspected discharge signal, its three-dimensional features, including frequency range parameters, phase distribution feature vector, and amplitude change rate, are obtained;

[0047] The three-dimensional features of the suspected discharge signal are used to detect the defect type through the GIS insulation defect diagnosis model. Based on the defect type output probability, it is determined whether the suspected discharge signal is an interference signal or a real discharge signal.

[0048] If the output probability of the defect type is still determined to be a suspected discharge signal, then dynamic verification is performed in conjunction with the amplitude change rate to obtain a valid verification result of the suspected discharge signal.

[0049] Secondly, embodiments of this application provide a variable-frequency narrowband phase correlation diagnostic device for GIS insulation defects, comprising:

[0050] The narrowband signal acquisition module is used to acquire the characteristic signals of the device under test based on the target frequency band and to acquire the narrowband signal stream.

[0051] The narrowband signal segmentation module is used to divide the narrowband signal stream into multiple continuous narrowband sub-bands and perform noise reduction processing on them to obtain effective narrowband sub-bands.

[0052] The effective signal identification module is used to construct a frequency-phase feature matrix based on the distribution of effective narrowband sub-bands on the power frequency cycle phase of the power system, and to identify effective discharge signals based on the frequency-phase feature matrix combined with the interference feature library.

[0053] The insulation defect detection module is used to integrate the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate. It uses a GIS insulation defect diagnosis model based on an improved random forest algorithm to diagnose and locate the insulation defect type of the equipment under test.

[0054] The beneficial effects of this application are:

[0055] 1. By constructing a multi-dimensional mapping database, it is not necessary to customize hardware for each type of equipment, thus avoiding the problem of needing a dedicated detection module due to different resonant frequencies caused by structural differences in different gas chambers (such as busbar chambers and circuit breaker chambers). This enables automatic identification of the target frequency band and dynamic adjustment of the sampling rate and filtering bandwidth accordingly, achieving universal adaptation for different types of GIS equipment. At the same time, it ensures that the collected signal is focused on the frequency band most likely to contain effective discharge information, improving the signal-to-noise ratio of the original signal, reducing the amount of data and computational complexity of subsequent processing, and avoiding invalid data processing.

[0056] 2. By constructing a structured frequency-phase feature matrix, explicit modeling of the phase regularity of discharge signals is achieved, making the effective discharge (with a fixed phase distribution) and interference (random phase) show significant differences in matrix shape (such as bright islands vs. global noise). The matrix recognition results are further verified by the interference feature library, avoiding the identification error of the effective discharge signal in the two-dimensional feature matrix due to sudden interference, and providing an objective and quantifiable criterion for accurate differentiation between the two in the future.

[0057] 3. By integrating multidimensional independent evidence and adopting a targeted optimized classification model, it is possible to effectively distinguish different defect types with some overlapping features (such as air gap discharge and surface discharge), overcoming the classification ambiguity problem caused by the single feature in traditional methods, so as to achieve accurate type identification and spatial positioning of insulation defects. Attached Figure Description

[0058] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0059] Figure 1 This is a flowchart of a variable frequency narrowband phase correlation diagnosis method for GIS insulation defects provided in an embodiment of this application.

[0060] Figure 2 A flowchart illustrating an effective narrowband sub-band partitioning method provided in this application embodiment.

[0061] Figure 3 A flowchart of an effective discharge signal identification method provided in this application embodiment.

[0062] Figure 4 This is a flowchart of a GIS insulation defect feature association diagnosis method provided in an embodiment of this application.

[0063] Figure 5 This is a structural diagram of a variable frequency narrowband phase correlation diagnostic device module for GIS insulation defects provided in an embodiment of this application. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely one preferred embodiment of this application and are only used to explain this application. They do not limit the scope of protection of this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] Example 1: As Figure 1 As shown, a variable frequency narrowband phase correlation diagnostic method for GIS insulation defects includes steps S1-S4, wherein:

[0066] S1. Obtain the characteristic signals of the device under test based on the target frequency band, and obtain the narrowband signal stream.

[0067] In an optional embodiment, step S1 includes:

[0068] Construct a multi-dimensional feature mapping database containing one-to-one correspondences between GIS component locations, defect types, sensitive frequency bands, and typical phases;

[0069] The signal spectrum distribution of the device under test is identified based on a multidimensional feature mapping database to determine the target frequency band range.

[0070] The signal acquisition frequency division parameters are dynamically adjusted based on the target frequency band range. The signal acquisition frequency division parameters include at least the sampling rate and the filtering bandwidth.

[0071] Based on the adjusted frequency division parameters, continuous analog signal acquisition is performed on the target frequency band interval, and the signal is divided into discrete narrowband signal streams by a digital filter.

[0072] In some embodiments, a broadband signal acquisition sensor (such as an ultra-high frequency sensor or an ultrasonic sensor) is installed at a designated location on the GIS equipment to be tested, such as at the housing corresponding to a basin insulator.

[0073] In some embodiments, the multidimensional feature mapping database includes typical GIS components such as busbar compartments, circuit breaker compartments, disconnector compartments, and instrument transformer compartments, as well as characteristic frequency ranges corresponding to common defect types such as internal air gaps, free metal particles, and suspended tips. Before inspection, a wideband pre-scan (scanning range covering 1kHz-1GHz) is used to obtain the signal spectrum distribution of the device under inspection. Based on the mapping database, the target frequency range is automatically identified, and the frequency division parameters (including sampling rate 10kSps-100MSps, filter bandwidth 1kHz-50kHz) are dynamically adjusted. High-precision signal acquisition is performed only on the target frequency band. A continuous analog signal is divided into discrete digital narrowband signal streams through a digital filter bank, achieving adaptive adaptation for different GIS component types and different defect types. Some feature mapping tables in the multidimensional feature mapping database are shown in Table 1.

[0074] Table 1 Feature Mapping Table of Multidimensional Feature Mapping Database

[0075] ;

[0076] The multidimensional feature mapping database supports multi-condition queries based on GIS component location + defect type, sensitive frequency band, and typical phase characteristics. For example, inputting "pot-type insulator + internal air gap" can quickly return its sensitive frequency band range of 300kHz~500kHz. It also reserves on-site data upload interface and manual input interface. When a new device is detected on-site, signal data can be uploaded to the cloud and the database will be automatically updated after laboratory verification.

[0077] In this embodiment, due to differences in geometry and materials, the electromagnetic resonant frequencies excited by internal defects in different GIS air chamber structures (such as busbars and circuit breakers) are different. The multidimensional feature mapping database pre-calibrates the sensitive frequency bands of defects corresponding to various components. Therefore, it can automatically match the most likely discharge frequency band of the device under test. Then, by dynamically adjusting the sampling rate and filtering bandwidth, the acquisition device can only receive signals within the target frequency band, which greatly reduces the coupling of out-of-band broadband noise and interference from adjacent frequency bands.

[0078] S2. Divide the narrowband signal stream into multiple continuous narrowband sub-bands and perform noise reduction processing on them to obtain effective narrowband sub-bands.

[0079] In an optional embodiment, combined with Figure 2 As shown, step S2 includes:

[0080] S21. Divide the narrowband signal stream into multiple consecutive narrowband sub-bands in the frequency domain, and number each narrowband sub-band.

[0081] S22. Determine the noise reduction threshold based on the statistical characteristics of the narrowband signal in each narrowband sub-band, and perform wavelet threshold filtering on the narrowband sub-band signal according to the noise reduction threshold to determine the candidate narrowband sub-band.

[0082] S23. Based on the pulse signal constraint, perform leakage signal verification on the candidate narrowband sub-frequency band. If there is a pulse signal in the candidate narrowband sub-frequency band and the pulse signal attribute meets the pulse standard, then the corresponding candidate narrowband sub-frequency band is taken as a valid narrowband sub-frequency band.

[0083] Specifically, within the identified target frequency range, multiple continuous narrow-band sub-bands are divided according to a preset bandwidth (10kHz or 20kHz), and wavelet threshold filtering algorithm is used to perform noise reduction processing on each sub-band.

[0084] In some embodiments, the partitioning window of the narrowband sub-band is determined based on the signal spectral density of the narrowband signal stream, wherein the signal spectral density characterizes regions with high signal concentration. For example, regions with high density are partitioned at 5 kHz (to accurately preserve features), and regions with low density are partitioned at 30 kHz (to reduce the number of sub-bands). For instance, a 5 kHz concentrated signal in the busbar cylinder is partitioned into one sub-band, and a dispersed signal is partitioned into two sub-bands.

[0085] In this embodiment, adaptive subband division adapts to the different effective frequency concentration area widths of different discharge signals, avoiding the problems of poor real-time performance caused by fixed bandwidth, dilution of signal characteristics at wide bandwidth, and loss of effective information after noise reduction.

[0086] As a specific implementation method, the target signal area is first divided into 8 small signal segments (narrow sub-bands): the large range of 50kHz-200kHz, which may have leaky signals, is divided into 8 consecutive small sub-bands with a width of 20kHz, equivalent to: 50-70kHz; 70-90kHz; 90-110kHz; 110-130kHz; 130-150kHz; 150-170kHz; 170-190kHz; 190-210kHz (the last segment slightly exceeds 200kHz to ensure complete coverage). This is to avoid the noise in a small segment masking the leaky signals in other segments during a time check. After being divided into small segments, the signal in each segment is purer, which is convenient for subsequent processing. The divided narrow sub-bands are numbered, with one number for each frequency range.

[0087] Furthermore, wavelet threshold filtering is applied to each of the eight sub-bands. For example, for the 50-70kHz sub-band, the magnitude (amplitude) of all signals within this sub-band is first statistically analyzed, and a noise reduction threshold is set (e.g., three times the average magnitude of the signals within the sub-band). Signals below this threshold are treated as noise and deleted; signals above this threshold are retained as candidate narrowband sub-bands (which are likely to be the pulses of leaky signals). Because the type and magnitude of noise in each sub-band are different, individual noise reduction can accurately remove the noise of each sub-band, preserving the characteristics of the leaky signal (e.g., the pulse shape of the leaky signal) better than noise reduction of the entire band.

[0088] Furthermore, after noise reduction, the characteristics of whether there are leaky signals in each of the eight sub-bands are determined based on pulse signal constraints. The pulse signal constraints include: whether there are pulse signals in the sub-band (leaky signals are sudden pulses, while noise is a continuous and stable signal), and whether the number and magnitude of the pulse signals meet the standards (e.g., ≥5 pulses within 1 minute, and the pulse peak value ≥ a preset threshold). For example, if only 3 of the 8 sub-bands meet the requirements (e.g., 50-70kHz, 90-110kHz, 150-170kHz), these 3 are valid sub-bands, and the remaining 5 are all noise and are directly discarded.

[0089] In this embodiment, real partial discharge pulses have statistical characteristics such as short duration, suddenness, and high-frequency attenuation, while residual noise or interference often exhibits a stable or non-pulse pattern. Therefore, by setting a wavelet threshold based on the sub-band's own statistical characteristics (such as variance and kurtosis), weak discharge components can be preserved and background noise suppressed. Furthermore, by setting time-domain constraints on the pulse signal (such as pulse width <1μs and rise time <100ns), non-pulse continuous interference (such as oscillations and harmonics) that remains after noise reduction can be eliminated, ensuring that the output effective sub-bands do indeed contain physically significant discharge events.

[0090] In this embodiment, by further dividing the target frequency band into sub-bands, it is possible to obtain the spectral shift or energy concentration differences that different defects may have in the same target frequency band. This allows for noise reduction based on the statistical characteristics of the sub-bands, enabling precise location of the effective discharge response frequency band within the target frequency band. This significantly improves the signal-to-noise ratio within a single sub-band, avoids phase distribution distortion caused by noise residue or non-pulse interference, and ensures the reliability of subsequent phase correlation analysis.

[0091] In an optional embodiment, before dividing the narrowband sub-bands, the narrowband signal is divided and judged based on the signal density and signal time-frequency continuity of the narrowband signal stream. When the signal density is greater than a preset threshold and the signal is continuous, the narrowband signal stream is divided into narrowband sub-bands.

[0092] In this embodiment, real partial discharges typically manifest as dense, continuous pulse clusters in the time-frequency domain, while random noise or sporadic interference exhibits low-density, discrete distribution characteristics. Therefore, by setting dual thresholds for signal density and time-frequency continuity, signal periods with potential discharge activity can be effectively identified, avoiding being misled by sudden, discontinuous noise signals, improving the rationality of sub-band division, and indirectly enhancing the effectiveness of subsequent noise reduction and screening. Based on this, sub-band division and noise reduction processing are performed, avoiding meaningless calculations on pure noise segments, reducing invalid data processing overhead, and preventing random fluctuations from being misjudged as valid frequency bands, ensuring that subsequent defect diagnosis focuses on high-value data segments.

[0093] S3. Construct a frequency-phase feature matrix based on the distribution of effective narrowband sub-bands on the power frequency cycle phase of the power system, and identify effective discharge signals based on the frequency-phase feature matrix combined with the interference feature library.

[0094] In an optional embodiment, combined with Figure 3 As shown, step S3 includes:

[0095] S31. Mark the effective narrowband sub-bands with effective sub-band identifiers, and divide the phase intervals and number the intervals according to the phase angle of the power system frequency cycle.

[0096] S32. Extract narrowband signal data based on the effective narrowband sub-frequency band, identify the power frequency period corresponding to each narrowband signal and assign it a corresponding phase interval number;

[0097] S33. Calculate the pulse amplitude of all narrowband signals within the target time period under each effective sub-band-phase combination, perform effective verification of the narrowband signals based on the pulse amplitude, and obtain the average pulse amplitude of the effective narrowband signals;

[0098] S34. The average pulse amplitude of the effective narrowband signal is used as a matrix element, and a frequency-phase feature matrix is ​​constructed with the effective narrowband sub-band number as the horizontal axis and the phase interval number as the vertical axis.

[0099] S35. Identify valid discharge signals based on the frequency-phase feature matrix combined with the interference feature library.

[0100] In some embodiments, the power system frequency period is 50Hz (period 20ms), and the phase angle range is 0°-360°, divided into 60° intervals, as shown in Table 2.

[0101] Table 2 Phase Interval

[0102] ;

[0103] The power frequency phase interval is divided into 60° intervals to balance statistical accuracy and computational efficiency, thereby achieving standardized statistics of the phase distribution of the discharge signal. The mean amplitude is used as the matrix element to eliminate invalid data, which can intuitively distinguish effective signals (high-brightness islands in a specific frequency band-phase), white noise (global noise floor), and narrowband interference (full-phase through horizontal stripes).

[0104] Furthermore, for each sub-band signal, the signal is divided into phase intervals, each signal pulse is labeled with a phase, and the amplitude of all signal pulses within 1 minute under each sub-band-phase combination is statistically analyzed. The average value is calculated to remove invalid data with amplitude < background signal mean + 2 times standard deviation.

[0105] In this embodiment, different types of insulation defects (such as free particles, air gaps, and floating potentials) have specific discharge phase preferences (such as positive half-cycle peak, negative half-cycle valley, bipolar symmetry, etc.) under the action of power frequency voltage. Therefore, by mapping the discharge pulses to the power frequency phase interval according to their occurrence time and jointly statistically analyzing them with the sub-band frequency bands, a two-dimensional distribution pattern with category specificity can be formed. Using the mean amplitude as the matrix element, the characterization of the high-energy response region is strengthened, and the contribution of low-amplitude noise is weakened, so that the effective discharge is presented as a local bright region in the matrix, while the interference is manifested as a low amplitude or random distribution across the entire domain.

[0106] In an optional embodiment, the step of identifying valid discharge signals based on the frequency-phase feature matrix combined with the interference feature library includes:

[0107] The pulse signal amplitude ratio of the effective sub-band is obtained based on the real-time frequency-phase feature matrix, and its overlap with the standard phase amplitude ratio is calculated.

[0108] The overlap is compared with a set overlap threshold. If the overlap is less than a first overlap threshold, it is determined to be an interference signal; if the overlap is greater than or equal to the first overlap threshold and less than a second overlap threshold, it is determined to be a suspected discharge signal; if the overlap is greater than or equal to the second overlap threshold, it is determined to be a valid discharge signal; the first overlap threshold is less than the second overlap threshold.

[0109] The effective discharge signal or the suspected discharge signal is compared with an interference feature library containing frequency-phase-duration information to perform sudden interference verification on the effective discharge signal and obtain the final effective discharge signal.

[0110] In some embodiments, the phase amplitude ratio of the effective sub-band in the measured matrix is ​​first extracted, and then the phase amplitude ratio of the corresponding standard template is extracted. The overlap ratio is calculated based on the overlap ratio calculation formula: overlap ratio = Σ(|measured ratio - template ratio|) × 100%. Two levels of overlap ratio thresholds are set: the first overlap ratio threshold is 60%, and the second overlap ratio threshold is 85%. If the overlap ratio is < 60%, it is determined to be an interference signal; if the overlap ratio is ≤ 60% and < 85%, it is determined to be a suspected discharge signal; if the overlap ratio is ≥ 85%, it is determined to be a valid discharge signal.

[0111] In some embodiments, narrowband segmented noise reduction targets broadband fixed noise (such as power grid harmonics). However, sudden narrowband interference may occur on-site (e.g., high-frequency electric sparks from nearby construction or instantaneous discharges from other equipment). The frequency of this interference may fall within the effective sub-band and may mimic the phase distribution of an effective signal, leading to misjudgment of the two-dimensional matrix matching as an effective signal. For example, when inspecting a busbar, a sudden 250MHz electric spark generated by nearby construction (effective sub-band 200MHz-300MHz, a typical sensitive frequency band for defects such as poor busbar joint contact) may occur, and it may happen to occur at the voltage peak phase (90°-120°). The two-dimensional (frequency band-phase) matrix matching similarity reaches 90%, but the system misjudges it as free metal particle discharge from the busbar, leading to unnecessary gas chamber opening for maintenance. This misjudgment of sudden interference increases maintenance costs and reduces staff trust in the system. Therefore, in this embodiment, the frequency-phase-duration features of common sudden interferences (such as construction electric sparks or instantaneous equipment vibrations) are pre-stored in an interference feature library. After two-dimensional feature matrix matching, the interference feature library is compared to eliminate sudden interference.

[0112] For suspected valid signals, a local high-frequency rescan is triggered to collect the signal in that frequency band again. Only if the phase distribution similarity of the two rescans reaches 90% is the signal considered valid, thus avoiding misjudgment due to single interference.

[0113] In some embodiments, frequency-phase two-dimensional matrix matching relies on a standard phase template calibrated in the laboratory. However, the field environment (temperature, humidity, aging of GIS components, and power grid voltage fluctuations) differs significantly from the laboratory environment, leading to a shift in the phase distribution of the actual discharge signal (e.g., in the laboratory, insulation gap discharge is concentrated at 120°-150°, while in the field, due to increased temperature, it shifts to 130°-160°). This causes the standard template to fail to match, increasing the misjudgment rate. Furthermore, changes in the field environment result in template mismatch, further reducing the utilization rate of phase features. Therefore, in practical applications, a 5-minute background signal without discharge is first acquired. Combined with the field temperature, humidity, and voltage values, the phase range of the standard template is automatically adjusted (e.g., for every 10° increase in temperature, the phase range shifts 10° to the right). For critical equipment, a normal discharge simulation signal is acquired during the initial test to generate a dedicated phase template for that equipment. Subsequent tests prioritize the use of this dedicated template for matching.

[0114] As a specific implementation method, assuming that a certain basin-type insulator needs to be tested, the specific operation can be as follows:

[0115] Preliminary on-site search information: Install the broadband partial discharge detection equipment on the corresponding GIS shell of the basin insulator. Do not rush to check for faults. First, collect 10 minutes of background signal, that is, the signal when the basin insulator is running normally. There are no partial discharge defects, only environmental electromagnetic interference and normal equipment operation noise (e.g., on-site temperature 35℃, humidity 55%, electromagnetic interference intensity of surrounding switchgear -60dBm).

[0116] Automatic calibration of fault timing patterns: The system stores laboratory standards: for the internal air gap discharge of a basin-type insulator, under normal power frequency conditions, the discharge phase is concentrated in the voltage absolute value rising segment (0°-90°, 180°-270°). Combined with the on-site temperature of 35°C, the system automatically calculates: the temperature is 10°C higher than the laboratory temperature of 25°C. According to the GIS equipment partial discharge phase temperature correction rule, "for every 10°C increase, the air gap discharge phase range expands 5° towards the voltage peak direction," the phase range should be adjusted to 5°-95° ​​and 185°-275° (equivalent to specifically expanding the original laboratory phase reference range to adapt to the changes in insulation medium characteristics caused by the on-site temperature).

[0117] Customized information for the equipment: The basin-type insulator is a key component connecting the GIS busbar and the circuit breaker. Simulated fault signals are collected (e.g., artificially creating a tiny air gap defect inside the insulator to simulate insulation degradation during actual operation). The actual discharge phase of the basin-type insulator is recorded as 10°-90° and 190°-270°, and the PRPD spectrum exhibits a typical "rabbit ear" shape. This specific phase range and spectrum feature are stored to mark the unique identification pattern of the internal air gap discharge of this GIS basin-type insulator. The next time this basin-type insulator is checked, the system does not need to perform temperature calibration; it directly retrieves its specific phase range of 10°-90° and 190°-270° and the "rabbit ear" shaped spectrum feature. By comparing the detected signal, as long as the signal phase falls within this range and the spectrum shape matches, it is determined to be a valid fault signal of internal air gap discharge of the insulator, preventing misjudgment due to on-site temperature fluctuations, surrounding electromagnetic interference, or slight aging of the insulator.

[0118] In this embodiment, by comparing the interference feature library, the coupling consistency of discharge in the three-dimensional space of "phase-amplitude-time" is utilized to effectively eliminate pseudo-discharge signals with only surface similarity. Multi-level criteria and spatiotemporal correlation verification jointly improve the ability to identify co-frequency and co-phase interference and reduce the misjudgment rate caused by the same frequency and phase overlap.

[0119] S4. By integrating the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate, a GIS insulation defect diagnosis model based on an improved random forest algorithm is used to diagnose and locate the insulation defect type of the equipment under test.

[0120] In an optional embodiment, combined with Figure 4 As shown, step S4 includes:

[0121] S41. Based on the phase interval number of each effective discharge signal, the distribution ratio of effective discharge signals in each phase interval is counted to obtain the phase distribution feature vector;

[0122] S42. Calculate the amplitude change rate per unit time based on the peak value of the effective discharge signal continuously monitored within the effective narrowband sub-band;

[0123] S43. Extract the frequency range parameters of each effective narrowband sub-band, and integrate them with the phase distribution feature vector and the amplitude change rate to form a structured three-dimensional feature parameter set;

[0124] S44. Obtain a standard defect feature library and use a random forest algorithm with feature importance weighted splitting and differentiating weights for different defect types as optimization objectives to establish a GIS insulation defect diagnosis model.

[0125] S45. Based on the GIS insulation defect diagnosis model, analyze the three-dimensional feature parameter set of the equipment to be tested, and output the insulation defect type and corresponding component location of the equipment to be tested.

[0126] In some embodiments, the amplitude change rate depends on the unit time window setting (e.g., calculated at 1 second / 10 seconds). Improper window settings can lead to numerical distortion (e.g., a short window misjudges sudden noise as a rapid amplitude change, while a long window masks the true rapid degradation). The calculation window can be automatically adjusted based on the signal pulse frequency (e.g., a 0.5-second window for dense pulses and a 5-second window for sparse pulses) to avoid distortion. Specifically, Poisson distribution detection can be used. If the pulse count per unit time conforms to a Poisson distribution (strong randomness), it is determined to be a floating potential discharge, and the calculation window is automatically shortened (e.g., 0.5 seconds). If the pulse intervals are uniform (strong periodicity), it is determined to be a fixed defect, and the window is extended (e.g., 5 seconds) to obtain a smooth trend. The amplitude change rate can reflect the development speed of the defect; for example, slow growth indicates slow defect expansion, and rapid fluctuations indicate particulate metal defects. This solves the problem that traditional diagnostics can only identify the defect type but cannot determine the development trend, thus improving the practical value of the detection results.

[0127] Specifically, since single features (such as phase or frequency only) often overlap between different defects (e.g., a 50kHz effective discharge signal and a 500kHz interference signal are in phase), isolated feature analysis methods cannot accurately distinguish defect types. The feature space formed by combination has higher-dimensional separability: frequency interval parameters determine the sensitive frequency band corresponding to the defect, adapting to the resonant frequency differences of different GIS components; phase distribution feature vectors distinguish between effective discharge signals and in-phase interference (effective signals have a fixed phase distribution pattern, while interference signals are irregular), and amplitude change rate reflects the dynamic trend of degradation; after the fusion of the three, even if single features overlap, accurate classification can be achieved through multi-dimensional combination. The improved random forest, through weighted splitting and sample weight optimization, prioritizes the use of high-discriminative features for decision-making and balances the identification sensitivity of various defects, thus achieving accurate classification even in feature overlap areas; at the same time, the frequency interval is associated with the "component-frequency band" mapping in the multi-dimensional mapping database, thereby achieving reverse location of the gas chamber where the defect is located.

[0128] Traditional random forest algorithms, which use equal weights to process samples and features, have two key drawbacks: first, they are weak at identifying defects with small sample sizes and high severity (such as metal tip discharge on the inner wall of an outer casing), easily masked by common defects with large sample sizes; second, they lack differentiated consideration of the discriminative power of different features, failing to highlight core features like phase distribution that contribute more to defect identification. An improved mechanism combining feature weighting and sample differentiation weights enhances the model's ability to identify key defects. Specifically, weights are allocated according to feature contribution (e.g., phase distribution 0.4, frequency range 0.35, amplitude change rate 0.25) to prioritize features with high discriminative power, improving classification specificity. Sample differentiation weights assign higher weights to defects with small sample sizes and high severity, addressing sample imbalance and improving the recall rate of rare, high-severity defects, thus preventing equipment failures due to missed detections.

[0129] In this embodiment, the three-dimensional feature parameter set overcomes the limitation of single feature overlap. Even if different defects have the same partial features, accurate classification can be achieved through multi-dimensional combination. The improved random forest algorithm introduces feature weighting and sample weight optimization, which can improve the model's sensitivity to key features, thereby improving the defect identification accuracy.

[0130] In an optional embodiment, the step of acquiring a standard defect feature library and establishing a GIS insulation defect diagnosis model using a random forest algorithm with feature importance-weighted splitting and differentiated weights for different defect types as optimization objectives includes:

[0131] Based on various GIS defect samples calibrated in the laboratory, the frequency range, phase distribution feature vector, and amplitude change rate of each type of defect were extracted to establish a standard defect feature library;

[0132] Calculate the classification contribution of the three-dimensional features in the standard defect feature library to the defect type, and determine the split weights of the frequency range, phase distribution feature vector and amplitude change rate based on the classification contribution.

[0133] The sample splitting weight for each type of defect sample is determined by linear weighting based on the number and severity of different defect samples;

[0134] Weighted embedding is performed on individual GIS defect samples based on the sample splitting weights, and the three-dimensional features of individual GIS defect samples are weighted based on the splitting weights. The weighted dataset is then divided into training and testing sets.

[0135] The single decision tree-based learner is trained based on the training set, and the basic GIS insulation defect diagnosis model generated by the integrated decision tree-based learner is obtained. The basic GIS insulation defect diagnosis model is then validated and optimized based on the test set to obtain the target GIS insulation defect diagnosis model.

[0136] In this embodiment, split weights are used to prioritize feature dimensions with high contribution when splitting nodes during training. For example, splitting is prioritized based on the concentration of phase distribution, and then the classification is refined by combining the number of effective subbands in the frequency range, improving the targeting of node splitting. Sample split weights are used to assign split weights to defect type samples. During training, samples with higher weights have a greater impact on the splitting of the decision tree, ensuring that the model can accurately identify such defects. For defects with a small sample size but high harm (such as metal tip discharge on the inner wall of the shell), a higher comprehensive weight is assigned, allowing these defects to have a greater influence when splitting decision tree nodes.

[0137] Specifically, the sample split weights are expressed as: ,in, Let be the number of samples of the i-th type of defect. The severity level of the i-th type of defect is determined by the degree of impact of the defect on GIS equipment and power system. For example, metal tip discharge is classified as high severity level and assigned a value of 1; air gap discharge is classified as medium severity level and assigned a value of 0.7. and These are the weighting coefficients, such as , Priority should be given to the balance of sample size, while also taking into account the severity of defects.

[0138] In this embodiment, since the distribution of on-site defect samples is uneven (e.g., few free particle samples but high hazard), and different features contribute differently to classification (e.g., phase distribution is usually more discriminative than amplitude change rate), the decision tree can be guided to prioritize the use of key physical features for segmentation by calculating feature importance and assigning greater splitting weight to high-contribution features. At the same time, higher training weights are assigned to minority or high-risk defect samples, so that the model pays more attention to these easily overlooked but important categories during the learning process, avoiding model bias towards the majority class due to sample imbalance. The weighted training model is validated and optimized on the test set to ensure that its generalization ability is not limited by specific equipment or calibration conditions.

[0139] In some embodiments, various GIS defect sample data calibrated in the laboratory are collected, covering common defect types such as free metal particles in busbar cylinders, air gaps inside basin insulators, and floating potential of conductive rods. Three-dimensional feature parameters—frequency range, phase distribution feature vector, and amplitude change rate—are extracted for each type of defect to establish a standard defect feature library. For example, the standard parameters for free metal particles are: frequency range 50kHz-200kHz (effective subband 60-80 / 100-120 / 160-180kHz), phase concentrated at 120°-150° (60%) and 240°-270° (35%), and amplitude change rate approximately 0.00083V / s. Simultaneously, valid defect data collected on-site are incorporated and supplemented to the feature library after laboratory verification to ensure data coverage of actual application scenarios.

[0140] Furthermore, feature parameters with different dimensions, such as amplitude change rate and phase proportion, are normalized to eliminate the influence of numerical range differences on the model. The 3σ principle is used to remove outlier data from the feature library, such as data with amplitude change rates exceeding ±3 standard deviations of the mean for similar defects, to avoid interfering with model training.

[0141] Furthermore, model parameters are initialized before model training, including the number of decision trees, the maximum depth of the decision trees, the minimum number of sample splits, and the number of random feature subsets. The number of random feature subsets is 1, 2, or 3, representing the feature dimensions (frequency range, phase distribution feature vector, and amplitude change rate) randomly selected for each split.

[0142] In some examples, sample splitting weights are directly assigned to individual samples. For instance, the sample weight for metal tip discharge on the inner wall of the casing (high hazard, few samples) is much higher than the sample weight for air gap discharge inside the basin insulator (medium hazard, many samples). Based on the contribution of three-dimensional features, the feature parameters in the samples are weighted, and the phase distribution features have the highest influence on the model after weighting. The weighted dataset is divided into training and test sets, and virtual samples are generated by perturbing the features of rare high-weight defect samples to ensure a balanced influence of various defect samples in the training set.

[0143] During training, the training set undergoes bootstrap sampling. High-weighted samples are selected more frequently during this process, ensuring that each decision tree's training samples contain a sufficient amount of high-hazard, low-sample defect data. Unsampled out-of-package (OOB) data is retained for subsequent model evaluation. After sampling, node splitting is performed based on feature contribution, and leaf node partitioning is guided by sample splitting weights to construct multiple independent improved decision trees. Specifically, when splitting nodes based on feature contribution, splitting features are prioritized from dimensions with high feature contribution, such as phase distribution features, then frequency interval features, and finally amplitude change rate features. When partitioning leaf nodes, sample weights influence the node's category determination. When high-weighted samples constitute a larger proportion of a node, that node is preferentially classified as the corresponding defect type, preventing rare defects from being "submerged" by the majority of samples.

[0144] Furthermore, a weighted voting strategy is employed to integrate the outputs of all decision trees. The voting weight of a single decision tree is positively correlated with its classification accuracy on out-of-package (OOB) data; the higher the accuracy of a decision tree, the greater its voting influence. For the input 3D feature parameter set, all decision trees output defect type prediction results, and the final diagnosis result is determined according to the number of weighted votes, while also outputting the confidence score (the proportion of the highest number of votes to the total number of votes). A test set is used to verify the model performance, including recall, overall defect identification accuracy, and the prediction error on out-of-package data, which must meet the corresponding preset thresholds. Based on the test results on the test set, the key parameters of the decision trees are adjusted, including the maximum depth, minimum number of sample splits, and number of random feature subsets. The optimal parameter combination is found using 5-fold cross-validation to reduce the risk of model overfitting.

[0145] In this embodiment, the model overcomes the shortcomings of the traditional random forest algorithm, solves the problems of sample imbalance and feature weight imbalance, and can simultaneously output the defect type, the corresponding component location, and the development trend. For example, by judging from the amplitude change rate, it can determine that the free metal particle discharge defect in the busbar cylinder is slowly expanding, guiding maintenance personnel to carry out recent maintenance, realizing the transformation from passive detection to proactive maintenance. It not only avoids the problems of difficult feature overlap and high interference misjudgment, but also achieves the dual goals of defect identification and trend prediction, ultimately realizing accurate diagnosis of GIS insulation defects in complex electromagnetic environments.

[0146] In an optional embodiment, the method further includes effectively verifying the suspected discharge signal, specifically including:

[0147] Based on the suspected discharge signal, its three-dimensional features, including frequency range parameters, phase distribution feature vector, and amplitude change rate, are obtained;

[0148] The three-dimensional features of the suspected discharge signal are used to detect the defect type through the GIS insulation defect diagnosis model. Based on the defect type output probability, it is determined whether the suspected discharge signal is an interference signal or a real discharge signal.

[0149] If the output probability of the defect type is still determined to be a suspected discharge signal, then dynamic verification is performed in conjunction with the amplitude change rate to obtain a valid verification result of the suspected discharge signal.

[0150] Specifically, the three-dimensional feature parameters of the suspected signal are input into the GIS insulation defect diagnosis model to obtain the probability of the defect type output by the model. If the output probability is greater than or equal to a preset threshold, it is determined to be a real discharge signal; if the probability is less than the preset threshold, it is determined to be an interference signal. If the model output probability is still within the suspected range, the suspected signal is monitored for an extended period (e.g., extended to 20 minutes) to analyze the stability of its amplitude change rate. If the amplitude change rate shows a stable growth or fluctuation trend, consistent with the change pattern of defect signals, it is determined to be a real discharge signal; if the amplitude change rate has no stable pattern and the signal duration is less than 0.1 seconds, consistent with the characteristics of sudden interference, it is determined to be an interference signal, and a secondary verification is completed.

[0151] In this embodiment, the phase distribution of some early defects or weak discharge signals may not overlap sufficiently with the standard template due to low signal-to-noise ratio, but their amplitude often shows a slow upward trend with insulation degradation, while the amplitude of interference signals usually fluctuates randomly or remains stable. Therefore, even if the phase matching degree is moderate, if the amplitude change rate shows a continuous increase, it can be confirmed that it is a real defect; conversely, if the model output probability is low and the amplitude has no trend, it can be excluded as interference. This mechanism utilizes the physical laws of defect evolution as the ultimate criterion to form a closed-loop verification of fuzzy signals.

[0152] In this embodiment, the amplitude change of the early weak discharge signal has a stable trend, while the sudden interference signal is irregular and short-lived. The two can be distinguished by dynamic verification of the amplitude change rate, which solves the problem that early weak discharge is easily misjudged as interference in traditional technology and improves the reliability of diagnosis. Through the dual guarantee of model re-judgment and dynamic trend verification, it is ensured that early weak defects are not easily ruled out, while preventing interference signals from being misidentified, which significantly reduces the false alarm rate and the missed alarm rate.

[0153] Based on the same inventive concept, this application also provides a variable frequency narrowband phase correlation diagnostic device for GIS insulation defects, corresponding to a variable frequency narrowband phase correlation diagnostic method for GIS insulation defects, such as... Figure 5 As shown, the device includes:

[0154] The narrowband signal acquisition module is used to acquire the characteristic signals of the device under test based on the target frequency band and to acquire the narrowband signal stream.

[0155] The narrowband signal segmentation module is used to divide the narrowband signal stream into multiple continuous narrowband sub-bands and perform noise reduction processing on them to obtain effective narrowband sub-bands.

[0156] The effective signal identification module is used to construct a frequency-phase feature matrix based on the distribution of effective narrowband sub-bands on the power frequency cycle phase of the power system, and to identify effective discharge signals based on the frequency-phase feature matrix combined with the interference feature library.

[0157] The insulation defect detection module is used to integrate the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate. It uses a GIS insulation defect diagnosis model based on an improved random forest algorithm to diagnose and locate the insulation defect type of the equipment under test.

[0158] In this embodiment, by limiting signal acquisition to the target frequency band, broadband noise pollution can be avoided during full-band acquisition, providing a high signal-to-noise ratio input for subsequent refined processing. By dividing the frequency band into sub-bands and reducing noise, the spectral components that may be excited by different defects can be further separated within the target frequency band, preventing weak discharge signals from being masked by interference, improving the detection signal-to-noise ratio and feature resolution. This allows for the construction of an accurate frequency-phase feature matrix using the physical coupling relationship between discharge and power frequency voltage, enabling the differentiation between effective discharge signals with fixed phase patterns and irregular interference in the effective narrowband sub-bands from a time-series perspective, thereby ensuring feature accuracy and reducing the false positive rate of pulse signals. Finally, by fusing three-dimensional features and using an improved random forest algorithm for defect diagnosis, the limitations of single feature detection can be overcome, achieving accurate diagnosis of defect type and location, and improving detection accuracy.

[0159] The above-described embodiments are preferred embodiments of this application and are not intended to limit the specific scope of this application. The scope of this application includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape, structure, and method of this application are within the protection scope of this application.

Claims

1. A variable-frequency narrowband phase correlation diagnostic method for GIS insulation defects, characterized in that: Includes the following steps: Based on the target frequency band, acquire the characteristic signals of the device under test and obtain the narrowband signal stream; The narrowband signal stream is divided into multiple consecutive narrowband sub-bands and noise reduction is performed on them to obtain effective narrowband sub-bands; A frequency-phase feature matrix is ​​constructed based on the distribution of effective narrowband sub-frequency bands in the power system's power frequency cycle phase. Effective discharge signals are identified based on this frequency-phase feature matrix combined with an interference feature library. Specifically, the effective narrowband sub-frequency bands are marked with effective sub-band identifiers, and phase intervals are divided and numbered according to the phase angle of the power system's power frequency cycle. Narrowband signal data is extracted from the effective narrowband sub-frequency bands, and the power frequency cycle corresponding to each narrowband signal is identified and its corresponding phase interval number is assigned to extract the effective narrowband signals from the effective narrowband sub-frequency bands. The average pulse amplitude of the effective narrowband signals in the effective narrowband sub-frequency bands is used as matrix elements to construct the frequency-phase feature matrix. The steps for identifying effective discharge signals include: The pulse signal amplitude ratio of the effective sub-band is obtained based on the real-time frequency-phase feature matrix, and its overlap with the standard phase amplitude ratio is calculated. The overlap is compared with a set overlap threshold. If the overlap is less than a first overlap threshold, it is determined to be an interference signal; if the overlap is greater than or equal to the first overlap threshold and less than a second overlap threshold, it is determined to be a suspected discharge signal; if the overlap is greater than or equal to the second overlap threshold, it is determined to be a valid discharge signal; the first overlap threshold is less than the second overlap threshold. The effective discharge signal or the suspected discharge signal is compared with an interference feature library containing frequency-phase-duration to perform burst interference verification on the effective discharge signal and obtain the final effective discharge signal. By integrating the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate, a GIS insulation defect diagnosis model based on an improved random forest algorithm is used to diagnose and locate the insulation defect type of the equipment under test. Specifically, the distribution ratio of the effective discharge signal in each phase range is statistically analyzed based on the phase range number of each effective discharge signal to obtain the phase distribution feature vector.

2. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 1, characterized in that: The step of acquiring the characteristic signal of the device under test based on the target frequency band and acquiring the narrowband signal stream includes: Construct a multi-dimensional feature mapping database containing one-to-one correspondences between GIS component locations, defect types, sensitive frequency bands, and typical phases; The signal spectrum distribution of the device to be detected is identified based on a multidimensional feature mapping database to determine the target frequency band range. The signal acquisition frequency division parameters are dynamically adjusted based on the target frequency band range. The signal acquisition frequency division parameters include at least the sampling rate and the filtering bandwidth. Based on the adjusted frequency division parameters, continuous analog signal acquisition is performed on the target frequency band interval, and the signal is divided into discrete narrowband signal streams by a digital filter.

3. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 1, characterized in that: The process of dividing the narrowband signal stream into multiple consecutive narrowband sub-bands and performing noise reduction processing on them to obtain effective narrowband sub-bands includes: The narrowband signal stream is divided into multiple consecutive narrowband sub-bands in the frequency domain, and each narrowband sub-band is numbered. A noise reduction threshold is determined based on the statistical characteristics of the narrowband signal in each narrowband sub-band. Wavelet threshold filtering is then applied to the narrowband sub-band signal based on the noise reduction threshold to determine candidate narrowband sub-bands. Based on pulse signal constraints, the candidate narrowband sub-frequency bands are checked for leaks. If a pulse signal exists in the candidate narrowband sub-frequency band and the pulse signal attributes meet the pulse standard, then the corresponding candidate narrowband sub-frequency band is taken as a valid narrowband sub-frequency band.

4. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 3, characterized in that: Before dividing the narrowband sub-bands, the narrowband signal is divided and judged based on the signal density and signal time-frequency continuity of the narrowband signal stream. When the signal density is greater than a preset threshold and the signal is continuous, the narrowband signal stream is divided into narrowband sub-bands.

5. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 3, characterized in that: The process involves constructing a frequency-phase feature matrix based on the distribution of effective narrowband sub-bands on the power system's power frequency cycle phase, and identifying effective discharge signals based on this frequency-phase feature matrix combined with an interference feature library, including: Statistically analyze the pulse amplitude of all narrowband signals within the target time period under each effective sub-band-phase combination, effectively verify the narrowband signals based on the pulse amplitude, and obtain the average pulse amplitude of the effective narrowband signals; The average pulse amplitude of the effective narrowband signal is used as a matrix element, and a frequency-phase feature matrix is ​​constructed with the effective narrowband sub-band number as the horizontal axis and the phase interval number as the vertical axis. The effective discharge signal is identified based on the frequency-phase feature matrix combined with the interference feature library.

6. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 1, characterized in that: The method integrates the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate. It then employs a GIS insulation defect diagnosis model based on an improved random forest algorithm to diagnose and locate the insulation defect type of the equipment under test, including: The amplitude change rate per unit time is calculated based on the peak value of the effective discharge signal continuously monitored within the effective narrowband sub-band. Extract the frequency range parameters of each effective narrowband sub-band, and integrate them with the phase distribution feature vector and the amplitude change rate to form a structured three-dimensional feature parameter set; A standard defect feature library was obtained, and a random forest algorithm with feature importance weighted splitting and differentiating weights for different defect types as optimization objectives was used to establish a GIS insulation defect diagnosis model. Based on the GIS insulation defect diagnosis model, the three-dimensional feature parameter set of the equipment to be tested is analyzed, and the insulation defect type and corresponding component location of the equipment to be tested are output.

7. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 6, characterized in that: The acquisition of the standard defect feature library involves establishing a GIS insulation defect diagnosis model using a random forest algorithm with feature importance-weighted splitting and differentiated weights for different defect types as optimization objectives, including: Based on various GIS defect samples calibrated in the laboratory, the frequency range, phase distribution feature vector, and amplitude change rate of each type of defect were extracted to establish a standard defect feature library; Calculate the classification contribution of the three-dimensional features in the standard defect feature library to the defect type, and determine the split weights of the frequency range, phase distribution feature vector and amplitude change rate based on the classification contribution. The sample splitting weight for each type of defect sample is determined by linear weighting based on the number and severity of different defect samples; Weighted embedding is performed on individual GIS defect samples based on the sample splitting weights, and the three-dimensional features of individual GIS defect samples are weighted based on the splitting weights. The weighted dataset is then divided into training and testing sets. The single decision tree-based learner is trained based on the training set, and the basic GIS insulation defect diagnosis model generated by the integrated decision tree-based learner is obtained. The basic GIS insulation defect diagnosis model is then validated and optimized based on the test set to obtain the target GIS insulation defect diagnosis model.

8. The variable frequency narrowband phase correlation diagnosis method for GIS insulation defects according to claim 6, characterized in that: The method also includes effective verification of suspected discharge signals, specifically including: Based on the suspected discharge signal, its three-dimensional features, including frequency range parameters, phase distribution feature vector, and amplitude change rate, are obtained; The three-dimensional features of the suspected discharge signal are used to detect the defect type through the GIS insulation defect diagnosis model. Based on the defect type output probability, it is determined whether the suspected discharge signal is an interference signal or a real discharge signal. If the output probability of the defect type is still determined to be a suspected discharge signal, then dynamic verification is performed in conjunction with the amplitude change rate to obtain a valid verification result of the suspected discharge signal.

9. A variable-frequency narrowband phase correlation diagnostic device for GIS insulation defects, applicable to the variable-frequency narrowband phase correlation diagnostic method for GIS insulation defects as described in any one of claims 1-8, characterized in that: include: The narrowband signal acquisition module is used to acquire the characteristic signals of the device under test based on the target frequency band and to acquire the narrowband signal stream. The narrowband signal segmentation module is used to divide the narrowband signal stream into multiple continuous narrowband sub-bands and perform noise reduction processing on them to obtain effective narrowband sub-bands. The effective signal identification module is used to construct a frequency-phase feature matrix based on the distribution of effective narrowband sub-frequency bands on the phase of the power system's power frequency cycle. Based on this matrix and an interference feature library, effective discharge signals are identified. Specifically, the effective narrowband sub-frequency bands are marked with effective sub-band identifiers, and phase intervals are divided and numbered according to the phase angle of the power system's power frequency cycle. Narrowband signal data is extracted from the effective narrowband sub-frequency bands. The power frequency cycle corresponding to each narrowband signal is identified and its corresponding phase interval is numbered to extract the effective narrowband signals from the effective narrowband sub-frequency bands. The average pulse amplitude of the effective narrowband signals in the effective narrowband sub-frequency bands is used as matrix elements to construct the frequency-phase feature matrix. The steps for identifying effective discharge signals include: The pulse signal amplitude ratio of the effective sub-band is obtained based on the real-time frequency-phase feature matrix, and its overlap with the standard phase amplitude ratio is calculated. The overlap is compared with a set overlap threshold. If the overlap is less than a first overlap threshold, it is determined to be an interference signal; if the overlap is greater than or equal to the first overlap threshold and less than a second overlap threshold, it is determined to be a suspected discharge signal; if the overlap is greater than or equal to the second overlap threshold, it is determined to be a valid discharge signal; the first overlap threshold is less than the second overlap threshold. The effective discharge signal or the suspected discharge signal is compared with an interference feature library containing frequency-phase-duration to perform burst interference verification on the effective discharge signal and obtain the final effective discharge signal. The insulation defect detection module is used to integrate the frequency range parameters of each effective narrowband sub-band, the phase distribution feature vector of the effective discharge signal, and the amplitude change rate. It uses a GIS insulation defect diagnosis model based on an improved random forest algorithm to diagnose and locate the insulation defect type of the device under test. Specifically, the phase distribution feature vector is obtained by statistically analyzing the distribution ratio of the effective discharge signal in each phase range based on the phase range number of each effective discharge signal.

Images