Surge arrester integrated detection method and device

Abstract

The present application relates to the technical field of power transformation equipment detection, and provides a lightning arrester integrated detection method and device, comprising: collecting original leakage current signals and appearance images of the lightning arrester; performing adaptive harmonic suppression processing on the original leakage current signals to obtain real leakage current signals; inputting the real leakage current signals into a pre-trained reconstruction model to obtain a state diagnosis result, wherein the reconstruction model adopts a bidirectional long short-term memory network autoencoder with residual connection; inputting the appearance images into a pre-trained recognition model to output a defect recognition result, wherein the recognition model adopts an end-to-end architecture with non-maximum suppression removed; after single-point detection is completed based on the above steps, the internal device connection line of the bearing platform is retained, only the external high-voltage lead and the grounding wire are removed, and the whole bearing platform is transferred for performing multi-point detection operation. The problems of low detection efficiency, insufficient standardization and safety of lightning arrester in the related art are solved.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing technology, and in particular to an integrated testing method and device for surge arresters. Background Technology

[0002] Zinc oxide surge arresters are core overvoltage protection devices in power system substations, and their insulation performance and leakage current parameters directly determine the safety of power grid operation. Currently, substation surge arrester testing mainly relies on traditional split-type testing equipment, manual wiring, and on-site readings. This approach has many technical shortcomings, considering existing research and field application status.

[0003] The relevant technical solutions generally employ decentralized testing instruments. Before testing, various equipment such as DC high-voltage generators, microammeters, insulation tools, and power supply equipment need to be transported one by one. The placement, wiring, grounding, and connection of equipment are all done manually. The equipment is scattered, the wiring is cumbersome, and the larger the substation area and the more equipment under test, the longer the operation takes and the lower the work efficiency. Some technical solutions attempt to implement intelligent testing, but the problems of non-standardized work processes and manual data processing still exist, affecting the overall testing efficiency.

[0004] Meanwhile, traditional operations often involve haphazard instrument placement, uneven ground that can cause high-voltage equipment to tip over, and the easy loss or dropping of insulated tools, resulting in poor on-site operational standardization. In terms of safety, traditional testing power supplies rely solely on plugging and unplugging to switch power on and off, lacking a clear electrical disconnect point, which does not meet the requirements of electrical safety work regulations. Furthermore, they lack leakage protection and voltage monitoring functions, increasing the risk of equipment damage and electric shock due to overvoltage and leakage. Additionally, traditional microammeters often use wired or ordinary wireless transmission modes, which are prone to data interruption, have low transmission success rates under signal interference, and exhibit poor measurement accuracy in high-voltage, strong electromagnetic environments, requiring repeated measurements and increasing the duration of high-voltage operations, thus creating safety hazards.

[0005] In summary, no effective solution has yet been proposed to address the issues of low testing efficiency, insufficient standardization, and inadequate safety in surge arrester technology. Summary of the Invention

[0006] The present invention provides an integrated testing method and apparatus for surge arresters, which at least solves the problems of low testing efficiency, insufficient standardization and safety of surge arresters in related technologies.

[0007] This invention provides an integrated testing method for surge arresters, comprising the following steps: acquiring the original leakage current signal and appearance image of the surge arrester; performing adaptive harmonic suppression processing on the original leakage current signal to obtain the true leakage current signal; inputting the true leakage current signal into a pre-trained reconstruction model to obtain a state diagnosis result, wherein the reconstruction model employs a bidirectional long short-term memory network autoencoder with residual connections for calculating reconstruction error and quantifying anomaly degree; inputting the appearance image into a pre-trained recognition model to output a defect recognition result, wherein the recognition model employs an end-to-end architecture with non-maximum suppression removed; after completing single-point testing based on the above steps, retaining the internal equipment connection lines of the carrying platform, removing only the external high-voltage leads and grounding wires, and transferring the entire carrying platform for multi-point testing operations.

[0008] As an optional approach, adaptive harmonic suppression processing is performed on the original leakage current signal to obtain the true leakage current signal. This includes: filtering the original leakage current signal based on a hardware filtering architecture to obtain a preprocessed signal; performing a fast Fourier transform on the preprocessed signal to extract real-time harmonic amplitudes; adaptively iteratively updating the harmonic filtering thresholds for each harmonic based on the historical harmonic statistical mean and the real-time harmonic amplitudes; performing harmonic suppression processing on the original leakage current signal based on the harmonic filtering thresholds to obtain a preliminary corrected time-domain signal; and performing phase distortion compensation correction processing on the preliminary corrected time-domain signal to obtain the true leakage current signal.

[0009] As an optional approach, the formula for adaptively iteratively updating the threshold of each harmonic filter is expressed as: ; In the formula, T h ( k ) indicates the first k Real-time dynamic suppression threshold for subharmonics α Represents the iterative weight coefficients. A k_avg Indicates the historical number within the sliding time window k Average amplitude of the second harmonic A k_real This indicates the FFT identification obtained in the current sampling period. k Real-time amplitude of subharmonics.

[0010] As an optional solution, the above hardware filtering architecture includes a multi-stage RC low-pass filter circuit, each stage of which is equipped with a high-frequency ferrite bead. Before filtering the original leakage current signal based on the hardware filtering architecture, the method further includes setting the cutoff frequency of the multi-stage RC low-pass filter circuit.

[0011] As an alternative approach, the mathematical model for the original leakage current signal is as follows: ; In the formula, This represents the original leakage current signal. This represents the actual leakage current signal. Indicates harmonic interference components. This represents random electromagnetic white noise components.

[0012] As an optional approach, before inputting the aforementioned real leakage current signal into the pre-trained reconstruction model, the method further includes: establishing a first initial model based on a bidirectional long short-term memory network encoder, a residual connection layer, and a unidirectional long short-term memory network decoder; and training the first initial model in an unsupervised manner based on samples of normal operating conditions of the surge arrester to obtain the aforementioned reconstruction model, wherein the model loss function includes a mean square error term, a Kohlbek-Leibler divergence term, and a time series regularization term.

[0013] As an optional approach, the aforementioned real leakage current signal is input into a pre-trained reconstruction model to obtain a state diagnosis result, including: calculating the reconstruction error corresponding to the aforementioned real leakage current signal based on the aforementioned reconstruction model, which is used to characterize the difference between the aforementioned real leakage current signal and the normal operating condition signal; calculating anomaly score based on Mahalanobis distance and the aforementioned reconstruction error, wherein the aforementioned state diagnosis result includes the aforementioned anomaly score, and the aforementioned anomaly score characterizes the degree of anomaly in the operation state of the surge arrester.

[0014] As an optional approach, before inputting the aforementioned appearance image into the pre-trained recognition model, the method further includes: establishing a second initial model based on the backbone feature extraction network, the multi-scale bidirectional feature fusion network, and the detection-segmentation dual-branch output head; training the second initial model based on the frozen pre-training stage and the full-parameter unfreezing fine-tuning stage to obtain the aforementioned recognition model; wherein, in the aforementioned frozen pre-training stage, pre-training weights from a general dataset are loaded, the parameters of the backbone feature extraction network are frozen, and only the multi-scale bidirectional feature fusion network and the detection-segmentation dual-branch output head are trained; in the aforementioned full-parameter unfreezing fine-tuning stage, all network weights are unlocked, and a cosine annealing dynamic learning rate decay strategy is used for training, the formula for adaptive decay of the learning rate with each iteration round is expressed as: ; In the formula, lr epoch The learning rate represents the current iteration round. lr min This represents the minimum threshold learning rate. lr init This represents the initial maximum learning rate. epoch Indicates the current iteration round number.E total This indicates the total number of iterations.

[0015] This invention also provides an integrated testing method for surge arresters, comprising: a data acquisition module for acquiring the original leakage current signal and appearance image of the surge arrester; a harmonic suppression module for performing adaptive harmonic suppression processing on the original leakage current signal to obtain a true leakage current signal; a diagnostic module for inputting the true leakage current signal into a pre-trained reconstruction model to obtain a state diagnostic result, wherein the reconstruction model employs a bidirectional long short-term memory network encoder with residual connections for calculating reconstruction error and quantifying anomaly degree; an identification module for inputting the appearance image into a pre-trained identification model and outputting a defect identification result, wherein the identification model employs an end-to-end architecture with non-maximum suppression removed; and a carrying platform integrating all modules and all testing equipment, used to retain internal equipment connection lines and remove only external high-voltage leads and grounding wires after completing single-point testing, and transfer the entire system to perform multi-point testing operations.

[0016] As an optional solution, the data acquisition module includes an image acquisition unit and a DC microammeter; the image acquisition unit is used to acquire an external image of the surge arrester; the DC microammeter includes a connected meter head and a handheld meter, the meter head adopts an arc-shaped shell, the meter head has a built-in anti-interference circuit and a Bluetooth transmission unit, and the handheld meter has a built-in current calculation unit; the DC microammeter is used to obtain the original leakage current signal of the surge arrester.

[0017] The integrated surge arrester testing method provided by this invention involves: acquiring the original leakage current signal and appearance image of the surge arrester; performing adaptive harmonic suppression processing on the original leakage current signal to obtain the true leakage current signal; inputting the true leakage current signal into a pre-trained reconstruction model to obtain the state diagnosis result, wherein the reconstruction model uses a bidirectional long short-term memory network autoencoder with residual connections for calculating reconstruction error and quantifying anomaly degree; inputting the appearance image into a pre-trained recognition model to output the defect recognition result, wherein the recognition model uses an end-to-end architecture with non-maximum suppression removed; after completing single-point testing based on the above steps, retaining the internal equipment connection lines of the carrier platform, and only removing the external high-voltage leads and grounding wires, the entire carrier platform is transferred for multi-point testing operations. A standardized, integrated, and intelligent surge arrester testing operation system can be constructed based on an integrated mobile carrier platform, intrinsically safe power switch, and Bluetooth smart micro-ammeter system. Through integrated equipment layout, multiple safety protections, precise Bluetooth wireless transmission, and automatic current calculation, this system enables rapid deployment, safe operation, intelligent data acquisition, accurate calculation, and convenient relocation for surge arrester testing in substations. It improves testing efficiency and operational safety, and is adaptable to complex substation environments and batch testing scenarios involving large numbers of surge arresters. This addresses the issues of low efficiency, insufficient standardization, and inadequate safety in surge arrester testing in related technologies. Attached Figure Description

[0018] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the steps of an integrated detection method for surge arresters in an embodiment of the present invention.

[0020] Figure 2 This is a framework diagram of an improved LSTM deep temporal learning model in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of a surge arrester integrated detection device in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the overall architecture of the mobile surge arrester integrated testing platform in an embodiment of the present invention.

[0023] Figure 5a This is a schematic diagram of the structure of a multifunctional power switch in an embodiment of the present invention. Figure 1 .

[0024] Figure 5b This is a schematic diagram of the structure of a multifunctional power switch in an embodiment of the present invention. Figure 2 .

[0025] Figure 6 This is a schematic diagram of a DC high-voltage microammeter structure based on Bluetooth transmission in an embodiment of the present invention.

[0026] Figure 7 This is a schematic diagram of the improved YOLO26 defect intelligent and accurate identification principle in an embodiment of the present invention.

[0027] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0028] The above figures include the following reference numerals: 1. 10kV test specimen placement hole; 2. Grounding electrode; 3. Directional front wheel; 4. DC high voltage generator slot; 5. Insulating rod placement area; 6. Control console placement area; 7. Foldable handle; 8. Miscellaneous items placement area; 9. Brakeable omnidirectional rear wheel. Detailed Implementation

[0029] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0030] Currently, the testing of surge arresters in substations mainly relies on traditional split-type testing equipment, manual wiring, and on-site readings. Related technical solutions generally employ decentralized testing instruments, requiring the transport of various equipment such as DC high-voltage generators, microammeters, insulation tools, and power supply devices before testing. Manual placement, wiring, grounding, and connection are then performed. This process is cumbersome due to the scattered equipment and complex wiring; the larger the substation area and the more equipment under test, the longer the operation takes and the lower the efficiency. While some technical solutions attempt intelligent testing, the non-standardized workflow and manual data processing remain, impacting overall testing efficiency.

[0031] Meanwhile, traditional operations often involve haphazard instrument placement, uneven ground that can cause high-voltage equipment to tip over, and the easy loss or dropping of insulated tools, resulting in poor on-site operational standardization. In terms of safety, traditional testing power supplies rely solely on plugging and unplugging to switch power on and off, lacking a clear electrical disconnect point, which does not meet the requirements of electrical safety work regulations. Furthermore, they lack leakage protection and voltage monitoring functions, increasing the risk of equipment damage and electric shock due to overvoltage and leakage. Additionally, traditional microammeters often use wired or ordinary wireless transmission modes, which are prone to data interruption, have low transmission success rates under signal interference, and exhibit poor measurement accuracy in high-voltage, strong electromagnetic environments, requiring repeated measurements and increasing the duration of high-voltage operations, thus creating safety hazards.

[0032] Therefore, referring to Figure 1 As shown, the present invention provides an integrated detection method for surge arresters, including steps S101 to S105.

[0033] Step S101: Acquire the original leakage current signal and appearance image of the surge arrester.

[0034] Step S102: Perform adaptive harmonic suppression processing on the original leakage current signal to obtain the real leakage current signal.

[0035] Step S103: Input the above real leakage current signal into the pre-trained reconstruction model to obtain the state diagnosis result. The reconstruction model adopts a bidirectional long short-term memory network autoencoder with residual connection to perform reconstruction error calculation and anomaly degree quantification.

[0036] Step S104: Input the above appearance image into the pre-trained recognition model and output the defect recognition result. The above recognition model adopts an end-to-end architecture that removes non-maximum suppression.

[0037] Step S105: After completing the single-point detection based on the above steps, retain the internal equipment connection lines of the bearing platform, remove only the external high-voltage lead and grounding wire, and transfer the bearing platform as a whole for multi-point detection operations.

[0038] To address the current challenges in substation zinc oxide surge arrester testing, such as low equipment integration, inadequate safety protection systems, large data errors in strong electromagnetic environments, weak early latent fault identification capabilities, and low efficiency in inspection and relocation, this invention creates a hardware-integrated protection system, an interference-adaptive suppression, and a multi-dimensional coupling detection method for accurate identification. This breaks through the technical bottlenecks of traditional surge arrester testing, which involve the separation of electrical parameter acquisition and visual defect identification, high reliance on manual labor, poor operational standardization, and difficulty in balancing testing accuracy and safety.

[0039] In terms of hardware integration, an integrated mobile high-efficiency platform, an intrinsically safe multi-functional power switch, and a Bluetooth intelligent DC microammeter system are used to build a field operation carrier, realizing the integrated storage and fixed deployment of a complete set of surge arrester testing instruments and auxiliary tools. The platform is based on the theoretical derivation of the air breakdown field strength formula, sets standardized test sample safety distances, and is equipped with a power switch that integrates a visually clear disconnection point of the disconnect switch, leakage protection, overcurrent protection, and real-time digital voltage monitoring—a four-fold protection mechanism that conforms to power safety work procedures and eliminates electrical safety hazards from the power supply head. It is equipped with a high-voltage anti-interference circuit and a stable Bluetooth transmission architecture, solving the defects of traditional wireless detection data such as easy interruption, packet loss, and large errors, ensuring the stability and reliability of detection data transmission.

[0040] At the algorithmic intelligent detection level, multiple core intelligent algorithms are integrated to construct a comprehensive and accurate diagnostic system. An improved adaptive harmonic suppression algorithm is employed to dynamically remove interfering harmonic components and accurately restore the true resistive leakage current data of the surge arrester. Simultaneously, based on the current coupling characteristics of series surge arresters, a high-precision differential calculation is independently derived, and an improved LSTM time-series feature reconstruction model is introduced to encode, reconstruct, and analyze errors in continuously acquired leakage current time-series data. This deeply mines subtle time-series feature distortions caused by moisture and aging in the surge arrester, enabling early prediction of equipment faults. Finally, defect detection is achieved using the YOLO26 end-to-end intelligent image model, equipped with a MuSGD optimizer and STAL feature enhancement strategy. This allows for accurate identification of various defects in surge arresters under complex substation backgrounds, including cracks, damage, contamination, and aging. Automatic defect classification, location, and severity determination are performed, improving the efficiency of batch inspections of multiple devices and large-area substations.

[0041] In summary, the method provided by the embodiments of this invention is applicable to routine preventive testing, condition monitoring, and fault diagnosis of zinc oxide surge arresters in substations. It can leverage an integrated mobile platform, intrinsically safe power switches, and Bluetooth smart micro-ammeter systems to construct a standardized, integrated, and intelligent surge arrester testing system. Through integrated equipment layout, multiple safety protections, precise Bluetooth wireless transmission, and automatic current calculation, it solves the problems of traditional testing equipment being scattered, having cumbersome wiring, being difficult to move, posing numerous safety hazards, having large data transmission errors, and having low efficiency in manual calculations. This enables rapid deployment, safe operation, intelligent data acquisition, accurate calculation, and convenient relocation for substation surge arrester testing, improving testing efficiency and operational safety, and adapting to various complex substation sites and batch testing scenarios involving a large number of surge arresters.

[0042] As an alternative approach, the mathematical model for the original leakage current signal is as follows: ; In the formula, This represents the original leakage current signal. This represents the actual leakage current signal. Indicates harmonic interference components. This represents random electromagnetic white noise components.

[0043] Specifically, to quantitatively distinguish between the effective leakage component and various types of interference noise, a mathematical model of the superposition of the original leakage current collected by the microammeter at the substation is first established. This model accurately characterizes the sources, distribution patterns, and waveform characteristics of various interferences, providing a theoretical basis for the hierarchical interference suppression algorithm. The microammeter collects and outputs the original current signal in real time. It is a linear superposition of three types of signals: the actual leakage current of the surge arrester, the harmonic interference of the power system, and the random electromagnetic noise in space. The mathematical model of the original leakage current signal is the basic signal superposition model.

[0044] Original leakage current signal This refers to the real-time raw current collected by the microammeter, in μA, which is the unprocessed raw time-domain sampling sequence after analog-to-digital conversion.

[0045] Real leakage current signal The effective component of the actual leakage current of the surge arrester, measured in μA, includes the resistive leakage fundamental frequency and the weak nonlinear distortion component caused by moisture / aging. It is the core effective signal for fault diagnosis.

[0046] Harmonic interference components This refers to the harmonic interference components caused by phase-to-phase coupling of the three-phase bus, frequency conversion equipment in the station, and disturbances from reactive power compensation devices. The unit is μA, which are concentrated in the low-frequency band of the power frequency and overlap with the spectrum of the effective leakage signal.

[0047] Random electromagnetic white noise components This refers to the random electromagnetic white noise component caused by substation spatial corona, switching operations, and wireless communication, measured in μA, primarily consisting of high-frequency pulse noise.

[0048] For harmonic interference components Based on the power harmonic distribution standards of distribution networks and substations, the coupling harmonics within the substation are mainly composed of the fundamental frequency, the 3rd, 5th, and 7th odd harmonics, while the amplitude of even harmonics is extremely low and negligible. A trigonometric series expansion of the harmonic components is performed to accurately quantify the superposition pattern of multiple harmonics. ; In the formula, k For harmonic orders, in field engineering, it is only necessary to identify the 1st, 3rd, 5th and 7th harmonics to cover more than 95% of the coupled interference energy; For the first k Real-time amplitude of subharmonic interference, in μA, dynamically changes with load and bus voltage fluctuations within the station. This is the power frequency angular frequency, and the value is uniformly taken as 314 rad / s for a 50 Hz power frequency system; For the first k The initial phase angle of the subharmonic, in rad, is determined by the degree of capacitive coupling between the three phases and will cause overall phase shift distortion of the leakage current waveform. n To determine the highest harmonic identification order, this embodiment takes... It balances computational efficiency with the integrity of interference suppression.

[0049] The mathematical model clearly distinguishes two types of interference characteristics: random noise is concentrated in the high-frequency band, while harmonic interference and effective signal coexist in the low-frequency band. A layered interference suppression design is employed, using hardware preprocessing to filter out high-frequency random noise. Software frequency domain analysis combined with dynamic iterative thresholding to remove low-frequency harmonics Finally, phase compensation is used to eliminate coupled phase shifts, and the pure effective components are accurately solved. This addresses the technical challenges of traditional algorithms, such as their inability to differentiate between different layers and simultaneously suppress multiple interferences.

[0050] As an optional approach, adaptive harmonic suppression processing is performed on the original leakage current signal to obtain the true leakage current signal. This includes: filtering the original leakage current signal based on a hardware filtering architecture to obtain a preprocessed signal; performing a fast Fourier transform on the preprocessed signal to extract real-time harmonic amplitudes; adaptively iteratively updating the harmonic filtering thresholds for each harmonic based on the historical harmonic statistical mean and the real-time harmonic amplitudes; performing harmonic suppression processing on the original leakage current signal based on the harmonic filtering thresholds to obtain a preliminary corrected time-domain signal; and performing phase distortion compensation correction processing on the preliminary corrected time-domain signal to obtain the true leakage current signal.

[0051] It can specifically eliminate capacitive coupling interference between three phases in substations and improve the accuracy of leakage current detection data.

[0052] As an optional solution, the above hardware filtering architecture includes a multi-stage RC low-pass filter circuit, each stage of which is equipped with a high-frequency ferrite bead. Before filtering the original leakage current signal based on the hardware filtering architecture, the method further includes setting the cutoff frequency of the multi-stage RC low-pass filter circuit.

[0053] Specifically, the microammeter head incorporates a high-precision, high-voltage anti-interference hardware circuit, which first performs primary filtering on the original current signal to remove high-frequency random electromagnetic noise. I noi ( t The hardware filtering architecture employs a multi-stage RC low-pass filter and a ferrite bead shielding composite filtering scheme. Two stages of RC passive low-pass circuits are connected in series before the leakage current signal enters the analog-to-digital sampling chip. Each stage of the circuit is equipped with a high-frequency ferrite bead to suppress pulse spike noise, and a unified hardware cutoff frequency is set. .

[0054] Frequency characteristic design logic: The distortion component of the actual leakage current of the surge arrester and the frequencies of the 1st to 7th harmonics are all ≤350Hz. The hardware cutoff frequency of 100Hz can completely filter out corona pulses above 100Hz, high-frequency noise during switching operations, and radio frequency interference, and will not significantly attenuate the 50Hz power frequency fundamental wave and the core low-frequency components of the 3rd 150Hz harmonic.

[0055] ; The multi-level hardware preprocessing architecture provided in this embodiment can filter out more than 80% of random high-frequency white noise in the field, eliminate sampling data jumps and spectral clutter interference caused by instantaneous impulse noise, reduce the spectral floor noise during subsequent FFT frequency domain analysis, avoid clutter masking weak harmonic amplitude characteristics, and provide high signal-to-noise ratio preprocessed time domain data for accurate identification of harmonic components.

[0056] Considering that the signal after hardware preprocessing only has effective leakage components and low-frequency harmonic interference, and the two time-domain waveforms are completely coupled and cannot be directly separated, this embodiment uses Fast Fourier Transform (FFT) to map the time-domain signal to the frequency domain space, so as to independently extract the amplitude and phase parameters of each frequency harmonic and complete the accurate positioning of interference harmonics.

[0057] For length of N Discrete preprocessed time-domain sequences I pre ( t Perform an FFT transformation; the transformation calculation formula is as follows: ; The system selects a fixed number of sampling points. Balancing spectral resolution with the real-time computing efficiency of the embedded terminal, the 1024 sampling points achieve a spectral resolution of 0.488Hz, accurately distinguishing the 50Hz fundamental wave, the 150Hz 3rd harmonic, the 250Hz 5th harmonic, and the 350Hz 7th harmonic without spectral aliasing or harmonic component consolidation. Furthermore, the 1024-point FFT can be completed in milliseconds by the embedded hardware floating-point unit, meeting the real-time requirements for continuous millisecond-level acquisition of leakage current.

[0058] FFT output frequency domain sequence X ( n It contains complex amplitude information for each frequency point. Taking the modulus of the complex sequence yields the corresponding harmonic amplitude. A k The initial phase of the harmonic is obtained by taking the argument. φ k It can independently separate the four core interference harmonic components: the fundamental frequency, the 3rd, 5th, and 7th harmonics, and output the independent amplitude and phase parameters for each harmonic group. A k, φ k Compared to traditional time-domain moving average filtering and fixed-order digital low-pass filtering, FFT frequency domain identification can accurately distinguish between effective leakage signals and harmonic interference in the same frequency band. It will not misjudge the weak low-frequency distortion components caused by moisture in the surge arrester as harmonics and filter them out, thus solving the engineering problem of effective signal loss.

[0059] As an optional approach, the formula for adaptively iteratively updating the threshold of each harmonic filter is expressed as: ; In the formula, T h ( k ) indicates the first k Real-time dynamic suppression threshold for subharmonics α Represents the iterative weight coefficients. A k_avg Indicates the historical number within the sliding time window k Average amplitude of the second harmonic A k_real This indicates the FFT identification obtained in the current sampling period. k Real-time amplitude of subharmonics.

[0060] Specifically, traditional harmonic suppression technology uses a globally fixed filter threshold, with the threshold parameter calibrated offline based on historical steady-state data from the substation. When load switching occurs, bus voltage fluctuates, or interphase coupling strength changes, the real-time harmonic amplitude... Significant fluctuations may occur: when the harmonic amplitude increases, the fixed threshold filtering is incomplete, and residual harmonic interference causes falsely high leakage current. When the harmonic amplitude decreases, the fixed threshold filtering is excessive, erasing the weak nonlinear leakage characteristics caused by the aging of the surge arrester.

[0061] To address this deficiency, this embodiment provides a dynamic iterative harmonic suppression threshold model, which adaptively iteratively updates the filter threshold for each harmonic based on the historical harmonic statistical mean and the real-time harmonic amplitude. The formula for adaptively iteratively updating the filter threshold for each harmonic is the core iterative calculation formula.

[0062] No. k Real-time dynamic suppression threshold for subharmonics T h ( k The unit of ) is μA.

[0063] Iterative weight coefficients A uniform value of 0.8 is used to give higher weight to the historical harmonic statistical mean, thus avoiding severe threshold fluctuations caused by single instantaneous harmonic abrupt changes.

[0064] Regarding the history within the sliding time window k Average amplitude of subharmonic Ak_avg The sliding window length is set to 1 minute, and the corresponding time is synchronized. Figure 2 The average value of historical abnormal scores within the dynamic threshold calculation section.

[0065] The FFT identification obtained in the current sampling period k Real-time amplitude of subharmonic A k_real ,correspond Figure 2 Real-time input features within the dynamic threshold calculation module.

[0066] Figure 2 The dynamic threshold calculation section has a built-in sensitivity adjustment coefficient and iterative weights. The linkage mechanism allows for online adjustment of thresholds and adaptive response speed based on the substation monitoring accuracy requirements; the module output threshold is fed back to the front-end wavelet energy spectrum feature extraction link, forming a closed-loop optimization mechanism.

[0067] After the dynamic threshold iteration update is completed, for each harmonic amplitude exceeding the threshold T h ( k The interference components are completely stripped, and only the low-frequency steady-state true leakage current base component below the threshold is retained. After harmonic stripping, a preliminary corrected time-domain signal is obtained: ; It can achieve real-time matching between harmonic interference intensity and filtering threshold, and can accurately filter out excessive harmonics under load fluctuation and coupling strength change scenarios, while fully preserving the weak leakage distortion characteristics of the surge arrester itself, taking into account both interference suppression effect and fault diagnosis sensitivity.

[0068] Furthermore, the three-phase busbar and the surge arrester varistor exhibit phase-to-phase capacitive coupling. Coupled harmonics cause overall phase shift and waveform distortion in the leakage current waveform. Even after removing harmonic amplitude interference, the phase shift still leads to errors in the calculation of resistive leakage current, affecting the accuracy of the quantitative assessment of the surge arrester's aging degree. The technical solution provided in this embodiment employs a phase distortion compensation and correction algorithm to eliminate the phase distortion error caused by capacitive coupling.

[0069] Using the 50Hz power frequency voltage sampling waveform within the station as the reference phase source, the phase difference Δ between the preprocessed leakage signal and the reference power frequency signal is calculated in real time. φ The distortion compensation correction amount Δ is dynamically generated based on the phase difference. I ( φ ), superimposed on the initial signal after harmonic stripping I temp ( t The final output is a high-precision, true leakage current signal with no distortion, harmonics, or high-frequency noise. ; Correction amount Δ I ( φ The dynamic assignment of phase difference in real time can offset the overall waveform shift and local distortion caused by three-phase capacitive coupling, solve the problem of long-term monitoring data drift caused by neglecting phase distortion in traditional monitoring schemes, and output a pure and true leakage current sequence to provide the wavelet multi-scale decomposition and bidirectional LSTM autoencoder feature reconstruction link in the attached figure, thus completing the efficient data acquisition and harmonic interference suppression of the whole process.

[0070] The technical solution provided in this embodiment also performs in-depth analysis and early warning diagnosis of leakage current.

[0071] In terms of data acquisition and classification labeling, the high-precision current transformer acquires the simulated leakage current signal of the surge arrester under all operating conditions in real time. After the simulated signal is denoised and amplitude conditioned by a signal conditioning module composed of an amplifier circuit and a first-order low-pass filter circuit, it is converted into a discrete digital time sequence signal by an analog-to-digital converter to obtain the original leakage current time sequence.

[0072] Set sampling frequency f s =2kHz, a single sample contains N =1024 discrete sampling points, covering 10 50Hz power frequency cycles, ensuring complete preservation of the periodic and instantaneous distortion characteristics of the leakage current. The original discrete-time signal of the leakage current is defined as: X ( t )= x 1, x 2,..., x N , t =1,2,..., N ,in, x t For the first t The instantaneous value of leakage current at each sampling moment, in mA; N This represents the total number of sampling points per sample.

[0073] In signal preprocessing and time-frequency feature extraction based on wavelet transform, surge arrester leakage current is a non-stationary time-series signal, superimposed with environmental electromagnetic noise and equipment clutter interference. Directly inputting it into a deep learning model would reduce the accuracy of feature learning. The technical solution provided in this embodiment uses wavelet multi-scale decomposition to complete signal denoising, and combines continuous wavelet transform to extract time-frequency energy features to construct the model input feature sequence.

[0074] First, wavelet multi-scale decomposition and signal denoising affect the original signal. X ( t )conduct J Layer wavelet multiresolution decomposition, the decomposition formula is: ; In the formula, j For wavelet decomposition, this embodiment takes... J =5; d j ( t ) is the first j Layer wavelet detail coefficients correspond to the high-frequency components of the signal, including noise and transient pulse anomaly information; a j ( t ) is the first j The wavelet approximation coefficients correspond to the low-frequency trend components of the signal and reflect the main operating characteristics of the leakage current.

[0075] This embodiment uses the db4 wavelet basis for decomposition, which possesses excellent time-frequency localization characteristics and is suitable for non-stationary leakage current signals containing abrupt changes and distortions. After decomposition, threshold denoising is performed on the high-frequency detail coefficients to filter out random environmental noise and retain the effective high-frequency distortion components caused by the fault, resulting in a denoised, clean leakage current signal. X '( t )).

[0076] Secondly, continuous wavelet transform and time-frequency energy feature extraction are performed on the denoised signal to extract joint time-frequency features. The calculation formula for continuous wavelet transform is as follows: ; In the formula, For the mother wavelet function, this embodiment uses the db4 wavelet. The complex conjugate of the mother wavelet; a The scaling parameter controls the wavelet scaling, which is inversely proportional to the signal frequency. b The translation parameter controls the sliding position of the wavelet in the time domain; CWT ( a , b )) represents the wavelet transform coefficients at the corresponding scale and time-domain location.

[0077] The frequency band energy characteristics are calculated based on wavelet transform coefficients. These energy characteristics can intuitively reflect the energy distribution of leakage current in different frequency bands. Moisture, aging, and breakdown faults in surge arresters will all change the energy proportion of each frequency band. j Formulas for calculating signal energy at each decomposition scale: ; In the formula, CWT j ( k ) is the first j The first scale k Wavelet coefficients for discrete sampling points; E jThis represents the frequency band energy value at the corresponding scale. The energy features at all scales are arranged temporally to construct the time-frequency feature sequence of the model input: E = E 1, E 2,..., E j This sequence integrates the time-domain variation of leakage current with the frequency-domain energy distribution, and serves as input data for the improved LSTM model.

[0078] As an optional approach, before inputting the aforementioned real leakage current signal into the pre-trained reconstruction model, the method further includes: establishing a first initial model based on a bidirectional long short-term memory network encoder, a residual connection layer, and a unidirectional long short-term memory network decoder; and training the first initial model in an unsupervised manner based on samples of normal operating conditions of the surge arrester to obtain the aforementioned reconstruction model. The model loss function includes a mean squared error term, a Kohlbek-Leibler divergence term, and a time series regularization term. This approach enables in-depth mining of the leakage current time series characteristics and accurate identification of early latent faults in the surge arrester.

[0079] As an optional approach, the aforementioned real leakage current signal is input into a pre-trained reconstruction model to obtain a state diagnosis result, including: calculating the reconstruction error corresponding to the aforementioned real leakage current signal based on the aforementioned reconstruction model, which is used to characterize the difference between the aforementioned real leakage current signal and the normal operating condition signal; calculating anomaly score based on Mahalanobis distance and the aforementioned reconstruction error, wherein the aforementioned state diagnosis result includes the aforementioned anomaly score, and the aforementioned anomaly score characterizes the degree of anomaly in the operation state of the surge arrester.

[0080] Specifically, the technical solution provided in this embodiment makes two core improvements on the traditional LSTM autoencoder: replacing the unidirectional LSTM encoder with a dual LSTM to simultaneously capture forward and backward bidirectional temporal dependencies; and introducing residual connections between the encoder and decoder to alleviate the gradient vanishing problem during deep model training. The overall model consists of three parts: a bidirectional LSTM encoder, a residual connection layer, and a unidirectional LSTM decoder, and is trained using an unsupervised method based on normal samples.

[0081] Regarding the basic LSTM unit computation principle, the LSTM unit solves the problem of long-term dependency loss in traditional recurrent neural networks through three gate control structures: input gate, forget gate, and output gate. The state calculation formulas for each unit are as follows: Forget gate: controls the retention ratio of historical unit states, expressed by the formula: ; Input gate: Controls the input ratio of the current time-series features, expressed by the formula: ; Candidate unit state: The temporary feature state at the current time, expressed by the formula: ; Unit state update: merging historical states and current candidate states, expressed by the formula: ; Output gate and hidden state output: Outputs the final time series characteristics at the current time step, expressed by the formula: ; In the formula, e t for t The time-frequency energy characteristics of the input at any given time; W f , W i、 W c、 W o ) represents the weight matrix for each access control system. b f , b i、 b c、 b o Bias vector; σ Here, is the Sigmoid activation function, and tanh is the hyperbolic tangent activation function; This refers to the element-wise multiplication operation of matrices; f t , i t、 o t These represent the output values ​​of the forget gate, input gate, and output gate, respectively. c t This is the state of the LSTM cell. h t This is the hidden state output by the unit.

[0082] In bidirectional LSTM encoders, traditional unidirectional LSTMs can only utilize historical timing information and cannot combine subsequent timing features, resulting in insufficient global feature extraction of the continuous waveform of leakage current. The technical solution provided in this embodiment uses a forward LSTM and a backward LSTM to form a bidirectional encoder, extracting features from the forward and reverse timing directions respectively, and then splicing and fusing the bidirectional hidden states.

[0083] Forward LSTM hidden states: ; Backward LSTM hidden states: ; Final encoded features after bidirectional fusion: ; In the formula, Forward LSTM unit, For backward LSTM units; H t for t High-dimensional temporal features after bidirectional encoding of time.

[0084] Considering that LSTM networks with deeply stacked residual connection layers are prone to gradient vanishing and model degradation, this embodiment adds a residual connection between the encoder output and the decoder input to fuse the original input features with the encoded features. The formula is as follows: ; In the formula, The fused features after residual connection are used as input to the decoder. This is the residual mapping matrix, used to match the dimensions of the original features and the encoded features.

[0085] Regarding the unidirectional LSTM decoder, a single-layer unidirectional LSTM is used, based on coding fusion features. Reconstruct the original time-frequency feature sequence to realize the data reconstruction function of the autoencoder, and reconstruct the computational formula: ; In the formula, for t Energy characteristics after time-reconstruction; LSTMd is the decoder LSTM unit; W d , b d These are the decoder weight matrix and bias vector, respectively; c t-1 This represents the unit state of the decoder at the previous moment.

[0086] Finally, the complete reconstructed feature sequence is obtained: .

[0087] Regarding the model loss function, to ensure reconstruction accuracy, feature distribution regularity, and temporal continuity, this embodiment provides a multi-constraint joint loss function that integrates mean squared error, KL divergence (Kourbeck-Leibler divergence term), and temporal regularization term: ; Mean square error term : To measure the deviation between the original features and the reconstructed features, and to ensure the reconstruction effect.

[0088] KL divergence term Constraining the distribution of encoded features improves the model's generalization ability.

[0089] Temporal regularization terms It forces adjacent timing characteristics to be continuous, adapting to the smooth change characteristics of leakage current.

[0090] These are weighting coefficients used to balance the proportion of various losses. During the model training phase, only normal operating condition samples of the surge arrester are input, and the weights and bias parameters are iteratively optimized until the loss function converges, thus completing model training.

[0091] After the model training is completed, all network parameters are fixed, and the leakage current feature sequence collected in real time is input. The degree of deviation of the sample from the normal pattern is characterized by reconstruction error. Mahalanobis distance is introduced to replace the traditional Euclidean distance to eliminate the interference of correlation between features and accurately quantify the abnormal level.

[0092] Specifically, the reconstruction error is calculated using the L2 norm to determine the reconstruction error between the original features and the reconstructed features of a single sample: ; The larger the reconstruction error value, the greater the difference between the current leakage current timing characteristics and normal operating conditions, and the higher the probability of equipment failure.

[0093] Euclidean distance does not consider the correlation between feature dimensions and is easily affected by data distribution, while Mahalanobis distance relies on the mean and covariance matrix of the reconstruction error of normal samples, thus eliminating dimensional coupling interference. The formula for calculating outlier scores based on Mahalanobis distance is: ; In the formula, SE is the anomaly score of the current sample; μ e This is the mean vector of reconstruction errors for all normal samples during the training phase; The covariance matrix of the reconstruction error of normal samples. It is its inverse matrix.

[0094] The anomaly score (SE) directly reflects the degree of abnormality in the operation of the surge arrester; the higher the score, the more severe the fault.

[0095] As an optional approach, before inputting the aforementioned appearance image into the pre-trained recognition model, the method further includes: establishing a second initial model based on a backbone feature extraction network, a multi-scale bidirectional feature fusion network, and a detection-segmentation dual-branch output head; training the second initial model based on a frozen pre-training stage and a full-parameter thawing fine-tuning stage to obtain the aforementioned recognition model. This enables high-precision, low-latency, and strong anti-interference intelligent recognition of various types of defects in surge arresters.

[0096] In the aforementioned frozen pre-training phase, pre-trained weights from a general dataset are loaded, the parameters of the backbone feature extraction network are frozen, and only the multi-scale bidirectional feature fusion network and the detection-segmentation dual-branch output heads are trained. In the aforementioned full-parameter unfreezing fine-tuning phase, all network weights are unlocked, and a cosine annealing dynamic learning rate decay strategy is used for training. The formula for adaptive decay of the learning rate with each iteration is expressed as: ; In the formula, lr epoch The learning rate represents the current iteration round. lr min This represents the minimum threshold learning rate. lr init This represents the initial maximum learning rate. epoch Indicates the current iteration round number. E total This indicates the total number of iterations.

[0097] For example, YOLO26 is used for intelligent and accurate defect identification. YOLO26 is a lightweight instance segmentation integrated model customized and optimized for deployment at the power edge. It continues the single-stage end-to-end detection paradigm of the YOLO series, removes redundant and complex post-processing logic, and achieves an optimal trade-off between model parameter size, inference speed, and small target accuracy. The overall architecture is divided into three core layers: the Backbone feature extraction network, the Neck multi-scale bidirectional feature fusion network, and the Head detection-segmentation dual-branch output head. The complete structural logic is as follows.

[0098] The backbone network employs a lightweight convolution stacked set adaptive channel attention embedding design. The YOLO26 backbone uses a lightweight basic module with depthwise separable convolution stacks to construct a stepped downsampling structure, extracting shallow texture details and deep global semantic features from the surge arrester image step by step. An adaptive channel attention (ACAM) module is embedded at the end of each downsampling module to autonomously and dynamically allocate feature channel weights, strengthening weak feature responses such as cracks, minor damage, and local electrolytic erosion spots, and suppressing invalid background feature weights such as substation steel frames, weeds, adjacent power equipment, and lighting noise, thereby reducing the probability of false background detection from the source.

[0099] Suppose that the dimension of the output feature map of a certain convolutional layer is... ,in, C The total number of channels. H , W These represent the height and width of the feature map, respectively. The derivation of the Adaptive Channel Attention (ACAM) calculation formula is as follows.

[0100] Global spatial mean pooling compression is used to aggregate single-channel features globally, resulting in a one-dimensional representation vector for each channel. ; In the formula, Z c For the first c Channel global mean representation, F c ( i,j ) is the first c Channel coordinates ( i , j (Position pixel value)

[0101] Dual-channel nonlinear mapping learning uses two lightweight 1×1 convolutional layers to perform channel compression and expansion, and introduces Sigmoid activation to generate adaptive weight coefficients for each channel. w c : ; In the formula, W 1 represents the channel compression convolution weight. W 2 represents the channel-dilated convolution weights; It is a ReLU nonlinear activation function; Using the Sigmoid activation function, with constrained weight range Feature adaptive recalibration involves multiplying the original features with the corresponding channel weights channel by channel to achieve feature enhancement and noise reduction.

[0102] The backbone network employs a 4-level stepped downsampling scheme, outputting feature maps at four different scales. This adapts to four vastly different arrester defect scales: ultra-large area contamination, medium-sized skirt damage, fine surface cracks, and micron-level electrolytic erosion spots. This addresses the technical challenge of poor adaptability of single-scale models to defects of varying sizes. Furthermore, the entire backbone network replaces traditional standard convolutions with depthwise separable convolutions, significantly reducing the computational cost of standard convolutions. ; Total computational cost of depthwise separable convolution: ; Operational compression ratio: ; In the convolution kernel K In scenario 3, the theoretical computational cost of convolution is reduced by more than 75%, significantly reducing the computational overhead of the model and meeting the real-time inference deployment requirements of mobile devices and portable inspection terminals.

[0103] The Neck module employs a bidirectional, multi-scale feature fusion structure. Addressing the issue that shallow features of surge arresters include crack edges, damaged outlines, and subtle texture details, while deep features contain semantic information about contaminated areas, overall aging, and large-area defects, directly separating shallow and deep features can cause a sharp drop in accuracy for small defect recognition. The YOLO26 Neck module utilizes a PANet bidirectional upsampling and downsampling fusion architecture, fusing features from four layers of the backbone network at multiple scales to construct a bidirectional feature flow path: top-down semantic enhancement and bottom-up detail supplementation. The top-down path uses upsampling to progressively match the dimensions of deep, high-semantic features to shallow features, adding and fusing them element-by-element to supplement global discriminative semantics for shallow, small-defect features and mitigate texture misjudgments caused by illumination distortion. The bottom-up path uses convolutional downsampling to compress redundant shallow details and supplement contour boundary information upwards, improving the completeness of segmentation for large-scale contamination and missing skirts.

[0104] Let the two layers of features involved in the fusion be the deep upsampling features. F up Shallow homology characteristics F low The fused feature expression is: ; This represents a pixel-by-pixel matrix addition and fusion operation.

[0105] To suppress channel and feature redundancy during the fusion process, a lightweight bottleneck layer is added after each fusion to perform feature filtering and compression, eliminating invalid fusion noise and avoiding gradient vanishing and slow model convergence caused by multi-scale superposition. Compared with YOLO's native unidirectional FPN structure, this bidirectional fusion structure improves the recall rate of detecting 0.5~3mm micro-surface cracks by more than 12%, effectively solving the common problem of missed detection of small target defects in complex substation scenarios.

[0106] Regarding the dual-branch parallel output head, the bounding box regression and classification prediction detection branches complete target location regression and defect multi-classification based on the anchor frame mechanism, and preset five types of typical arrester defect labels: surface cracks, skirt damage, surface contamination, insulation aging, and localized electrical erosion defects. Three sets of prior anchor frames are preset for each grid, and three sets of prediction vectors are output: target confidence. conf Category probability P class Bounding box offset t x , t y , t w , t h Boundary box correction calculation formula: ; In the formula, c x , c y The coordinates of the top left corner of the grid; p w , p h Preset anchor frame width and height; b x , b y , b w , b h These are the corrected parameters for the actual bounding box.

[0107] Total loss function for detection branches L det Loss from coordinate regression L box Classification loss L cls Confidence loss L conf Weighted composition: ; The CIoU loss (Complete Intersection over Union loss) is used to replace the traditional MSE coordinate loss, and constraints on overlapping area, center distance, and aspect ratio are introduced to improve the convergence speed and positioning accuracy of the bounding box. The classification uses cross-entropy loss to suppress the class imbalance problem and solve the training bias problem caused by the fact that the number of dirty samples in the field is much greater than that of crack and defect samples.

[0108] Then, the segmentation branch is multiplied by the detection branch fusion feature output using a pixel-level instance mask. This feature is then restored to the original image resolution through continuous deconvolution upsampling. Pixel-by-pixel binary classification is used to determine whether a pixel belongs to a defect region, and a single-channel mask matrix is ​​output. The segmentation loss expression is as follows: ; In the formula, L mask The overall loss function for instance splitting branches. For pixel-by-pixel binary classification, cross-entropy loss, p i To predict pixel probability, q i The actual pixel value of the label. To minimize the smoothing coefficient and prevent the denominator from being zero, the mask output allows for precise calculation of the total area, maximum length, and average width of defect pixels, enabling quantitative classification of defects and overcoming the limitation of traditional target detection methods that can only select boxes and cannot provide quantitative evaluation.

[0109] A two-stage transfer learning training model is adopted, which greatly reduces the difficulty of training with small samples.

[0110] During the pre-training freeze phase, pre-trained weights from the COCO general large dataset are loaded, the Backbone network parameters are frozen, and only the Neck and Head branches are trained. Initial learning rates are set... l r =1 10 -3 The process involves 120 iterations to quickly adapt to the distribution of defects in surge arresters.

[0111] During the full parameter unfreezing and fine-tuning phase, all network weights are unlocked, and a cosine annealing dynamic learning rate decay strategy is adopted, with the learning rate adaptively decaying with each iteration round: ; In the formula, lr init The initial maximum learning rate, lr min The minimum threshold learning rate, epoch This represents the current iteration round number. E total The total number of iterations is set to 240 to prevent oscillations and divergence in the later stages and to improve the upper limit of model accuracy.

[0112] Regarding the defect classification and fusion diagnosis, relying on the pixel-level mask data output by YOLO26, combined with camera calibration and ranging parameters to convert physical dimensions, the geometric parameters of various defects are automatically calculated: maximum defect length L, defect projected area S, and defect percentage of umbrella skirt surface area R.

[0113] Based on the power equipment maintenance regulations, thresholds for judging defects of three levels are set: general, major, and emergency.

[0114] General defects: minor shallow cracks, slight dirt accumulation, parameters not exceeding limits, do not affect insulation operation, only for record-keeping and tracking; Major defects: crack depth extension, moderate contamination accumulation, localized damage to the umbrella skirt, and a slow deterioration trend in leakage current. A short-term maintenance plan should be developed.

[0115] Emergency defects: through-cracks, large-area defects, severe electrolytic aging, high risk of insulation failure, immediately arrange power outage and replacement.

[0116] Quantitative calculation formula: ; In the formula, S defect Measured area of ​​the defect ，S umbrella To correspond to the reference surface area of ​​the umbrella skirt, the defect level is automatically quantified and output, thus eliminating the drawbacks of manual subjective classification.

[0117] When YOLO26 detects obvious external cracks, damage, or severe contamination, and the improved bidirectional LSTM model determines that the leakage current reconstruction error exceeds the standard and there is a trend of moisture aging, the root cause of the surge arrester insulation degradation is determined to be surface current distortion induced by external external damage. If there are no obvious visible defects in the appearance inspection, but the leakage current timing characteristics are continuously abnormal and the Mahalanobis distance exceeds the limit, it is determined that there is latent moisture or internal aging, and the fault is determined to be internal latent defects such as valve plate moisture or internal seal failure. If the appearance defects are minor and all leakage current indicators are within the normal range, the equipment is determined to be in good condition and the regular inspection cycle can continue. By complementing each other with electrical and visual information, the shortcomings of simple electrical inspection cannot be located in external causes, and the limitations of single visual recognition cannot be judged in terms of internal insulation status, which greatly improves the comprehensiveness and reliability of the diagnostic conclusions.

[0118] like Figure 3 As shown in the figure, the present invention also provides an integrated testing device for surge arresters, including a data acquisition module 31, a harmonic suppression module 32, a diagnostic module 33, an identification module 34, and a support platform.

[0119] The data acquisition module 31 is used to acquire the original leakage current signal and appearance image of the surge arrester.

[0120] The harmonic suppression module 32 is used to perform adaptive harmonic suppression processing on the original leakage current signal to obtain the real leakage current signal.

[0121] The diagnostic module 33 is used to input the above-mentioned real leakage current signal into the pre-trained reconstruction model to obtain the state diagnosis result. The above-mentioned reconstruction model adopts a bidirectional long short-term memory network encoder with residual connection to perform reconstruction error calculation and anomaly degree quantification.

[0122] The recognition module 34 is used to input the above appearance image into a pre-trained recognition model and output the defect recognition result. The above recognition model adopts an end-to-end architecture that removes non-maximum suppression.

[0123] The platform integrates all modules and testing equipment. After completing a single-point test, it retains the internal equipment connection lines, removes only the external high-voltage leads and grounding wires, and transfers the entire system to perform multi-point testing operations.

[0124] As an optional solution, the aforementioned data acquisition module includes an image acquisition unit and a DC microammeter. The image acquisition unit is used to acquire an image of the surge arrester's appearance. The DC microammeter includes a connected meter head and a handheld meter. The meter head has an arc-shaped outer shell and incorporates anti-interference circuitry and a Bluetooth transmission unit. The handheld meter incorporates a current calculation unit. The DC microammeter is used to acquire the surge arrester's original leakage current signal.

[0125] For example, such as Figure 4 As shown, the aforementioned support platform is a mobile, high-efficiency platform with overall dimensions of 1000mm in length, 800mm in width, and 900mm in handrail height. It features four 150mm diameter high-polymer rubber wheels at the four corners of the bottom, with directional front wheels and braked, steerable rear wheels, adaptable to complex substation terrain. The platform integrates nine functional zones, including three sets of 10kV test sample placement holes spaced 200mm apart, a dedicated grounding electrode, a DC high-voltage generator slot, an insulating rod placement area, a control console placement area, and a miscellaneous items placement area. 100mm protective railings are installed on both sides, allowing for integrated, secure storage of all testing instruments and preventing equipment tipping and tool drops. The platform is equipped with foldable hinged handles, allowing for bidirectional positioning and balancing loading / unloading with high-voltage safety distance protection.

[0126] The platform integrates an intrinsically safe multi-functional power switch. The switch is made of flame-retardant engineering plastic shell and measures 25.5cm × 4.6cm × 8.6cm. Internally, it contains a series disconnect switch, a residual current device (RCD), and an overcurrent protection module, and a parallel digital voltmeter, forming a four-fold safety protection system of visible disconnect switch opening, RCD protection, overcurrent protection, and real-time voltage monitoring. The disconnect switch can form a clear electrical disconnect point, which meets the requirements of power safety regulations and avoids electrical safety risks from the power supply head.

[0127] The matching Bluetooth smart DC microammeter system consists of a stainless steel meter head and a handheld meter. The meter head is equipped with a 6mm / 8mm dual-specification interface, a built-in 9V rechargeable battery and a high-precision high-voltage anti-interference circuit. The arc-shaped shell eliminates corona interference and transmits data wirelessly via Bluetooth, solving the problems of easy interruption and unstable signal in traditional wireless transmission of meters.

[0128] Specifically, the platform architecture design is the core foundation for ensuring the stability, load-bearing capacity, and adaptability of the mobile test platform for surge arresters. This architecture design is based entirely on the on-site test conditions of substations. Through extensive field research, equipment size measurement, workflow streamlining, and mechanical load verification, the overall architecture has been finalized and optimized. The overall architecture adopts an integrated plate-type load-bearing structure, taking into account structural strength, ease of movement, and functional integration.

[0129] In terms of overall size design, considering the overall footprint of the complete set of 10kV surge arrester testing equipment, the minimum width of the substation working passage, and the vehicle transport clearance standards, the final platform dimensions were determined to be 1000mm long and 800mm wide. Compared with traditional simple test carts, this size increases the load-bearing area by more than 40%, allowing the complete set of testing equipment to be stored at once without disassembly and transport. At the same time, the 900mm handrail height conforms to ergonomic design, adapting to the normal height of workers standing, pushing, and operating the equipment, effectively reducing physical fatigue during long-term work and improving work comfort.

[0130] The platform's support and mobility structure employs a four-corner wheel layout. The bottom load-bearing frame is welded from thickened galvanized steel pipes, ensuring strong overall load-bearing capacity, resistance to deformation, corrosion, and oxidation. It can withstand harsh outdoor operating environments, including exposure to sun and rain, and alternating high and low temperatures. Four wheels, uniformly 150mm in diameter, are evenly distributed at the four corners of the platform's bottom. This wheel diameter has undergone force testing and can easily traverse substation ground gaps, minor bumps, and gravel obstacles, avoiding the problems of small wheels causing jamming and excessive bumps. The wheels are made of high-polymer rubber, which, compared to ordinary rubber wheels, offers moderate hardness, wear resistance, pressure resistance, excellent insulation, and quiet vibration reduction. It does not produce harsh noise during operation and avoids the high-voltage safety hazards caused by conductive metal wheels, fully meeting the insulation safety requirements of high-voltage testing.

[0131] In terms of architectural and functional design, the platform features a flat, integrated structure with no hollow or weak areas. Its overall load-bearing capacity reaches 300kg, fully supporting the weight of a complete set of equipment, including a DC high-voltage generator, control console, test materials, and surge arresters under test. The platform has pre-installed mounting slots for protective railings around its perimeter, and 100mm high protective railings are integrated on both sides. The railing height has been optimized through multiple tests to effectively prevent small items such as insulated rods, screws, insulating pads, and test lines from falling, without obstructing equipment operating space or interfering with test wiring operations. The platform also has pre-installed mounting positions for handrails and grounding electrodes at the rear, and pre-installed wheel mounting bases and slots at the bottom. All mounting structures utilize an integrated pre-embedded design, avoiding later drilling and modification that could damage the platform's structural strength and ensuring the overall stability and integrity of the architecture.

[0132] Meanwhile, the architecture design fully considers storage and safety adaptation requirements. A hinged handle mounting base is installed at the rear of the platform. The base is reinforced with thickened steel plates, capable of withstanding repeated folding and pushing / pulling mechanical wear, ensuring a sturdy and stable structure. The overall architecture layout is neat and the center of gravity is centrally located, ensuring overall stability after equipment placement and preventing issues such as unilateral weight distribution, tilting, or tipping. This structural design guarantees the stability and safety of testing operations, making it perfectly adaptable to various testing scenarios, including indoor and outdoor substations, rooftops, and near switchgear.

[0133] Regarding the main functional components of the platform, based on the standardized operating procedures for 10kV surge arrester field testing, and combined with the structural form, function, and operating habits of the testing equipment and auxiliary materials, each component performs its own function and cooperates with each other to form an integrated, standardized, and streamlined testing operation system. This solves the problems of traditional testing equipment being cluttered, inconvenient to access, easily damaged, and disorganized in operation. Each functional component is customized according to the actual size of the equipment and the operational requirements, making it highly adaptable.

[0134] 10kV Test Specimen Placement Assembly: A custom-designed placement structure specifically for 10kV distribution surge arrester testing. It directly matches the specifications of the bottom fixing bolts on 10kV surge arresters, eliminating the need for temporary support structures during testing. The surge arrester under test can be directly inserted into the bolt holes for secure installation, quickly placing the test specimen and shortening test preparation time. Three sets of bolt holes are evenly arranged in a straight line, with the center-to-center distance between adjacent bolt holes strictly set at 200mm, ensuring the safety and stability of simultaneous testing of multiple test specimens.

[0135] Platform-specific grounding electrode assembly: Two sets of dedicated grounding electrodes are symmetrically arranged on both sides of the platform. The grounding electrodes are made of galvanized copper, which has excellent conductivity, corrosion resistance, and stable contact, and can be quickly connected with dedicated grounding wires. During testing, the entire platform is reliably grounded through the grounding electrodes, realizing zero potential on the platform and eliminating the risk of electric shock and discharge caused by static electricity accumulation and floating potential during high-voltage testing. It is a core component to ensure the inherent safety of high-voltage testing.

[0136] Dedicated slot assembly for DC high-voltage generator: Conventional test sites often have uneven surfaces, gravel accumulation, and slope deviations. Relying on the equipment's built-in rotatable and extendable feet for support can easily lead to tilting, swaying, and tipping, causing equipment damage and high-voltage leakage risks. This design utilizes a dedicated embedded slot structure. The slot dimensions precisely match the high-voltage generator base, enabling complete locking and positioning of the equipment. Pre-drilled bolt holes allow for rigid fixation of the equipment to the platform, resolving issues of slippage, tipping, and swaying, and ensuring stable operation of the high-voltage generator.

[0137] Insulating rod placement area component: Designed with a long, dedicated storage slot, the slot adopts a closed limiting design, which can neatly store and fix various insulating rods, avoiding problems such as insulation layer wear, bending deformation, and staining caused by random placement of insulating rods, ensuring that the insulation performance of insulating tools meets the standards, while facilitating quick access during operation and return to the correct position after use, thus standardizing the operation process.

[0138] Console Placement Area Component: A custom-designed, sunken placement area with precisely sized dimensions to fit the console body, allowing for stable, embedded placement and preventing the console from sliding or shifting during operation. The placement area includes space for operation windows and wiring, ensuring uninterrupted button operation, data viewing, and wiring connections, while also providing protection and stability to prevent damage from external impacts.

[0139] Foldable hinged handle assembly: Utilizing a high-strength metal hinge structure, it achieves bidirectional folding and positioning locking functions, allowing for flexible switching of states depending on the work scenario. During transport and loading, it folds forward for storage, fitting snugly against the platform surface and minimizing overall space occupation, making it suitable for loading various maintenance vehicles. During high-voltage testing, it folds backward to increase the safe operating distance, avoiding the risk of discharge from high-voltage equipment to the platform and operators, balancing convenient storage with operational safety.

[0140] Intrinsically safe power switch assembly: It has four functions: knife switch on / off, leakage protection, overcurrent protection, and voltage monitoring. It provides a safe, stable, and compliant power supply guarantee for the entire set of test equipment, avoids electrical safety risks from the power source, and is compatible with the stringent power safety regulations for high-voltage testing.

[0141] Bluetooth Smart DC Microammeter System Components: As the core component for test data acquisition, it solves the problems of unstable signals, easy data loss, and accuracy interference of traditional metering equipment through anti-interference structural design and wireless Bluetooth transmission technology, ensuring accurate, real-time, and effective test data and improving the quality of test detection.

[0142] Regarding the dedicated slots on the support platform, a design with a 200mm spacing between adjacent slots has been verified through high-voltage electric field simulation tests. This design effectively resolves high-voltage crosstalk and arc discharge issues between adjacent samples during testing, ensuring both operational efficiency for simultaneous testing of multiple samples and robust high-voltage safety measures through structural dimensions. The slots feature an embedded, recessed design, ensuring a stable center of gravity for the sample after it is fixed, preventing tilting or swaying. This guarantees sample stability during testing and avoids data deviations caused by sample displacement.

[0143] The dedicated slot for DC high voltage generators adopts a fully enclosed limiting slot structure design. The slot has reserved special bolt fixing holes, and the equipment can be horizontally fixed through the slot without manual leveling, reducing the difficulty of operation preparation, and avoiding safety accidents such as instrument damage and high voltage leakage caused by equipment tipping over.

[0144] The dedicated slot for insulating rods adopts a long, enclosed limiting slot design. The slot specifications are customized according to the length and diameter parameters of commonly used insulating rods, discharge rods, and operating rods in the test. Multiple limiting buckles are set in the slot, which can fix multiple insulating rods in layers and categories. The slot opening fits the shape of the rod, while ensuring the standardization of the test operation.

[0145] The dedicated slot for the control console adopts a sunken integrated design. The depth, length and width of the slot are adapted to the body of the test control console. The embedded placement can lower the center of gravity of the control console. The edges of the slot are rounded and polished, while providing sufficient space for wiring and operation. The staff can directly complete the control console parameter setting, data reading and wiring connection on the platform without the need for hand-held or elevated operation, which improves the convenience of operation.

[0146] The miscellaneous items storage slots adopt a partitioned open structure, dividing the storage areas into different sizes to accommodate small auxiliary materials of different specifications such as screws, terminals, insulating pads, data cables, and spare parts. This enables classified storage and easy access, solving the problems of messy stacking, loss, and cumbersome retrieval of small test parts in traditional testing, improving work efficiency, and maintaining a standardized and tidy testing site.

[0147] Regarding the safety of the platform, a 100mm protective fence forms a closed-loop protection. The equipment's dedicated slots and grooves are all designed with limiting and fixing features, allowing all testing instruments and samples to be fixed at specific points. The high-polymer rubber insulated wheel sets have excellent insulation performance, which can isolate potential ground conductivity hazards.

[0148] In terms of high-voltage safety, the scientifically designed 200mm test specimen spacing eliminates the problems of high-voltage discharge and crosstalk between adjacent surge arresters. The foldable handle with bidirectional positioning design unfolds backward during testing, effectively increasing the safe distance between high-voltage equipment and operators / platform body. Dedicated grounding electrodes on both sides of the platform ensure reliable three-point grounding of the equipment, platform, and earth, quickly releasing static electricity and residual voltage accumulated during the test. Simultaneously, dedicated slots on the insulating rods effectively protect the integrity of insulating tools, ensuring that insulation protection performance meets standards.

[0149] In terms of electrical safety, the integrated intrinsically safe multi-functional power switch constructs a four-fold electrical protection system. The physical knife switch can form a clear electrical disconnect point, allowing for direct power cut-off during operation, maintenance, wiring, and debugging, eliminating the risk of hidden live current. The leakage current protector can cut off power in milliseconds in case of circuit leakage, line damage, or equipment failure. Real-time digital voltage monitoring allows staff to keep track of the power supply status in real time, promptly detect voltage anomalies and power supply faults, and avoid potential safety hazards in advance.

[0150] In terms of operational safety, the platform is highly ergonomically designed to reduce operator fatigue and minimize the risk of human error; the rear wheel brake locking function can fix the platform's position; the split-type Bluetooth micro-ammeter enables wireless remote data reading, allowing workers to operate away from high-voltage equipment, further increasing the safety distance and reducing the risk of close-range high-voltage operations.

[0151] In terms of data security, the Bluetooth smart DC microammeter is equipped with a high-precision anti-interference circuit, which can shield the signal interference from high-voltage electric fields and on-site electromagnetic equipment, avoid data fluctuations, distortions, and disconnections, ensure that test data is accurate, effective, and traceable, and eliminate test misjudgments and equipment hazard omissions caused by data anomalies.

[0152] Regarding the platform's ease of use, it employs a differentiated wheel layout, with the front wheels being directional and the rear wheels featuring omnidirectional brakes and steering. Equipped with 150mm large-diameter high-polymer rubber wheels, it offers low resistance, flexible steering, and excellent maneuverability, allowing a single person to easily push, steer, and position the equipment without the need for multiple people to move it. Once in position, the equipment can be quickly locked and secured using the rear wheel foot brake.

[0153] In terms of storage and transportation, the foldable hinged handle supports bidirectional folding and switching, eliminating the need to disassemble instruments and equipment for overall transportation. The platform's nine functional zones enable integrated storage of all test equipment. All equipment, tools, and auxiliary materials have their designated places, ensuring neat storage and easy access. After the test, equipment can be quickly returned to its place and the site can be tidied up, significantly shortening the time required to complete the work.

[0154] In terms of test operation, the dedicated slots all adopt a quick-fix structure that requires no debugging. 10kV surge arresters can be directly plugged in and fixed, and DC high-voltage generators can be directly bolted in, reducing test preparation time by more than 60%. The platform layout conforms to the test operation process, with smooth workflow. Workers can stand and complete the entire process, including sample fixing, equipment startup, parameter adjustment, data reading, and line connection, without bending over, tiptoeing, or repeatedly moving equipment, making operation comfortable, efficient, and convenient. The Bluetooth wireless micro-ammeter system eliminates the constraints of wired lines, eliminating the need for complicated wiring. It allows for remote handheld data reading, offering flexible operation while avoiding the problems of tangled, damaged, or obstructed working space caused by wired lines.

[0155] In summary, the technical solution provided in this embodiment adopts hardware-integrated mobile safety operation technology, which integrates dedicated slot layout, standardized safety distance, quadruple power protection, and full-area grounding system to achieve integrated operation of equipment deployment, safety protection, and rapid relocation; the entire system adopts wire-free rapid relocation technology to achieve overall relocation with test line, improving the efficiency of batch inspection operations.

[0156] like Figure 5a and Figure 5b As shown, this embodiment also provides a multi-functional power switch to address the issue that power conversion equipment used in field testing often relies on plugging and unplugging plugs to disconnect the power, lacking a fixed and visible physical disconnection point. This fails to meet the core requirements of power safety regulations, and frequent plugging and unplugging can cause plug wear, poor contact, prolonged operation time, and significant safety hazards. For example, H1 is 4.5cm, H2 is 8.6cm, L is 25.5cm, and W is 4.6cm, with uniform overall dimensions.

[0157] The main circuit of the gate panel integrates a visualized single-throw disconnector structure, abandoning the traditional plug-and-play power-off mode. The mechanical opening and closing action of the disconnector creates a compliant and clearly visible physical disconnection point, making the disconnector's status intuitive and easy to understand, in line with safety regulations. This avoids the repeated plugging and unplugging of power plugs, preventing secondary safety issues such as plug wear, loose connections, and leakage, thus improving the efficiency and safety of power testing operations.

[0158] By designing a multi-linked safety protection architecture, the leakage current protector and overcurrent protection module are integrated into the main circuit of the socket and connected in series for linkage protection. This allows for rapid disconnection of the main circuit power supply in the event of leakage, overload, or overcurrent faults. Simultaneously, a digital voltmeter connected in parallel with the main circuit enables real-time dynamic monitoring of power parameters, promptly identifying abnormal operating conditions such as high or low voltage, and voltage fluctuations.

[0159] To address the issues of scattered component layout, messy wiring, low integration, bulky size, and unstable installation in current multi-functional power switches and test power supply equipment, a modular integrated circuit layout design is adopted. This design integrates core safety components such as disconnect switches, leakage current protectors, and overcurrent protection modules in series on the main circuit. Voltmeters are connected in parallel for independent monitoring, and multiple power outlets are uniformly connected to the main circuit, forming a standardized and integrated internal circuit structure. This layout effectively reduces the overall size of the equipment, avoids cross-circuit interference and short-circuit faults, improves equipment operational stability, and facilitates later maintenance and troubleshooting, thereby reducing equipment operation and maintenance costs.

[0160] Addressing the challenges of dusty, complex, and frequently moving electrical equipment at power construction sites, coupled with stringent fire safety requirements, traditional power strips often utilize ordinary plastic shells. These shells offer poor flame retardancy and impact resistance, are prone to aging and damage, and lack a fixed mounting structure, making them susceptible to displacement and falling during operation, thus limiting their practicality. In contrast, this new device features a high-strength, flame-retardant engineering plastic unibody shell, offering excellent flame retardancy, insulation, aging resistance, and impact resistance, making it suitable for various complex construction site environments. Furthermore, its compact and lightweight design, with regular overall dimensions, small size, and light weight, allows for direct and stable placement on the work platform without additional fixing structures, balancing portability and stability. This makes it suitable for various scenarios such as mobile testing, temporary power supply, and equipment debugging at power sites.

[0161] The DC high-voltage generator is started to output the test voltage, and the microammeter head collects the total leakage current of the surge arrester in real time. Leakage current in the next section Electromagnetic noise is filtered through a high-voltage anti-interference circuit; simultaneously, an improved harmonic suppression method is used to dynamically eliminate capacitive coupling interference between three phases in the substation, stripping away harmonic distortion components to restore accurate leakage current data. High-precision data is then stably transmitted to a handheld terminal via Bluetooth. Utilizing proprietary intelligent current difference calculation technology, the leakage current between the upper and lower sections of the surge arrester is automatically and accurately calculated using theoretically derived formulas, avoiding errors from manual calculations.

[0162] Regarding the improved harmonic suppression method, the three-phase AC operating voltage in the field can introduce power frequency fundamental, 3rd, 5th, and 7th odd harmonic interference into the DC test circuit through phase-to-phase capacitive coupling. Simultaneously, the operation of power electronic devices such as switches and corona discharge during the high-voltage generator's voltage boosting process generates high-frequency pulse harmonic interference, causing the leakage current signal collected by the microammeter to be mixed with a large amount of distortion components, resulting in drift, fluctuation, and distortion of the test data. This embodiment provides an FFT frequency domain identification adaptive dynamic harmonic interference suppression algorithm, based on harmonic component separation modeling, dynamic threshold iteration, and fundamental distortion compensation principles, to achieve accurate suppression of harmonic interference under complex substation operating conditions.

[0163] First, the data preprocessed using multi-stage RC low-pass filtering is subjected to Fast Fourier Transform (FFT) to accurately extract the amplitude and phase parameters of each harmonic, thus achieving precise localization of interfering harmonics. Abandoning the fixed-weight design, a harmonic fluctuation gradient factor is innovatively introduced. σ A variable step size dynamic weight iterative model is constructed, which adaptively adjusts the iteration step size according to the real-time fluctuation intensity of the harmonic amplitude, greatly improving the algorithm's dynamic adaptability.

[0164] Define harmonic wave gradient factor σ Used to characterize the scene k The real-time fluctuation intensity of subharmonic interference is determined by constructing variable step-size weighting coefficients based on gradient factors. α ( σ _ k ), replacing traditional fixed α The hyperbolic tangent function is chosen as the iteration function to achieve smooth adaptive adjustment of the weights. After iterative calculation of the new harmonic suppression threshold with dynamic weights, harmonic interference components exceeding the threshold are fully stripped, while the low-frequency steady-state true leakage current component is retained, resulting in a preliminary effective signal after harmonic suppression.

[0165] To address the issues of frequent wireless connection interruptions and low data transmission success rates in DC high-voltage microammeters, and to shorten testing time and reduce safety risks, this embodiment provides a Bluetooth-based DC high-voltage microammeter. The DC high-voltage microammeter mainly consists of two parts: a meter head and a handheld meter. The meter head uses a stainless steel casing, making it robust, reliable, lightweight, and durable. The rounded casing effectively eliminates corona interference. For example, the interface aperture is available in 8mm and 6mm sizes, providing versatility. A built-in 9V rechargeable battery ensures long battery life. Internally, a high-precision anti-interference circuit is used, ensuring measurement accuracy is unaffected under high voltage. The display screen structure is as follows... Figure 6 As shown, the display screen uses an LCD 4.5-digit digital display, making the readings intuitive. The handheld meter has a plastic structure, making it lightweight and durable. It includes a current measurement circuit to effectively obtain the leakage current of the lower section of the surge arrester. The total test current is transmitted to the handheld meter via Bluetooth, and the meter has an automatic calculation program to calculate the difference and obtain the leakage current of the upper section of the surge arrester.

[0166] The method provided in this embodiment utilizes unsupervised learning to train and improve the LSTM deep learning model, enabling it to learn only the temporal and time-frequency characteristic distribution of leakage current under normal operating conditions of the surge arrester. When the surge arrester experiences latent faults such as dampness, aging, or partial breakdown, the leakage current waveform and frequency domain energy distribution become distorted, the deviation between the fault samples and the normal characteristic distribution learned by the model increases, and the data reconstruction error rises significantly. By combining Mahalanobis distance to quantify the degree of anomaly, the operating status of the equipment is determined through dynamic thresholds, thus completing the early latent fault identification.

[0167] The overall implementation process is divided into five modules: leakage current time-series data acquisition module, signal preprocessing and time-frequency feature extraction module, improved bidirectional LSTM-deep learning time-series model module, reconstruction error and Mahalanobis distance anomaly quantification module, and dynamic threshold state discrimination and fault output module. Based on the dynamic threshold judgment, it outputs latent fault warnings such as normal / slight moisture / performance aging / local breakdown. The overall principle block diagram is as follows. Figure 7 As shown.

[0168] The lightweight YOLO26 end-to-end model is used to identify, classify, locate and segment various defects such as cracks, damage, dirt and aging on the surface of surge arresters in real time. Relying on the advantages of the model's strong anti-interference, high accuracy and low latency, the complex background and light interference of the substation are filtered out to accurately determine the type and level of appearance defects of the surge arrester and realize intelligent assessment of appearance status.

[0169] YOLO26 is an end-to-end lightweight integrated detection and segmentation network. Its overall structure consists of three parts: a backbone feature extraction network, a multi-scale feature fusion network, and a detection and segmentation head. It abandons the complex post-processing structure of the traditional YOLO series and achieves the optimal balance between inference efficiency and detection accuracy.

[0170] The backbone network employs a lightweight stacked convolutional structure with an embedded adaptive channel attention mechanism to automatically enhance defect feature weights and suppress background noise features, adapting to the extraction of small target defects in complex substation environments. The Neck module uses a multi-scale bidirectional feature fusion structure to fuse shallow texture features (cracks, damaged edges) with deep semantic features (aging, large-area contamination), solving the detection challenge of varying arrester defect sizes and significant feature differences. The Head module adopts a parallel structure of detection and segmentation, simultaneously outputting defect category, confidence level, location coordinates, and pixel-level segmentation mask, achieving integrated multi-task output. By fusing electrical timing analysis results with visual defect recognition results, a comprehensive state diagnosis conclusion for the arrester is formed. After single-point inspection is completed, only the high-voltage test lead and platform grounding wire are removed; there is no need to disassemble the internal connection wires of the equipment. Relying on a mobile platform, the entire site can be quickly transferred with the lines, enabling rapid completion of iterative inspection of all arresters in the entire substation.

[0171] In summary, the technical solutions provided by the embodiments of this invention establish a multi-dimensional technical coupling system, which includes a high-voltage detection safety distance electric field coupling model for surge arresters, an interphase capacitive harmonic interference adaptive suppression model, a bidirectional LSTM leakage current timing feature reconstruction model, and a YOLO26 multi-task defect visual perception model. This system quantitatively analyzes the impact of complex and strong electromagnetic environment disturbances in substations, equipment structural layout parameters, leakage current timing distortion characteristics, and appearance defect texture characteristics on the safety, data accuracy, and reliability of surge arrester detection. It provides a fully intelligent integrated detection strategy that integrates hardware-integrated safety deployment, dynamic electromagnetic interference correction, timing latent fault discovery, end-to-end visual defect recognition, and rapid iterative transfer without wire disconnection. This achieves full coverage of high-voltage safety protection, precise interference error removal, early latent fault prediction, high-precision identification of multiple appearance defects, and efficient batch inspection transfer during routine surge arrester testing in substations.

[0172] By tracking the dynamic correction process of harmonic interference suppression, the evolution law of time-series current reconstruction error, the iterative characteristics of YOLO26 defect identification accuracy, and the platform's mobile operation flow status, this study grasps the state changes and error evolution mechanism of surge arrester testing operations under complex substation scenarios with strong electromagnetic fields, multiple devices, and complex sites. It combines high-voltage operation's all-dimensional safety protection capabilities, high-precision anti-interference performance of electrical testing, the ability to uncover hidden time-series faults, strong robust visual defect identification performance, and efficient on-site operation adaptation capabilities, providing a brand-new integrated technical solution for substation zinc oxide surge arrester safety standardization testing, high-precision parameter detection, all-dimensional intelligent status diagnosis, and batch rapid inspection operations.

[0173] The present invention also provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of the present invention.

[0174] The present invention also provides a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of the embodiments of the present invention.

[0175] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of this invention.

[0176] refer to Figure 8 This is a structural block diagram of an electronic device, either a server or a client, according to an embodiment of the present invention. It is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0177] like Figure 8As shown, the electronic device includes a computing unit 801, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. The RAM 803 may also store various programs and data required for the operation of the electronic device. The computing unit 801, the read-only memory ROM 802, and the random access memory RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0178] Multiple components in the electronic device are connected to I / O interface 805, including: input unit 806, output unit 807, storage unit 808, and communication unit 809. Input unit 806 can be any type of device capable of inputting information into the electronic device. Input unit 806 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 807 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 808 may include, but is not limited to, disks and optical discs. Communication unit 809 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0179] The computing unit 801 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 802 and / or communication unit 809. In some embodiments, the computing unit 801 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0180] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0181] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0182] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more". The descriptions of terms such as "first", "second", etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features.

[0183] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are subject to strict compliance with relevant laws, regulations, and regulatory requirements in their collection, storage, use, processing, transmission, provision, and disclosure, and adhere to the principles of legality, legitimacy, necessity, and good faith. The acquisition of relevant information and data is premised on the user's explicit consent or other legitimate reasons, and a clear and convenient authorization management approach is provided to the user, allowing the user to independently choose to consent, withdraw consent, or refuse to provide relevant information. For functions that rely on user information, if the user does not authorize or withdraws authorization, the corresponding technical function cannot be implemented, and the technical solution of this invention is not applicable in this scenario.

[0184] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0185] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0186] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for integrated testing of surge arresters, characterized in that, Includes the following steps: Acquire raw leakage current signals and external images of the surge arrester; The original leakage current signal is subjected to adaptive harmonic suppression processing to obtain the real leakage current signal; The actual leakage current signal is input into the pre-trained reconstruction model to obtain the state diagnosis result. The reconstruction model uses a bidirectional long short-term memory network autoencoder with residual connection to perform reconstruction error calculation and anomaly degree quantification. The appearance image is input into a pre-trained recognition model, which outputs a defect recognition result. The recognition model adopts an end-to-end architecture that removes non-maximum suppression. After completing the single-point testing based on the above steps, retain the internal equipment connection lines of the carrier platform, remove only the external high-voltage leads and grounding wires, and transfer the entire carrier platform for multi-point testing operations.

2. The method according to claim 1, characterized in that, Adaptive harmonic suppression processing is performed on the original leakage current signal to obtain the true leakage current signal, including: The original leakage current signal is filtered based on a hardware filtering architecture to obtain a preprocessed signal. Perform a Fast Fourier Transform on the preprocessed signal to extract the real-time harmonic amplitude; The filter threshold for each harmonic is adaptively and iteratively updated based on the historical harmonic statistical mean and the real-time harmonic amplitude. Based on the harmonic filtering thresholds of each order, the original leakage current signal is subjected to harmonic suppression processing to obtain a preliminary corrected time-domain signal. The initial corrected time-domain signal is subjected to phase distortion compensation correction processing to obtain the actual leakage current signal.

3. The method according to claim 2, characterized in that, The formula for adaptively iteratively updating the threshold of each harmonic filter is expressed as: ； In the formula, T h ( k ) indicates the first k Real-time dynamic suppression threshold for subharmonics α Represents the iterative weight coefficients. A k_avg Indicates the historical number within the sliding time window k Average amplitude of the second harmonic A k_real This indicates the FFT identification result obtained in the current sampling period. k Real-time amplitude of subharmonics.

4. The method according to claim 2, characterized in that, The hardware filtering architecture includes a multi-stage RC low-pass filter circuit, and each stage of the RC low-pass filter circuit is equipped with a high-frequency ferrite bead. Before filtering the original leakage current signal based on the hardware filtering architecture, the method further includes setting the cutoff frequency of the multi-stage RC low-pass filter circuit.

5. The method according to claim 1, characterized in that, The mathematical model of the original leakage current signal is: ； In the formula, This represents the original leakage current signal. This represents the actual leakage current signal. Indicates harmonic interference components. This represents random electromagnetic white noise components.

6. The method according to claim 1, characterized in that, Before inputting the actual leakage current signal into the pre-trained reconstruction model, the method further includes: A first initial model is established based on a bidirectional long short-term memory network encoder, a residual connection layer, and a unidirectional long short-term memory network decoder. Based on the normal operating condition samples of the surge arrester, the first initial model is trained in an unsupervised manner to obtain the reconstructed model, wherein the model loss function includes a mean square error term, a Kolb-Leibler divergence term, and a time series regularization term.

7. The method according to claim 1, characterized in that, The actual leakage current signal is input into the pre-trained reconstruction model to obtain state diagnosis results, including: The reconstruction error corresponding to the actual leakage current signal is calculated based on the reconstruction model, which is used to characterize the difference between the actual leakage current signal and the normal operating condition signal. Anomaly scores are calculated based on Mahalanobis distance and the reconstruction error, wherein the state diagnosis results include the anomaly scores, which characterize the degree of anomaly in the operation state of the surge arrester.

8. The method according to claim 1, characterized in that, Before inputting the appearance image into the pre-trained recognition model, the method further includes: A second initial model is established based on the backbone feature extraction network, the multi-scale bidirectional feature fusion network, and the detection-segmentation dual-branch output head. The second initial model is trained based on the frozen pre-training stage and the full parameter unfreezing fine-tuning stage to obtain the recognition model; In the frozen pre-training phase, the pre-training weights of the general dataset are loaded, the parameters of the backbone feature extraction network are frozen, and only the multi-scale bidirectional feature fusion network and the detection-segmentation dual-branch output head are trained. In the full parameter unfreezing and fine-tuning stage, all network weights are unlocked, and a cosine annealing dynamic learning rate decay strategy is used for training. The formula for adaptive decay of the learning rate with each iteration is expressed as: ； In the formula, lr epoch The learning rate represents the current iteration round. lr min This represents the minimum threshold learning rate. lr init This represents the initial maximum learning rate. epoch Indicates the current iteration round number. E total This indicates the total number of iterations.

9. An integrated testing device for surge arresters, characterized in that, include: The data acquisition module is used to acquire the original leakage current signal and appearance image of the surge arrester; The harmonic suppression module is used to perform adaptive harmonic suppression processing on the original leakage current signal to obtain the real leakage current signal. The diagnostic module is used to input the real leakage current signal into the pre-trained reconstruction model to obtain the state diagnostic result. The reconstruction model adopts a bidirectional long short-term memory network encoder with residual connection to perform reconstruction error calculation and anomaly degree quantification. The recognition module is used to input the appearance image into a pre-trained recognition model and output the defect recognition result, wherein the recognition model adopts an end-to-end architecture that removes non-maximum suppression; The platform integrates all modules and testing equipment. After completing a single-point test, it retains the internal equipment connection lines, removes only the external high-voltage leads and grounding wires, and transfers the entire system to perform multi-point testing operations.

10. The apparatus according to claim 9, characterized in that, The data acquisition module includes an image acquisition unit and a DC microammeter; The image acquisition unit is used to acquire external images of the surge arrester; The DC microammeter includes a connected meter head and a handheld meter. The meter head has an arc-shaped outer shell and a built-in anti-interference circuit and Bluetooth transmission unit. The handheld meter has a built-in current calculation unit. The DC microammeter is used to acquire the original leakage current signal of the surge arrester.

Images