A ring main unit partial discharge on-line monitoring and fault diagnosis method

Abstract

The application discloses a kind of ring main unit partial discharge on-line monitoring and fault diagnosis method, belong to electric power equipment state monitoring technical field;Utilize ultrasonic, transient ground voltage and ultrahigh frequency sensor array synchronous acquisition signal, wherein transient ground voltage and ultrahigh frequency sensor integrated detection circuit, convert high frequency into low frequency envelope, solve bandwidth conflict;Signal is after adaptive filtering, experience modal decomposition gets intrinsic mode function component, extracts multi-scale feature fusion, inputs deep confidence network and diagnoses fault type.Denoising uses normalized relative threshold singular value decomposition and mutual information constraint;Deep confidence network introduces class interval penalty to improve separability, combined with finite element simulation and maximum mean difference screening expansion sample;Through synchronous time service and phase-locked loop tame clock and encode time stamp ensure synchronization;Based on gengbeier distribution dynamic threshold control false alarm, comprehensive discharge capacity and so on Index weighted evaluation insulation state.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring technology, specifically a method for online monitoring and fault diagnosis of partial discharge in ring main units. Background Technology

[0002] Ring main units (RMS), as core node equipment in the field of distribution network automation, undertake critical functions such as power distribution, line protection, and load transfer in modern urban power distribution networks. Compared with traditional open substations, RMS have significant advantages such as compact structure, small footprint, and high operational reliability, and have been widely used in power supply systems of residential communities, industrial parks, and urban commercial complexes. With the continuous expansion of urban power distribution networks and the increasing demands of users for power supply reliability, the safe and stable operation of RMS directly affects the service quality and service life of the entire power distribution network. However, during long-term operation, the internal insulation structure of RMS inevitably experiences insulation performance degradation due to the combined effects of multiple stress factors, including electrical, thermal, chemical, and mechanical factors. Partial discharge, as a subcritical state characterization before electrical breakdown of insulating materials, often indicates potential defects within the insulation system, such as bubbles, impurities, and surface discharge channels. Therefore, implementing online partial discharge monitoring of RMS to accurately capture early signals of insulation degradation has significant engineering practical value and economic significance for preventive maintenance and avoiding sudden power outages.

[0003] In the development of partial discharge detection technology, ultrasonic detection, electrical pulse current method, and ultra-high frequency detection method constitute the three main technical paths. Ultrasonic detection is based on the principle that transient mechanical stress is generated during partial discharge. An ultrasonic sensor mounted on the equipment casing receives the acoustic signal generated by the discharge, thereby locating the discharge position and assessing its intensity. This method has the advantages of strong resistance to electromagnetic interference and no need for electrical connection to a high-voltage circuit. However, it also has technical limitations such as severe attenuation of sound waves in solid media and limited sensitivity. Especially for weak early partial discharge signals, its detection effectiveness often fails to meet the requirements of engineering applications. The electrical pulse current method obtains the pulse current signal generated by partial discharge through a coupling device. This method has a relatively mature technical system in signal time-domain feature extraction and discharge quantity calibration, and is currently one of the most widely used methods in offline partial discharge detection. However, the electrical pulse current method faces strong power frequency harmonic interference in online monitoring scenarios. Balancing the reliability of signal extraction with anti-interference capability has remained a technical bottleneck in this field. Ultra-high frequency (UHF) detection utilizes the characteristic that partial discharge processes generate UHF electromagnetic radiation of hundreds of megahertz to gigahertz. A broadband antenna sensor receives the radiated signal, thereby achieving highly sensitive detection of partial discharges. This method exhibits significant advantages in sensitivity and anti-interference capability, but it also faces technical challenges such as complex signal propagation paths, difficulty in accurately modeling attenuation characteristics, and the challenge of quantifying discharge.

[0004] In the intelligent processing of partial discharge signals, traditional threshold alarm methods rely on pre-set discharge quantity or discharge frequency thresholds for single-dimensional discrimination. Their discrimination logic is overly simplistic and unable to adapt to the differentiated characteristics of various operating conditions and insulation aging stages. In recent years, with the gradual penetration of machine learning technology into the field of power equipment condition monitoring, fault diagnosis models built based on algorithms such as support vector machines, random forests, and convolutional neural networks have begun to be applied in partial discharge pattern recognition. However, existing intelligent diagnostic methods generally rely on manually set feature parameters in the feature engineering stage. The selection of these features is often highly subjective and empirical, making it difficult to comprehensively and systematically characterize the inherent complexities of partial discharge signals. Furthermore, the difficulty in obtaining training samples is also a key factor restricting the generalization ability of intelligent diagnostic models. The scarcity of partial discharge fault samples in actual engineering leads to a significant decrease in the model's recognition accuracy when facing new fault types.

[0005] More importantly, existing online partial discharge monitoring systems for ring main units (RNBs) generally face challenges in coordinating signal acquisition and data transmission during actual deployment. RNBs are typically installed in space-constrained distribution rooms with complex and variable internal electromagnetic environments, making it difficult to guarantee the optimality of sensor placement. This results in significant randomness and uncertainty in the time and frequency domains of the acquired signals. Furthermore, to achieve a comprehensive understanding of the RNB's operating status, the monitoring system often needs to simultaneously acquire signal data from multiple physical quantities, such as ultrasonic signals, transient ground voltage signals, and ultra-high frequency electromagnetic signals. The sampling synchronization between different sensor channels directly affects the effectiveness of subsequent signal fusion processing. Under actual operating conditions, limitations in sensor hardware performance and data transmission bandwidth often make it difficult to simultaneously achieve synchronous acquisition and real-time transmission of multi-channel signals. In addition, distinguishing partial discharge signals from background noise remains a technical challenge. Environmental electromagnetic interference and interference signals generated during normal equipment operation overlap with the actual partial discharge signals in the frequency domain, posing a significant technical challenge to the accurate extraction of discharge signals and the correct identification of fault types.

[0006] In summary, how to achieve high-sensitivity and high-reliability acquisition of partial discharge signals from ring main units under complex electromagnetic environments and spatial constraints, how to establish a quantitative mapping relationship between multi-dimensional signal characteristics and insulation degradation, and how to overcome sample limitations to construct an intelligent fault diagnosis model with strong generalization ability have become key challenges and urgent technical problems facing those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a method for online monitoring and fault diagnosis of partial discharge in ring main units, in order to solve the problems of poor signal acquisition reliability, insufficient multi-channel synchronization, weak anti-interference ability, and limited generalization ability of intelligent diagnostic models in the existing online monitoring and fault diagnosis technologies for partial discharge in ring main units mentioned in the background art.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A method for online monitoring and fault diagnosis of partial discharge in a ring main unit includes the following steps:

[0010] The partial discharge signal inside the ring main unit is synchronously acquired using a distributed multiphysics field sensor array. The sensor array includes an ultrasonic sensor, a transient ground voltage sensor, and an ultra-high frequency sensor. The transient ground voltage sensor and the ultra-high frequency sensor each integrate a detection circuit to convert the high-frequency raw signals they receive into low-frequency envelope signals that reflect the outer envelope of the discharge pulse energy.

[0011] The acquired signals are subjected to adaptive filtering for noise reduction and standardization preprocessing to obtain the preprocessed signal;

[0012] Empirical mode decomposition is performed on the preprocessed signal to obtain multiple intrinsic mode function components;

[0013] Multi-scale features of each intrinsic mode function component are extracted and fused to form a fused feature vector.

[0014] The fused feature vectors are input into the trained deep belief network fault diagnosis model, which outputs the partial discharge fault type.

[0015] According to the above technical solution, the detection circuit performs nonlinear rectification and low-pass filtering on the received signal to remove the high-frequency carrier and extract the low-frequency envelope signal; the signals collected by each sensor are sampled at the same sampling frequency, and the bandwidth of the low-frequency envelope signal is less than half of the sampling frequency.

[0016] According to the above technical solution, the steps of adaptive filtering and denoising include:

[0017] The original signal is divided into continuous and partially overlapping analysis frames, and a Hankel matrix is ​​constructed for each frame of data.

[0018] Singular value decomposition is performed on the Hankel matrix, and a normalized relative threshold based on the maximum singular value is adaptively determined according to the signal-to-noise ratio of the signal, and the principal singular values ​​that satisfy the relative threshold are retained.

[0019] The matrix is ​​reconstructed using the retained principal singular values ​​and their corresponding singular vectors, then restored to a one-dimensional signal by anti-diagonal averaging, and adjacent frames are spliced ​​together using the overlapping addition method to obtain the denoised signal.

[0020] According to the above technical solution, empirical mode decomposition adopts a dual screening stopping criterion: when the number of extreme points of the current residual component is less than or equal to 2, and the mutual information between the residual component and the preprocessed signal is lower than a preset threshold, the decomposition is terminated.

[0021] Mutual information The calculation method is as follows:

[0022]

[0023] in, For residual components, The input signal is standardized. These are the amplitude quantization box indices for the residual component and the input signal, respectively. For joint probability distribution, and This represents a marginal probability distribution.

[0024] According to the above technical solution, multi-scale features include time-domain statistical features, frequency-domain statistical features, and empirical mode decomposition features;

[0025] Time-domain statistical characteristics include peak factor, impulse factor, waveform factor, margin factor, and correlation coefficient between adjacent intrinsic mode function components;

[0026] Frequency domain statistical characteristics include frequency domain centroid, frequency domain variance, frequency domain entropy, and dominant frequency;

[0027] Empirical mode decomposition features include the energy distribution ratio of each intrinsic mode function component, the ratio of adjacent energies, and the average period.

[0028] Feature fusion includes: using a recursive feature elimination method to filter all extracted features, and then serially concatenating the filtered features to form a fused feature vector.

[0029] According to the above technical solution, the deep belief network fault diagnosis model is composed of multiple restricted Boltzmann machines stacked together, and its training process includes unsupervised layer-by-layer pre-training and supervised fine-tuning.

[0030] In the supervised fine-tuning phase, the loss function used is... Introduce an inter-class margin penalty term based on the cross-entropy loss:

[0031]

[0032] in, For real labels, To predict probabilities, and For the classifier corresponding to the first Class and First The weight vector of the fault class, This represents the interval penalty coefficient.

[0033] Based on the above technical solution, the training samples of the fault diagnosis model are enhanced in the following ways:

[0034] Finite element simulation models based on physical mechanisms are established for different discharge types, and simulation signals are generated.

[0035] The simulated signal is superimposed with background noise at different signal-to-noise ratios to obtain a synthetic sample;

[0036] Calculate the maximum mean difference between the synthetic sample and the real collected sample, and only include the synthetic sample with the maximum mean difference below a preset threshold into the training sample library.

[0037] According to the above technical solution, the implementation of synchronous data acquisition includes:

[0038] Each sensor acquisition unit uses the GPS precise time synchronization protocol to discipline the sampling clock, so that the sampling clocks between channels are synchronized;

[0039] Data frames are transmitted via industrial Ethernet that supports the IEEE 1588 precision time protocol, and timestamps are carried in the data frames to achieve time alignment of multi-channel signals.

[0040] Based on the above technical solution, adaptive noise threshold estimation is used when detecting partial discharge signals:

[0041] Background noise signals are collected, and a Gumbel distribution model is established to fit the probability distribution of noise peaks. The cumulative distribution function is as follows:

[0042]

[0043] in, For position parameters, The scale parameter is used; the noise threshold is calculated based on the Gumbel distribution model and the preset signal level; the position parameter and scale parameter are dynamically updated using a sliding window to track changes in background noise; the peak value of the real-time acquired signal is compared with the current noise threshold, and if it exceeds the noise threshold, it is determined that there is a partial discharge signal.

[0044] According to the above technical solution, after outputting the partial discharge fault type, an insulation degradation assessment step is also included:

[0045] Discharge quantity, number of discharges, discharge phase distribution, and discharge frequency are extracted as evaluation indicators;

[0046] The weight coefficients of each evaluation indicator were determined using the analytic hierarchy process (AHP).

[0047] The discharge quantity was mapped using an exponential function, the number of discharges was mapped using a logarithmic function, the discharge phase distribution was mapped using an Euclidean distance metric, and the discharge frequency was mapped using a linear function to obtain scores for each index.

[0048] The insulation degradation comprehensive score is obtained by weighting and summing the scores of each indicator based on the weighting coefficients, and the insulation degradation level is classified according to the comprehensive insulation degradation score.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This invention solves the bandwidth conflict and spectral aliasing problems caused by integrating a detection circuit at the front end of a multi-physics sensor, converting high-frequency transient ground voltage and ultra-high-frequency signals into low-frequency envelope signals. This fundamentally addresses the issue of bandwidth conflict and spectral aliasing during unified sampling of multi-channel signals, ensuring waveform fidelity in the digital domain. Furthermore, it employs adaptive singular value decomposition denoising based on normalized relative thresholds and empirical mode decomposition constrained by mutual information to achieve high-fidelity reconstruction of weak discharge signals under complex electromagnetic environments. Multi-scale features are extracted from the time domain, frequency domain, and intrinsic mode structure, providing a comprehensive feature space for fault identification. The deep belief network introduces an inter-class interval penalty term during training, increasing the separability of different fault types in the feature space and improving the ability to distinguish similar fault modes and the reliability of diagnostic confidence. Simultaneously, it utilizes a deep belief network based on… The finite element simulation of the physical mechanism generates discharge signals, and the training sample library is expanded after screening by the maximum mean difference statistical test, which overcomes the constraint of the scarcity of actual fault samples and enhances the model's generalization ability to unknown working conditions. Each acquisition unit tames the local sampling clock through the synchronization timing module and digital phase-locked loop, and encodes the absolute timestamp into the sampling data frame, establishing a precise time correlation between multi-channel signals at the signal acquisition source, providing a strict time reference for multi-physics field fusion analysis. On this basis, the background noise peak is modeled using the Gumbel extreme value distribution and the threshold is dynamically updated to theoretically control the false alarm probability. The differential mapping and weighted evaluation of multi-dimensional indicators such as discharge quantity, discharge number, discharge phase distribution and discharge frequency are comprehensively performed to achieve objective quantitative classification of insulation status, providing an intuitive basis for operation and maintenance decisions. Attached Figure Description

[0051] Figure 1 This is a flowchart of the online partial discharge monitoring and fault diagnosis method of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] like Figure 1As shown, this invention provides a method for online monitoring and fault diagnosis of partial discharge in ring main units. This method synchronously collects partial discharge signals inside the ring main unit using a distributed multiphysics sensor array. After preprocessing, multi-scale features are extracted and fused. Finally, a deep belief network is used to achieve intelligent diagnosis of fault types. The technical solution of this invention will be described in detail below with reference to specific embodiments.

[0055] The overall architecture of the distributed multiphysics sensor array includes three types: ultrasonic sensors, transient ground voltage sensors, and ultra-high frequency sensors. These sensors are arranged in a three-dimensional distributed manner along the interior of the ring network cabinet, forming a multi-dimensional signal acquisition network covering the internal space of the cabinet. In a specific embodiment of this invention, the ultrasonic sensor is a piezoelectric sensor with a center frequency of 40kHz. This sensor uses lead zirconate titanate piezoelectric ceramic as the sensitive element, with a resonant frequency of 40kHz±2kHz, a sensitivity better than -85dBre1V / μPa, a directivity angle of ±30°, and an operating temperature range of -40℃ to 85℃. The ultrasonic sensors are arranged at three different heights on the inner wall of the cabinet, at heights of 200mm, 500mm, and 800mm from the bottom of the cabinet. One ultrasonic sensor is placed at each height, and the sensor probes are in close contact with the inner wall of the cabinet through a coupling agent to ensure effective transmission of the acoustic signal. This three-dimensional layered arrangement allows the ultrasonic sensor array to cover partial discharge acoustic emission sources at different heights within the cabinet. Furthermore, by analyzing the arrival time difference and amplitude attenuation relationship between the sensor signals at the three different heights, the spatial location range of the discharge source can be preliminarily inferred. The core of the three-dimensional distributed layout lies in constructing a spatially complete signal observation matrix. Assuming there is a local discharge source at any point within the cabinet, the propagation time of its radiated acoustic signal to the i-th ultrasonic sensor is... The sensor array then collects signals. It can be represented as:

[0056]

[0057] in The attenuation coefficient is... As the source signal, This is additive noise. By performing generalized cross-correlation processing on the three height position signals, the time difference of arrival can be estimated, providing spatial constraint information for subsequent positioning. In addition, the non-coplanar arrangement of the ultrasonic sensors in three-dimensional space can effectively avoid spatial filtering effects on the signal transmission path, improving the consistency of detection sensitivity for power sources at different azimuth angles inside the cabinet.

[0058] The transient ground voltage sensor is a capacitively coupled sensor, consisting of a metal plate electrode and a matching capacitor with a coupling capacitance of 20pF. It has a frequency response range of 3MHz to 100MHz, an input impedance of 1MΩ, and a sensitivity of 1mV / mV. The sensor is positioned at the cabinet's grounding bolt and connected to the cabinet's grounding busbar via a grounding wire, effectively coupling the transient ground voltage signal to the cabinet surface. To match the 3MHz to 100MHz transient ground voltage signal with the subsequent uniform sampling rate of 10MHz, the sensor integrates a broadband charge envelope detection circuit. This circuit performs nonlinear rectification and low-pass filtering on the high-frequency transient ground voltage signal, removing the high-frequency carrier wave and extracting a low-frequency transient ground voltage envelope signal with a bandwidth below 5MHz, reflecting the transient ground voltage pulse energy. This low-frequency envelope signal is then output for further processing, thus achieving a sampling bandwidth that covers the characteristic frequency band of the transient ground voltage signal at the equivalent envelope level. In a specific embodiment of the invention, the sensor is installed at a height of 150mm from the bottom of the cabinet at the grounding bolt. This location is the main conduction path of the cabinet's grounding current, which is beneficial for the effective acquisition of transient ground voltage signals. The transient ground voltage signal essentially reflects the transient voltage change induced on the cabinet surface by the electromagnetic waves generated by partial discharge. Its propagation path is closely related to the distributed parameters of the cabinet's grounding system. Placing the sensor at the grounding bolt maximizes the coupling of the common-mode voltage signal excited by the discharge pulse current injected into the grounding system. Furthermore, the grounding impedance at this location is relatively concentrated, which is beneficial for establishing a deterministic signal coupling model. According to transmission line theory, the cabinet grounding wire can be equivalent to having distributed inductance at high frequencies. and distributed capacitance The transmission line has a transient ground voltage sensor sampling voltage at its terminal. Since the influence of distributed parameters on the voltage division ratio within the 3MHz to 100MHz characteristic frequency band is suppressed by high-frequency impedance matching, its equivalent circuit model can be simplified to: sensor coupling capacitance... Grounding capacitance The sensor output voltage is formed by a capacitor voltage divider network:

[0059]

[0060] in This is to account for the potential rise in the grounding wire caused by partial discharge. By selecting an appropriate coupling capacitor value (20pF), the voltage division ratio is kept flat within the frequency band of interest, ensuring the linearity of signal coupling.

[0061] The ultra-high frequency (UHF) sensor employs a planar ultra-wideband antenna with a bandwidth of 300MHz to 3GHz. This antenna is fabricated using planar printing technology, with a RogersRT / duroid5880 substrate and a dielectric constant of 2.2. The antenna dimensions are 50mm × 50mm × 1.6mm, and the VSWR is less than 2.5 within the operating frequency band, with a gain better than 3dBi. The UHF sensor is positioned directly opposite the observation window of the cabinet, with the sensor probe facing the internal cavity of the cabinet, to receive high-frequency electromagnetic radiation signals generated by partial discharge. To match the 300MHz to 3GHz UHF signal with the subsequent 10MHz low-frequency sampling rate, the UHF sensor integrates a high-frequency charge envelope detection circuit. This circuit performs nonlinear rectification and low-pass filtering on the received UHF electromagnetic radiation signal, removing the high-frequency carrier and extracting the low-frequency envelope signal, which reflects the energy envelope of the discharge pulse and has a bandwidth below 5MHz. This low-frequency envelope signal is then output for further processing.

[0062] Example 2

[0063] In a specific embodiment of the invention, the ultra-high frequency sensor is fixed to the observation window frame, and the angle between the probe's central axis and the normal direction of the observation window plane is less than 15°, ensuring effective reception of electromagnetic wave signals. The ultra-high frequency sensor at the observation window utilizes the cabinet window as a coupling channel for electromagnetic waves, effectively receiving ultra-high frequency electromagnetic waves propagating within the cabinet. Because partial discharge pulses have extremely fast rise times (typically sub-nanoseconds to several nanoseconds), the spectrum of their radiated electromagnetic waves can extend to several GHz, and the ultra-high frequency sensor has high receiving sensitivity in this frequency band. The signal power received by the antenna... With the power of the discharge source The relationship is given by the Friesian transport equation:

[0064]

[0065] in This represents the equivalent antenna gain of the discharge channel. For UHF antenna gain, For wavelength, For propagation distance. In the complex multipath environment inside the cabinet, the reflection and diffraction effects of the metal wall also need to be considered. The actual received signal is a vector synthesis of the direct wave and multiple reflected waves. Therefore, the sensor position should be selected to avoid standing wave nodes.

[0066] The signal preprocessing module performs adaptive filtering, denoising, and standardization on the acquired raw signals (including ultrasonic signals, low-frequency transient ground voltage envelope signals after envelope detection, and pulse envelope signals after ultra-high frequency sensor detection). In the adaptive filtering and denoising step, a signal denoising method based on singular value decomposition (SVD) is used. This method constructs the raw signal into a Hankel matrix, then performs SVD on the matrix, determines the number of effective singular values ​​based on a singular value energy ratio threshold, and reconstructs the denoised signal. In a specific embodiment of this invention, the raw signal sampling frequency is 10MHz, the sampling duration is 20ms, and the number of data points processed each time is 200,000. To meet real-time processing requirements and ensure matrix numerical stability, the raw signal is divided into continuous and partially overlapping analysis frames, each with a length of [missing information]. There are 100 sampling points, with 512 sampling points overlapping between frames. For each frame of data, a Hankel matrix is ​​constructed. Its number of rows With column number satisfy Optimize selection based on minimizing the condition number. and :when At that time, take , At this time, the matrix The condition number is close to 1, which is beneficial for the accurate separation of the signal subspace and the noise subspace by singular value decomposition.

[0067] Singular value decomposition is performed on the Hankel matrix constructed for each frame of data to obtain the singular value vector. The maximum singular value is adaptively determined based on the signal-to-noise ratio of the signal. Normalized relative threshold and noise truncation: When the signal-to-noise ratio is below 10dB, it indicates that the noise subspace is energy-intensive and dense. To avoid introducing too much high-frequency noise into the reconstruction, the relative threshold is set to 0.05 (i.e., retaining the noise level that meets the requirements). The singular values ​​are selected, and at most the first 20 principal singular values ​​are retained; when the signal-to-noise ratio is between 10dB and 20dB, the relative threshold is set to 0.01 (i.e., retaining the singular values ​​that meet the requirements). The system selects the singular values ​​and retains a maximum of 40 singular values. When the signal-to-noise ratio is higher than 20dB, it indicates that the background noise is extremely low, and the relative threshold is set to 0.001 (i.e., retaining the values ​​that meet the requirements). The normalized singular values ​​are selected, and at most the first 60 singular values ​​are retained to restore richer details of the partial discharge signal. The theoretical basis for the adaptive threshold setting is that the magnitude of the normalized singular values ​​corresponds to the relative energy proportion of each component in the signal subspace, while the normalized singular values ​​in the noise subspace are relatively uniform and have lower amplitudes. A normalized energy proportion accumulation function is defined. The essence of threshold selection is to find an inflection point. , making when hour Rapid rise, when hour The increase is slow. A lower signal-to-noise ratio (SNR) means a smaller proportion of the signal component in the total energy and a larger noise component. Therefore, the page needs to achieve truncation denoising by setting a higher normalized relative threshold (0.05) and limiting the retention of a smaller number (20) of principal singular values. When the SNR is high, the signal component energy is concentrated and the details are complete. A lower normalized relative threshold (0.001) can be set, and a larger number (60) of singular values ​​can be retained to fully preserve the details of weak signals. This adaptive mechanism eliminates the interference of the absolute magnitude of the signal amplitude on the threshold selection, avoiding the under-denoising or over-denoising problems that may occur when processing signals with different SNRs using fixed threshold denoising. The first few values ​​that meet the above adaptive relative threshold conditions are retained. The singular values ​​and their corresponding left and right singular vectors are used to reconstruct the denoised signal. The reconstruction process is achieved through matrix multiplication.

[0068]

[0069] in and They are respectively the front A matrix composed of left and right singular vectors. For matrix transpose operator, For the front A diagonal matrix composed of singular values. Then reconstruct the matrix. The signal is reconstructed into a one-dimensional signal by averaging along the anti-diagonal. Adjacent frames are smoothly stitched together using an overlapping and adding method to eliminate inter-frame discontinuities. After processing using this method, the signal-to-noise ratio of the reconstructed signal is improved by an average of more than 15 dB compared to the original signal.

[0070] The signal standardization step employs the Z-score standardization method, which normalizes the mean of the denoised signal to zero and the variance to one. Specifically, for each sampled channel, the mean of that channel's signal is calculated. and standard deviation Then for each sample value According to the formula Standardization is performed. In a specific embodiment of the present invention, for the ultrasound signal channel, the mean signal after standardization is... With a standard deviation of 0.998, standardization adjusts the dynamic range of the signal to the [-3, 3] interval, effectively eliminating dimensional differences between different sensor channels. The essence of standardization is to remove the differences in first- and second-order statistics of the signal, ensuring that each channel signal has an equal numerical scale in subsequent feature extraction, avoiding feature shifts caused by differences in sensor sensitivity and gain. For the j-th sensor channel, its standardized output... It follows a standard normal distribution approximation, and this distribution normalization helps improve the stability of gradient propagation during subsequent training of deep belief networks.

[0071] The multi-scale feature extraction method based on empirical mode decomposition first performs empirical mode decomposition on the standardized signal, decomposing the signal into several intrinsic mode function components and a residual component. The Empirical Mode Decomposition (EMD) method employed in this invention introduces an improved screening stopping criterion to avoid mode aliasing and over-decomposition problems that may occur in standard EMD. Specifically, in one embodiment of this invention, the termination condition for decomposition iteration is no longer based solely on the number of extrema or energy, but rather employs a dual criterion: the number of extrema of the current residual component is less than or equal to 2, and the mutual information between the residual component and the original signal is below a set threshold (e.g., 0.05 bits). The introduction of mutual information effectively distinguishes residual components containing valid information from pure noise residuals, preventing useful signal components from being incorrectly retained in the residual components and thus not being decomposed. Mutual Information The calculation formula is:

[0072]

[0073] in For residual components, The input signal is standardized. These are the amplitude quantization box indices for the residual component and the input signal, respectively. For joint probability distribution, and The probability distribution is marginal. When the mutual information is below a threshold, the statistical dependence between the residual components and the original signal is very weak, and the decomposition can be terminated. Using the improved method of this invention, 8 to 12 intrinsic mode function components are obtained through decomposition.

[0074] Then, for each intrinsic mode function component, the time-domain statistical characteristics, frequency-domain statistical characteristics, and empirical mode decomposition characteristics are calculated respectively.

[0075] Time-domain statistical characteristics include five parameters: peak factor, impulse factor, waveform factor, margin factor, and correlation coefficient. Taking the peak factor as an example, its calculation formula is as follows:

[0076]

[0077] in The maximum value of the absolute value of the component. This represents the total number of sampling points in a single frame of signal. For the first component Each sample value reflects the pulse characteristics of the signal. Different local discharge types result in varying sharpness and duration of the discharge pulse, leading to differentiated peak factors. The formula for calculating the pulse factor is:

[0078]

[0079] This characteristic is sensitive to pulsed signals, and its value is relatively large in continuous pulsed discharge modes such as corona discharge. The formula for calculating the waveform factor is:

[0080]

[0081] This feature reflects the degree of distortion in the signal waveform and is related to both pulse repetition rate and pulse width. The formula for calculating the margin factor is:

[0082]

[0083] This feature amplifies the contribution of samples with larger amplitudes in the signal, and has good distinguishability for discharge modes with irregular waveforms, such as surface discharge. Correlation coefficient The formula used to measure the linear correlation between two signals is:

[0084]

[0085] in The first eigenmode function component of the current intrinsic mode function Each sample value, For the next eigenmode function component adjacent to it, the first Each sample value, and These are the mean values ​​of the two components. In a specific embodiment of the invention, the correlation coefficient is used to characterize the similarity between adjacent intrinsic mode function components, and the gradient of the similarity change can be used as a basis for fault type identification. The above five time-domain statistical features characterize the waveform morphology of the intrinsic mode function components from different statistical perspectives, constituting a multi-dimensional descriptive space for the shape of the discharge pulse.

[0086] Frequency domain statistical characteristics include four parameters: frequency domain centroid, frequency domain variance, frequency domain entropy, and dominant frequency. The formula for calculating the frequency domain centroid is:

[0087]

[0088] in For the first Frequency values ​​of each frequency point This represents the power spectral density value at the corresponding frequency point. This refers to the number of effective points in the spectrum. In a specific embodiment of the invention, the power spectral density is obtained by performing a Fast Fourier Transform on the intrinsic mode function components, resulting in 1024 spectral points. The calculated frequency domain centroid range is 5kHz to 1500kHz. This characteristic reflects the center position of the signal's frequency distribution. Different discharge types have different energy concentration frequency bands; for example, levitation discharge typically has a higher frequency domain centroid. The formula for calculating the frequency domain variance is:

[0089]

[0090] This characteristic reflects the dispersion of the signal frequency distribution; broadband discharges exhibit a larger frequency domain variance. The formula for calculating the frequency domain entropy is:

[0091]

[0092] in For the normalized power spectral probability distribution, it is stipulated that when hour, Frequency domain entropy quantifies the complexity of the spectrum based on information theory principles. The entropy value increases in the presence of harmonic components or modulation sidebands. The dominant frequency is the frequency value corresponding to the maximum power spectral density, directly indicating the most dominant frequency component in the signal. These four frequency domain features constitute a complete descriptive vector in the frequency domain space, allowing the capture of the frequency domain characteristics of the discharge signal from four dimensions: overall spectral shape (centroid, variance), information content (entropy), and dominant component (dominant frequency).

[0093] Empirical mode decomposition features include three parameters: the energy distribution ratio of intrinsic mode function components, the energy ratio between adjacent intrinsic mode function components, and the average period of the intrinsic mode function components. The formula for calculating the energy distribution ratio is:

[0094]

[0095] in For the first The energy of each intrinsic mode function component is obtained by calculating the component variance. This represents the total number of intrinsic mode function components obtained from the decomposition. In a specific embodiment of the present invention, the energy proportion of the first intrinsic mode function component is typically 0.45 to 0.65, indicating that the main energy of the partial discharge signal is concentrated in the low-frequency components, and this proportion exhibits a regular shift with the discharge type. The formula for calculating the energy ratio is:

[0096]

[0097] This characteristic reflects the energy decay between adjacent components. Theoretical analysis shows that for pulsed discharges, energy decays rapidly over time, resulting in a large ratio between adjacent energies; for continuous discharges, energy decays slowly, with a ratio close to 1. The formula for calculating the average period is:

[0098]

[0099] in The signal sampling frequency, For the first The number of zero-crossing points of each intrinsic mode function component, and the average period corresponding to the characteristic time scale of that component. These three EMD features describe the multi-scale properties of the signal from three dimensions: energy distribution, energy decay, and characteristic time scale, making up for the inadequacy of traditional time-frequency domain features in reflecting the intrinsic time-scale structure of the signal.

[0100] The multi-dimensional feature fusion step first employs a feature selection method based on recursive feature elimination to screen the extracted time-domain statistical features, frequency-domain statistical features, and empirical mode decomposition features. In a specific embodiment of the invention, the initial feature dimension is 85 dimensions. The feature selection algorithm adopts a backward recursive feature elimination strategy, eliminating features with the smallest contribution to classification in each iteration. The feature contribution is quantified by training a classifier on the current feature subset and evaluating the importance of the features. Specifically, the importance of each feature is measured by the average impurity reduction index based on random forest. The iteration terminates when the feature dimension decreases to 25 dimensions or the classification accuracy decreases by more than 2%. During the recursive feature elimination process, the classification performance of each feature subset is recorded, and features that perform stably under different classifiers are preferentially retained to enhance the robustness of the selected features. After feature selection, 25 key features are retained, including peak factor, impulse factor, waveform factor, margin factor, frequency domain centroid, frequency domain variance, frequency domain entropy, dominant frequency, energy distribution ratio of the first 5 intrinsic mode function components, adjacent energy ratio, and average period, etc. Then, a sequential feature fusion method is used to concatenate the selected features one by one, forming a 25-dimensional fused feature vector. Mathematically, the sequential fusion method is equivalent to constructing an augmented feature vector in a high-dimensional feature space. This vector contains discriminative information from features of each subclass, providing a multi-dimensional input representation for subsequent deep belief networks.

[0101] The fault diagnosis model based on deep belief networks consists of several stacked restricted Boltzmann machines (RBMs). Each RBM contains one visible layer and one hidden layer, with a fully connected structure between the visible and hidden layers. In a specific embodiment of the invention, the deep belief network contains four RBMs, with 25 nodes in the visible layer and 100, 80, 60, and 40 nodes in the hidden layer, respectively, for a total of five layers. The activation function for each node is the sigmoid function, with the function expression as follows: This invention introduces an adaptive network depth determination mechanism based on standard deep belief networks. Specifically, the number of layers in the Restricted Boltzmann Machine (RBM) is not fixed beforehand, but is automatically determined by monitoring the rate of decrease in reconstruction error during layer-by-layer pre-training. The method involves calculating the reconstruction error of the input data for each newly added RBM layer. ,like If the decrease in depth compared to the previous layer is less than 5%, it is considered that the newly added layer no longer contributes significantly to feature learning, and the number of layers is stopped to avoid overfitting caused by an excessively deep network. In this embodiment, the final number of layers is 5 after adaptive determination.

[0102] The training process of the fault diagnosis model is divided into two stages: unsupervised pre-training and supervised fine-tuning. In the unsupervised pre-training stage, the contrastive divergence algorithm is used to train the deep belief network layer by layer. Specifically, the CD-1 algorithm is used for parameter learning of the Restricted Boltzmann Machine (RBM), with a learning rate of 0.01, a momentum coefficient of 0.5, a weight decay coefficient of 0.0002, and a maximum training epoch of 200. In the training of each RBM, the positive pass phase calculates the hidden layer activation probability based on the input data, and the negative pass phase updates the network parameters through reconstruction error. The reconstruction error is calculated using the following formula: The energy function of a restricted Boltzmann machine is defined as:

[0103]

[0104] in The visible layer state vector, The hidden layer state vector, To connect the weight matrix, and These are the bias vectors for the visible and hidden layers, respectively. The contrastive divergence algorithm updates parameters using the approximate gradient of the KL divergence.

[0105]

[0106] in Represents the expected value under the data distribution. This represents the expected value of the model distribution after one reconstruction step. In a specific embodiment of the present invention, the training time of the first restricted Boltzmann machine is approximately 45 seconds, and the reconstruction error decreases from the initial 0.85 to 0.12, indicating that the model has effectively learned the first-order statistical structure of the input data.

[0107] In the supervised fine-tuning phase, the backpropagation algorithm is used to optimize the parameters of the pre-trained deep belief network. The learning rate is set to 0.001, the batch size to 32, and the maximum training epochs to 500. Weight updates are performed using stochastic gradient descent with a driving term, as shown in the formula:

[0108]

[0109] in For learning rate, ,in One-hot encoding of the real label. To predict probabilities, and to further improve the model's ability to distinguish fault types, this invention introduces an inter-class margin penalty term into the cross-entropy loss function. The modified total loss function is as follows:

[0110]

[0111] in This is the interval penalty coefficient. and These are the first two classes in the softmax classifier. Class and First The penalty term, which assigns weight vectors to different fault types, forces the weight vectors of different fault types to move away from each other in the feature space, thereby increasing the separability between classes. In a specific embodiment of the invention, the training time during the fine-tuning phase is approximately 3 minutes, and the accuracy on the validation set is improved from 65% after pre-training to 92%.

[0112] The softmax classifier classifies the output of the deep belief network to obtain the partial discharge fault type diagnosis result. The classification probability is calculated using the following formula:

[0113]

[0114] in This is the output feature vector of the deep belief network. This represents the total number of fault types. In a specific embodiment of the present invention, the fault types include five types: corona discharge, surface discharge, suspension discharge, air gap discharge, and internal discharge. The value is 5. Each fault type corresponds to a weight vector. and bias The classifier outputs a probability value for each fault type, selects the type with the highest probability as the diagnostic result, and outputs the corresponding confidence value. To calibrate the output confidence value and make it more accurately reflect the probability of diagnostic accuracy, this invention further employs a temperature scaling method for post-processing the softmax output. The temperature scaling operation is as follows:

[0115]

[0116] in The temperature parameter was optimized by minimizing the negative log-likelihood loss on the validation set. After calibration, the expected calibration error (ECE) between the confidence level and the actual accuracy of the diagnostic model output can be reduced to below 0.05, effectively improving the reliability of the diagnostic results.

[0117] A training sample augmentation method based on fault physics mechanisms is used to expand the training sample library and solve the problem of scarce fault samples in practical engineering. This method establishes a simulation model based on the physical generation mechanism of partial discharge, uses the simulation to generate synthetic samples, and performs feature annotation. In a specific embodiment of the invention, the corona discharge simulation uses a needle-plate electrode model with a needle electrode radius of 0.5 mm and a plate electrode of 100 mm diameter. The electrode spacing is 20 mm, and the electric field calculation uses an axisymmetric finite element model with a mesh size of 5000 elements. The surface discharge simulation uses a rod-plate electrode model with a rod electrode of 10 mm diameter and 100 mm length, tilted at 30°. The levitation discharge simulation uses a suspended metal particle model with metal particles of 2 mm diameter and adjustable position. The particles and surrounding conductors form a floating potential conductor, and their charging and discharging process is described by a combined physical mechanism of field emission and micro-discharge. The air gap discharge simulation uses a bubble sandwich model with a bubble diameter of 5 mm and a sandwich thickness of 1 mm. The initial electrons for discharge inside the bubble are provided by background radiation and wall detrapping. The internal discharge simulation uses an internal hole model with a hole diameter of 8 mm, located 10 mm deep inside the insulation material. The conductivity of the hole wall changes dynamically with the number of discharges, simulating the increase in conductivity during the insulation degradation process.

[0118] The electric field calculations for each simulation model were performed by solving the Poisson equation using the finite element method.

[0119]

[0120] in Where is the dielectric constant. Let be the electric field intensity vector. Let be the space charge density. The sound field calculation uses the finite-difference time-domain method to solve the sound wave equation:

[0121]

[0122] in For sound pressure, The speed of sound is given. In the corona discharge simulation, discharge is triggered when the electric field strength reaches 3 kV / mm. The calculated rise time of the discharge current pulse is 5 ns, the full width at half maximum (FWHM) is 20 ns, and the pulse repetition frequency is related to the applied voltage frequency. The spectrum range of ultra-high frequency electromagnetic signals is 300 MHz to 3 GHz, and the spectrum range of acoustic signals is 20 kHz to 200 kHz. To ensure the consistency of the characteristic distribution between the simulated synthetic samples and the actual collected samples, this invention introduces the Maximum Mean Difference (MMD) index to evaluate the quality of the generated samples. The MMD calculation formula is:

[0123]

[0124] in For the class of unit ball functions in the reproducing kernel Hilbert space, To synthesize the sample set, For the real sample set, and These are the corresponding sample feature vectors. and These represent the number of synthetic samples and the number of real samples, respectively. When the MMD value is below a preset threshold, it indicates that the synthetic samples are statistically indistinguishable from the real samples and can be included in the training sample library; otherwise, the simulation parameters need to be adjusted and the samples regenerated. This quality screening mechanism ensures the effectiveness and representativeness of the augmented samples.

[0125] During the synthesis sample generation process, the simulated signal and the background noise signal are superimposed according to a set signal-to-noise ratio (SNR). In a specific embodiment of the invention, the SNR is set to five levels: 0dB, 5dB, 10dB, 15dB, and 20dB. The number of enhanced samples at each level is three times the number of actual fault samples. For each fault type actually collected, 50 samples are generated, resulting in a total of 750 enhanced samples. The background noise signal is extracted from the data collected during the normal operation of the ring mains unit, with a duration of 100 seconds, and includes various components such as electromagnetic interference from switching operations, power grid carrier signal interference, and environmental electromagnetic noise. The SNR is defined as:

[0126]

[0127] in To simulate the average power of the signal, The average power of the background noise is denoted as . Before superposition, the background noise signal is segmented and its statistical characteristics (mean, variance, power spectral density) are ensured to be consistent with the original long-term background noise.

[0128] The multi-channel signal synchronous acquisition and transmission method employs a sampling clock synchronization scheme based on the GPS precise time synchronization protocol and a data transmission scheme based on the IEEE 1588 precise time protocol. In the sampling clock synchronization scheme, each sensor acquisition unit embeds a high-precision GPS receiver module with a receiving sensitivity better than -160dBm and a time accuracy better than 20ns. The GPS receiver module extracts precise time stamps and 1PPS pulse signals from satellite signals, and the sampling clock of the acquisition unit is disciplined based on the 1PPS signal. The discipline process is implemented using a digital phase-locked loop (DPLL), and its loop filter transfer function is designed as follows:

[0129]

[0130] in For proportional gain, For integral gain, For unit delay operators, Let be the complex variable of the Z-transform. The loop bandwidth of the phase-locked loop is set to 0.1Hz to ensure sufficient noise suppression and stable phase tracking. Through this discipline mechanism, the internal clock frequency of each acquisition unit is locked to the long-term stability of GPS time, while utilizing the low phase noise characteristics of the local temperature-controlled crystal oscillator to maintain short-term stability, thus achieving synchronization between the sampling clock and GPS time. In a specific embodiment of the invention, the sampling time deviation between channels is controlled within 100 nanoseconds to meet the requirements of multi-physics signal fusion processing. The mathematical model of the synchronization error can be described as:

[0131]

[0132] in This is due to GPS receiver time calculation error. The propagation delay of the antenna cable can be reduced to less than 1 nanosecond by antenna cable delay calibration.

[0133] The data transmission scheme uses industrial Ethernet as the transmission medium and supports the IEEE 1588 precision time protocol. Data frames include a timestamp field with an accuracy better than 50 nanoseconds. The protocol stack employs a dual-step mode to ensure timestamp accuracy. In dual-step mode, after the master clock sends a Sync message, it immediately sends a Follow_Up message containing the Sync message's sending time. The slave clock calculates the clock offset based on the arrival times of the two messages. The clock offset calculation uses the least squares method to filter multiple measurement results; the mathematical model is as follows:

[0134]

[0135] in Master clock transmission time, To receive time from the clock, This refers to propagation delay. Propagation delay is measured using a delayed request-response mechanism:

[0136]

[0137] in To send a delay request time from the clock, The master clock receives the delay request time. Data transmission delay jitter is less than 10 microseconds, the network bandwidth is 1000Mbps, and the single-channel data transmission rate is approximately 80Mbps. In a specific embodiment of the invention, a master-slave clock architecture is adopted, with the master clock deployed in the control room and the slave clocks deployed in each sensor acquisition unit. The time synchronization period between the master and slave clocks is 1 second, which is much smaller than the clock drift time constant, ensuring the maintenance of synchronization accuracy.

[0138] The adaptive noise threshold estimation method is based on statistical signal processing theory and employs a noise threshold estimation method based on the Gumbel distribution. The Gumbel distribution is a model in extreme value theory describing the asymptotic distribution of the maximum value of independent and identically distributed random variables, suitable for modeling the statistical characteristics of background noise peaks. In a specific embodiment of this invention, background noise signals are first collected under normal operating conditions of the ring main unit for 30 minutes at a sampling frequency of 10MHz. Then, statistical analysis is performed on the background noise signals to extract their probability distribution parameters. Specifically, the background noise signal is divided into 1000 data segments of 30,000 points each, and the maximum value of each data segment is calculated to form an extreme value sequence. The extreme value sequence is then fitted with a Gumbel distribution, and its probability density function is... The cumulative distribution function is The location parameters of the Gumbel distribution are estimated using the maximum likelihood estimation method. and scale parameters The maximum likelihood estimate is solved using the likelihood function:

[0139]

[0140] right and Taking partial derivatives and setting them to zero yields the estimated equation, which is then solved using numerical iteration. The length of the extreme value sequence is given. Based on the Gumbel distribution fitting results, the noise threshold at a given confidence level is calculated. In a specific embodiment of the invention, the confidence level is set to 99%, corresponding to the cumulative probability. The value is 0.99, and the formula for calculating the noise threshold is:

[0141]

[0142] This threshold theoretically guarantees that the probability of the background noise peak exceeding the threshold is only 1% when there is no discharge signal, thus keeping the false alarm rate at an acceptable low level.

[0143] To adapt to the slowly changing statistical characteristics of background noise over time, this invention further introduces a sliding window mechanism to dynamically update the Gumbel distribution parameters. Specifically, the window length is set to 30 minutes, and the window slides every 10 minutes. The extreme value sequence is updated using newly acquired background noise data, and the parameters are re-estimated. and The noise threshold is updated in real time. The weighted calculation formula for parameter updates uses an exponential forgetting factor:

[0144]

[0145] in and Estimate parameters for the current window. and These are historical parameters. This dynamic update mechanism allows the noise threshold to track changes in the background electromagnetic environment, maintaining stable detection performance. The peak value of the real-time acquired signal is compared with the noise threshold; if the signal peak value exceeds the threshold, a partial discharge signal is determined to exist; otherwise, it is considered background noise. In actual testing, this method has a false alarm rate of less than 1% for background noise and a detection probability of over 95% for partial discharge signals.

[0146] The insulation degradation assessment method is based on multi-dimensional feature fusion results and employs a multi-index comprehensive evaluation model to assess the degree of insulation degradation. In a specific embodiment of this invention, the feature parameters include four indices: discharge quantity, discharge frequency, discharge phase distribution, and discharge rate. These four indices comprehensively reflect the insulation state from four perspectives: discharge intensity, discharge frequency, discharge timing, and discharge rate. The correlation between the indices is verified for consistency using Kendall's coefficient of harmony to ensure the rationality of the index system.

[0147] An exponential function is used to map the discharge level to the degree of degradation. The mapping formula is as follows:

[0148]

[0149] in The discharge quantity is expressed in pC. The physical basis of the exponential mapping lies in the approximately exponential relationship between the aging degree of the insulation material and the discharge quantity. When the discharge quantity is small, the insulation degradation is in a slow development stage; when the discharge quantity exceeds a certain threshold, the degradation rate accelerates sharply. The exponential function can appropriately describe this nonlinear relationship. A logarithmic function is used to map the number of discharges to the degree of degradation. The mapping formula is:

[0150]

[0151] in The number of discharges represents the degree of degradation. The logarithmic function mapping reflects the "diminishing marginal returns" effect, meaning that when the number of discharges is low, increasing the number of discharges has a significant impact on the degree of degradation; however, when the number of discharges is already high, the contribution of new discharges to the incremental degradation decreases, which is consistent with the phenomenon of insulation cumulative damage saturation. The Euclidean distance is used to measure the relationship between the discharge phase distribution and the degree of degradation. The measured phase distribution is compared with a standard phase distribution, which is a typical phase distribution of a brand-new device under rated voltage. A larger distance value indicates a more severe degree of degradation. The formula for calculating the Euclidean distance is:

[0152]

[0153] in and The measured and reference phase histograms are respectively at the 1st... Phase intervals The value, This represents the total number of phase intervals. A linear function mapping is used between discharge frequency and degradation degree, with the mapping formula being: ,in This represents the discharge frequency, measured in cycles per second. Linear mapping is suitable for discharge frequency, a relatively direct indicator of discharge activity density.

[0154] The weight coefficients of each feature parameter are determined using the analytic hierarchy process (AHP). In a specific embodiment of this invention, the judgment matrix is ​​constructed using the nine-scale method, and the elements of the judgment matrix... Indicators relative to indicators The importance of, satisfying Based on research into the mechanism of partial discharge's impact on insulation degradation, the discharge quantity is the most direct indicator characterizing the energy of a single discharge and the degree of insulation damage, therefore it is given the highest weight; the number of discharges and the discharge frequency reflect the cumulative effect, and are given the next highest weight; the discharge phase distribution reflects the statistical characteristics of the discharge, and is given a relatively low weight. The constructed judgment matrix is ​​as follows:

[0155]

[0156] Calculate its largest eigenvalue Consistency indicators Random consistency ratio (in For the number of indicators, (This is the average random consistency index), satisfying the consistency requirements. After mathematical calculation and normalization, the feature vector yields a strictly matched judgment matrix with the following weight allocation: discharge quantity weight 0.54, discharge frequency weight 0.22, discharge phase distribution weight 0.12, and discharge frequency weight 0.12. The formula for calculating the comprehensive insulation degradation score is:

[0157]

[0158] The scoring range is from 0 to 100 points. Based on the scoring results, the degree of insulation degradation is divided into four levels: a score of 80 or above is considered normal, 60 to 80 is slightly degraded, 40 to 60 is moderate, and below 40 is severe. This scoring system provides maintenance personnel with intuitive quantitative data on insulation status, supporting condition-based maintenance decisions.

[0159] To verify the technical effectiveness of the method of this invention, a certain type of ring main unit was tested under laboratory conditions. The test object was a 10kV ring main unit, internally equipped with three-phase high-voltage cables and vacuum circuit breakers, with SF6 gas as the insulating medium. Five fault types were set up for testing: corona discharge was simulated by needle electrodes, surface discharge was simulated by rod electrodes, suspended discharge was simulated by metal particles, air gap discharge was simulated by bubble interlayers, and internal discharge was simulated by internal holes. Ten test samples were set for each fault type, for a total of 50 test samples. The fault diagnosis results using the method of this invention were as follows: the diagnostic accuracy rate was 95% for corona discharge test samples; 93% for surface discharge test samples; 96% for suspended discharge test samples; 91% for air gap discharge test samples; and 95% for internal discharge test samples; the overall average diagnostic accuracy rate of the test samples reached 94%.

[0160] The comparative study employed a traditional Fourier transform feature extraction method and a backpropagation (BP) neural network diagnostic model for fault diagnosis. Fourier transform feature extraction only extracted the frequency domain features of the signal, including four parameters: dominant frequency, frequency centroid, frequency variance, and frequency entropy, resulting in a 4-dimensional feature set. The BP neural network used a 3-layer structure: 4 nodes in the input layer, 20 nodes in the hidden layer, and 5 nodes in the output layer. Training parameters were set as follows: learning rate of 0.01, maximum training iterations of 1000, and training termination condition of an error less than 0.001. The comparative study also used 50 test samples for diagnostic testing. The diagnostic results were as follows: corona discharge diagnostic accuracy was 75%, surface discharge diagnostic accuracy was 70%, suspension discharge diagnostic accuracy was 78%, air gap discharge diagnostic accuracy was 65%, internal discharge diagnostic accuracy was 72%, and the overall average diagnostic accuracy of the test samples was only 72%.

[0161] Test data shows that the average diagnostic accuracy of the method of this invention is 94%, with diagnostic accuracies of 95%, 93%, 96%, 91%, and 95% for the five fault types, respectively, significantly better than the average diagnostic accuracy of 72% of the comparative method. The diagnostic accuracy of the method of this invention is significantly higher than that of the comparative method for all fault types. The main reason is that this invention uses a multi-physics field sensor array to acquire multi-dimensional signals (especially solving the high reproducibility sampling of ultra-high frequency signals and transient ground voltage signals through hardware envelope detection), uses empirical mode decomposition to extract multi-scale features of signals, uses a deep belief network to learn deep feature representations using an unsupervised pre-training mechanism, and uses a training sample augmentation method based on fault physics mechanisms to expand the training sample library.

[0162] Furthermore, the performance of each technical module in the method of this invention was verified. In the performance test of the adaptive filtering and denoising module, a noisy signal with a signal-to-noise ratio (SNR) of 5dB was used for testing. The SNR of the signal before denoising was 5dB. After processing with the adaptive singular value decomposition (SVD) denoising method of this invention, the SNR of the signal was improved to 22dB, demonstrating a significant denoising effect. In the performance test of the feature extraction module, the feature dimension extracted using the empirical mode decomposition (EMD) method was 85 dimensions. After feature selection, 25 key features were retained. The classification accuracy decreased by only 1.5% after feature selection, effectively achieving feature dimensionality reduction and redundancy elimination.

[0163] In the performance testing of the deep belief network diagnostic model, comparative experiments were conducted on the network structure parameters. The results show that the deep belief network composed of 4 layers of restricted Boltzmann architecture exhibits the best performance, with an 8% improvement in accuracy compared to the 3-layer structure and a 3% decrease in accuracy compared to the 5-layer structure. Too few layers lead to insufficient feature learning, while too many layers result in overfitting. In the performance testing of the training sample augmentation method, the addition of augmented samples increased the number of training samples from 250 to 1000, and the model validation set accuracy increased from 82% to 94%, an improvement of 12%. The effectiveness of the training sample augmentation method is also reflected in the model's generalization ability to unknown operating conditions: when using only the original training samples, the model's accuracy in identifying discharge signals under newly applied voltage levels was 75%; after adding augmented samples, the accuracy increased to 90%, indicating that the augmented samples generated by physical simulation effectively cover the range of feature variations caused by changes in operating conditions.

[0164] In the performance testing of the multi-channel signal synchronous acquisition and transmission method, an oscilloscope was used to measure the sampling time deviation between each channel. Test results show that after using GPS precise time synchronization, the sampling time deviation between channels is 85 nanoseconds, meeting the design requirement of 100 nanoseconds. After using the IEEE 1588 precise time protocol for data transmission, the timestamp accuracy of the data frame is 42 nanoseconds, and the data transmission delay jitter is 8 microseconds, meeting the design requirements.

[0165] In the performance test of the adaptive noise threshold estimation method, extreme value statistical analysis was performed on a 30-minute background noise signal. The estimated location parameter of the Gumbel distribution was 2.3 mV, and the estimated scale parameter was 0.8 mV. Under a 99% confidence level, the calculated noise threshold was 6.0 mV. Using this threshold to detect the test signal, the false alarm rate was 0.8% and the detection probability was 96.5% under static conditions. By dynamically updating the threshold, after 24 hours of continuous monitoring, the false alarm rate remained below 1.2%, and the detection probability remained above 95%, verifying the effectiveness of the adaptive mechanism under long-term operating conditions.

[0166] In the performance test of the insulation degradation assessment method, a comprehensive insulation degradation score was calculated for each of the five fault types. The test results showed that corona discharge scored 72 points, belonging to the slight level; surface discharge scored 65 points, also belonging to the slight level; floating discharge scored 78 points, also belonging to the slight level; air gap discharge scored 55 points, belonging to the moderate level; and internal discharge scored 48 points, also belonging to the moderate level. The assessment results are consistent with the actual insulation degradation state, indicating that the assessment method has high accuracy.

[0167] The technical solution of this invention achieves multi-dimensional synchronous acquisition of partial discharge signals from ring main units (RNBs) through a distributed multi-physics sensor array. It effectively suppresses background electromagnetic interference through adaptive filtering and denoising, extracts multi-scale features of the signal through empirical mode decomposition, realizes intelligent fault type diagnosis through deep belief networks, expands the training sample library through training sample enhancement, ensures signal synchronization through multi-channel synchronous acquisition and transmission, achieves effective signal detection through adaptive noise threshold estimation, and realizes quantitative evaluation of insulation status through insulation degradation assessment. The entire technical solution constitutes a complete online monitoring and fault diagnosis system for partial discharge in RNBs, which can effectively improve the safety and reliability of RNB operation.

[0168] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0169] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for online monitoring and fault diagnosis of partial discharge in a ring main unit, characterized in that: Includes the following steps: Step S1: A distributed multiphysics sensor array is used to synchronously acquire partial discharge signals inside the ring main unit. The sensor array includes an ultrasonic sensor, a transient ground voltage sensor, and an ultra-high frequency sensor. The transient ground voltage sensor and the ultra-high frequency sensor each integrate a detection circuit to convert the high-frequency raw signals they receive into low-frequency envelope signals that reflect the outer envelope of the discharge pulse energy. The detection circuit performs nonlinear rectification and low-pass filtering on the received signals to remove the high-frequency carrier and extract the low-frequency envelope signal. The signals acquired by each sensor are sampled at the same sampling frequency, and the bandwidth of the low-frequency envelope signal is less than half of the sampling frequency. Step S2: Perform adaptive filtering, noise reduction, and standardization preprocessing on the acquired signals to obtain the preprocessed signal; Step S3: Perform empirical mode decomposition on the preprocessed signal to obtain multiple intrinsic mode function components; wherein, the empirical mode decomposition adopts a dual screening stopping criterion: when the number of extreme points of the current residual component is less than or equal to 2, and the mutual information between the residual component and the preprocessed signal is lower than a preset threshold, the decomposition is terminated. Mutual information The calculation method is as follows: in, For residual components, The input signal is standardized. These are the amplitude quantization box indices for the residual component and the input signal, respectively. For joint probability distribution, and It represents a marginal probability distribution; Step S4: Extract the multi-scale features of each intrinsic mode function component and perform feature fusion to form a fused feature vector; Step S5: Input the fused feature vector into the trained deep belief network fault diagnosis model and output the partial discharge fault type. The deep belief network fault diagnosis model is composed of multiple restricted Boltzmann machines stacked together. Its training process includes unsupervised layer-by-layer pre-training and supervised fine-tuning. In the supervised fine-tuning phase, the loss function used is... Introduce an inter-class margin penalty term based on the cross-entropy loss: in, For real labels, To predict probabilities, and For the classifier corresponding to the first Class and First The weight vector of the fault class, This represents the interval penalty coefficient.

2. The method for online monitoring and fault diagnosis of partial discharge in a ring main unit according to claim 1, characterized in that: The steps of adaptive filtering denoising include: The original signal is divided into continuous and partially overlapping analysis frames, and a Hankel matrix is ​​constructed for each frame of data. Singular value decomposition is performed on the Hankel matrix, and a normalized relative threshold based on the maximum singular value is adaptively determined according to the signal-to-noise ratio of the signal, and the principal singular values ​​that satisfy the relative threshold are retained. The matrix is ​​reconstructed using the retained principal singular values ​​and their corresponding singular vectors, then restored to a one-dimensional signal by anti-diagonal averaging, and adjacent frames are spliced ​​together using the overlapping addition method to obtain the denoised signal.

3. The method for online monitoring and fault diagnosis of partial discharge in a ring main unit according to claim 1, characterized in that: Multi-scale features include time-domain statistical features, frequency-domain statistical features, and empirical mode decomposition features; Time-domain statistical characteristics include peak factor, impulse factor, waveform factor, margin factor, and correlation coefficient between adjacent intrinsic mode function components; Frequency domain statistical characteristics include frequency domain centroid, frequency domain variance, frequency domain entropy, and dominant frequency; Empirical mode decomposition features include the energy distribution ratio of each intrinsic mode function component, the ratio of adjacent energies, and the average period. Feature fusion includes: using a recursive feature elimination method to filter all extracted features, and then serially concatenating the filtered features to form a fused feature vector.

4. The method for online monitoring and fault diagnosis of partial discharge in a ring main unit according to claim 1, characterized in that: The training samples for the fault diagnosis model are enhanced in the following ways: Finite element simulation models based on physical mechanisms are established for different discharge types, and simulation signals are generated. The simulated signal is superimposed with background noise at different signal-to-noise ratios to obtain a synthetic sample; Calculate the maximum mean difference between the synthetic sample and the real collected sample, and only include the synthetic sample with the maximum mean difference below a preset threshold into the training sample library.

5. The method for online monitoring and fault diagnosis of partial discharge in a ring main unit according to claim 1, characterized in that: The implementation of synchronous data acquisition includes: Each sensor acquisition unit uses the GPS precise time synchronization protocol to discipline the sampling clock, so that the sampling clocks between channels are synchronized; Data frames are transmitted via industrial Ethernet that supports the IEEE 1588 precision time protocol, and timestamps are carried in the data frames to achieve time alignment of multi-channel signals.

6. The method for online monitoring and fault diagnosis of partial discharge in a ring main unit according to claim 1, characterized in that: When detecting partial discharge signals, an adaptive noise threshold estimation is used: Background noise signals are collected, and a Gumbel distribution model is established to fit the probability distribution of noise peaks. The cumulative distribution function is as follows: in, For position parameters, The scale parameter is used; the noise threshold is calculated based on the Gumbel distribution model and the preset signal level; the position parameter and scale parameter are dynamically updated using a sliding window to track changes in background noise; the peak value of the real-time acquired signal is compared with the current noise threshold, and if it exceeds the noise threshold, it is determined that there is a partial discharge signal.

7. The method for online monitoring and fault diagnosis of partial discharge in a ring main unit according to claim 1, characterized in that: After outputting the partial discharge fault type, an insulation degradation assessment step is also included: Discharge quantity, number of discharges, discharge phase distribution, and discharge frequency are extracted as evaluation indicators; The weight coefficients of each evaluation indicator were determined using the analytic hierarchy process (AHP). The discharge quantity was mapped using an exponential function, the number of discharges was mapped using a logarithmic function, the discharge phase distribution was mapped using an Euclidean distance metric, and the discharge frequency was mapped using a linear function to obtain scores for each index. The insulation degradation comprehensive score is obtained by weighting and summing the scores of each indicator based on the weighting coefficients, and the insulation degradation level is classified according to the comprehensive insulation degradation score.

Images