A method and system for separating partial discharge signals of a cable branch box

Abstract

The application discloses a cable branch box partial discharge signal separation method and system, and belongs to the field of power system distribution network insulation state monitoring. The method comprises the following steps: performing multi-dictionary pattern component analysis decomposition on the collected partial discharge signal to obtain each component and decomposition residual; calculating the residual energy ratio, and inputting the residual into a bistable random resonance system for weak signal enhancement when the residual exceeds the threshold; superimposing the enhanced signal back to the pulse component channel to perform secondary decomposition iteration and calculate the signal-to-noise ratio gain; constructing the historical interference component sparse coefficient vector into a meta-learning task through density clustering; performing model-independent meta-learning double-loop optimization, and dynamically determining the learning rate of the outer loop by the signal-to-noise ratio gain through an S-shaped mapping function; and finally outputting the discharge type classification result and the early warning. The application realizes the adaptive separation of weak partial discharge signals in a dynamic and complex electromagnetic environment through the nonlinear mapping closed-loop driving of specific internal parameters among the three algorithms.

Description

Technical Field

[0001] This application belongs to the field of insulation status monitoring of power system distribution networks, and specifically relates to a method and system for separating partial discharge signals in cable branch boxes. Background Technology

[0002] Cable distribution boxes are key equipment in urban medium-voltage power distribution networks, enabling the distribution of electrical energy from main cable lines to branch lines. During long-term operation, weak points in the insulation inside the box may cause partial discharge, requiring early detection through online monitoring.

[0003] Current partial discharge monitoring systems for cable branch boxes typically employ a cascaded processing architecture: signal denoising / separation → weak signal enhancement → pattern recognition and classification. The signal separation stage often utilizes methods such as wavelet transform, empirical mode decomposition, or morphological component analysis; the weak signal enhancement stage may employ nonlinear methods such as stochastic resonance; and the pattern recognition stage may utilize methods such as deep learning. However, in this cascaded architecture, each processing stage operates as an independent module, lacking a parameter-level feedback mechanism: the signal separation module cannot automatically adjust the operating parameters of subsequent enhancement modules based on separation quality; the enhancement effect of the signal enhancement module cannot automatically drive the classification module to adjust its learning behavior; and the classification module cannot automatically update its adaptation strategy based on changes in the historical signal environment.

[0004] This modular, parameter-independent processing architecture leads to two drawbacks. First, the operating parameters of each module require manual pre-setting or offline calibration, making it impossible to track the dynamic changes in the electromagnetic environment at the cable branch box (the type and intensity of interference sources continuously change with weather conditions, load status, and the operating conditions of nearby equipment). This causes a sharp degradation in system performance under non-calibrated conditions, resulting in a high false alarm rate. Second, there is no quality feedback channel between processing stages. When signal separation is insufficient (meaningful weak components still exist in the residuals), subsequent stages cannot perceive this information and make corresponding adjustments. When the signal enhancement effect is uncertain, the classification model cannot adjust its learning conservatism accordingly, making it prone to overfitting to noise or under-responding to real new patterns.

[0005] The technical essence of the above problem lies in the fact that the existing processing architecture lacks nonlinear mapping and closed-loop driving relationships for the specific internal parameters between algorithms. The various algorithm modules cannot form an organic whole with mutually driving parameters to achieve fully adaptive operation in dynamic environments. This is the fundamental problem hindering the online monitoring of cable branch box partial discharge from the laboratory to large-scale maintenance-free engineering deployment. Summary of the Invention

[0006] In view of the above technical problems, this application provides a method and system for separating partial discharge signals in cable branch boxes. By establishing a nonlinear mapping closed-loop driving relationship of specific internal parameters among three algorithms—morphological component analysis, adaptive stochastic resonance, and model-independent element learning—the three algorithms, which originally operated independently, form an organic whole with mutually driving parameters, realizing fully adaptive operation of the processing link. This eliminates the dependence of each processing link on manual parameter adjustment and achieves adaptive separation of weak partial discharge signals in dynamic and complex electromagnetic environments without manual intervention.

[0007] According to one aspect of this application, a method for separating partial discharge signals in a cable branch box is provided, comprising: segmenting and windowing the partial discharge signal collected by the cable branch box to form a signal window; performing morphological component analysis decomposition based on multiple dictionaries on the signal window, using pulse dictionary, periodic dictionary and transient dictionary to sparsely represent the corresponding morphological components respectively, and obtaining the sparse coefficient vector of each component and the decomposition residual through block coordinate relaxation iteration; calculating the ratio of the decomposition residual energy to the original signal energy as the residual energy ratio; when the residual energy ratio exceeds a preset threshold, inputting the residual signal into a bistable stochastic resonance system for enhancement, wherein the potential well depth parameter in the Langevin equation of the bistable stochastic resonance system is determined by the over-limit amplitude of the residual energy ratio through a power function mapping, and the potential well spacing parameter is determined by the standard deviation of the residual signal, so that the barrier height and the amplitude of the signal to be enhanced are aligned. The process involves: matching the stochastic resonance output signal back to the pulse component channel for secondary decomposition iteration of morphological component analysis, separating the enhanced weak partial discharge pulses, and calculating the signal-to-noise ratio (SNR) gain before and after stochastic resonance processing; combining the sparse coefficient vectors of interference components obtained from morphological component analysis of historical time periods as interference fingerprints, dividing them into multiple interference pattern categories through density clustering, with each category forming a learning task for model-independent meta-learning; performing a dual-loop optimization of model-independent meta-learning, where the inner loop updates the task-specific parameters for each task with a small number of support sample gradients, and the outer loop's meta-learning rate is dynamically determined by the SNR gain through a sigmoid mapping function, with a higher SNR gain resulting in a higher meta-learning rate; inputting the pulse components obtained from the secondary decomposition iteration into the meta-learning-optimized classification network, outputting the discharge type classification result, and triggering an early warning when the discharge is determined to be internal partial discharge and the confidence level exceeds a threshold.

[0008] In some embodiments of this application, the morphological component analysis decomposition based on multiple dictionaries includes: using a Dirac dictionary as the pulse dictionary to sparsely represent the partial discharge pulse components with isolated spike morphology using a set of unit pulse basis functions; using a discrete cosine transform dictionary as the periodic dictionary to sparsely represent the periodic narrowband interference components of communication carriers and higher harmonics; and using a Gabor atom dictionary as the transient dictionary to sparsely represent the damped oscillation transient interference components of switching operations. The signal window is modeled as the linear superposition of the corresponding components of the above three dictionaries and the sum of the residuals. The objective function is to minimize the sum of the l1 norms of each sparse coefficient vector, and the constraint is that the l2 norm of the reconstruction error does not exceed an allowable threshold. The sparse coefficient vectors of each component are iteratively optimized alternately through block coordinate relaxation until convergence.

[0009] In some embodiments of this application, the potential well depth parameter of the bistable stochastic resonance system is determined as follows: A power-law nonlinear mapping of the difference between the residual energy ratio and a preset threshold is superimposed on a preset base parameter. The power exponent is greater than zero and less than one, such that a larger residual energy ratio results in a deeper potential well. The potential well spacing parameter is determined by dividing the potential well depth parameter by the square of the product of the residual signal standard deviation and the scale matching coefficient, such that the potential well spacing given by the square root of the ratio of the potential well depth parameter to the potential well spacing parameter is within the range of 1-3 times the residual signal amplitude.

[0010] The bistable stochastic resonance system uses the fourth-order Runge-Kutta method to numerically solve the Langevin equation, with a solution step size no greater than 1 / 10 of the sampling period.

[0011] In some embodiments of this application, the learning rate of the outer loop element is dynamically determined by the signal-to-noise ratio gain through a sigmoid mapping function, including: using the base learning rate multiplied by the sigmoid mapping function value as the outer loop element learning rate, where the independent variable of the sigmoid mapping function is the product of the mapping slope parameter and the difference between the signal-to-noise ratio gain and the gain threshold. When the signal-to-noise ratio gain is lower than the gain threshold, the output value of the sigmoid mapping function approaches zero, the learning rate of the outer loop element approaches zero, and the classification network parameters remain stable; when the signal-to-noise ratio gain is significantly higher than the gain threshold, the learning rate of the outer loop element approaches the base learning rate, and the classification network quickly adapts to new discharge modes.

[0012] In some embodiments of this application, the periodic component sparse coefficient vector and the transient component sparse coefficient vector obtained after morphological component analysis decomposition of each signal window are concatenated to form the interference fingerprint vector for that time period. Using the cosine distance between interference fingerprint vectors as a similarity measure, a density-based clustering method is employed to classify the historically accumulated interference fingerprint vectors into several interference pattern categories. When the distance between the interference fingerprint vector of a new time period and any existing category exceeds the neighborhood radius, a new interference pattern category is created, and a corresponding meta-learning task is established, allowing the task distribution to continuously expand with environmental changes.

[0013] In some embodiments of this application, the preset threshold is determined as follows: during the power outage maintenance of the cable branch box, pure background noise signals are collected, the same morphological component analysis decomposition is performed on multiple pure noise signal windows respectively, and the residual energy ratio of each window is calculated. The mean of the obtained residual energy ratio sequence plus twice the standard deviation is used as the preset threshold.

[0014] In some embodiments of this application, the number of supporting samples for each task is three to five, the number of inner loop gradient update steps is 1 to 3, and the inner loop learning rate is a fixed value. The classification network is a lightweight convolutional network, containing three convolutional layers and one fully connected output layer. The total number of network parameters is controlled within 50,000. The input is a phase-resolved statistical map of the pulse components (a two-dimensional map with 36 phase intervals × 20 amplitude intervals), and the output is the probability distribution of each discharge type. After manual confirmation, the classification results are added to the supporting sample set of the corresponding task as newly labeled samples, realizing online incremental task updates.

[0015] According to another aspect of this application, a partial discharge signal separation system for a cable branch box is provided, comprising: a signal acquisition module configured to simultaneously acquire partial discharge signals via a high-frequency current sensor installed on the grounding wire of the cable branch box and a transient ground voltage sensor on the outer shell of the box, and perform analog-to-digital conversion and segmented windowing to form a signal window; a morphological component analysis module configured to perform multi-dictionary sparse decomposition based on pulse dictionary, periodic dictionary and transient dictionary on the signal window to obtain each morphological component and decomposition residual, and calculate the residual energy ratio; and an adaptive stochastic resonance module configured to be activated when the residual energy ratio exceeds a preset threshold, receive the residual signal and perform weak signal enhancement through a bistable nonlinear system, wherein the potential well depth parameter is determined by the residual... The energy ratio exceeding the limit is determined by a power function mapping, and the potential well spacing parameter is determined by the residual signal standard deviation. The iterative re-decomposition module is configured to feed back the output signal of the adaptive stochastic resonance module to the pulse component channel to perform secondary morphological component analysis decomposition and calculate the signal-to-noise ratio gain. The task construction module is configured to combine the sparse coefficient vectors of interference components from different historical time periods as interference fingerprints and divide them into multiple interference mode categories through density clustering. Each category is constructed as a meta-learning task. The meta-learning classification module is configured to perform a double-loop optimization of model-independent meta-learning. The outer loop meta-learning rate is dynamically determined by the signal-to-noise ratio gain through a sigmoid mapping function, and the pulse components obtained from the secondary decomposition are input into the optimized classification network to output the discharge type and warning information.

[0016] In some embodiments of this application, the morphological component analysis module and the adaptive stochastic resonance module are coupled unidirectionally through the residual energy ratio parameter; the adaptive stochastic resonance module and the iterative re-decomposition module are coupled in a closed-loop iterative manner through enhanced signal feedback; the iterative re-decomposition module and the meta-learning classification module are coupled through the signal-to-noise ratio gain parameter to achieve dynamic adjustment of the learning rate; and the task construction module and the meta-learning classification module are coupled through the interference fingerprint clustering results to achieve dynamic updating of task distribution.

[0017] In some embodiments of this application, the system is deployed on an edge computing terminal, the morphological component analysis module and the adaptive stochastic resonance module perform windowed short-time processing in units of signal windows, the meta-learning classification module performs only 1-3 steps of gradient update during the inference phase, and the length of the signal window covers at least 20 power frequency cycles.

[0018] In summary, the beneficial effects of this application are as follows: This application utilizes a real-time adaptive adjustment mechanism for the parameters of the stochastic resonant potential well through morphological component analysis and residual energy ratio. This mechanism enables the operating parameters of the stochastic resonant system to automatically match the amplitude characteristics of weak partial discharge pulses in the current signal environment. Without manual parameter adjustment, the system can achieve optimal enhancement under different interference intensities, thereby increasing the detection sensitivity of weak partial discharge signals at the pC level by more than an order of magnitude and significantly reducing the early defect false alarm rate.

[0019] This application utilizes a multi-dictionary morphological decomposition framework to separate the physical mechanisms of signal components by taking advantage of the essential differences in sparse representation between pulsed partial discharge signals and periodic and transient interference. It does not rely on the characteristic differences of signals in a single domain, and overcomes the technical obstacle of three-dimensional coupling between useful signals and interference in the time, frequency, and phase domains, thereby reducing the false alarm rate of the online monitoring system to an acceptable level.

[0020] This application establishes a direct feedback correlation between signal enhancement credibility and model learning aggressiveness by controlling the S-shaped mapping of the meta-learning outer loop learning rate through stochastic resonance signal-to-noise ratio gain. This avoids overfitting of the model to noise and ensures a rapid response to real new discharge modes, enabling the system to adapt to new interference environments with only three to five field-labeled samples.

[0021] This application employs a mechanism to dynamically construct the distribution of meta-learning tasks through historical interference fingerprint clustering. This ensures that the training task set of the meta-learning model continuously reflects the diversity and dynamic changes of actual interference patterns in the field, enabling automatic tracking and adaptation of classification capabilities to non-stationary environments without the need for periodic manual retraining. The morphological component analysis decomposition and stochastic resonance numerical solution described in this application are both windowed short-time operations. The meta-learning inference stage requires only a few gradient updates, keeping the overall computational complexity within the computing power range of edge computing hardware, thus meeting the engineering constraints of real-time performance and low power consumption for cable branch box online monitoring systems.

[0022] Terminology Explanation Morphological Component Analysis (MCA) is a signal decomposition method derived from image processing and sparse representation theory. It separates each morphological component by designing redundant dictionaries of sparse representations for signal components with different morphological features and using the l1 norm minimization solution.

[0023] Stochastic Resonance (SR): A nonlinear physical phenomenon in which a suitable amount of noise energy is transferred to a weak signal through the potential well structure of a nonlinear bistable system, thereby enhancing the signal-to-noise ratio of the weak signal output.

[0024] Model-Agnostic Meta-Learning Meta-Learning (MAML): A meta-learning framework that learns a model with optimal initial parameters for a new task through a double loop optimization, enabling it to quickly adapt to a new task with only a very small number of new task samples and through a few steps of gradient updates.

[0025] The Langevin Equation is a dynamic equation that describes the dynamics of a particle driven by random forces in a potential field. In stochastic resonance systems, it is used to describe the motion of the system's state variables in a bistable potential well.

[0026] Dirac Dictionary: A sparse representation dictionary composed of unit impulse basis functions (i.e., discrete counterparts of the Dirac delta function at each location), which has the best sparse representation capability for signals with isolated spike-like shapes.

[0027] Discrete Cosine Transform Dictionary (DCTDictionary): An orthogonal transform dictionary composed of cosine basis functions of different frequencies, which has sparse representation capability for periodic and quasi-periodic signals.

[0028] Gabor Atom Dictionary: A collection of time-frequency atoms composed of Gaussian window-modulated sinusoidal signals, which has good sparse representation ability for damped oscillating transient signals.

[0029] Block Coordinate Relaxation (BCR) is an iterative algorithm for solving multivariate optimization problems. It gradually approaches the global optimum by alternately fixing some variables and optimizing the rest.

[0030] High Frequency Current Transformer (HFCT): A ring-shaped current transformer installed on the grounding wire of a cable, which obtains the discharge signal by detecting the high frequency current pulses generated by partial discharge.

[0031] Transient Earth Voltage (TEV): A capacitively coupled sensor mounted on the surface of a metal casing. It obtains discharge information by detecting the transient voltage signal induced on the metal casing by partial discharge.

[0032] Runge-Kutta Method: A high-precision numerical solution method for ordinary differential equations. The fourth-order Runge-Kutta method achieves fourth-order accuracy in each step of the calculation through a weighted combination of four slope estimates.

[0033] The sigmoid function is a monotonically increasing smooth function that maps the real number field to the interval between zero and one. It has the characteristic of rapidly switching around a threshold and saturating far from the threshold.

[0034] Signal-to-noise ratio (SNR) gain: The ratio of the output SNR to the input SNR after signal processing, used to quantify the signal enhancement effect.

[0035] Potential well depth: The energy difference between the top of the potential barrier and the bottom of the potential well in the potential function of a bistable system, which determines the energy threshold that the system needs to overcome to transition between two steady states.

[0036] Potential well spacing: The distance between the two stable equilibrium points of a bistable system determines the amplitude range of the system's output signal.

[0037] Residual energy ratio: The ratio of the energy of the residual signal after morphological component analysis decomposition to the total energy of the original signal. It is used to assess the completeness of the decomposition and to detect possible abnormal components in the residual.

[0038] Interference fingerprint: A feature vector formed by concatenating the sparse coefficient vectors of various interference components obtained from morphological component analysis decomposition, used to characterize the compositional features of interference patterns in a signal environment at a certain time period.

[0039] Phase-Resolved Partial Discharge (PRPD) is a two-dimensional graph that statistically plots partial discharge pulses according to their corresponding power frequency phase angles. It is an important feature for identifying partial discharge types.

[0040] Density-based clustering is a clustering method that does not predetermine the number of clusters and divides clusters based on the connectivity of data point density. It can discover clusters of arbitrary shapes and identify noise points.

[0041] Edge Computing Terminal: An embedded computing device deployed at the data acquisition site, featuring low power consumption and real-time processing capabilities, used to complete signal analysis tasks at the data source end. Attached Figure Description

[0042] Figure 1 A flowchart illustrating the overall process of separating partial discharge signals in a cable branch box according to an embodiment of this application; Figure 2A schematic diagram of the three sets of deep parameter coupling relationship architectures provided in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the principle of multi-dictionary morphological component analysis and decomposition provided in the embodiments of this application. Figure 4 This is a schematic diagram illustrating the adaptive stochastic resonant potential well parameter determination and signal enhancement provided in an embodiment of this application. Figure 5 A flowchart of the random resonance enhancement signal feedback iterative re-decomposition provided in the embodiments of this application; Figure 6 This is a schematic diagram illustrating the dynamic update of the distribution of interference fingerprint clustering and meta-learning tasks provided in an embodiment of this application. Figure 7 This is a module architecture diagram of a cable branch box partial discharge signal separation system provided in an embodiment of this application. Detailed Implementation

[0043] The technical solutions of this application are described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may also be implemented in other ways different from those described herein, therefore the scope of protection of this application is not limited to the specific embodiments below.

[0044] The partial discharge signal separation method for cable branch boxes provided in this application comprises deep parameter coupling of three algorithms: morphological component analysis, adaptive stochastic resonance, and model-independent element learning, which form a closed-loop driving relationship through nonlinear mapping of specific internal parameters. For example... Figure 2 As shown, the three sets of coupling relationships are as follows: morphological component analysis residual energy ratio drives stochastic resonance potential well parameters (coupling relationship one), stochastic resonance signal-to-noise ratio gain drives meta-learning outer loop learning rate (coupling relationship two), and morphological component analysis interference fingerprint drives meta-learning task distribution (coupling relationship three).

[0045] like Figure 1 As shown, the processing flow of this method includes steps S1 to S8.

[0046] In step S1, multi-channel signal acquisition and preprocessing are performed.

[0047] In some embodiments of this application, step S1 may include at least one of steps S11 to S13.

[0048] In step S11, a high-frequency current sensor installed on the grounding wire of the cable branch box synchronously acquires the high-frequency current pulse signal generated by partial discharge. In some embodiments of this application, the high-frequency current sensor is an open-close ferrite core structure with a bandwidth range of 500kHz to 30MHz, and is installed on the grounding copper busbar or grounding lead of each outgoing cable inside the cable branch box. Simultaneously, a transient ground voltage sensor installed on the outer surface of the metal casing of the cable branch box acquires the transient voltage signal induced on the casing by the discharge. In some embodiments of this application, the transient ground voltage sensor is a planar capacitive coupling structure with a bandwidth range of 3MHz to 100MHz, and is magnetically attached to the metal panel on the side wall of the box. Both sensors are triggered for sampling by the same clock source, achieving multi-channel time synchronization.

[0049] In step S12, the analog signal acquired by the sensor is converted from analog to digital to form a discrete time sequence x(n). In some embodiments of this application, the analog-to-digital conversion sampling rate is not less than 100 MS / s and the quantization accuracy is not less than 12 bits to ensure effective sampling of the rising edge of the partial discharge pulse (usually several nanoseconds to tens of nanoseconds).

[0050] In step S13, the discrete-time sequence x(n) is segmented and windowed to form signal windows. In some embodiments of this application, the window length covers at least 20 power frequency cycles (400ms for a 50Hz system), and a Hanning window is used to suppress spectral leakage. A 50% overlap rate is set between adjacent windows to avoid missing signals at window boundaries. Each signal window serves as the basic computational unit for subsequent morphological component analysis and stochastic resonance processing.

[0051] The embodiments described above in this application ensure that the subsequent signal separation algorithm obtains complete and time-aligned input data through multi-channel synchronous acquisition and reasonable signal preprocessing, covering a sufficient length of power frequency cycle so that the phase-resolved characteristics of the discharge signal can be presented.

[0052] In step S2, morphological component analysis decomposition based on multiple dictionaries is performed.

[0053] like Figure 3 As shown, in some embodiments of this application, step S2 may include at least one of steps S21 to S24.

[0054] In step S21, a multi-dictionary system is constructed. For each signal window x(n), it is modeled as a superposition of pulse components, periodic components, transient components, and residuals.

[0055] Specifically, three types of redundant dictionaries are constructed: the first type is the Dirac dictionary, which consists of unit pulse basis functions at various positions along the signal length dimension. It has sparse representation capability for isolated peak-shaped partial discharge pulse components, meaning that any isolated pulse can be accurately represented by only one dictionary atom. The second type is the discrete cosine transform dictionary, which consists of cosine basis functions of different frequencies. It has sparse representation capability for periodic narrowband interference signals of communication base station carriers and high-order harmonics of frequency conversion equipment. The third type is the Gabor atom dictionary, which consists of Gaussian modulated sinusoidal signals with different center frequencies, bandwidths, and time positions. It has sparse representation capability for interference signals of transient pulses during switching operations and damped oscillation patterns of transient oscillations in cable lines.

[0056] In some embodiments of this application, the construction parameters of the Gabor atomic dictionary are as follows: the center frequency takes 64 values ​​logarithmically uniformly within the bandwidth range of the high-frequency current sensor (500kHz to 30MHz); the bandwidth parameter (standard deviation of the Gaussian window) takes 4 values ​​for each center frequency, so that the time-frequency resolution covers an order of magnitude range from narrowband to broadband; the time position is uniformly shifted within the signal window in steps of twice the standard deviation of the Gaussian window.

[0057] For a 400ms window and the above parameter settings, the total number of atoms in the Gabor atom dictionary is approximately 3 to 5 times the length of the signal window (i.e., 120,000,000 to 200,000,000 atoms). In the block coordinate relaxation iteration, the sparse approximation of the corresponding components in the Gabor dictionary is achieved through a matching pursuit method. That is, in each iteration, only a few atoms with the largest absolute value of the inner product with the current residual are selected (the number of selections does not exceed 100), and the inner product is calculated in batches using the analytical structure of Gaussian-modulated sinusoidal atoms through fast convolution. There is no need to perform explicit matrix multiplication operations on all atoms, so that the computational cost of this step is sublinearly related to the dictionary size, which meets the computing power constraints of the edge computing terminal.

[0058] In step S22, an optimization model for morphological component analysis is established. The signal window x(n) is represented as the sum of the weighted composite components of the sparse coefficients corresponding to each dictionary plus the residual. The objective function is to minimize the sum of the l1 norms of the sparse coefficient vectors of each component, and the constraint is that the l2 norm of the signal reconstruction error does not exceed the allowable threshold ε. In some embodiments of this application, the value of the allowable threshold ε is determined according to the noise floor level of the sensor system, and is usually set to the amplitude level corresponding to 1% to 5% of the signal window energy.

[0059] In step S23, the above optimization problem is solved using a block coordinate relaxation iterative algorithm. In each iteration, this algorithm sequentially fixes the sparse coefficients of the other two components, performs a single-dictionary sparse approximation on the current component, for example, by using hard or soft thresholding to shrink the Dirac dictionary and the Discrete Cosine Transform dictionary, and by using the matching-tracking greedy atom selection described in step S21 for the Gabor atom dictionary. Both are specific implementations of single-dictionary sparse approximation. Then, the residual is updated, and the three components are processed iteratively until convergence.

[0060] In some embodiments of this application, the convergence criterion is that the l2 norm of the change in each sparse coefficient vector between two adjacent iterations is less than a preset precision (e.g., the relative change is less than 0.1%), or the maximum number of iterations is reached (e.g., 50 times).

[0061] In step S24, the decomposition results are obtained. After iterative convergence, three sets of sparse coefficient vectors and corresponding signal component reconstructions are output: the impulse component is obtained by multiplying the Dirac dictionary with its sparse coefficient vector, the periodic component is obtained by multiplying the discrete cosine transform dictionary with its sparse coefficient vector, and the transient component is obtained by multiplying the Gabor atom dictionary with its sparse coefficient vector. The residual is the original signal minus the sum of the above three components.

[0062] The embodiments described above in this application utilize a multi-dictionary morphological component analysis framework to achieve signal separation at the physical mechanism level by leveraging the differences in sparse representations of different morphological signals under their respective dictionaries. This approach is fundamentally different from the unified threshold strategy of traditional wavelet denoising and can specifically address the coexistence problem of multi-source and multi-morphological interference in the operating environment of cable branch boxes.

[0063] In step S3, the residual energy ratio is calculated and the random resonance activation condition is determined.

[0064] In some embodiments of this application, step S3 may include at least one of steps S31 to S33.

[0065] In step S31, the ratio of the energy of the residual signal r(n) after morphological component analysis decomposition to the energy of the original signal window x(n) is calculated as the residual energy ratio ρ. The physical meaning of the residual energy ratio ρ is: the proportion of signal energy in the total energy that still cannot be effectively expressed by any dictionary atom after sufficient sparse decomposition of the three types of dictionaries.

[0066] In step S32, a preset threshold for the residual energy ratio is determined. In some embodiments of this application, the preset threshold is determined as follows: during power outage maintenance of the cable branch box or during periods when there is known no discharge activity, pure background noise signals are collected; the same multi-dictionary morphological component analysis decomposition is performed on multiple pure noise signal windows (no less than 100 windows) and the residual energy ratio of each window is calculated; the mean of the resulting residual energy ratio sequence plus twice the standard deviation is used as the preset threshold ρ. th This threshold setting method ensures that the probability of random resonance being falsely triggered in a purely noisy environment is less than 2.5%.

[0067] In step S33, the residual energy ratio ρ of the current signal window is compared with the preset threshold ρ. th Compare them. When ρ is greater than ρ th When the residual contains anomalous components that cannot be sparsely represented by the existing dictionary (e.g., weak partial discharge pulses suspected of being buried by strong noise), the subsequent stochastic resonance enhancement process is triggered. When ρ is not greater than ρ... th If the residual is determined to contain only random noise components, random resonance processing is not triggered, and the pulse components obtained in step S2 are directly sent to the subsequent classification process.

[0068] The embodiments described above in this application use the residual energy ratio as a parameter to transmit information to the stochastic resonance system through morphological component analysis, thereby achieving conditional triggering coupling between the two algorithms. This avoids the waste of computational resources and the possible introduction of pseudo-enhancements caused by performing stochastic resonance operations unnecessarily.

[0069] In step S4, adaptive stochastic resonance enhancement of weak components is performed.

[0070] like Figure 4 As shown, in some embodiments of this application, step S4 may include at least one of steps S41 to S44.

[0071] In step S41, a mathematical model of the bistable stochastic resonance system is established. The bistable stochastic resonance system is described by the Langevin equation, which describes the motion of the system's state variables in a symmetric bistable potential function. The Langevin equation takes the form that the time derivative of the system's state variables equals the negative gradient of the potential function plus the input signal.

[0072] In this application, the Langevin equation is expressed in a normalized time form, meaning that the system parameters a and b in the equation already include a time scale factor, allowing the equation to be directly discretized in numerical solution using the solution step size h as the time unit. Specifically, the discretized form of the Langevin equation is a fourth-order Runge-Kutta scheme, where each iteration corresponds to a physical time h (not greater than 1 / 10 of the sampling period). The numerical values ​​of the system parameters a and b have been scaled according to this time step, so that the terms on the right-hand side of the equation directly have the signal amplitude dimension (e.g., mV), while the time dimension is reflected in the stepping process of the Runge-Kutta iteration.

[0073] The potential function is a symmetric double-well structure in the form of a fourth-order polynomial, determined by the potential well parameters a and b. The potential function also includes the potential well depth ΔU and the potential well spacing x. m They are respectively:

[0074] ΔU=a 2 / (4b); x m =sqrt(a / b); Where a is the linear recovery coefficient and b is the nonlinear saturation coefficient.

[0075] In step S42, the potential well parameters of the stochastic resonance system are determined based on the residual energy ratio driven by morphological component analysis. This is the first set of core relationships in the deep parameter coupling of this application (hereinafter referred to as coupling relationship one). The potential well parameter a is determined by the overlimit amplitude of the residual energy ratio through a power function mapping:

[0076] a = a0·[1+k1·(ρ-ρ) th ) α1 ] Where a0 is the basic potential well parameter, k1 is the gain coefficient, ρ is the current residual energy ratio, and ρ th As a preset threshold, α1 is a power exponent with a value ranging from 0 to 1, excluding itself. α1 being less than 1 ensures a sublinear mapping relationship, meaning that when the residual energy ratio moderately exceeds the limit, the potential well depth increases rapidly, while when it significantly exceeds the limit, the increase tends to level off, thus avoiding excessive amplification of the potential well parameters that could lead to system instability.

[0077] The potential well parameter b is determined by constraining the relationship between the potential well spacing and the residual signal amplitude: b=a / (σ residual ·c1) 2 Where σ residual Let be the standard deviation of the residual signal r(n), and c1 be the scale matching coefficient. In some embodiments of this application, c1 ranges from 1 to 3, such that the potential well spacing x m =σ residual• c1 satisfies the optimal matching condition of stochastic resonance within the range of 1 to 3 times the amplitude of the residual signal, that is, the potential well spacing and the amplitude of the weak signal to be detected are on the same order of magnitude.

[0078] With the above two parameters determined, the potential well depth ΔU can be further expressed as: ΔU=a×(σ residual ·c1) 2 / 4 This expression shows that the potential well depth is simultaneously affected by the residual energy ratio (via parameter a) and the residual signal amplitude statistics (via σ). residual The joint regulation of ) enables real-time adaptive driving of the morphological component analysis decomposition quality to the stochastic resonance operating point.

[0079] In some embodiments of this application, the basic potential well parameter a0 ranges from 0.5 to 2.0, the gain coefficient k1 ranges from 1.0 to 10.0, and the power exponent α1 ranges from 0.3 to 0.7. The specific values ​​of the parameters are determined through offline calibration based on the typical noise level at the cable branch box installation site.

[0080] It should be noted that the matching of the potential barrier height and the amplitude of the signal to be enhanced mentioned here refers to a functional matching within the framework of stochastic resonance theory, rather than numerical equality. Specifically, the potential well spacing x m Set as the standard deviation σ of the residual signal residual c1 times (1 to 3 times) to ensure the system's output amplitude range (within the two steady-state equilibrium points ± x m The amplitude statistical characteristics of the residual signal (jumping between) are on the same order of magnitude as those of the residual signal.

[0081] When the amplitude of the weak partial discharge pulse is A signal Less than σ residual When the signal is submerged in noise (a prerequisite for triggering stochastic resonance), noise components can frequently drive the system to jump between the two potential wells. When a weak partial discharge pulse is superimposed on the noise, it changes the statistical distribution of the timing of the jumps, making the jumps tend to synchronize with the pulse occurrence time, thus appearing in the system output at an x-order interval. m The magnitude jump indicates the time of occurrence of the pulse event, realizing the transition from A... signal Order of magnitude to x m The magnitude of the amplitude is amplified. The potential well depth ΔU is determined by the joint parameters a and b, controlling the floor jump rate under pure noise drive, and avoiding an excessively high floor jump rate that drowns out useful pulse synchronization jump events.

[0082] In step S43, the residual signal r(n) is applied as the input signal to the bistable stochastic resonance system with the parameters already determined, and the Langevin equation is numerically solved using the fourth-order Runge-Kutta method. In some embodiments of this application, the numerical solution step size is no greater than 1 / 10 of the sampling period to ensure the stability and accuracy of the numerical integration. At a typical sampling rate of 100 MS / s for partial discharge monitoring of cable branch boxes, the sampling period is 10 ns, and the numerical solution step size is set to 1 ns.

[0083] In the 10 Runge-Kutta iterations within each sampling interval (10 ns), the input signal r(n) is processed in a zero-order hold manner, that is, the signal value r(n) of the nth sampling point is used as the input in all 10 integration steps between the nth sampling point and the (n+1)th sampling point (including the intermediate time evaluation required for k2 and k3 in the fourth-order Runge-Kutta method).

[0084] The rationale for this approach lies in the fact that the duration of the partial discharge pulse (typically tens to hundreds of nanoseconds) is much greater than the 10 ns sampling interval, the signal amplitude variation within a single sampling interval is limited, and the approximation error introduced by the zero-order hold is negligible. Furthermore, the initial condition of the state variable x(t) of the bistable stochastic resonance system is set to x(0) = 0, meaning the system's initial state is at an unstable equilibrium point at the top of the potential barrier, causing the system to fall into a potential well on its own under noise-driven conditions within the initial few sampling points.

[0085] In step S44, the output signal y of the stochastic resonance system is obtained. SR (n). When the optimal condition for random resonance is met, the weak partial discharge pulse component buried by strong noise in the residual is enhanced by the noise energy transfer effect, and in the output signal y SR (n) represents a pulse event with an amplitude significantly higher than the noise floor.

[0086] The embodiments described above in this application use a mechanism that directly drives the adaptive adjustment of stochastic resonance potential well parameters through morphological component analysis of residual energy ratio. This overcomes the fundamental limitation in traditional stochastic resonance applications where potential well parameters need to be manually searched or pre-set. The system can automatically select the optimal operating point based on the characteristics of the current signal environment, thus achieving real-time adaptive signal enhancement capability to environmental changes.

[0087] In step S5, a random resonance enhancement signal feedback morphological component analysis re-decomposition is performed.

[0088] like Figure 5 As shown, in some embodiments of this application, step S5 may include at least one of steps S51 to S53.

[0089] In step S51, the random resonance output signal y SR(n) Superimpose the pulse component channel back. Specifically, superimpose the pulse component obtained from the initial decomposition in step S2 with y SR (n) are added together to form the updated signal.

[0090] In step S52, a second decomposition iteration of morphological component analysis is performed on the updated signal. The second decomposition iteration uses the same three-dictionary system (Dirac dictionary, discrete cosine transform dictionary, and Gabor atom dictionary) and the same block coordinate relaxation iterative algorithm as step S2 to perform morphological component analysis decomposition on the updated signal.

[0091] At this point, since the random resonance has enhanced the weak pulse components that were originally buried in the residual to a higher amplitude level, the Dirac dictionary’s ability to represent these enhanced pulses sparsely is significantly improved (because the amplitude of the enhanced pulse exceeds the sparse approximation threshold in the block coordinate relaxation iteration), so that the weak partial discharge pulses that were originally missed in the residual can be effectively separated and included in the pulse components.

[0092] Meanwhile, a small number of periodic or transient components that may be mixed in the random resonance output are captured and stripped by the corresponding periodic and transient dictionaries during the secondary decomposition process, ensuring that the signal finally classified into pulse components has higher purity.

[0093] In step S53, the signal-to-noise ratio gain G before and after stochastic resonance processing is calculated. SNR The signal-to-noise ratio gain is defined as the random resonant output signal y. SR The ratio of the signal-to-noise ratio of r(n) to the signal-to-noise ratio of the input residual signal r(n):

[0094] G SNR =SNR out / SNR in ; The signal-to-noise ratio is calculated using the ratio of the pulse peak value to the root mean square value of the background.

[0095] In some embodiments of this application, SNR out For y SR (n) pulse peak value and y SR (n) The ratio of the root mean square value of background fluctuation, SNR in G is the ratio of the amplitude of the impulse component in r(n) (determined through posterior analysis) to the root mean square value of the background fluctuations in r(n). SNR A value greater than 1 indicates that random resonance has achieved effective signal enhancement, G SNR The larger the value, the more significant the enhancement effect.

[0096] In some embodiments of this application, the SNR in Real-time estimation is achieved in the following ways: Since the signal gain of a stochastic resonant system under optimal matching conditions can be derived theoretically from the potential well parameters, the estimation of the input signal amplitude can be obtained by inversely calculating the output signal amplitude and the theoretical system gain. Specifically, under the adiabatic approximation of a bistable system, when the amplitude of the weakly periodic input signal is much smaller than the barrier height, the system output amplitude approximately satisfies a linear amplification relationship with the input amplitude, and the amplification factor can be estimated by the potential well parameters a and b and the noise intensity D through the Kramose transition rate theory.

[0097] Therefore, SNR in It can be estimated through the following steps: First, detect y. SR (n) Pulse events exceeding the preset detection threshold (2-3 times the root mean square value of the output background fluctuation, typically 2 times, to balance the detection capability for medium-amplitude pulse events and the suppression of false alarms) and their peak amplitude A are determined. out ; Then, based on the current potential well parameters and the residual background noise intensity, the theoretical signal gain factor G of the bistable system is calculated. system ; The amplitude of the input pulse was then estimated to be A. in =A out / G system ; Final calculation of SNR in =A in / σ residual .

[0098] In some embodiments of this application, when a reliable SNR cannot be obtained through the above theoretical estimation... in In such cases, a simplified alternative can be adopted: directly using SNR. out As G SNR An approximate measure (i.e., let SNR) in (The reference value is 1), at which point the gain threshold G of the sigmoid mapping function is... th The corresponding adjustment is to SNR. out The baseline value. Although this simplified scheme reduces G... SNR While maintaining physical accuracy, it does not affect the adaptive adjustment function of the following coupling relationship 2. The S-shaped mapping function can still achieve gradient control of the meta-learning rate based on the relative magnitude of the output signal-to-noise ratio.

[0099] The embodiments described above in this application form a closed-loop processing chain of decomposition-enhancement-re-decomposition by iteratively feeding the stochastic resonance output signal to the morphological component analysis. This enables both the morphological component analysis and stochastic resonance algorithms to achieve incremental improvements in decomposition capabilities through iterative feedback. Simultaneously, the signal-to-noise ratio gain G calculated in step S53... SNR As a quantitative indicator for measuring the enhancement effect of stochastic resonance, it has become a related variable driving the adjustment of parameters in subsequent meta-learning algorithms.

[0100] In step S6, interference pattern map construction and meta-learning task distribution update are performed.

[0101] like Figure 6 As shown, in some embodiments of this application, step S6 may include at least one of steps S61 to S63.

[0102] In step S61, interference fingerprint vectors for each time period are constructed.

[0103] The periodic component sparse coefficient vector and the transient component sparse coefficient vector obtained after morphological component analysis decomposition of each signal window are concatenated to form the interference fingerprint vector for that time period. In some embodiments of this application, the dimension of the periodic component sparse coefficient vector is determined by the number of atoms in the discrete cosine transform dictionary (usually the same as the signal window length), and the dimension of the transient component sparse coefficient vector is determined by the number of atoms in the Gabor atom dictionary (usually several times the signal window length). The concatenated interference fingerprint vector characterizes the joint mode features of periodic and transient interference in the current signal environment.

[0104] It should be noted that the sparse coefficient vector of the pulse component is not included in the interference fingerprint because the pulse component contains the partial discharge signal to be identified itself.

[0105] In step S62, density-based clustering is performed on the historically accumulated interference fingerprint vectors. Using the cosine distance between interference fingerprint vectors as a similarity measure, a density-based clustering method is employed to classify them into several interference pattern categories. In some embodiments of this application, the neighborhood radius parameter of the clustering is adaptively determined based on the average cosine distance of the fingerprint vector set, and the minimum cluster sample size is set to 3, so that the clustering results can reflect the true distribution of interference patterns. Each clustered interference pattern category represents a specific combination of electromagnetic environments (e.g., a specific combination of communication base station interference and transient interference from adjacent switchgear during a certain period).

[0106] In step S63, the task distribution for model-independent meta-learning is constructed and updated based on the clustering results. Each interference pattern category corresponds to a meta-learning task, where the support set of the task consists of partially discharged and non-partially discharged labeled samples accumulated through manual verification or historical judgment under that category, and the query set consists of new samples to be judged. When the cosine distance between the interference fingerprint vector of a new time period and any existing category exceeds the neighborhood radius, the system creates a new interference pattern category and a corresponding new meta-learning task, so that the task distribution continues to expand with the dynamic changes of the electromagnetic environment. In some embodiments of this application, at the initial stage of system operation, no less than 5 initial task categories are preset (based on data collected during different time periods during debugging). As the online running time accumulates, the number of task categories gradually increases to cover more diverse combinations of interference environments.

[0107] The above embodiments of this application dynamically transform the historical decomposition results (sparse coefficient vector of interference components) of morphological component analysis into the task distribution of meta-learning, realizing the third deep coupling relationship between morphological component analysis and model-independent meta-learning (hereinafter referred to as coupling relationship three). This enables the training task set of the meta-learning model to continuously and automatically reflect the diversity and dynamic evolution of the actual interference environment on site, which is fundamentally different from the practice of manually pre-constructing a fixed task set in traditional methods.

[0108] In step S7, adaptive classification with model-independent meta-learning is performed.

[0109] In some embodiments of this application, step S7 may include at least one of steps S71 to S73.

[0110] In step S71, the inner loop optimization of model-independent meta-learning is performed. For each task constructed in step S6, the classification network parameters are updated with a few steps of gradient using a small number of labeled samples from its support set to obtain task-specific parameters for that task (i.e., that perturbation mode category).

[0111] In some embodiments of this application, the number of supporting samples for each task is 3-5, the number of inner loop gradient update steps is 1-3, and the inner loop learning rate is a preset fixed value (e.g., 0.01).

[0112] The loss function L used in the inner and outer loop optimizations T (f_θ) is the cross-entropy loss function, specifically defined as: ; Where S T For the sample set of task T (the inner loop uses the support set, and the outer loop uses the query set), |S T | represents the number of samples in the sample set, C represents the number of discharge type categories (C=4 in this embodiment), y i,c For sample x i The one-hot encoding of the c-th component of the true label (when sample x) i When the true category is class c, y i,c =1, otherwise y i,c =0), p i,c The c-th component of the output layer of the classification network Softmax, i.e., the network's response to sample x. i The predicted probability of belonging to class c.

[0113] The classification network is a lightweight convolutional network structure.

[0114] In some embodiments of this application, the lightweight convolutional network comprises three convolutional layers and one fully connected output layer: The first convolutional layer contains 16 convolutional kernels of size 3×3, using the ReLU activation function, followed by 2×2 max pooling; The second convolutional layer contains 32 convolutional kernels of size 3×3, using the ReLU activation function, followed by 2×2 max pooling; The third convolutional layer contains 64 3×3 convolutional kernels, using the ReLU activation function, followed by global average pooling to compress the feature map into a one-dimensional vector. The fully connected output layer maps this vector to an output with the number of discharge types as the dimension, which is then normalized to a probability distribution using Softmax. The total number of network parameters is kept below 50,000 to meet the storage and inference speed constraints of edge computing terminals.

[0115] The input is a phase-resolved statistical spectrum of the pulse components. This spectrum plots the pulse events as a 36×20 pixel two-dimensional grayscale spectrum according to the power frequency phase angle position (360° is divided into 36 equal intervals) and the pulse amplitude (20 equal logarithmic intervals from the minimum detection amplitude to the maximum pulse amplitude within the window). The pixel value is the normalized result of the count value of the pulse event in the corresponding interval.

[0116] The network output is the probability distribution of each discharge type (internal insulation defect partial discharge, external corona interference, switching transient interference, and no discharge).

[0117] In step S72, outer loop optimization of model-independent meta-learning is performed. The meta-parameters are updated using the sum of query set losses from all tasks. Coupling relationship two is introduced: the outer loop meta-learning rate β. meta The signal-to-noise ratio gain G calculated in step S5 SNR Dynamically determined using a sigmoid mapping function:

[0118] β meta =β0·sigmoid(λ·(G SNR -G th )); Where β0 is the basic learning rate, G th λ is the gain threshold, λ is the mapping slope parameter, and sigmoid is the sigmoid mapping function.

[0119] The physical meaning of this coupling relationship is: when the signal-to-noise ratio gain G SNR Significantly greater than the gain threshold G th This indicates that stochastic resonance successfully enhanced meaningful weak signal components from residual noise (a new partial discharge mode discovery with high confidence), the output value of the sigmoid mapping function approaches 1, and the learning rate β of the outer loop element... metaApproaching the base learning rate β0, the classification network rapidly adjusts its parameters with larger step sizes to incorporate newly discovered signal patterns; When the signal-to-noise ratio gain G SNR Below or close to the gain threshold G th When this occurs, it indicates that the stochastic resonance has not been effectively enhanced (the residual does not contain a meaningful weak signal or the enhancement has failed), the output value of the sigmoid mapping function approaches 0, and the learning rate β of the outer loop element... meta When the values ​​approach 0, the classification network parameters hardly update, maintaining the stability of the current state to avoid overfitting to noise.

[0120] In some embodiments of this application, the base learning rate β0 ranges from 0.001 to 0.01, and the gain threshold G... th The value of the signal-to-noise ratio gain of the output of the stochastic resonance system under pure noise input conditions is determined (generally about 1.0 to 1.2). The value of the mapping slope parameter λ ranges from 5 to 20 and is determined based on the sensitivity of the change in the meta-learning rate at the gain threshold to satisfy the preset gradient condition.

[0121] In step S73, after the meta-parameter update is completed, a classification network optimized by meta-learning is obtained. In some embodiments of this application, when the system is in online inference mode, only 1-3 steps of inner loop gradient update corresponding to the task that best matches its interference fingerprint are needed to obtain a task-specific classification network for the current environment.

[0122] The embodiments described above establish a direct quantitative correlation between signal enhancement credibility and model learning aggressiveness by dynamically adjusting the S-shaped mapping of the meta-learning rate through signal-to-noise ratio gain. This coupling relationship ensures that the parameter update magnitude of the meta-learning model is tied to the credibility of the newly discovered signal: rapid learning when credibility is high, and conservative waiting when credibility is low.

[0123] In some embodiments of this application, the initial meta-parameters θ0 of the classification network are obtained as follows: During the debugging phase before the system is officially deployed, an initial training task set is constructed by combining historical data collected at the cable branch box installation site under different operating conditions and standard samples generated by known discharge models in the laboratory.

[0124] Specifically, the construction of the initial training task set includes the following steps: During the commissioning period after the cable branch box is installed (usually 1-2 weeks), signals are continuously collected and morphological component analysis is performed at different times (covering daytime load peak and nighttime low load periods) and under different meteorological conditions to obtain no less than 5 initial interference mode categories as initial tasks. The labeled samples for each initial task can be obtained in the following ways: (a) Using a laboratory standard partial discharge simulator, discharge signals of known type and known magnitude are generated on cable accessories of the same type. The standard PRPD spectrum of each discharge type is obtained as a positive sample by the acquisition and processing of the same type of sensor. (b) The pulse component PRPD spectrum during the no-discharge activity period confirmed during debugging is used as negative samples of the "no-discharge" type. The classification network is trained offline using standard model-independent meta-learning using the above initial task set: multiple rounds (usually 500 to 2000 rounds) of outer loop iteration are performed on all tasks. In each round, a batch of tasks is randomly sampled from the task set to perform inner loop optimization and outer loop meta-parameter update until the meta-loss converges.

[0125] The network parameters obtained after training are the initial meta-parameters θ0, and the system then enters online operation mode. As online operation time accumulates and new labeled samples are added, the initial meta-parameters are continuously optimized.

[0126] In step S8, classification decision and early warning output are performed.

[0127] In some embodiments of this application, step S8 may include at least one of steps S81 to S83.

[0128] In step S81, the pulse components obtained from the secondary decomposition iteration of the current signal window in step S5 are converted into a phase-resolved statistical spectrum, which is then input into the task-specific classification network optimized by meta-learning in step S7 to obtain the probability distribution output of each discharge type. In some embodiments of this application, the discharge types include, but are not limited to: internal insulation defect partial discharge (including subtypes such as air gap discharge, surface discharge, and electrical tree discharge), external corona interference, switching transient interference, and no discharge (containing only background noise).

[0129] In step S82, an early warning decision is made based on the probability distribution output of the classification network.

[0130] An alert is triggered when the system determines that the discharge is internal and the corresponding probability (confidence level) exceeds a preset confidence threshold. In some embodiments of this application, the confidence threshold is set to 0.85, meaning that an alert signal is only generated when the classification network determines the presence of internal partial discharge with a confidence level of more than 85%, thus achieving a balance between sensitivity and false alarm rate.

[0131] In step S83, the classification results are updated. After manual confirmation (or automatic confirmation after a certain observation period with no contradictory evidence), the classification results are added as new labeled samples to the support sample set of the corresponding interference mode task, realizing online incremental task updates. As the system runs for a longer period, the support sample set for each task gradually increases, and the classification accuracy of the meta-learning model continues to improve.

[0132] The above embodiments of this application use an online incremental feedback update mechanism to enable the entire signal separation system to form a closed loop of acquisition-decomposition-enhancement-classification-feedback. The system's classification capability continues to evolve over time without the need for periodic shutdowns for retraining.

[0133] The following is a comprehensive explanation of the collaborative working mechanism of the three sets of deep parameter coupling relationships in this application.

[0134] 1. The mechanism of action of coupling relationship one (morphological component analysis → stochastic resonance) is as follows: The residual energy ratio ρ after morphological composition analysis decomposition serves as a quantitative indicator of the decomposition quality. It drives the potential well parameters a and b of the stochastic resonant bistable system in real time via a power function mapping. The formula for this coupling is:

[0135] a = a0·[1+k1·(ρ-ρ) th ) α1 ]; b=a / (σ residual ·c1) 2 ; This coupling ensures that the operating point of the stochastic resonance system automatically adapts to the amplitude characteristics of weak components in the current signal environment, achieving enhancement effects under different interference intensities without external manual intervention.

[0136] 2. The mechanism of the second coupling relationship (stochastic resonance → meta-learning) is: the signal-to-noise ratio gain G before and after stochastic resonance processing. SNR The learning rate β of the outer loop for learning model-independent components is dynamically determined using a sigmoid mapping function. meta The formula for this coupling is:

[0137] β meta =β0·sigmoid(λ·(G SNR -G th )); This coupling establishes a quantitative feedback channel between the confidence level of the signal enhancement effect and the magnitude of model parameter updates, ensuring that the model's learning behavior is correlated with the real existence of new signals in the environment.

[0138] 3. The mechanism of coupling relationship three (morphological component analysis → meta-learning task distribution) is as follows: the sparse coefficient vectors of interference components obtained from morphological component analysis at various historical time periods are dynamically constructed into the meta-learning task distribution after clustering. This coupling is achieved through the following relationship: each unique interference fingerprint cluster category defines an independent meta-learning task, and the interference fingerprint is constructed by concatenating the sparse coefficient vectors of periodic and transient components obtained from morphological component analysis at each time period. As morphological component analysis continues to run and identifies new combinations of interference patterns, the task space of meta-learning is continuously enriched and expanded.

[0139] The closed-loop mechanism formed by the above three sets of coupling relationships is as follows: the decomposition results of morphological component analysis simultaneously output driving signals in two directions, namely, the residual energy ratio drives the stochastic resonance system parameters downward (coupling relationship one), and the sparse coefficient vector of the interference component drives the construction of the meta-learning task distribution laterally (coupling relationship three); the enhancement effect (signal-to-noise ratio gain) of stochastic resonance drives the learning rate regulation of meta-learning upward (coupling relationship two); the classification network optimized by meta-learning outputs the classification results, which are updated after feedback confirmation, thereby affecting the subsequent inner loop optimization process. The three sets of couplings together constitute an adaptive signal separation framework with a hierarchical feedback structure.

[0140] The following describes the complete workflow of the method described in this application through a specific engineering application scenario.

[0141] Consider a 10kV cable branch box installed in an underground utility tunnel. This branch box receives one main cable and branches into four sub-cables. A communication base station antenna is located 20 meters above the box, and a ring main unit is located 50 meters away. During a certain operating period, a T-joint inside the branch box develops a small air gap at its interface due to long-term operation, resulting in a weak partial discharge with an amplitude of approximately 5 pC. Simultaneously, the following interferences exist: the communication base station generates periodic narrowband electromagnetic interference at frequencies of 900MHz and 1800MHz; the adjacent ring main unit experiences transient damped oscillating pulses due to load switch switching approximately every 30 minutes; and crosstalk coupling between the parallel power cables laid in the underground utility tunnel creates broadband background noise.

[0142] It should be noted that the 900MHz and 1800MHz radio frequency carriers of the communication base station are located outside the sensor bandwidth and sampling frequency band (Nyquist frequency 50MHz). Their impact on the monitoring channel is manifested as stable periodic narrowband components falling into the acquisition frequency band after coupling path and front-end nonlinearity. For the sake of brevity, this application uses the original carrier frequency of the interference source to refer to the corresponding narrowband interference component formed in the acquisition frequency band. The narrowband components captured and sparsely represented by the discrete cosine transform dictionary are these in-band narrowband components.

[0143] In this scenario, the amplitude of the 5pC partial discharge pulse in the signal collected by the high-frequency current sensor is approximately 2mV, while the amplitude of the communication base station interference is approximately 10mV, the peak value of the switching transient interference is approximately 50mV, and the root mean square value of the background noise is approximately 5mV. The partial discharge signal is completely buried under multi-source interference.

[0144] The system first performs segmented windowing processing on the acquired signal in step S1.

[0145] In step S2, morphological component analysis decomposition is performed: the discrete cosine transform dictionary captures the periodic narrowband interference of 900MHz and 1800MHz and classifies it into the periodic component, and the Gabor atom dictionary captures the switch transient decay oscillation and classifies it into the transient component. However, since the amplitude of the 5pC partial discharge pulse is much lower than the noise floor (about 0.4 times the root mean square value of the noise), the Dirac dictionary cannot effectively identify the weak pulse under the current threshold, and its energy remains in the decomposition residual.

[0146] In step S3, the residual energy ratio is calculated. Since the residual includes weak partial discharge pulse energy and some incompletely separated broadband noise energy, the residual energy ratio ρ exceeds the preset threshold ρ. th This triggers random resonance processing.

[0147] In step S4, the potential well parameters a and b are adaptively determined by the residual energy ratio and the standard deviation of the residual signal, so that the potential well spacing matches the amplitude characteristics of the weak partial discharge pulse. The broadband noise in the residual acts as the driving noise source of the random resonance, and its energy is partially transferred to the weak pulse component that matches the potential well spacing through the nonlinear potential well structure, thereby enhancing the 5pC partial discharge pulse.

[0148] In step S5, the enhanced signal is fed back to morphological component analysis for secondary decomposition. At this point, the Dirac dictionary successfully classifies the enhanced partial discharge pulse into the pulse component. The calculated signal-to-noise ratio gain G... SNR It is approximately 3.5 (meaning the signal-to-noise ratio has been improved by about 3.5 times).

[0149] In step S6, the interference fingerprint of the current signal window (composed of the coefficient amplitudes of 900MHz and 1800MHz in the periodic component and the time-frequency characteristics of the switching transient in the transient component) is matched with a certain interference pattern category obtained by historical clustering (this category corresponds to the combination pattern of "communication base station interference plus switching transient interference").

[0150] In step S7, the meta-learning outer loop learning rate β meta By G SNR =3.5 is determined by the sigmoid mapping function, since G SNR Significantly higher than the gain threshold G th (Approximately 1.1), the output value of the sigmoid mapping function is close to 1, β meta Approaching the baseline learning rate β0, the classification network quickly adapts to the newly discovered weak partial discharge pattern with a relatively large learning rate. Finally, in step S8, the classification network determines with 92% confidence that the pulse component contains a partial discharge signal indicating an internal insulation defect, triggering an early warning.

[0151] The following describes the system implementation of the method in this application.

[0152] like Figure 7As shown, the cable branch box partial discharge signal separation system provided in this application includes: The module includes a signal acquisition module, a morphological component analysis module, an adaptive stochastic resonance module, an iterative re-decomposition module, a task construction module, and a meta-learning classification module.

[0153] The signal acquisition module is configured to simultaneously acquire partial discharge signals through a high-frequency current sensor installed on the grounding wire of the cable branch box and a transient ground voltage sensor on the outer shell of the box, and perform analog-to-digital conversion and segmented windowing to form a signal window.

[0154] In some embodiments of this application, the hardware of the signal acquisition module includes a high-frequency current sensor (bandwidth 500kHz to 30MHz), a transient ground voltage sensor (bandwidth 3MHz to 100MHz), a high-speed analog-to-digital converter (sampling rate not less than 100MS / s, quantization accuracy not less than 12 bits), and a clock synchronization circuit.

[0155] The morphological component analysis module is configured to perform multi-dictionary sparse decomposition based on pulse, period, and transient dictionaries on the signal window to obtain each morphological component and the decomposition residual, and to calculate the residual energy ratio. The morphological component analysis module and the adaptive stochastic resonance module are unidirectionally coupled via the residual energy ratio parameter: the morphological component analysis module transmits the calculated residual energy ratio to the adaptive stochastic resonance module as input for determining its potential well parameters.

[0156] The adaptive stochastic resonance module is configured to activate when the residual energy ratio exceeds a preset threshold. It receives the residual signal and amplifies it weakly using a bistable nonlinear system. Its potential well depth parameter is determined by mapping the excess magnitude of the residual energy ratio using a power function, and the potential well spacing parameter is determined by calibrating the standard deviation of the residual signal. The adaptive stochastic resonance module and the iterative re-decomposition module achieve closed-loop iterative coupling through enhanced signal feedback: the output signal of the adaptive stochastic resonance module is transmitted to the iterative re-decomposition module to perform quadratic morphological decomposition.

[0157] The iterative refactoring module is configured to feed back the output signal of the adaptive stochastic resonance module to the pulse component channel to perform secondary morphological component analysis decomposition and calculate the signal-to-noise ratio (SNR) gain. The iterative refactoring module and the meta-learning classification module are coupled through a dynamic adjustment of the learning rate via the SNR gain parameter: the SNR gain calculated by the iterative refactoring module is passed to the meta-learning classification module as the input for determining its outer loop learning rate.

[0158] The task construction module is configured to combine the sparse coefficient vectors of interference components from different historical time periods as interference fingerprints, and then divide them into multiple interference pattern categories using density clustering. Each category constitutes a meta-learning task. The task construction module and the meta-learning classification module are coupled through the interference fingerprint clustering results to achieve dynamic updates of the task distribution: the task distribution output by the task construction module is used by the meta-learning classification module as the task set for inner and outer loop optimization.

[0159] The meta-learning classification module is configured to perform a dual-loop optimization of model-independent meta-learning. The outer loop meta-learning rate is dynamically determined by the signal-to-noise ratio gain through a sigmoid mapping function. The pulse components obtained from the secondary decomposition are input into the optimized classification network to output the discharge type and warning information.

[0160] In some embodiments of this application, to meet the real-time processing constraints of edge computing terminals, one or more combinations of the following computational optimization strategies may be adopted: (1) The fast iterative shrinking threshold algorithm is used to replace the standard block coordinate relaxation iteration for morphological component analysis decomposition. The fast transformation property of the dictionary matrix (Dirac dictionary is an identity matrix operation, discrete cosine transform dictionary supports fast cosine transform, Gabor atom dictionary supports inner product batch calculation based on fast convolution and matching pursuit greedy atom selection) reduces the amount of computation in each iteration. (2) An adaptive step size strategy is adopted for the numerical solution of stochastic resonance. The step size is increased in the interval where the signal amplitude changes slowly (e.g., 1 / 2 of the sampling period), and a fine step size (1 / 10 of the sampling period) is used only in the interval near the detection of a pulse event, so that the average step size is significantly greater than the minimum step size, thereby reducing the total number of computation steps. (3) The NEON vectorized instruction set of the ARM processor is used to parallelize and accelerate the fourth slope calculation of the Runge-Kutta method; (4) At the signal window level, a pipelined parallel strategy is adopted. While performing morphological component analysis and decomposition in the current window, stochastic resonance and meta-learning inference in the previous window are performed simultaneously, taking advantage of the parallel capabilities of multi-core processors.

[0161] Through the above optimization strategies, the system can achieve near real-time operation with a single-window processing latency not exceeding 80% of the window duration on a quad-core ARM Cortex-A series processor with a main frequency of no less than 1.5GHz. For more stringent real-time requirements, the stochastic resonance numerical solution module can be deployed on an FPGA coprocessor to obtain hardware acceleration.

[0162] The modules in the above system embodiments of this application together constitute a deep parameter coupling architecture, forming an organic whole with multi-layer feedback and parameter mutual drive, realizing adaptive and collaborative separation of weak partial discharge signals in complex electromagnetic environments.

Claims

1. A method for separating partial discharge signals in a cable branch box, comprising: The partial discharge signal collected from the cable branch box is segmented and windowed to form a signal window; The signal window is subjected to morphological component analysis decomposition based on multiple dictionaries. The corresponding morphological components are sparsely represented by pulse dictionary, periodic dictionary and transient dictionary respectively. The sparse coefficient vector of each component and the decomposition residual are obtained by block coordinate relaxation iteration. The ratio of the decomposed residual energy to the original signal energy is calculated as the residual energy ratio. When the residual energy ratio exceeds a preset threshold, the residual signal is input into a bistable stochastic resonance system for enhancement. The potential well depth parameter in the Langevin equation of the bistable stochastic resonance system is determined by the over-limit amplitude of the residual energy ratio through a power function mapping, and the potential well spacing parameter is determined by the standard deviation of the residual signal to match the potential barrier height with the amplitude of the signal to be enhanced. The random resonance output signal is superimposed back onto the pulse component channel to perform a second decomposition iteration of morphological component analysis, separate the enhanced weak partial discharge pulse, and calculate the signal-to-noise ratio gain before and after random resonance processing. The sparse coefficient vectors of the interference components obtained from the morphological component analysis of different historical periods are combined as interference fingerprints, which are then divided into multiple interference pattern categories by density clustering. Each category is used as a learning task for model-independent meta-learning. The model-independent meta-learning is optimized by a double loop. The inner loop updates the gradient of each task with a small number of support samples to obtain the task-specific parameters. The meta-learning rate of the outer loop is dynamically determined by the signal-to-noise ratio gain through the sigmoid mapping function. The higher the signal-to-noise ratio gain, the higher the meta-learning rate. The pulse components obtained from the second decomposition iteration are input into a classification network optimized by meta-learning, and the discharge type classification result is output. When it is determined to be internal partial discharge and the confidence level exceeds the threshold, an early warning is triggered.

2. The method according to claim 1, wherein, The multi-dictionary-based morphological component analysis decomposition includes: The Dirac dictionary is used as the pulse dictionary, and the set of unit pulse basis functions is used to sparsely represent the partial discharge pulse components with isolated spike morphology; the discrete cosine transform dictionary is used as the periodic dictionary, and the periodic narrowband interference components of communication carrier and higher harmonics are sparsely represented; the Gabor atom dictionary is used as the transient dictionary, and the damped oscillation transient interference components of switching operation are sparsely represented. The signal window is modeled as the sum of the linear superposition of the corresponding components of the three dictionaries and the residual. The objective function is to minimize the sum of the l1 norms of each sparse coefficient vector, and the constraint is that the l2 norm of the reconstruction error does not exceed the allowable threshold. The sparse coefficient vectors of each component are alternately optimized through block coordinate relaxation iteration until convergence.

3. The method according to claim 1, wherein, The methods for determining the potential well parameters of the bistable stochastic resonance system include: The potential well depth parameter is determined by superimposing a power-law nonlinear mapping amount of the difference between the residual energy ratio and the preset threshold on a base of preset basic parameters. The power exponent is greater than zero and less than one, so that the larger the residual energy ratio, the deeper the potential well. The potential well spacing parameter is determined by dividing the potential well depth parameter by the square of the product of the residual signal standard deviation and the scale matching coefficient, such that the potential well spacing given by the square root of the ratio of the potential well depth parameter to the potential well spacing parameter is within the range of 1 to 3 times the residual signal amplitude. The bistable stochastic resonance system uses the fourth-order Runge-Kutta method to numerically solve the Langevin equation, with a solution step size no greater than 1 / 10 of the sampling period.

4. The method according to claim 1, wherein, The learning rate of the outer loop element is dynamically determined by the signal-to-noise ratio gain through a sigmoid mapping function, including: The learning rate of the outer loop element is obtained by multiplying the base learning rate by the value of the sigmoid mapping function. The independent variable of the sigmoid mapping function is the product of the mapping slope parameter and the difference between the signal-to-noise ratio gain and the gain threshold. The gain threshold is determined based on the statistical mean of the signal-to-noise ratio gain of the output of the stochastic resonance system under pure noise input conditions; the mapping slope parameter is determined based on the sensitivity of the change in the meta-learning rate at the gain threshold to satisfy a preset gradient condition. When the signal-to-noise ratio gain is lower than the gain threshold, the output value of the S-shaped mapping function approaches zero, the learning rate of the outer loop element approaches zero, and the classification network parameters remain stable; when the signal-to-noise ratio gain is significantly higher than the gain threshold, the learning rate of the outer loop element approaches the base learning rate, and the classification network quickly adapts to the new discharge mode.

5. The method according to claim 1, wherein, The method of combining the sparse coefficient vectors of interference components obtained from morphological component analysis of historical time periods as interference fingerprints, and classifying them into multiple interference pattern categories through density clustering, includes: The periodic component sparse coefficient vector and transient component sparse coefficient vector obtained after morphological component analysis decomposition of each signal window are concatenated to form the interference fingerprint vector for that time period. Using the cosine distance between interfering fingerprint vectors as a similarity measure, a density-based clustering method is used to divide the historically accumulated interfering fingerprint vectors into several interfering pattern categories. The neighborhood radius parameter of the clustering is adaptively determined based on the average cosine distance of the fingerprint vector set. When the distance between the interference fingerprint vector of a new time period and any existing category exceeds the neighborhood radius, a new interference pattern category is created and a corresponding meta-learning task is created, so that the task distribution continues to expand as the environment changes.

6. The method according to claim 1, wherein, The preset threshold is determined as follows: during the power outage maintenance of the cable branch box, pure background noise signals are collected, and the same morphological component analysis decomposition is performed on multiple pure noise signal windows respectively, and the residual energy ratio of each window is calculated. The mean of the obtained residual energy ratio sequence plus twice the standard deviation is used as the preset threshold.

7. The method according to claim 1, wherein, The inner loop updates the gradient of each task with a small number of support samples to obtain task-specific parameters, including: the number of support samples for each task is three to five, the number of inner loop gradient update steps is 1 to 3, and the inner loop learning rate is a fixed value; the classification network is a lightweight convolutional network, containing three convolutional layers and one fully connected output layer, the input is the phase-resolved statistical spectrum of the pulse components, and the output is the probability distribution of each discharge type; the classification results are manually confirmed and then added to the support sample set of the corresponding task as newly labeled samples, realizing online incremental task updates.

8. A partial discharge signal separation system for a cable branch box, used to implement the method as described in any one of claims 1-7, comprising: The signal acquisition module is configured to synchronously acquire partial discharge signals through a high-frequency current sensor installed on the grounding wire of the cable branch box and a transient ground voltage sensor on the outer shell of the box, and perform analog-to-digital conversion and segmented windowing to form a signal window; The morphological component analysis module is configured to perform multi-dictionary sparse decomposition based on pulse dictionary, periodic dictionary and transient dictionary on the signal window to obtain each morphological component and decomposition residual, and calculate the residual energy ratio; An adaptive stochastic resonance module is configured to be activated when the residual energy ratio exceeds a preset threshold, receive the residual signal and perform weak signal enhancement through a bistable nonlinear system, and its potential well depth parameter is determined by the excess amplitude of the residual energy ratio through a power function mapping, and the potential well spacing parameter is determined by the standard deviation of the residual signal. The iterative re-decomposition module is configured to feed back the output signal of the adaptive stochastic resonance module to the pulse component channel to perform secondary morphological component analysis decomposition and calculate the signal-to-noise ratio gain; The task construction module is configured to combine the sparse coefficient vectors of interference components from different historical time periods as interference fingerprints and divide them into multiple interference pattern categories through density clustering. Each category is constructed as a meta-learning task. The meta-learning classification module is configured to perform a dual-loop optimization of model-independent meta-learning. The outer loop meta-learning rate is dynamically determined by the signal-to-noise ratio gain through a sigmoid mapping function. The pulse components obtained from the secondary decomposition are input into the optimized classification network to output the discharge type and warning information.

9. The system according to claim 8, wherein, The morphological component analysis module and the adaptive stochastic resonance module are coupled unidirectionally through the residual energy ratio parameter. The adaptive stochastic resonance module and the iterative re-decomposition module are coupled in a closed-loop iterative manner through enhanced signal feedback. The iterative re-decomposition module and the meta-learning classification module are coupled through the signal-to-noise ratio gain parameter to achieve dynamic adjustment of the learning rate. The task construction module and the meta-learning classification module are coupled through the interference fingerprint clustering results to achieve dynamic updating of task distribution.

10. The system according to claim 8, wherein, The system is deployed on an edge computing terminal. The morphological component analysis module and the adaptive stochastic resonance module perform windowed short-time processing in units of signal windows. The meta-learning classification module performs only 1-3 steps of gradient update during the inference phase. The length of the signal window covers at least 20 power frequency cycles.

Images