Fault detection method for electrical high and low voltage equipment based on fault self-diagnosis

Abstract

The application discloses a fault detection method for electrical high and low voltage equipment based on fault self-diagnosis, relates to the technical field of electrical equipment detection, and comprises the following steps: obtaining equipment state data sets to be detected, performing conditional independence test on standardized data through a causal inference model, identifying potential causal dependence relationship between electrical parameters and physical state, and constructing a multi-level causal network containing hidden variables and explicit variables. Typical fault modes are introduced into the network as intervention factors to simulate fault propagation, path hypotheses are generated, real-time high-frequency monitoring data are combined for verification and screening, fault root component positioning is performed, the influence degree of the fault on the observation signal is quantified, and finally a diagnosis report containing fault components, modes, propagation paths and influence ranges is generated. The method can accurately represent the causal correlation of equipment state, autonomously deduce the fault propagation path, realize fault root positioning and influence quantification, and optimize the fault diagnosis effect of electrical high and low voltage equipment.

Description

Technical Field

[0001] This invention belongs to the field of electrical equipment testing technology, specifically a fault detection method for high and low voltage electrical equipment based on fault self-diagnosis. Background Technology

[0002] Existing fault detection methods for high and low voltage electrical equipment mainly rely on threshold comparison of electrical parameters, statistical correlation analysis, and shallow model matching. By collecting monitorable signals such as equipment voltage, current, and temperature, real-time data is matched with preset thresholds and historical fault samples to complete the identification and judgment of equipment faults. Some technical solutions will establish the correspondence between monitoring signals and fault manifestations through simple correlation models to complete basic fault identification.

[0003] This type of detection method can only capture the superficial correlation between monitored variables and cannot distinguish the real causal relationship between electrical parameters and the internal physical state of the equipment. It is easily affected by environmental interference, signal fluctuations and other factors, which can lead to misjudgment. The analysis of fault propagation path depends entirely on human experience. It is impossible to autonomously deduce the fault transmission process through data, it is difficult to accurately locate the root component of the fault, and it is also impossible to quantify the degree of effect of the fault on different observed signals. The diagnostic results can only reflect the fault symptoms and cannot fully present the fault transmission logic and the scope of impact.

[0004] It is necessary to rely on data to identify the potential causal dependence between electrical parameters and physical state, build a multi-level network structure that adapts to the operating logic of the equipment, introduce fault-related factors into the network structure to deduce the propagation path, and combine real-time monitoring data to complete path verification and screening, so as to realize the location of the root cause of the fault and the quantification of the degree of fault impact, and finally form a complete fault diagnosis content. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a fault detection method for high and low voltage electrical equipment based on fault self-diagnosis, including: Obtain the status dataset of the high- and low-voltage electrical equipment to be tested; Based on the causal inference model, conditional independence tests are performed on standardized state datasets to identify potential causal dependencies between electrical parameters and physical states. Based on the identified potential causal dependencies, a multi-level causal network is constructed, with the internal physical state of the device as the latent variable and the observable electrical and physical signals as the explicit variables. In the multi-level causal network, known typical failure modes are introduced as intervention factors to simulate the failure propagation process and generate failure propagation path hypotheses. The high-frequency monitoring data collected in real time is compared and verified with the status of key nodes in the fault propagation path hypothesis to screen out potential fault paths with high confidence. Based on the selected high-confidence potential fault paths, the root cause component of the fault is located, and the impact of the fault on different observed signals in multi-source monitoring data is quantified. Based on the quantitative results of the impact and the root cause component location information, a diagnostic report is generated that includes the faulty component, fault mode, propagation path and scope of impact.

[0006] Furthermore, the acquisition of the status dataset of the electrical high- and low-voltage equipment to be tested includes: Collect multi-source monitoring data of the electrical high and low voltage equipment to be tested. The multi-source monitoring data includes vibration signals, temperature data, current waveforms, and partial discharge signals. The collected multi-source monitoring data is time-stamped and cleaned to remove abnormal noise and missing signal fragments, forming a standardized state dataset. The process of aligning and cleaning the collected multi-source monitoring data with timestamps, removing abnormal noise and missing signal fragments, and forming a standardized state dataset specifically includes: Raw vibration signals, temperature data, current waveforms, and partial discharge signals with their respective sampling times are acquired from different sensor acquisition channels. Using a global clock as a reference, an interpolation method is used to align the sampling time points of all original signals to a unified time series; For each time-aligned signal, a filter matching the signal characteristics is applied: a bandpass filter is used for vibration signals, a moving average filter is used for temperature data, a notch filter is used for current waveforms, and wavelet thresholding is used for partial discharge signals. Missing segments in the filtered signal are detected. For short-term missing segments, linear interpolation of data from the time points before and after the corresponding signal is used to fill them in. For long-term missing segments, all signal data corresponding to the time period will be marked as invalid and excluded from subsequent analysis. The filled signal data are resampled at uniform time intervals and combined into a standardized state dataset with the same time dimension but different signal dimensions.

[0007] Furthermore, the causal inference model-based approach involves performing conditional independence tests on the standardized state dataset to identify potential causal dependencies between electrical parameters and physical states. Specifically, this includes: From the standardized state dataset, current harmonic components, voltage RMS value, and three-phase unbalance are selected as key electrical parameter variables, and vibration dominant frequency amplitude, temperature gradient, and partial discharge pulse frequency are selected as key physical state variables. On the joint dataset consisting of key electrical parameter variables and key physical state variables, perform an independence test based on conditional mutual information to calculate the independence of each pair of variables given a subset of other variables. Record the variable pairs that pass the independence test, that is, the variable pairs that have conditional dependencies, and form a set of conditional dependencies; For each pair of variables in the set of conditional dependencies, the order of signal transmission is determined by analyzing the time delay cross-correlation between them. The variable that is ahead in time is initially marked as the cause, and the variable that is lagging behind is initially marked as the effect. Integrate all variable pairs with causal orientation to form a preliminary potential causal dependency graph; Using known knowledge of device operation mechanisms as constraints, the potential causal dependency graph is modified and verified. Causal edges that violate physical mechanisms are removed and necessary causal connections that conform to device operation logic are added, forming a causal inference model for constructing a multi-level causal network.

[0008] Furthermore, based on the identified potential causal dependencies, a multi-level causal network is constructed, with the internal physical state of the device as the latent variable and observable electrical and physical signals as the explicit variables. Specifically, this includes: Define a set of latent variables characterizing the core state of the equipment, including the degree of winding deformation, the degree of insulation aging, the degree of contact wear, the cooling efficiency, and the degree of mechanical looseness; Based on the preliminary potential causal dependency diagram, the observable key electrical parameter variables and key physical state variables are used as explicit observation indicators of latent variables to establish an observation model from latent variables to explicit variables. Using the structural causal model framework and based on the knowledge of equipment operation mechanism, structural equations are established among latent variables to describe the causal relationship between latent variables; By merging the structural equations between latent variables with the observational model from latent variables to manifest variables, a complete multi-level causal network structure containing latent and manifest layers is formed. Using a standardized state dataset, the posterior probability distribution of all latent variables in a multi-level causal network structure is estimated through variational inference.

[0009] Furthermore, known typical failure modes are introduced as intervention factors into the multi-level causal network to simulate the failure propagation process and generate failure propagation path hypotheses, specifically including: Establish a library of typical fault modes, including but not limited to winding turn short circuit, partial insulation breakdown, poor contact, heat sink blockage and bearing wear; Select a failure mode from the typical failure mode library, and set the hidden variable state corresponding to the failure mode to an "abnormal" state in the multi-level causal network. This operation is regarded as a deterministic intervention. In the multi-level causal network after intervention, the causal propagation simulation algorithm is run to iteratively calculate the probability change of its influence on the state of downstream latent variables, starting from the latent variable being intervened and along the causal edges between latent variables. When the probability of an abnormal state of a downstream latent variable exceeds a preset threshold, the corresponding latent variable is activated, its state is also considered "abnormal", and the influence continues to propagate to its downstream latent variables until no new latent variables are activated. Record the propagation order and path from the root cause hidden variable to all activated abnormal hidden variables in this simulation, and form the fault propagation path hypothesis for the corresponding fault mode.

[0010] Furthermore, the step of comparing and verifying the real-time collected high-frequency monitoring data with the key node states in the fault propagation path hypothesis to screen out potential fault paths with high confidence specifically includes: From each fault propagation path hypothesis, we extract latent variables that are in the middle of the propagation path and have clearly observable signals, as key nodes for verification; Determine the type of observable signal corresponding to each key verification node and its characteristic value range under normal and abnormal states; Initiate high-frequency monitoring of the equipment, collect and verify the latest observation signal segments corresponding to key nodes, and extract the corresponding feature values ​​from the signal segments in real time; The real-time extracted feature values ​​are compared with the preset range of normal and abnormal state feature values ​​to determine whether the actual state of the verification key node matches the state predicted in the fault propagation path hypothesis. The state matching rate of all key verification nodes in a fault propagation path hypothesis is calculated. If the matching rate is higher than a set threshold, the corresponding fault propagation path hypothesis is marked as a potential fault path with high confidence.

[0011] Furthermore, the step of locating the root cause component of the failure based on the selected high-confidence potential fault paths, and simultaneously quantifying the impact of the failure on different observed signals in multi-source monitoring data, specifically includes: For each high-confidence potential failure path, analyze its initial abnormal hidden variable, which directly corresponds to a root cause component of the failure. In the multi-level causal network, the average causal effect value on all observable electrical parameter variables and key physical state variables is calculated when the anomalous latent variable is in an "abnormal" state. The average causal effect value is compared with the historical normal fluctuation range of the corresponding observed variable to calculate the relative impact intensity of the target fault on each observed variable, that is, the percentage of the change of the observed variable to its own normal fluctuation range. The relative impact intensity of the target failure on all key observed variables is summarized and sorted by intensity to form the characteristic impact spectrum of the corresponding failure mode.

[0012] Furthermore, based on the quantification results of the impact degree and the root cause component location information, a diagnostic report is generated that includes the faulty component, fault mode, propagation path, and scope of impact, specifically including: All high-confidence potential failure paths are classified according to the root component corresponding to their initial anomaly hidden variable; From similar faulty components, the potential fault path with the highest verification key node state matching rate is selected as the fault evolution path of the corresponding faulty component; Translate all the hidden variables involved in the fault evolution path from the terminology of the internal physical state of the equipment into the names of specific mechanical or electrical components; Based on the characteristic impact spectrum of the fault mode, describe the intensity level of the impact of this fault on different monitoring signals; Finally, the faulty component name, corresponding typical fault mode, fault evolution path description, and monitoring signal influence intensity level are integrated to generate a structured fault self-diagnosis report.

[0013] Furthermore, in the multi-level causal network after intervention, a causal propagation simulation algorithm is run to iteratively calculate the probability change of its influence on the state of downstream latent variables, starting from the intervened latent variable and proceeding along the causal edges between latent variables. Specifically, this includes: The state probability distribution of the latent variable to be intervened is forcibly changed from the posterior probability distribution to a deterministic distribution representing complete anomaly; Starting with the latent variable being intervened as the starting node, traverse all directed causal edges originating from the starting node in the multi-level causal network; For the currently activated anomalous latent variable, along each of its directed causal edges, calculate the incremental change in the state probability of its adjacent downstream latent variables based on the causal effect strength defined in the structural equation. The original posterior probability distribution of the downstream latent variables is superimposed with the incremental change to obtain its updated state probability distribution. Check the updated state probability distribution. If the probability of abnormal state exceeds the preset activation threshold, mark the downstream hidden variable as a new abnormal activation node and add it to the subsequent propagation traversal process until no new node is activated.

[0014] Furthermore, the step of extracting latent variables with clearly observable signals that are located in the middle of each fault propagation path hypothesis as key verification nodes specifically includes: For the activation sequence of latent variables recorded in the fault propagation path hypothesis, remove the root fault latent variable that is at the beginning of the sequence; In the remaining sequence of latent variables, those latent variables that correspond to weak observed signals, whose features are difficult to extract, or that are easily affected by environmental interference are removed. From the sequence of latent variables that have been eliminated, select the most frequently occurring latent variables in order as candidate key nodes for verification. Check each candidate verification key node to ensure that the corresponding observed signal type has been monitored and that the characteristic value range of its normal and abnormal states has been predefined in historical data; The final selected candidate verification key nodes are the verification key nodes used for actual comparison.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By testing the conditional independence of variables in a standardized state dataset using a causal inference model, the potential causal dependencies between electrical parameters and physical states are identified. Based on these dependencies, a multi-level causal network is constructed, with the internal physical state of the equipment as the latent variable and observable electrical and physical signals as the explicit variables. This network can remove false correlations in the monitoring data, clearly delineate the intrinsic correlation logic between electrical parameters and the physical state of the equipment, and the hierarchical network structure conforms to the transmission form of the internal physical state of high and low voltage electrical equipment. It restores the true correlation mechanism of the equipment operating state variables, avoids interference from irrelevant interference signals on the correlation determination of state variables, and makes the representation of equipment state conform to the actual physical operation law.

[0016] Introducing known typical failure modes as intervention factors into a multi-level causal network, this method simulates the failure propagation process and generates failure propagation path hypotheses. It compares and verifies the real-time high-frequency monitoring data with the key node states in these hypotheses, filtering out high-confidence potential failure paths. Based on the path information, it locates the root cause components of the failure and quantifies the impact of the failure on different observed signals in multi-source monitoring data. Based on the location information and quantification results, it generates a diagnostic report containing the failure components, failure modes, propagation paths, and scope of impact. This method can autonomously deduce the failure propagation process without relying on human experience, eliminate unreasonable path hypotheses through real-time data verification, identify the core components of the failure, and intuitively present the differences in the effects of the failure on various observed signals. This ensures that the fault diagnosis fully covers the core information of the failure and related content on its transmission and impact. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating the steps of the fault detection method for high and low voltage electrical equipment based on fault self-diagnosis as described in this invention. Figure 2 A flowchart for causal inference and identification of potential causal dependencies; Figure 3 A flowchart generated based on the fault propagation path assumption. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention discloses a fault detection method for high and low voltage electrical equipment based on fault self-diagnosis. This method achieves precise fault location and propagation analysis by constructing a model reflecting the internal causal mechanism of the equipment. The overall implementation scheme of this method includes the following steps: acquiring a state dataset of the high and low voltage electrical equipment to be tested; based on a causal inference model, performing conditional independence tests on the standardized state dataset to identify potential causal dependencies between electrical parameters and physical states; constructing a multi-level causal network based on the identified potential causal dependencies, with the internal physical state of the equipment as a latent variable and observable electrical and physical signals as explicit variables; introducing known typical fault modes as intervention factors into the multi-level causal network to simulate the fault propagation process and generate fault propagation path hypotheses; comparing and verifying the real-time high-frequency monitoring data with the key node states in the fault propagation path hypotheses to screen out high-confidence potential fault paths; locating the root cause component of the fault based on the screened high-confidence potential fault paths, and quantifying the degree of impact of the fault on different observed signals in multi-source monitoring data; and generating a diagnostic report containing the faulty component, fault mode, propagation path, and scope of impact based on the quantification results of the impact and the root cause component location information.

[0020] In one embodiment of the present invention, multi-source monitoring data of the high- and low-voltage electrical equipment under test is collected. The multi-source monitoring data includes vibration signals, temperature data, current waveforms, and partial discharge signals. The collected multi-source monitoring data undergoes timestamp alignment and data cleaning to remove abnormal noise and missing signal segments, forming a standardized state dataset. Specifically, raw vibration signals, temperature data, current waveforms, and partial discharge signals with their respective sampling times are acquired from different sensor acquisition channels. Using a global clock as a reference, interpolation methods are used to align the sampling time points of all raw signals to a unified time series. For each type of time-aligned signal, a filter matching the signal characteristics is applied: a bandpass filter for vibration signals, a moving average filter for temperature data, a notch filter for current waveforms, and wavelet threshold denoising for partial discharge signals. Missing segments in the filtered signals are detected. For short-term missing segments, linear interpolation of data before and after the corresponding signal time points is used to fill in the missing segments. For long-term missing segments, all signal data corresponding to the corresponding time period are marked as invalid and excluded from subsequent analysis. The filled signal data are resampled at a unified time interval and combined to form a standardized state dataset with the same time dimension but different signal dimensions.

[0021] In practical implementation, the status dataset of the high- and low-voltage electrical equipment under test is obtained through the following method: Multi-source monitoring data of the equipment is collected, including vibration signals, temperature data, current waveforms, and partial discharge signals. The collected multi-source monitoring data undergoes timestamp alignment and data cleaning to remove abnormal noise and missing signal segments, forming a standardized status dataset. Raw vibration signals, temperature data, current waveforms, and partial discharge signals with their respective sampling times are acquired from different sensor acquisition channels. Using the time signal output from a high-precision global clock source as a reference, a linear interpolation method is employed to align the sampling time points of all raw signals to a unified, fixed-interval time series. In practice, a bandpass filter with a preset specific passband is applied to the time-aligned vibration signal to retain the characteristic frequency band related to mechanical vibration. A moving average filter with a fixed time window is applied to the time-aligned temperature data to smooth measurement noise. A notch filter for power frequency and its specific harmonics is applied to the time-aligned current waveform to eliminate power grid background interference. A threshold denoising method based on wavelet transform is applied to the time-aligned partial discharge signal to separate partial discharge pulses from background white noise.

[0022] In some embodiments, the detection of missing segments in the filtered signal is achieved by setting a threshold for counting consecutive sampling points. When the number of consecutive invalid or zero sampling points is lower than this threshold, it is determined to be a short-term missing segment. For short-term missing segments, linear interpolation is performed using data from valid time points before and after the missing segment corresponding to the signal type. The linear interpolation process can be understood as follows: Among them, symbols Indicates a point in time The filled signal value, sign and These represent the timestamps of the nearest valid data points before and after the missing segment, respectively. and This represents the original valid signal value at the corresponding time point. When the number of consecutive invalid sampling points exceeds the threshold, it is determined to be a long-term missing data point. All types of signal data blocks within the corresponding time period will be marked as invalid data segments and excluded from subsequent analysis. In specific implementation, the filled signal data of various types are resampled at a uniform, preset time interval. The resampling process uses an anti-aliasing filter combined with a sampling rate conversion algorithm. Finally, the vibration signal, temperature data, current waveform, and partial discharge signal are combined into a standardized state dataset matrix with the row dimension representing time points and the column dimension representing different signal channels.

[0023] It is understood that in some embodiments, the timestamp-aligned interpolation method is not limited to linear interpolation; cubic spline interpolation can also be used to adapt to signal types with more drastic changes. Optionally, the passband range of the bandpass filter used for the vibration signal is set according to the natural frequency range of the rotor or winding of the monitored high- and low-voltage electrical equipment. Optionally, in the wavelet threshold denoising method used for the partial discharge signal, the selection of the wavelet basis function is determined based on the waveform characteristics of the partial discharge pulse, and the threshold selection rule adopts a general threshold criterion. In specific implementations, the standardized state dataset is stored in memory or a database in matrix form. Each row of the matrix corresponds to a unified time point, and each column of the matrix corresponds to a processed signal sequence, namely, a vibration signal sequence, a temperature data sequence, a current waveform sequence, and a partial discharge signal sequence.

[0024] In one embodiment of the present invention, see [reference] Figure 2From a standardized state dataset, current harmonic components, effective voltage values, and three-phase imbalance are selected as key electrical parameter variables, while vibration dominant frequency amplitude, temperature gradient, and partial discharge pulse frequency are selected as key physical state variables. On the joint dataset composed of these key electrical parameter and physical state variables, an independence test based on conditional mutual information is performed. The independence of each pair of variables under given subsets of other variables is calculated, and the variable pairs that pass the independence test—i.e., those with conditional dependencies—are recorded, forming a set of conditional dependencies. For each pair of variables in the set of conditional dependencies, the time-delay cross-correlation between them is analyzed to determine the order of signal transmission. Variables that are time-leading are initially labeled as causes, and those that are lagging are initially labeled as effects. All causal pairs are integrated to form a preliminary potential causal dependency graph. Using known equipment operating mechanisms as constraints, the potential causal dependency graph is revised and verified. Causal edges that violate physical mechanisms are removed, and necessary causal connections that conform to the equipment operating logic are added, forming the final causal inference model used to construct a multi-level causal network.

[0025] A set of latent variables characterizing the core internal state of the equipment is defined, including winding deformation, insulation aging, contact wear, cooling efficiency, and mechanical looseness. Based on a preliminary latent causal dependency graph, observable key electrical parameters and key physical state variables are used as explicit observation indicators of the latent variables, establishing an observation model from latent to explicit variables. Using a structural causal model framework and based on knowledge of the equipment's operating mechanism, structural equations are established among the latent variables to describe their causal relationships. The structural equations among the latent variables and the observation model from latent to explicit variables are merged to form a complete multi-level causal network structure containing latent and explicit layers. Using a standardized state dataset, variational inference methods are used to estimate the posterior probability distribution of all latent variables in the multi-level causal network structure.

[0026] In practical implementation, based on a causal inference model, conditional independence tests are performed on standardized state datasets to identify potential causal dependencies between electrical parameters and physical states. This process is implemented as follows: From the standardized state dataset, current harmonic components, effective voltage values, and three-phase imbalance are selected as key electrical parameter variables, while vibration dominant frequency amplitude, temperature gradient, and partial discharge pulse frequency are selected as key physical state variables. On the joint dataset composed of key electrical parameter variables and key physical state variables, an independence test based on conditional mutual information is performed to calculate the independence of each pair of variables given other subsets of variables. The core of the conditional mutual information independence test is to evaluate the mutual information between variables X and Y given a set of variables Z. Its value is calculated using the following formula: Among them, symbols Indicates a given set of condition variables At that time, variable With variables Conditional mutual information values, symbols , , Representing variables respectively , , Specific values, symbols Representing variables , , The joint probability distribution estimate, sign , , This represents the estimated value of the corresponding conditional probability distribution. When the calculated conditional mutual information value is lower than a significance threshold determined by statistical testing, it is determined that variables X and Y are independent under given condition Z. The variable pairs that pass the independence test are recorded, i.e., the variable pairs with conditional dependencies, forming a set of conditional dependencies.

[0027] In some embodiments, for each pair of variables in the conditional dependency set, the order of signal transmission is determined by analyzing the time-delay cross-correlation between the variable pairs. The variable that is ahead in time is initially labeled as the cause, and the lagging variable is initially labeled as the effect. All variable pairs with causal orientation are integrated to form a preliminary potential causal dependency graph. Known knowledge of device operating mechanisms is used as a constraint to revise and verify the preliminary potential causal dependency graph. In specific implementations, the revision operation includes removing causal connections that are impossible in terms of device operating mechanisms and adding causal connections that are mechanistically certain but may not have been fully discovered by data testing, forming the final causal inference model used to construct a multi-level causal network.

[0028] It is understandable that when constructing a multi-level causal network, a set of latent variables characterizing the core state of the equipment needs to be defined, including the degree of winding deformation, insulation aging, contact wear, cooling efficiency, and mechanical looseness. Based on a preliminary latent causal dependency graph, observable key electrical parameters and key physical state variables are used as explicit observation indicators of latent variables, establishing an observation model from latent variables to manifest variables. Using a structural causal model framework, based on knowledge of equipment operation mechanisms, structural equations are established between latent variables to describe the causal influence relationships between them. These structural equations can be expressed as linear or nonlinear functions. In some embodiments, the structural equations between latent variables are merged with the observation model from latent variables to manifest variables to form a complete multi-level causal network structure containing latent and manifest layers. Optionally, using a standardized state dataset, the posterior probability distribution of all latent variables in the multi-level causal network structure is estimated using variational inference methods. Variational inference approximates the true posterior distribution by maximizing the lower bound of evidence. In practice, the structural parameters of the multi-level causal network and the parameters of the observation model are obtained through joint learning. The final network can characterize the probability generation mechanism from the internal latent variable state to the external observable signal.

[0029] In one embodiment of the present invention, see [reference] Figure 3 A typical fault mode library is established, including but not limited to winding turn-to-turn short circuits, partial insulation breakdown, poor contact, heat sink blockage, and bearing wear. A fault mode is selected from the library, and in the multi-level causal network, the corresponding latent variable state is set to an "abnormal" state; this operation is considered a deterministic intervention. On the intervened multi-level causal network, a causal propagation simulation algorithm is run. Starting from the intervened latent variable, along the causal edges between latent variables, the probability change of its influence on the downstream latent variable state is iteratively calculated. The propagation order and path from the root cause latent variable to all activated abnormal latent variables are recorded in this simulation, forming the fault propagation path hypothesis for the corresponding fault mode.

[0030] The specific operation process of the causal propagation simulation algorithm is as follows: The state probability distribution of the latent variable to be intervened is forcibly set from its posterior probability distribution to a deterministic distribution representing complete anomalousness. Starting with the latent variable to be intervened, all directed causal edges originating from the starting node in the multi-level causal network are traversed. For the currently activated anomalous latent variable, along each directed causal edge, the incremental change in the state probability of its adjacent downstream latent variables is calculated according to the causal effect strength defined in the structural equation. The original posterior probability distribution of the downstream latent variable is superimposed with the incremental change to obtain its updated state probability distribution. The updated state probability distribution is checked; if its anomalous state probability exceeds a preset activation threshold, the downstream latent variable is marked as a new activated anomalous node and added to the subsequent propagation traversal process until no new nodes are activated.

[0031] In practical implementation, known typical fault modes are introduced as intervention factors into the multi-level causal network to simulate the fault propagation process and generate fault propagation path hypotheses. This process is implemented as follows: a typical fault mode library is established, which includes, but is not limited to, winding turn-to-turn short circuits, partial insulation breakdown, poor contact, heat sink blockage, and bearing wear. The typical fault mode library defines the hidden variables of the core internal state of the equipment directly corresponding to each typical fault mode. Refer to Table 1, which shows the contents of a typical fault mode library.

[0032] Table 1: Typical Failure Modes and Corresponding Hidden Variables A typical failure mode is selected from the typical failure mode library. In the multi-level causal network, the state of the latent variable corresponding to the typical failure mode is set to an "abnormal" state. This operation is considered a deterministic intervention. On the multi-level causal network after the intervention, a causal propagation simulation algorithm is run. Starting from the intervened latent variable, along the causal edges between latent variables, the probability change of its influence on the state of downstream latent variables is iteratively calculated. When the abnormal state probability of a downstream latent variable exceeds a preset threshold, the corresponding latent variable is activated, its state is also considered "abnormal," and the influence continues to propagate to its downstream latent variables until no new latent variables are activated. The propagation order and path from the root cause latent variable to all activated abnormal latent variables in this simulation are recorded, forming the failure propagation path hypothesis for the corresponding typical failure mode.

[0033] In some embodiments, the specific operation of the causal propagation simulation algorithm is as follows: The state probability distribution of the latent variable to be intervened is forcibly set from its posterior probability distribution to a deterministic distribution representing complete anomalousness. Starting with the latent variable to be intervened, all directed causal edges originating from the starting node in the multi-level causal network are traversed. For the currently active anomalous latent variable, along each directed causal edge, the incremental change in the state probability of its adjacent downstream latent variables is calculated according to the causal effect strength defined in the structural equation. The update of the state probability can be expressed as follows: Among them, symbols Representing downstream hidden variables Updated state probability distribution, symbol Representing downstream hidden variables The original posterior probability distribution, sign Indicates the currently active anomalous hidden variable to downstream hidden variables The causal effect strength coefficient, obtained from the structural equation, is shown in the symbol... This represents the anomalous latent variables caused by intervention. The state offset is calculated by superimposing the original posterior probability distribution of the downstream latent variable with the incremental change to obtain the updated state probability distribution of the downstream latent variable. The updated state probability distribution is then checked; if its abnormal state probability exceeds a preset activation threshold, the downstream latent variable is marked as a new activated abnormal node and added to the subsequent propagation traversal process until no new nodes are activated.

[0034] It is understandable that deterministic intervention sets the state probability vector of the selected latent variable as a one-hot encoded vector. For example, the state of the latent variable representing the "degree of winding deformation" is directly set from [normal probability = 0.9, abnormal probability = 0.1] to [normal probability = 0.0, abnormal probability = 1.0]. In specific implementation, the causal propagation simulation algorithm uses a queue data structure to manage all currently active anomalous latent variable nodes. The algorithm starts from the queue containing the initial intervention node, processes each node in the queue sequentially, calculates its propagation impact on downstream nodes, and adds newly activated nodes to the tail of the queue until the queue is empty, at which point the simulation ends. Optionally, the preset activation threshold is a configurable parameter whose value reflects the confidence level required to classify a latent variable as "abnormal". In some embodiments, when recording simulation results, not only the sequence of activated latent variables is recorded, but also the propagation delay and probability change magnitude on each causal path are recorded. In specific implementation, the above intervention and simulation process is executed independently once for each typical fault mode in the typical fault mode library, thereby generating a set of fault propagation path hypotheses covering multiple typical fault modes.

[0035] In one embodiment of the present invention, latent variables with clearly observable signals located in the middle of each fault propagation path hypothesis are extracted as verification key nodes. The type of observable signal corresponding to each verification key node and its feature value range under normal and abnormal states are determined. High-frequency monitoring of the equipment is initiated, and the latest observed signal segments corresponding to the verification key nodes are collected, and the corresponding feature values ​​are extracted from the signal segments in real time. The real-time extracted feature values ​​are compared with preset normal and abnormal state feature value ranges to determine whether the actual state of the verification key node matches the state predicted in the fault propagation path hypothesis. The state matching rate of all verification key nodes in a fault propagation path hypothesis is calculated. If the matching rate is higher than a set threshold, the corresponding fault propagation path hypothesis is marked as a high-confidence potential fault path.

[0036] The extraction process for key verification nodes is as follows: The latent variable activation sequence recorded in the fault propagation path hypothesis is processed by removing the root cause fault latent variable at the beginning of the sequence. From the remaining latent variable sequence, those corresponding to weak observed signals, difficult-to-extract features, or susceptible to environmental interference are removed. From the processed latent variable sequence, the most frequently occurring latent variables are selected sequentially as candidate key verification nodes. Each candidate key verification node is checked to ensure that its corresponding observed signal type has been monitored and that the characteristic value range of its normal and abnormal states has been predefined in historical data. The final selected candidate key verification nodes are then used as the key verification nodes for actual comparison.

[0037] In practice, the real-time high-frequency monitoring data is compared and verified with the key node states in the fault propagation path hypothesis to screen out high-confidence potential fault paths. This process is implemented as follows: From each fault propagation path hypothesis, latent variables with clearly observable signals in the middle of the propagation path are extracted as verification key nodes. The observable signal type corresponding to each verification key node and its characteristic value range under normal and abnormal states are determined. High-frequency monitoring of the equipment is activated, and the latest observed signal segments corresponding to the verification key nodes are collected and the corresponding characteristic values ​​are extracted from the signal segments in real time. The real-time extracted characteristic values ​​are compared with the preset normal and abnormal state characteristic value ranges to determine whether the actual state of the verification key node matches the state predicted in the fault propagation path hypothesis. The state matching rate of all verification key nodes in a fault propagation path hypothesis is calculated. If the matching rate is higher than a set threshold, the corresponding fault propagation path hypothesis is marked as a high-confidence potential fault path.

[0038] The extraction process for key verification nodes is as follows: The latent variable activation sequence recorded in the fault propagation path hypothesis is processed by removing the root cause fault latent variable at the beginning of the sequence. From the remaining latent variable sequence, those corresponding to weak observed signals, difficult-to-extract features, or susceptible to environmental interference are removed. From the processed latent variable sequence, the most frequently occurring latent variables are selected sequentially as candidate key verification nodes. Each candidate key verification node is checked to ensure that its corresponding observed signal type has been monitored and that the characteristic value range of its normal and abnormal states has been predefined in historical data. The final selected candidate key verification nodes are used as the key verification nodes for actual comparison. Table 2 shows the key verification node extraction results for the "winding inter-turn short circuit" fault propagation path hypothesis.

[0039] Table 2. Verification Key Node Extraction Table In some embodiments, the state matching rate is calculated by statistically analyzing the state judgment results of all verification key nodes in a fault propagation path hypothesis. The calculation relationships are as follows: Among them, symbols Represents the state matching rate, symbol This represents the total number of key verification nodes selected in the fault propagation path hypothesis, with the symbol [symbol missing]. Indicates the index of the verification key node, symbol The fault propagation path assumption is given for the first... The predicted state ("normal" or "abnormal") of each key verification node is represented by a symbol. This indicates the first [number] determined through real-time monitoring data. The actual state of each key verification node, symbol It is a comparison function, when and Output 1 if the states match, otherwise output 0. This is based on the calculated state matching rate. When the value is higher than a preset threshold, the fault propagation path is assumed to be marked as a potential fault path with high confidence.

[0040] It is understandable that determining the characteristic value range of normal and abnormal states requires based on historical monitoring data from long-term equipment operation. In specific implementations, for the observed signal of vibration dominant frequency amplitude, its normal state range may be defined as the 95% confidence interval of the amplitude distribution in historical data, while the abnormal state range is defined as the value range exceeding this confidence interval. After high-frequency monitoring is activated, the system continuously collects vibration signals and calculates the dominant frequency amplitude within the current time window in real time. This calculated value is compared with a preset range to determine the actual state of the "vibration dominant frequency amplitude," a key verification node. Optionally, for observed signals such as "partial discharge pulse frequency," the characteristic value range of its abnormal state may be set as an absolute threshold, for example, a pulse count exceeding 100 times per minute is considered an abnormal state. In some embodiments, the real-time characteristic value extraction and comparison process is executed cyclically at a fixed period to ensure that the verification of fault propagation path hypotheses reflects the latest operating state of the equipment. In specific implementations, all hypotheses of potential fault paths not marked as high confidence will be temporarily shelved or moved to the next verification cycle.

[0041] In one embodiment of the present invention, for each high-confidence potential fault path, its initial anomalous latent variable is analyzed, whereby the anomalous latent variable directly corresponds to a root cause component of the fault. In the multi-level causal network, the average causal effect value on all observable electrical parameter variables and key physical state variables is calculated when the anomalous latent variable is in an "abnormal" state. The average causal effect value is compared with the historical normal fluctuation range of the corresponding observed variable to calculate the relative impact intensity of the target fault on each observed variable, i.e., the percentage of change in the observed variable relative to its own normal fluctuation range. The relative impact intensity of the target fault on all key observed variables is summarized and sorted by intensity to form the characteristic impact spectrum of the corresponding fault mode.

[0042] All high-confidence potential fault paths are categorized according to the root component corresponding to their initial anomaly latent variables. From the fault components of the same type, the potential fault path with the highest verification key node state matching rate is selected as the fault evolution path for the corresponding fault component. All latent variables involved in the fault evolution path are translated from internal equipment physical state terms into specific mechanical or electrical component names. Combining the characteristic influence spectrum of the fault mode, the intensity level of the fault's impact on different monitoring signals is described. Finally, the fault component name, corresponding typical fault mode, fault evolution path description, and monitoring signal influence intensity level are integrated to generate a structured fault self-diagnosis report.

[0043] In practical implementation, based on the selected high-confidence potential fault paths, the root cause component of the fault is located, and the impact of the fault on different observed signals in multi-source monitoring data is quantified. This process is implemented as follows: For each high-confidence potential fault path, the anomalous latent variable at the beginning of the high-confidence potential fault path is analyzed. The anomalous latent variable directly corresponds to a root cause component of the fault. In the multi-level causal network, the average causal effect value on all observable electrical parameter variables and key physical state variables is calculated when the anomalous latent variable is in an "abnormal" state. The average causal effect value is compared with the historical normal fluctuation range of the corresponding observed variable to calculate the relative impact intensity of the target fault on each observed variable, which is the percentage of the change in the observed variable relative to its own normal fluctuation range. The relative impact intensity of the target fault on all key observed variables is summarized and sorted by intensity to form the characteristic impact spectrum of the corresponding fault mode.

[0044] Based on the quantification results of the impact degree and the root cause component location information, a diagnostic report containing the faulty component, fault mode, propagation path, and scope of impact is generated. This process is implemented as follows: All high-confidence potential fault paths are classified according to the root cause components corresponding to their initial abnormal latent variables. From the faulty components of the same type, the potential fault path with the highest verification key node state matching rate is selected as the fault evolution path of the corresponding faulty component. All latent variables involved in the fault evolution path are translated from the equipment's internal physical state terminology into specific mechanical or electrical component names. Combining the characteristic impact spectrum of the fault mode, the impact intensity level of this fault on different monitoring signals is described. Finally, the faulty component name, corresponding typical fault mode, fault evolution path description, and monitoring signal impact intensity level are integrated to generate a structured fault self-diagnosis report.

[0045] In some embodiments, the average causal effect value is calculated by intervening in the causal network model and observing changes in the predicted distribution. In specific implementations, the relative influence strength... The calculation follows the following relationship: Among them, symbols Indicates the effect of target fault on observed variables The relative intensity of the impact, symbol This represents the observed variable obtained through multi-level causal network calculations when the anomalous latent variable corresponding to the root component is intervened into an "abnormal" state. The average change of the predicted value relative to its normal baseline value, sign Represents observed variables The range of fluctuations in observed values ​​during historical normal operation is typically defined as the standard deviation or the width of a specific percentile interval of observed values ​​over a period of time. The calculated relative influence strength... It is a dimensionless percentage value used to quantify the relative impact of a fault on a specific observed signal.

[0046] It is understandable that the characteristic influence spectrum of a failure mode is a spectrum ordered by relative influence intensity. The list is sorted from largest to smallest, with each item recording the name of the observed variable and its corresponding relative influence strength value. In practice, the translation of latent variable names follows a predefined mapping table, which maps internal state terms such as "winding deformation degree" and "insulation aging degree" to specific component names such as "high-voltage winding phase A" and "main insulating bushing". Optionally, the description of the fault evolution path is formed by connecting a series of translated component names in the order of activation in the fault propagation simulation. In some embodiments, the monitoring signal influence strength level is based on the relative influence strength. Divide the numerical range, for example, define For "high influence", As a "medium influence", The impact is defined as "low". In practice, the structured fault self-diagnosis report is output in a specified electronic document format. The report content clearly lists the located faulty component, the inferred typical fault mode, the detailed fault evolution path, and the impact intensity level of each monitoring signal.

[0047] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A fault detection method for high and low voltage electrical equipment based on fault self-diagnosis, characterized in that, Includes the following steps: Obtain the status dataset of the high- and low-voltage electrical equipment to be tested; Based on the causal inference model, conditional independence tests are performed on standardized state datasets to identify potential causal dependencies between electrical parameters and physical states. Based on the identified potential causal dependencies, a multi-level causal network is constructed, with the internal physical state of the device as the latent variable and the observable electrical and physical signals as the explicit variables. In the multi-level causal network, known typical failure modes are introduced as intervention factors to simulate the failure propagation process and generate failure propagation path hypotheses. The high-frequency monitoring data collected in real time is compared and verified with the status of key nodes in the fault propagation path hypothesis to screen out potential fault paths with high confidence. Based on the selected high-confidence potential fault paths, the root cause component of the fault is located, and the impact of the fault on different observed signals in multi-source monitoring data is quantified. Based on the quantitative results of the impact and the root cause component location information, a diagnostic report is generated that includes the faulty component, fault mode, propagation path and scope of impact.

2. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, The acquisition of the status dataset of the electrical high and low voltage equipment to be tested includes: Collect multi-source monitoring data of the electrical high and low voltage equipment to be tested. The multi-source monitoring data includes vibration signals, temperature data, current waveforms, and partial discharge signals. The collected multi-source monitoring data is time-stamped and cleaned to remove abnormal noise and missing signal fragments, forming a standardized state dataset. The process of aligning and cleaning the collected multi-source monitoring data with timestamps, removing abnormal noise and missing signal fragments, and forming a standardized state dataset specifically includes: Raw vibration signals, temperature data, current waveforms, and partial discharge signals with their respective sampling times are acquired from different sensor acquisition channels. Using a global clock as a reference, an interpolation method is used to align the sampling time points of all original signals to a unified time series; For each time-aligned signal, a filter matching the signal characteristics is applied: a bandpass filter is used for vibration signals, a moving average filter is used for temperature data, a notch filter is used for current waveforms, and wavelet thresholding is used for partial discharge signals. Missing segments in the filtered signal are detected. For short-term missing segments, linear interpolation of data from the time points before and after the corresponding signal is used to fill them in. For long-term missing segments, all signal data corresponding to the time period will be marked as invalid and excluded from subsequent analysis. The filled signal data are resampled at uniform time intervals and combined into a standardized state dataset with the same time dimension but different signal dimensions.

3. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, The aforementioned causal inference model involves performing conditional independence tests on standardized state datasets to identify potential causal dependencies between electrical parameters and physical states. Specifically, this includes: From the standardized state dataset, current harmonic components, voltage RMS value, and three-phase unbalance are selected as key electrical parameter variables, and vibration dominant frequency amplitude, temperature gradient, and partial discharge pulse frequency are selected as key physical state variables. On the joint dataset consisting of key electrical parameter variables and key physical state variables, perform an independence test based on conditional mutual information to calculate the independence of each pair of variables given a subset of other variables. Record the variable pairs that pass the independence test, that is, the variable pairs that have conditional dependencies, and form a set of conditional dependencies; For each pair of variables in the set of conditional dependencies, the order of signal transmission is determined by analyzing the time delay cross-correlation between them. The variable that is ahead in time is initially marked as the cause, and the variable that is lagging behind is initially marked as the effect. Integrate all variable pairs with causal orientation to form a preliminary potential causal dependency graph; Using known knowledge of device operation mechanisms as constraints, the potential causal dependency graph is modified and verified. Causal edges that violate physical mechanisms are removed and necessary causal connections that conform to device operation logic are added, forming a causal inference model for constructing a multi-level causal network.

4. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, Based on the identified potential causal dependencies, a multi-level causal network is constructed, with the internal physical state of the device as the latent variable and observable electrical and physical signals as the explicit variables. Specifically, this includes: Define a set of latent variables characterizing the core state of the equipment, including the degree of winding deformation, the degree of insulation aging, the degree of contact wear, the cooling efficiency, and the degree of mechanical looseness; Based on the preliminary potential causal dependency diagram, the observable key electrical parameter variables and key physical state variables are used as explicit observation indicators of latent variables to establish an observation model from latent variables to explicit variables. Using the structural causal model framework and based on the knowledge of equipment operation mechanism, structural equations are established among latent variables to describe the causal relationship between latent variables; By merging the structural equations between latent variables with the observational model from latent variables to manifest variables, a complete multi-level causal network structure containing latent and manifest layers is formed. Using a standardized state dataset, the posterior probability distribution of all latent variables in a multi-level causal network structure is estimated through variational inference.

5. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, In the multi-level causal network, known typical failure modes are introduced as intervention factors to simulate the failure propagation process and generate failure propagation path hypotheses, specifically including: Establish a library of typical fault modes, including but not limited to winding turn short circuit, partial insulation breakdown, poor contact, heat sink blockage and bearing wear; Select a failure mode from the typical failure mode library, and set the hidden variable state corresponding to the failure mode to an "abnormal" state in the multi-level causal network. This operation is regarded as a deterministic intervention. In the multi-level causal network after intervention, the causal propagation simulation algorithm is run to iteratively calculate the probability change of its influence on the state of downstream latent variables, starting from the latent variable being intervened and along the causal edges between latent variables. When the probability of an abnormal state of a downstream latent variable exceeds a preset threshold, the corresponding latent variable is activated, its state is also considered "abnormal", and the influence continues to propagate to its downstream latent variables until no new latent variables are activated. Record the propagation order and path from the root cause hidden variable to all activated abnormal hidden variables in this simulation, and form the fault propagation path hypothesis for the corresponding fault mode.

6. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, The process of comparing and verifying the real-time collected high-frequency monitoring data with the key node states in the fault propagation path hypothesis to screen out potential fault paths with high confidence specifically includes: From each fault propagation path hypothesis, we extract latent variables that are in the middle of the propagation path and have clearly observable signals, as key nodes for verification; Determine the type of observable signal corresponding to each key verification node and its characteristic value range under normal and abnormal states; Initiate high-frequency monitoring of the equipment, collect and verify the latest observation signal segments corresponding to key nodes, and extract the corresponding feature values ​​from the signal segments in real time; The real-time extracted feature values ​​are compared with the preset range of normal and abnormal state feature values ​​to determine whether the actual state of the verification key node matches the state predicted in the fault propagation path hypothesis. The state matching rate of all key verification nodes in a fault propagation path hypothesis is calculated. If the matching rate is higher than a set threshold, the corresponding fault propagation path hypothesis is marked as a potential fault path with high confidence.

7. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, The process of locating the root cause component of the fault based on the selected high-confidence potential fault paths, and quantifying the impact of the fault on different observed signals in multi-source monitoring data, specifically includes: For each high-confidence potential failure path, analyze its initial abnormal hidden variable, which directly corresponds to a root cause component of the failure. In the multi-level causal network, the average causal effect value on all observable electrical parameter variables and key physical state variables is calculated when the anomalous latent variable is in an "abnormal" state. The average causal effect value is compared with the historical normal fluctuation range of the corresponding observed variable to calculate the relative impact intensity of the target fault on each observed variable, that is, the percentage of the change of the observed variable to its own normal fluctuation range. The relative impact intensity of the target failure on all key observed variables is summarized and sorted by intensity to form the characteristic impact spectrum of the corresponding failure mode.

8. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 1, characterized in that, Based on the quantification results of the impact degree and the root cause component location information, a diagnostic report is generated that includes the faulty component, fault mode, propagation path, and scope of impact, specifically including: All high-confidence potential failure paths are classified according to the root component corresponding to their initial anomaly hidden variable; From similar faulty components, the potential fault path with the highest verification key node state matching rate is selected as the fault evolution path of the corresponding faulty component; Translate all the hidden variables involved in the fault evolution path from the terminology of the internal physical state of the equipment into the names of specific mechanical or electrical components; Based on the characteristic impact spectrum of the fault mode, describe the intensity level of the impact of this fault on different monitoring signals; Finally, the faulty component name, corresponding typical fault mode, fault evolution path description, and monitoring signal influence intensity level are integrated to generate a structured fault self-diagnosis report.

9. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 5, characterized in that, The aforementioned method involves running a causal propagation simulation algorithm on the multi-level causal network after intervention. Starting from the intervened latent variable, iteratively calculating the probability change of its influence on the state of downstream latent variables along the causal edges between latent variables. Specifically, this includes: The state probability distribution of the latent variable to be intervened is forcibly changed from the posterior probability distribution to a deterministic distribution representing complete anomaly; Starting with the latent variable being intervened as the starting node, traverse all directed causal edges originating from the starting node in the multi-level causal network; For the currently activated anomalous latent variable, along each of its directed causal edges, calculate the incremental change in the state probability of its adjacent downstream latent variables based on the causal effect strength defined in the structural equation. The original posterior probability distribution of the downstream latent variables is superimposed with the incremental change to obtain its updated state probability distribution. Check the updated state probability distribution. If the probability of abnormal state exceeds the preset activation threshold, mark the downstream hidden variable as a new abnormal activation node and add it to the subsequent propagation traversal process until no new node is activated.

10. The fault detection method for high and low voltage electrical equipment based on fault self-diagnosis according to claim 6, characterized in that, The step of extracting latent variables with clearly observable signals in the middle of each fault propagation path hypothesis as key verification nodes specifically includes: For the activation sequence of latent variables recorded in the fault propagation path hypothesis, remove the root fault latent variable that is at the beginning of the sequence; In the remaining sequence of latent variables, those latent variables that correspond to weak observed signals, whose features are difficult to extract, or that are easily affected by environmental interference are removed. From the sequence of latent variables that have been eliminated, select the most frequently occurring latent variables in order as candidate key nodes for verification. Check each candidate verification key node to ensure that the corresponding observed signal type has been monitored and that the characteristic value range of its normal and abnormal states has been predefined in historical data; The final selected candidate verification key nodes are the verification key nodes used for actual comparison.

Images