Cooperative identification method and system for power line communication signal and arc fault

Abstract

The invention provides a cooperative identification method and system for a power line communication signal and an arc fault, and the method comprises the steps: firstly, constructing and training a double-flow communication fault sensor comprising a sensing early-warning sub-module and an active sensing sub-module, then carrying out the communication fault sensing based on the double-flow communication fault sensor, generating a communication fault sensing data set, and transmitting the communication fault sensing data set to a server; then, a graph neural network is adopted to carry out fault mode recognition on the communication fault sensing data set, a communication fault mode graph is generated, then, communication fault positioning is carried out based on the communication fault mode graph, a communication fault positioning data set is generated, and finally, actual feedback data of the communication fault positioning data set is obtained. And the system is updated according to the actual feedback data, so that the fault diagnosis rate and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and more specifically, to a method and system for the collaborative identification of power line communication signals and arc faults. Background Technology

[0002] In modern power systems, reliable and efficient communication is an important guarantee for ensuring safe and stable operation. However, communication failures caused by factors such as electric arcs, electromagnetic interference, equipment deterioration, and loose connections occur frequently. Moreover, communication failures are often characterized by strong concealment, intermittency, and high propagation. Traditional methods mostly rely on passive monitoring based on fixed thresholds, single-dimensional signal analysis, and isolated diagnosis based on expert rules, which often have certain limitations when dealing with communication failures.

[0003] On the one hand, the fault diagnosis process of traditional methods often remains at the level of correlation analysis based on statistics or simple rules, and mostly lacks in-depth exploration and analysis of the inherent causal logic of the fault and the actual propagation path in the physical topology; on the other hand, the diagnostic results are usually presented in the form of simple alarms or probabilities, often lacking quantitative integration of evidence from different sources and uncertainty assessment, which makes it difficult to trace the real root cause of the fault and is prone to misjudgment. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for the collaborative identification of power line communication signals and arc faults, the method comprising:

[0005] Construct and train a dual-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule;

[0006] Based on the dual-stream communication fault sensor, communication faults are detected, and a communication fault detection dataset is generated.

[0007] A graph neural network is used to perform fault pattern recognition on the communication fault perception dataset to generate a communication fault pattern map.

[0008] Based on the communication failure mode map, communication failure localization is performed, and a communication failure localization dataset is generated.

[0009] Obtain actual feedback data from the communication fault location dataset and update the system based on the actual feedback data.

[0010] Furthermore, embodiments of the present invention also provide a collaborative identification system for power line communication signals and arc faults, comprising:

[0011] A structural assembly module is used to construct and train a dual-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule.

[0012] A fault perception module is used to perceive communication faults based on the dual-stream communication fault sensor and generate a communication fault perception dataset.

[0013] The pattern recognition module is used to perform fault pattern recognition on the communication fault perception dataset using a graph neural network to generate a communication fault pattern map.

[0014] A fault location module is used to locate communication faults based on a communication fault mode map and generate a communication fault location dataset.

[0015] The feedback optimization module is used to acquire the actual feedback data of the communication fault location dataset and update the system based on the actual feedback data.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] The sensing and early warning submodule of the dual-stream communication fault sensor effectively filters out normal fluctuations in the power grid based on the constructed dynamic baseline and accurately captures early and weak fault signs, thereby issuing an early warning. Its active sensing submodule can automatically generate and execute safe and targeted active detection after the early warning is triggered, realizing the transformation from passively waiting for the fault to appear to actively stimulating fault characteristics, thus solving the problem of insufficient detection capability of traditional methods for intermittent and latent faults.

[0018] Meanwhile, by performing causal reasoning on multi-source data obtained through proactive exploration and associating it with the actual power grid physical topology, a communication fault mode map is generated. This map can clearly identify the most likely root cause of the fault and show the complete impact path of the root cause. Furthermore, by performing feature extraction and uncertainty quantification on the communication fault mode map, efficient and accurate analysis of the map is achieved, thereby improving the speed and accuracy of fault diagnosis. Attached Figure Description

[0019] Figure 1 This is a flowchart of the steps of a collaborative identification method for power line communication signals and arc faults according to the present invention.

[0020] Figure 2 This is a schematic diagram of a collaborative identification system for power line communication signals and arc faults according to the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart of the steps of a collaborative identification method for power line communication signals and arc faults according to the present invention. The following is a detailed description of the collaborative identification method for power line communication signals and arc faults.

[0022] Step S1: Construct and train a dual-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule.

[0023] Specifically, the active perception submodule consists of an exploration strategy generator and a response data collector. By constructing and training the two submodules, the dual-stream communication fault sensor is constructed.

[0024] In this embodiment, step S1 includes:

[0025] Step S1-1: Construct and train the perception and early warning submodule.

[0026] Specifically, the perception and early warning submodule consists of two early warning channels: a communication channel and an energy channel. The communication channel is used to monitor the spectral efficiency offset index and decodes and analyzes the spectral efficiency offset index using a hidden Markov model to obtain the corresponding communication channel monitoring data. The energy channel is used to perform short-time Fourier transform on the sampled current of the power line to generate a time-frequency spectrum matrix, and simultaneously combine it with a pre-trained attention network for feature analysis to generate an attention weight matrix.

[0027] Furthermore, the real-time frequency spectrum matrix is ​​weighted based on the attention weight matrix to obtain energy channel monitoring data. When the communication channel monitoring data and the energy channel monitoring data reach the early warning conditions, a fault early warning is triggered, and a corresponding fault early warning vector is generated.

[0028] In one possible embodiment, the perception and early warning submodule continuously monitors the operating status of the power grid and compares it with the corresponding dual-channel baseline to capture abnormal signs that deviate from the normal mode. The communication channel is used to monitor the quality of physical communication links such as power line carrier and wireless private network. Specifically, it monitors the composite index "spectral efficiency offset" in real time with a fixed time window (e.g., 10 milliseconds). The index can be calculated as: spectral efficiency offset = (actual throughput / theoretical Shannon capacity at the current signal-to-noise ratio) * (1 - bit error rate). The value of this index fluctuates between 0 and 1. The closer it is to 1, the closer the communication efficiency is to the ideal state.

[0029] For example, suppose that in a certain time window, the current actual throughput is 4.5 Mbps, the theoretical limit of the current signal-to-noise ratio is 5 Mbps, and the detected bit error rate is 0.01, then the spectral efficiency offset within the current time window is (4.5 / 5)*(1-0.01)=0.9*0.99=0.891. The system will continuously track this value.

[0030] An online learning Hidden Markov Model is used to model the time series of the spectral efficiency offset index. Specifically, it is assumed that the communication state of the system transitions between a finite number of hidden states such as excellent, fluctuating, and deteriorating, and the observed spectral efficiency offset value is generated by these states with a certain probability distribution. By using online algorithms such as the Baum-Welch algorithm with a sliding window to continuously update the state transition probability matrix and the observation probability distribution, a dynamic baseline corresponding to the index is constructed.

[0031] For example, suppose the communication state is abstracted into three hidden states: stable, fluctuating and deteriorating. Each state is assumed to have an observed value, namely the spectral efficiency offset, which follows a Gaussian distribution N (mean μ, variance σ²). In the initialization phase, the parameters corresponding to the hidden states are initially set based on historical normal data. For example, the mean of state S0 is initialized to 0.92 and the standard deviation is initialized to 0.05. Similarly, the transition probability matrix between each hidden state and the initial values ​​of the initial state distribution are set based on historical data.

[0032] The Baum-Welch algorithm is used for online parameter learning to achieve dynamic baseline modeling. This involves maintaining a fixed-length sliding window of length L to store the L most recent spectral efficiency offset observations. Whenever a new observation is acquired, it is added to the end of the window, and the old observation at the front is removed to maintain the window length. Next, based on the complete observation sequence within the current window, the algorithm recursively calculates the forward and backward probabilities for each state at each time step. Based on these probabilities, the probability of the system being in state i at time t (state occupancy probability γ_t(i)) and the probability of transitioning from state i to state j (state transition probability ξ_t(i,j)) are calculated. The mean of the new state observation distribution is updated by a weighted average of the old mean and the observations within the current window, where the weight of each observation is its corresponding state occupancy probability. Similarly, the variance and state transition probabilities are updated in the same way. This process is iteratively executed until the parameter change is less than a preset threshold or the maximum number of iterations is reached, thus completing the baseline construction.

[0033] The Viterbi algorithm is used to decode the most likely state sequence under a given observation sequence, and the log-likelihood probability of the entire observation sequence in the current state is obtained. The log-likelihood probability is obtained by summing the forward probabilities of all states at the final time and taking the logarithm. Then, the log-likelihood value is normalized by dividing by the sequence length, and the negative of the log-likelihood value is taken as the anomaly score. The larger the anomaly score, the more abnormal the sequence is. The proportion of the K most recent time moments in the deteriorated state is counted. A communication channel warning is triggered and a corresponding communication warning vector is generated if and only if the anomaly score and the proportion of the deteriorated state both exceed a preset threshold.

[0034] Understandably, the communication early warning vector includes at least four dimensions: anomaly status identifier, deviation, duration, and anomaly confidence, as well as a feature vector corresponding to the communication channel. The anomaly status identifier is defined as the identifier triggered when a communication channel early warning is activated. The deviation is expressed as (the difference between the current anomaly score and a preset threshold for anomaly scores) / (the difference between the historical maximum anomaly score and a preset threshold for anomaly scores). The duration is represented by the number of consecutive time windows in a deteriorated or fluctuating state from the first decoding of the deteriorated state by the Hidden Markov Model to the current moment, multiplied by the corresponding window size. The anomaly confidence is represented as the proportion of states in a deteriorated state in the last K moments. The feature vector corresponding to the communication channel includes at least spectral features, such as the current channel's power spectral density, dominant interference frequency, and its offset relative to the fundamental frequency; modulation domain features, such as the divergence of the constellation diagram, error vector amplitude, and phase noise variance; impulse response features, such as the root mean square delay spread and the shape parameters of the power delay spectrum extracted from channel estimation; and temporal features, such as the temporal sequence of spectral efficiency offset.

[0035] The energy channel is used to capture weak nonlinear distortions in the power grid current waveform. Specifically, a windowed short-time Fourier transform is applied to the high-frequency sampled current data to convert the one-dimensional time series into a two-dimensional time-spectrum matrix. In this matrix, the row vectors correspond to frequency components, the column vectors correspond to time windows, and the element values ​​represent the energy intensity of the signal at a specific time and frequency. The time-spectrum matrix can simultaneously describe the time-domain and frequency-domain characteristics of the signal, thereby enabling the capture of features such as instantaneous and non-stationary distortions like electric arcs.

[0036] It should be noted that, in order to locate the region corresponding to the fault feature in the broad time-frequency domain, a pre-trained attention neural network is introduced. This network uses the time-frequency spectrum of historical normal data and typical fault data such as electric arc and harmonic interference as the training set, and deep learning is used to train the attention neural network. Through training, the network can learn knowledge such as "energy accumulation in a specific high-frequency band near the current zero crossing point is a strong feature of electric arc" and encode it as a high weight value, thereby generating an attention weight matrix with the same dimension as the input time-frequency spectrum. Each element of the weight matrix represents the importance of fault detection at the corresponding time-frequency point.

[0037] For example, suppose a short-time Fourier transform analysis is performed on a 50Hz current signal containing an arc event to obtain the time spectrum. Based on a pre-trained attention neural network, the time spectrum is analyzed to generate a corresponding attention weight matrix. This attention weight matrix can be represented as outputting higher attention weights near each zero-crossing point (e.g., 10ms, 20ms...) on the time axis corresponding to the current arc event, and at frequency bands corresponding to odd harmonics (e.g., 150Hz, 250Hz...) or higher frequency noise on the frequency axis; while outputting lower weights near the current peak or fundamental frequency, thereby achieving the annotation of the time-frequency region that needs to be focused on.

[0038] The time-spectrum matrix obtained in real time from the energy channel is multiplied element-wise with the aforementioned attention weight matrix to achieve focused enhancement of the time spectrum. This aims to automatically strengthen the characteristics of fault-sensitive areas and suppress the energy contribution of normal operating conditions or irrelevant interference areas. By summing the amplitudes of all elements in this weighted time-spectrum matrix, a scalar value representing the attention-weighted distortion energy within the current time window is obtained. This scalar value is defined as the attention-weighted distortion energy value. Simultaneously, a sliding time window is used to obtain the time sequence corresponding to the attention-weighted distortion energy value.

[0039] For example, suppose that in the above example of an electric arc event, the corresponding attention weight matrix has an original spectral amplitude of 5 and an attention weight of 0.9 at the zero-crossing point and a frequency of 250Hz; and an original spectral amplitude of 100 and an attention weight of 0.01 at the non-zero-crossing point and a frequency of 50Hz. After weighting, the former contributes 4.5 and the latter contributes 1. After weighted summation, a scalar value is finally obtained, in which the former's contribution is more significant, meaning that the characteristics of the electric arc event in the time spectrum are more obvious. By weighting the spectrum, even if the fundamental energy is high, as long as its mode is normal, that is, the weight value in the attention weight matrix is ​​low, the scalar value will not be significantly pushed up. Conversely, even if the absolute value of the fault feature energy is not large, as long as it appears in the key time-frequency region, that is, the region with high weight value in the attention weight matrix, its corresponding feature will be effectively captured and amplified, thereby greatly improving the signal-to-noise ratio of the feature and the ability to identify the fault.

[0040] A dynamic baseline for attention-weighted distortion energy values ​​is constructed using a variable bandwidth kernel density estimation mechanism. Specifically, a fixed bandwidth is used for initial density pre-estimation, and then the bandwidth of the kernel function is adaptively adjusted according to the local density in the neighborhood of each data point. This allows for the use of a wider bandwidth at the tail of the sparse data distribution to ensure smoothness, and a narrower bandwidth in the peak region of the dense data to preserve details. It should be noted that each attention-weighted distortion energy value judged as normal by the subsequent early warning mechanism is included in a pre-constructed historical sample pool. Based on the data in this sample pool, the baseline is periodically recalculated, thereby achieving dynamic adaptive adjustment of the baseline.

[0041] For example, suppose 10,000 attention-weighted distortion energy values ​​under normal operating conditions are collected in the initial stage, with values ​​roughly between [0.05, 0.22]. Using fixed bandwidth kernel density estimation for preliminary estimation, it is found that most data are concentrated around 0.1, and there is also a small bulge around 0.2. When performing variable bandwidth kernel density estimation, a smaller bandwidth (such as 0.01) is allocated to the dense data points in the 0.1 region, making the estimated probability density curve sharp at this point; for the sparser data points around 0.2, a larger bandwidth (such as 0.03) is allocated, making the curve transition smoothly at this point, thus ensuring that the fitted probability density curve can more accurately describe the true distribution of the data.

[0042] The early warning mechanism of the energy channel is represented as a two-layer decision mechanism that integrates instantaneous anomaly assessment and sequence mutation detection. Specifically, a probability density function for historical data is generated based on a variable bandwidth kernel density estimation mechanism. This function describes the probability density of the attention-weighted distortion energy value taking any specific value under historical normal operating conditions. By performing a numerical integration operation on this probability density function from the observed value of the current attention-weighted distortion energy value to positive infinity, the right-tail probability of the current observation value is obtained. The right-tail probability represents the estimated probability of randomly observing an attention-weighted distortion energy value larger than the current observation value under historical normal operating conditions, i.e., the baseline state. The smaller the value, the lower the probability that the current observation value occurs in the normal distribution. This right-tail probability value is compared with a threshold preset by analyzing historical data. If it is lower than the threshold, it indicates that observing the current value under the current normal state is an event with an extremely low probability, thus marking the current observation value as an extreme anomaly candidate event.

[0043] A parallel Bayesian change point detection algorithm is used to continuously monitor the time series of attention-weighted distortion energy values. Specifically, the number of data points accumulated since the last statistical characteristic change is defined as the running length. A running length of zero indicates that a change has occurred at this moment. At each time step, the prior distribution of the current running length is predicted based on the posterior distribution of the running length at the previous time step. When new observation data is obtained, the likelihood of the data under each running length assumption is acquired. Combining the prior distribution and the likelihood, the posterior probability distribution of the current running length is updated using Bayes' theorem. The probability of a running length of zero is extracted from the posterior distribution. If this probability exceeds a preset threshold, it indicates an abnormal state.

[0044] An energy channel warning is triggered only when both the right-tail probability is below a preset threshold and the probability of a zero running length exceeds a preset threshold. A corresponding energy warning vector is then generated. This energy warning vector includes at least four dimensions: an anomaly status identifier, deviation, duration, and anomaly confidence, along with a corresponding energy channel feature vector. The anomaly status identifier indicates that an energy channel warning is triggered; the deviation is represented by the negative logarithm of the right-tail probability; the duration is the time from the start of the warning until its end; and the anomaly confidence is the geometric mean of the right-tail probability and the probability of a zero running length. The energy channel feature vector includes at least time-frequency domain features, such as the coordinates, energy percentage, and frequency concentration of the most concentrated time-frequency region extracted from the weighted time-frequency spectrum; harmonic features, such as the amplitude ratio and phase relationship of odd harmonics relative to the fundamental wave; transient features, such as distortion morphology parameters of the current waveform near the zero-crossing point, including dip depth, width, and asymmetry; and nonlinear features, such as the instantaneous frequency and amplitude modulation features of the signal extracted through Hilbert transform.

[0045] For example, suppose that based on historical normal data, kernel density estimation shows that the 99.9th percentile of the attention-weighted distortion energy value is 0.30, and the latest observed attention-weighted distortion energy value is 0.52. Assume that the calculated right-tail probability is 0.0003, which is less than the corresponding preset threshold of 0.001, thus satisfying the condition of an extreme anomaly. Simultaneously, the Bayesian change point detection algorithm continuously analyzes the observation sequence of the most recent attention-weighted distortion energy values. Assume that before incorporating the latest observation value of 0.52, the observation sequence came from a stable distribution with a mean of approximately 0.12, meaning no sudden changes occurred. When the extreme value of 0.52 is input, the probability of a statistical mutation occurring at the current moment suddenly increases to 0.72, exceeding the preset threshold of 0.5. Since both conditions of the energy channel are satisfied simultaneously, this is determined to be an anomaly caused by a persistent state change, rather than random noise. Therefore, an energy channel warning is triggered, and a corresponding warning vector is generated.

[0046] A fault warning vector is generated based on the communication warning vector and the energy warning vector. If only one of the communication channel warning or the energy channel warning is triggered, the corresponding warning vector is output as the fault warning vector. If both the communication channel warning and the energy channel warning are triggered, the communication warning vector and the energy warning vector are concatenated into a comprehensive vector and output as the fault warning vector.

[0047] It should be noted that the perception and early warning submodule uses normal historical data covering different loads and operating conditions to pre-train the communication channel, and combines gradient descent method to train the energy channel. Specifically, in the pre-built digital twin platform, various typical scenarios that the power grid may experience during long-term operation are simulated. For example, load fluctuations at different time scales such as day-night cycles and weekday and weekend modes, combined operation of various types of loads such as resistive, inductive, capacitive and nonlinear loads, and simulation of environmental factor changes such as the effect of simulated temperature on line parameters are generated. This generates time series data covering different loads and operating conditions. The time series data generated without faults and the normal historical data are used as the first training dataset, and the time series data generated with faults and the normal historical data are used as the second training dataset. The first and second training datasets are encapsulated into a training dataset. The training dataset contains at least the spectral efficiency offset and the corresponding attention-weighted distortion energy value of the communication channel in each training dataset.

[0048] Based on the spectral efficiency offset data in the training dataset, the Baum-Welch algorithm is used to pre-train the Hidden Markov Model of the communication channel. Specifically, the state semantics of the training data are defined and labeled, defining three states: stable S0, fluctuating S1, and deteriorating S2. The training data is then time-stamped and associated to train the communication channel to learn the corresponding characteristics. For example, state sequence segments are manually labeled before and after the simulation segment of a large motor starting up: {At the moment of starting up: from S0 to S1, after starting up: from S1 to S0}. Based on this state sequence segment, the model is guided to learn that the fluctuating state is a transient process that quickly returns to stability, rather than a continuous deterioration process. Through continuous training and learning, the state transition probability matrix output by the communication channel eventually realizes the empirical evidence such as "for fluctuating scenarios, the probability of the state jumping from S1 back to S0 is much higher than the probability of jumping to S2".

[0049] Based on the attention-weighted distortion energy values ​​in the first training dataset, the optimal global bandwidth parameter corresponding to the energy channel is determined simultaneously using maximum likelihood cross-validation. The attention-weighted distortion energy values ​​in the second training dataset are added to the training pool of the energy channel as an interference sample set. When re-estimating using the variable bandwidth kernel density estimation mechanism, the interference sample set will increase the local density of the energy channel in the interference and fault regions. This ensures that when encountering similar legitimate interference in actual operation, the calculated right-tail probability will be at a level that is "somewhat increased, but still within an acceptable range", thus avoiding false triggering caused by a single high threshold.

[0050] Step S1-2: Construct and train the active perception submodule.

[0051] Specifically, the active perception submodule consists of an exploration strategy generator and a response data collector. The exploration strategy generator takes a fault warning vector as input, uses a neural network to perform inference analysis on the fault warning vector, and generates a corresponding fault exploration strategy based on the analysis results. The response data collector is used to collect the response data generated when the fault exploration strategy is executed and generate a strategy response dataset. The strategy response dataset contains at least the response data corresponding to the communication channel and the energy channel, as well as the execution timestamp corresponding to the fault exploration strategy.

[0052] In one possible embodiment, the fault warning vector is input into a multi-branch neural network, which includes at least the following three branches: a basic feature processing branch, which is composed of a fully connected network, for example, a 4-dimensional to 16-dimensional to 32-dimensional fully connected layer, using the ReLU activation function, and this branch is used to process the three basic dimensions of deviation, duration, and anomaly confidence, analyzing the severity and urgency level of the anomaly; a temporal feature processing branch, which is composed of a one-dimensional convolutional neural network, for example, a network consisting of two convolutional layers with kernel sizes of 3 and 5, and channel numbers of 16 and 32, respectively, with batch normalization and ReLU activation after each convolutional layer, and finally connected to a global average pooling layer, and this branch is used to capture the local patterns and trends of temporal features in the feature vector; and a frequency domain feature processing branch, which has a similar structure to the temporal feature processing branch, and is used to independently process the frequency domain features of the feature vector, thereby capturing the pattern of frequency distribution.

[0053] The outputs of the three branches are concatenated along the feature dimension. The concatenated features are then fused through multiple fully connected layers, for example, from 128 dimensions to 64 dimensions to 32 dimensions. ReLU activation and Dropout regularization are used for deep fusion and abstraction. It should be noted that the number of neurons in the last fully connected layer is equal to the number of atomic policies in the preset exploration policy library. The Softmax activation function is used to convert the output of the last fully connected layer into a probability distribution vector, which can be represented as P=[p_1,p_2,...p_i,...p_M], where p_i represents the recommendation probability of selecting the i-th atomic policy under the current abnormal feature, and M is the size of the preset exploration policy library.

[0054] The recommendation probabilities in the probability distribution vector are sorted from high to low, and atomic strategies with recommendation probabilities exceeding a preset threshold are selected to generate a candidate strategy set. The candidate strategy set is then filtered and sorted according to predefined strategy screening rules, which include at least: diagnostic logic order rules, such as prioritizing "exploratory / scanning" strategies followed by "verification / focusing" strategies; physical constraints, such as avoiding executing two potentially interfering strategies too close in time; and fault hypothesis priority rules, such as prioritizing exploratory strategies with high recommendation probabilities. The filtered and sorted strategies are combined into an ordered sequence of exploratory strategies, and the total execution time is estimated. For example, if the warning features indicate... The presence of high-frequency narrowband interference accompanied by periodic voltage dips indicates a problem. The partial probability distribution vector output by the multi-branch neural network is: {Fine spectrum scanning: 0.88, Voltage dip synchronization phase measurement: 0.75, High-frequency carrier impedance measurement: 0.60, General harmonic analysis: 0.40}. Based on the strategy screening rules, the final exploration strategy is as follows: First, perform a fine spectrum scan to pinpoint the precise interference frequency. Then, perform a high-frequency carrier impedance measurement to assess the network impedance at the interference frequency and preliminarily determine the type of interference source. Finally, at the identified sensitive frequency, perform a voltage dip synchronization phase measurement to analyze the phase relationship between the dip and the interference, verifying whether it is a periodic disturbance caused by a switching power supply or other loads.

[0055] It should be noted that the atomic policies in the preset exploration policy library are all constrained by the corresponding security rules in the domain, and the atomic policies will have their parameters adjusted according to the incremental learning packages generated periodically to adapt to the actual power grid environment, thereby ensuring that the current communication environment will not be damaged when the exploration policy is executed.

[0056] Understandably, the active perception submodule trains the exploration strategy generator through a simulation training process based on meta-learning, involving injecting faults, executing strategies, and collecting responses. Specifically, in a digital twin environment, it simulates single faults, compound faults, and fault scenarios of different locations and degrees. For each scenario, the perception and early warning submodule records the complete feature vector it generates. For each fault scenario, a strategy evaluation loop is executed. Specifically, it iterates through each atomic strategy in the strategy library, executes it under safety constraints, collects the corresponding system response, and evaluates the utility of each atomic strategy. The utility evaluation dimensions include at least: excitation, i.e., the degree of matching between the strategy response and the expected response features; discriminability, i.e., the difference between the strategy's response in this fault scenario and its response in other fault scenarios; and safety, i.e., the magnitude of system disturbance caused by the strategy execution. By quantitatively analyzing the above three dimensions, the top k strategies with the best utility are selected as the expert strategies for this fault scenario. The training process uses multi-label cross-entropy loss as the loss function to encourage the exploration strategy generator to output high probabilities for expert strategies.

[0057] Step S2: Based on the dual-stream communication fault sensor, perform communication fault detection and generate a communication fault detection dataset.

[0058] Specifically, fault warnings are generated based on the perception and early warning submodule. When the warning conditions are met, a fault warning is triggered, and a corresponding fault warning vector is generated. The fault warning vector includes at least four dimensions: abnormal state identifier, deviation degree, duration, and abnormal confidence degree. The active perception submodule generates and executes a corresponding fault exploration strategy based on the fault warning vector and obtains a strategy response dataset. A deep neural network is used to analyze the strategy response dataset to generate a corresponding communication fault perception dataset.

[0059] In one possible embodiment, the policy response dataset contains multiple data streams with different physical meanings, sampling rates, and time starting points, such as current waveforms sampled at 1MHz, communication bit error rates sampled at 10kHz, and policy action markers with event stamps, etc. First, the policy response dataset is spatiotemporally aligned, and then a multi-branch feature extractor is used to extract features from the policy response dataset, generating a corresponding feature dataset. The multi-branch feature extractor includes at least a waveform branch, which uses a one-dimensional convolutional layer to extract local morphological features from the voltage and current waveforms, such as rising edge steepness, zero-crossing distortion, and specific frequencies. The oscillation envelope generates the corresponding waveform feature vector; the frequency branch uses a two-dimensional convolutional layer to extract features such as the shape and movement trajectory of energy accumulation regions from the short-time Fourier transform spectrum of the signal, generating the corresponding spectral feature vector; the temporal branch uses a long short-time memory network to process the temporal variation sequences of communication data such as bit error rate and signal-to-noise ratio, extracting degradation patterns such as bursts, gradual changes, and periodicity, generating the corresponding temporal feature vector; an attention mechanism is used to obtain the correlation matrix between the waveform feature vector, spectral feature vector, and temporal feature vector, and a weighted fusion is performed based on this matrix to generate a fused feature vector.

[0060] For example, suppose a fault detection strategy of "zero-crossing injection -2% depth depression" is implemented. The strategy response dataset corresponding to this strategy is a series of high-speed sampled current values. The convolution kernel of the waveform branch of the multi-branch feature extractor is activated at the moment when the depression occurs in the current value sequence and outputs features such as "depression width = 110μs". The spectrum branch extracts features such as "3rd harmonic energy increase of 8dB" from the spectrum before and after that moment. The attention weight between "ringing feature of depression trailing edge" and "instantaneous interference feature of communication received signal at 150kHz frequency" is calculated to be 0.9 through the attention mechanism, while the feature weight with the steady-state current value of the load is only 0.05. This indicates that the current abnormal mode is likely caused by high-frequency interference affecting communication through the coupling path, rather than by changes in the basic load. The generated fused feature vector will carry the above information, thus providing a data basis for subsequent analysis.

[0061] Regression analysis is performed on the fused feature vector based on two pre-trained parallel fully connected classification subnetworks to output fault attribute label vectors and fault perception feature vectors. The attribute label vectors are probability distributions and include at least fault types such as arcing, loosening, and harmonic sources; coupling types such as capacitive coupling, inductive coupling, and common-mode conduction; and severity levels such as slight, moderate, and severe. For example, an attribute label vector can be represented as [arc: 0.85, series coupling: 0.70, capacitive coupling: 0.60]. The fault perception features include at least the estimated core interference frequency caused by the fault, the equivalent interference power or signal-to-noise ratio degradation referred to the communication receiver, the rate at which the fault intensity decays with electrical distance, and the ratio of the observed response change to the specific strategy action. These four data points are sequentially encapsulated into a vector form of [characteristic frequency, equivalent interference power, spatial attenuation coefficient, response gain]. For example, a fault perception feature vector can be represented as [characteristic frequency: 125000 Hz, equivalent interference power: -45]. dBm, spatial attenuation coefficient: 0.5, response gain: 0.32], the spatial attenuation coefficient in this vector represents that for every unit increase in electrical distance, the fault intensity decreases synchronously by 0.5 units.

[0062] Finally, the fault warning vector, fault exploration strategy, attribute label vector, and fault perception feature vector are encapsulated into a communication fault perception dataset for output.

[0063] It should be noted that the deep neural network adopts a phased supervised learning training strategy. Its training objective is to enable the network to accurately parse the semantic attributes and quantitative features of the fault from the policy response dataset. The training data comes from samples generated by the digital twin environment, where each sample contains at least the policy response dataset and the corresponding ground truth label. The ground truth label consists of the ground truth values ​​of the fault attributes preset by the fault scenario and the ground truth values ​​of the fault perception features. The training process adopts end-to-end supervised learning, and its loss function is a multi-task loss consisting of the weighted cross-entropy loss of the classification subtask and the smoothed L1 loss of the regression subtask. In this way, the network is trained to learn accurate attribute discrimination and feature regression simultaneously.

[0064] Step S3: Use a graph neural network to perform fault mode recognition on the communication fault perception dataset to generate a communication fault mode map.

[0065] In this embodiment, step S3 includes:

[0066] Step S3-1: Perform fault mode identification on the communication fault perception dataset.

[0067] In this embodiment, step S3-1 includes:

[0068] Step S3-1-1: Construct a directed cause-effect graph for fault propagation.

[0069] Specifically, a cross-convergence mapping and a causal discovery algorithm based on transfer entropy are used to perform causal mining on the communication fault perception dataset to generate a directed causal graph of fault propagation. The nodes of the directed causal graph of fault propagation contain fault observation variables and the fault exploration strategy actions executed. The fault observation variables are represented as the attribute label vector and fault perception feature vector of the communication fault perception dataset. The directed edges represent the causal driving relationship between nodes, and the weights represent the causal strength.

[0070] In one possible implementation, a cross-convergence mapping is used to analyze the <fault observation variable, fault exploration strategy action> tuple. Specifically, it examines whether the dynamics of the other variable can be reconstructed from the historical state of one variable in the tuple, thereby determining whether there is a potential causal coupling relationship between the two. Undirected connections are established between tuples with causal coupling relationships. The propagation entropy between tuples with undirected connections is calculated, and the causal propagation direction is assigned to the undirected connection edges based on the propagation entropy. The normalized propagation entropy is then used as the edge weight of the causal edge, ultimately generating the corresponding directed causal graph of fault propagation.

[0071] For example, suppose the communication fault awareness dataset contains the following synchronization time-series data: At time t, a policy action D1 is executed: a test signal of 150kHz, +3dBm is injected. Subsequently, the observed variable "amplitude at characteristic frequency 150kHz" at monitoring point A jumps from -50dBm to -47dBm, while the observed variable "communication bit error rate" at monitoring point B increases from 10... -5 Deteriorated to 10 -3 The causal discovery algorithm first performs a cross-convergence mapping analysis on the variable pair <policy action D1, observed variable A_amplitude> and <observed variable A_amplitude, observed variable B_bit error rate> to confirm the existence of a dynamic coupling relationship between the two. Next, it calculates the propagation entropy and finds that the values ​​of TE_{D1 points to A_amplitude} and TE_{A_amplitude points to B_bit error rate} are significantly greater than the propagation entropy in the opposite direction. Therefore, the algorithm generates two directed edges in the undirected causal graph composed of multiple tuples: from the policy action node D1 to the observed variable node A, with a weight of 0.90; and from the observed variable node A to the observed variable node B, with a weight of 0.82. This constitutes a clear causal chain, that is, the actively injected test signal D1 leads to an increase in the interference amplitude at point A, and the change at point A further leads to the degradation of the communication quality at point B.

[0072] Step S3-1-2: Construct a fault propagation topology diagram.

[0073] Furthermore, a physical topology graph is constructed based on the real-time power grid topology, with monitoring nodes as vertices and electrical connections as edges. A spatiotemporal graph convolutional network is used to perform convolutional inference on the physical topology graph to generate corresponding spatiotemporal convolutional vectors. Based on the spatiotemporal convolutional vectors and a preset fault propagation and diffusion mechanism, the influence diffusion intensity of each node in the physical topology graph is quantified to generate an influence diffusion intensity dataset. Based on the influence diffusion intensity dataset, the physical topology graph is optimized to generate a fault diffusion topology graph.

[0074] In one possible embodiment, a physical topology graph is constructed based on the real-time topology connection relationship of the power grid, with each monitoring node as the vertex and the actual electrical connections between nodes, such as the closed state of cables, busbars, and circuit breakers, as the edges. The initial weight of the edges is set according to physical parameters such as electrical distance and line impedance.

[0075] A spatiotemporal graph is constructed based on temporal information from the physical topology graph and the communication fault perception dataset, such as the feature frequencies and equivalent interference power of each node at different times. A spatiotemporal graph convolutional network is used to perform deep reasoning on the spatiotemporal graph, aggregating the features of neighboring nodes in the spatial dimension of the graph. Specifically, in each convolutional layer, graph convolution is performed in the spatial dimension to aggregate the features of neighboring nodes, thereby capturing their spatial dependencies; one-dimensional convolution is used in the temporal dimension to capture the evolution trend of features. After multiple layers are stacked, a high-dimensional spatiotemporal convolutional vector is generated for each node in the physical topology graph. This vector deeply integrates the historical fault feature evolution pattern of the node itself, as well as its positional context information in the entire network topology structure.

[0076] For example, suppose there are three monitoring nodes A, B, and C in a power grid, connected in the order ABC. Over 10 consecutive time steps, the "equivalent interference power" of node A remains high and fluctuates. Node B subsequently exhibits similar fluctuations but with decreasing amplitude. Finally, node C shows a slight fluctuation. When the spatiotemporal graph convolutional network performs spatiotemporal convolution on node B, its graph convolution part aggregates features from the strong interference source neighbor A and the affected downstream neighbor C. Its temporal convolution part analyzes the sequence pattern of B's ​​own disturbance from zero to its decay. Finally, the spatiotemporal convolutional vector generated for node B incorporates the information that "as an intermediate node, it bears the propagation influence from node A and spreads it downstream to node C."

[0077] The spatiotemporal convolution vectors of each node in the physical topology graph are processed based on a pre-defined classification head. Specifically, firstly, the spatiotemporal convolution vectors are input into the shared layer of the classification head to obtain hidden features. Then, the hidden features of a node are fed into the fault source classification branch to obtain the probability that the node is a fault source. Simultaneously, for each directed edge in the graph, the hidden features of its source node and target node are concatenated, and the concatenated feature is input into the influence intensity prediction branch to obtain the influence intensity value of the directed edge. After traversing all nodes and all directed edges, an influence diffusion intensity dataset is generated, which contains the fault source probability of each node and the influence intensity of each directed edge. The physical topology graph is then optimized based on the influence diffusion intensity dataset.

[0078] For example, suppose nodes A, B, and C are connected sequentially in a physical topology graph, with corresponding feature vectors h_A, h_B, and h_C. The hidden feature of node A is calculated using the first branch of the classification head, yielding a fault source probability P_A = 0.85. Simultaneously, for the directed edge from A to B, concatenating h_A and h_B, the second branch calculates the influence strength I_{A to B} = 0.9. Similarly, I_{B to A} = 0.1, I_{B to C} = 0.8, and I_{C to B} = 0.05. The influence diffusion intensity dataset records the following for these three nodes: [Node probability: {A: 0.85, B: 0.2, C: 0.1}; Edge strength: {(A,B): 0.9, (B,A): 0.1, (B,C): 0.8, (C,B): 0.05}. The dataset shows that A is highly likely to be the source of the fault, and the impact of the fault mainly propagates strongly from A to B, and then relatively strongly from B to C. The back propagation strength is very weak, clearly outlining the causal transmission path of the fault spreading along the path from A to B to C. New enhanced attributes are assigned to the corresponding nodes: node A is assigned {fault source probability: 0.85, role: suspected source node}, node B is assigned {fault source probability: 0.2, role: propagation relay node}, and node C is assigned {fault source probability: 0.1, role: end-effect node}. Edge reconstruction is performed: edge (A, B) is transformed into two directed edges: {direction: A to B, weight: 0.9, label: dominant direction}, {direction: B to A, weight: 0.1, label: non-dominant direction}. Similarly, edge (B, C) is also transformed into a corresponding bidirectional edge.

[0079] It should be noted that the classification head consists of a shared low-level feature processor and two parallel output branches. The shared part consists of one or two fully connected layers, which are responsible for further refining the input spatiotemporal convolutional vectors into more abstract hidden features. The fault source classification branch is represented by mapping the hidden features to a 2D output through a fully connected layer, corresponding to the two categories of "is a fault source" and "is not a fault source", respectively. After normalization by the Softmax function, the output is the probability that the node is the initial fault source. The influence intensity prediction branch, for each directed edge in the physical topology graph, such as from node u to node v, concatenates the hidden feature vectors of nodes u and v to form a joint feature vector, which is then processed by a multilayer perceptron. Finally, the sigmoid activation function outputs a scalar between 0 and 1, representing the predicted intensity of the fault disturbance propagating from node u to node v.

[0080] The classification head and the spatiotemporal graph convolutional network are jointly trained. The training process is completed on a labeled dataset consisting of simulation and historical failure cases. Specifically, a multi-task joint loss function is used for supervised learning. The loss function is composed of a weighted sum of two parts: the cross-entropy loss for node failure source classification and the mean square error loss for edge influence intensity prediction. The cross-entropy loss is used to make the predicted probability approximate the true failure source label, and the mean square error loss is used to make the predicted intensity approximate the true propagation intensity label.

[0081] Step S3-2: Generate a communication fault mode map by fusing the data.

[0082] Specifically, a cross-graph attention mechanism is used to align the entity nodes of the fault propagation directed causal graph and the fault propagation topology graph, generating a corresponding cross-graph associated node set. The fault propagation directed causal graph and the fault propagation topology graph are iteratively optimized according to a dual-flow iterative reasoning mechanism that includes causal responsibility flow and spatial influence flow. Based on the optimized fault propagation directed causal graph and the fault propagation topology graph and the cross-graph associated node set, a communication fault mode graph is generated.

[0083] Understandably, the causal responsibility flow refers to obtaining the causal contribution of each node in the fault propagation directed causal graph to the observed fault, and optimizing the fault propagation topology graph based on the causal contribution; the spatial influence flow refers to obtaining the likelihood strength of the fault along the edge connection in the fault propagation topology graph, and optimizing the fault propagation directed causal graph based on the likelihood strength.

[0084] In one possible embodiment, the fault propagation directed causal graph and the fault propagation topology graph are input into a graph attention network to calculate the attention coefficient between any two nodes in the two graphs. This coefficient is determined by their feature similarity and their roles in the graph structure. Specifically, for each node in the causal graph, the attention weight between it and all nodes in the topology graph is calculated, and then the topology node with a weight exceeding a threshold is selected as its alignment target. Similarly, for each node in the topology graph, since there may be multiple observation variables on a device, each node in the topology graph may also be aligned with multiple nodes in the causal graph. By calculating the attention weight and filtering, a cross-graph associated node set is obtained, where each element is a vector consisting of a <causal graph node, topology graph node> tuple and the corresponding alignment confidence.

[0085] It should be noted that the graph attention network maps node features from the fault propagation directed causal graph and the fault propagation topology graph to a common semantic embedding space. In this space, the cosine similarity or dot product between the corresponding node feature vectors of the <causal graph node, topology graph node> tuple is calculated to obtain the original attention score. The original attention score is then normalized using the Softmax function to obtain the final attention coefficient. This attention coefficient represents the conditional probability that a given causal graph node points to the same entity as each topology graph node.

[0086] During the training phase, the graph attention network uses a large amount of historical fault data and simulation data as training datasets. The training datasets are labeled with the correct node correspondences, such as the variable "150kHz harmonic amplitude of monitoring point A" corresponding to "physical equipment: bus A". The network parameters are optimized by minimizing the cross-entropy loss between the predicted alignment and the true alignment. After training, for any causal graph node, an alignment pair is formed by selecting the node with the highest attention coefficient from the fault propagation topology graph, and the attention coefficient corresponding to the alignment pair is used as the corresponding alignment confidence.

[0087] Causal responsibility flow is used to quantify the causal contribution of each node in the directed causal graph of fault propagation to the observed fault, and to optimize the edge weights of the fault propagation topology graph based on this contribution. Specifically, firstly, the responsibility score of each node in the causal graph is obtained based on the directed edge weights and the node's own anomaly degree. The anomaly degree is generated based on the anomaly confidence in its corresponding fault warning vector. For example, suppose there are three nodes in the causal graph: A, B, and C, and their corresponding anomaly degree values ​​are 0.95, 0.7, and 0.6, respectively. The directed edges and causal strengths are: {AB, 0.8}, {AC, 0.6}, and a propagation probability of 0.85 is set. The propagation probability is expressed as an 85% probability of propagating along the current causal edge. Since A has no upstream node, its original responsibility score is 0.85 * 0.95 = 0.8075. Node B is a downstream node of A, so it receives the causal contribution from upstream node A, and its corresponding original responsibility score is 0.85 * (0.95 * 0.8) + 0.15 * 0.7 = 0.751. Similarly, the original responsibility score of node C is 0.5745. The above process is iteratively executed until the change in the original responsibility score of each node is less than the threshold or a fixed number of iterations is reached. The final converged original responsibility score is the responsibility score of each causal graph node.

[0088] By associating nodes across the graph, the responsibility scores of the causal graph nodes are propagated to the aligned topological graph nodes, thereby assigning corresponding initial responsibility scores to the topological graph nodes. Then, each directed edge in the topological graph is traversed, and the propagation strength of the edge is adjusted according to the responsibility scores of its starting and ending points. That is, if the starting point responsibility score is high and the ending point responsibility score is low, the weight of the edge is increased, indicating that the fault is more likely to propagate in this direction, and vice versa. For example, if the difference between the responsibility scores of the starting point and the ending point is greater than a preset threshold, the difference in responsibility scores is multiplied by a preset weight factor to obtain the corresponding adjusted weight value, and the adjusted weight value is added to the original weight of the corresponding edge.

[0089] The spatial influence flow obtains the propagation likelihood strength of each edge in the current fault propagation topology graph based on the edge weights and preset physical connection characteristic values. Then, for each directed edge in the directed causal graph of fault propagation, it finds the topology graph nodes that align with its start and end nodes in the fault propagation topology graph. If these two topology graph nodes are the same, i.e., the start and end points are the same, it is considered as internal coupling of the device, and a preset high spatial support strength value is directly assigned. This value is obtained by analyzing historical data. If they are different, it iterates through the topology graph to see if there is a path between the two, and takes the aggregated value of the propagation likelihood strength on the path as the spatial support strength of the causal edge, and adjusts the weight of the causal edge according to this spatial support strength.

[0090] For example, suppose there is a cable edge E1 in the topology graph from bus D1 to load cell D2. Its preset physical connection characteristic value is known to be 0.6, calculated based on the cable parameters. The responsibility score of device node D1 is 0.88. Assuming the propagation likelihood strength is calculated as a * preset physical connection characteristic value + (1-a) * responsibility score, in this example, a = 0.4, then the propagation likelihood strength of E1 is 0.4 * 0.6 + 0.6 * 0.88 = 0.768. Assume that edge A in the causal graph points to B, and its aligned topology graph nodes are D1 and D2. In the topology graph, there are two paths from D1 to D2: D1 directly connects to D2, with a corresponding propagation likelihood strength of 0.768, i.e., a spatial support strength of 0.768; D1 connects to D3 and then to D2, where the propagation likelihood of D1 connecting to D3 is... With a strength of 0.7, the propagation likelihood strength of the connection between D3 and D2 is 0.65. Assuming that the spatial support strength is calculated by multiplying the propagation likelihood strengths of the edges on the path in sequence, the spatial support strength of this path is 0.7 * 0.65 = 0.455. Taking the maximum value of the spatial support strengths corresponding to the two paths, the spatial support strength from D1 to D2 is 0.768. Assuming that the weight of the edge A pointing to B in the causal graph is 0.75, and the adjustment method of the edge weight is b * spatial support strength + (1-b) * edge weight, in this example, b = 0.6, then the adjusted edge weight is 0.6 * 0.768 + 0.4 * 0.75 = 0.7608. After adjustment, the edge weight is slightly enhanced, thus reflecting the strong support of the actual physical path.

[0091] The causal responsibility flow and spatial influence flow are executed alternately, iteratively performing multiple rounds of alternating optimization until the edge weight changes of the two graphs converge or reach a predetermined number of iterations. Using the physical nodes of the topological graph as the skeleton, the variable nodes of the causal graph are attached as attribute nodes to the corresponding physical nodes, forming a heterogeneous multi-layer graph. This graph contains two types of nodes: physical nodes and variable nodes, as well as two types of edges: physical connection edges and causal edges. All edges have optimized weights. The physical node with the highest responsibility score is identified from the heterogeneous multi-layer graph as the root cause candidate node. Starting from the root cause candidate node, traversal is performed along the high-weight physical edges and causal edges to generate one or more causal propagation chains. These causal propagation chains represent the complete path of the fault from its source to its manifestation. The recurring causal propagation chains are abstracted as fault mode nodes, such as high-frequency forward conduction interference, and specific instances are associated as node attributes of the node. Finally, a corresponding communication fault mode graph is constructed based on the heterogeneous multi-layer graph and the fault mode nodes.

[0092] Step S4: Based on the communication failure mode map, locate the communication failure and generate a communication failure location dataset.

[0093] Specifically, feature extraction is performed on the communication failure mode map to obtain the minimum sufficient causal subgraph and the dominant propagation path tree. A failure mode feature set is generated based on the minimum sufficient causal subgraph and the dominant propagation path tree. The failure mode feature set is then quantified using multilayer perceptron and DS evidence theory to generate a communication failure location dataset.

[0094] Understandably, the minimum sufficient causal subgraph is represented as the minimum set of nodes and edges consisting of the most likely root cause of the communication failure mode graph and the variables directly affected by it; the dominant propagation path tree is represented as the most likely physical propagation topology path of the failure; and the failure mode feature set includes at least the causal root cause score, the anomaly distribution fitting score, the inference confidence score, and the topological importance score.

[0095] In one possible implementation, firstly, all leaf variable nodes representing observed significant anomalies, such as excessive communication bit error rate, are identified from the communication failure mode graph. Then, starting from these leaf variable nodes, a reverse tracing is performed along the causal edges. A dynamic programming strategy is used to find the minimum ancestor node set that can cover more than 85% of the anomalous leaf nodes and the causal edges connected to it. This set must satisfy the characteristics that it can explain the target anomaly and that removing non-root nodes would destroy the anomaly explanation capability.

[0096] For example, suppose there are anomalous leaf nodes E1, E2, and E3 in the graph. Tracing back, we find that E1 and E2 are caused by node C1, and E3 is caused by node C2. Both C1 and C2 are caused by the root node R. We can select node R alone to cover the three anomalous leaf nodes E1, E2, and E3 through three paths: R-C1-E1 / E2 and R-C2-E3. Removing R cannot explain any anomalies. At this point, the extracted minimum ancestor node set only contains nodes R, C1, C2, E1, E2, E3, and the causal edges between the nodes, forming a minimal directed graph rooted at R that explains all key anomalies. This graph is the minimum sufficient causal subgraph.

[0097] Candidate root nodes are identified from the minimum sufficient causal subgraph, and corresponding physical nodes are selected from the physical layer subgraph of the communication failure mode graph and used as the root node of the tree. Then, the tree is recursively expanded outward along the physical connection edges. When selecting the next hop node during the expansion process, only the neighbor node pointed to by the edge with the highest propagation likelihood strength among all outgoing edges of the current node is selected and added to the tree until an edge with a propagation likelihood strength lower than a preset threshold is encountered or a preset maximum depth is reached, thereby forming a tree that shows the most important propagation path, i.e., the dominant propagation path tree.

[0098] For example: Suppose that node R corresponding to physical device "bus A" is determined as the root node. In the physical layer subgraph of the communication failure mode graph, R has three outgoing edges pointing to devices X, Y, and Z respectively, with propagation likelihood strengths of 0.9, 0.4, and 0.6 respectively. At this time, the edge with the highest strength is selected, that is, the edge pointing to X, and the node corresponding to device X is taken as the child node of the root node. Then, based on the outgoing edges of X, the edge with the highest strength and the corresponding node are selected again. Suppose the selected node is U, and the selected path is RXU. By iteratively executing the above process, a tree-like propagation path graph is finally formed.

[0099] It should be noted that when the preset maximum depth is reached, the iteration process will not end immediately. At this time, it will search for whether there is an edge whose propagation likelihood strength is lower than the preset threshold, and the starting node of the edge is located in the constructed dominant propagation path tree, while the neighbor node it points to is not included in the dominant propagation path tree, and the depth from the neighbor node to the root node has not reached the preset maximum depth. In this case, the iteration traversal will continue from the starting node of the edge.

[0100] Based on the extracted minimum sufficient causal subgraph and dominant propagation path tree, four feature indicators are calculated for the key nodes involved: causal root cause score, outlier distribution fitting score, inference confidence, and topological importance score, forming a fault mode feature set. Specifically, the causal root cause score is calculated by running the PageRank algorithm with the causal strength of the directed edges in the minimum sufficient causal subgraph as the weights of the transition probabilities, and using the normalized PageRank value corresponding to each node as its causal root cause score. For example, if node R is the only cause of C1 and C2, the corresponding edge weights... The initial PageRank values ​​are PR(R) = 1 / 3 ≈ 0.33, PR(C1) = 0.33, and PR(C2) = 0.33, respectively, with a damping factor d = 0.85. Each node's new PR value consists of two parts: the influence from other nodes and the equal share from random jumps. The calculation process for R's new PR value is as follows: since neither C1 nor C2 has an edge pointing to R, the influence from other nodes to R is 0. Because the damping factor d = 0.85, the sum of all node PR values ​​is 1 / 3. 5% is evenly distributed among the three nodes, i.e., (0.15 * total PR value) / 3 = (0.15 * 1) / 3 = 0.05. Therefore, PR_new(R) = 0 + 0.05 = 0.05. For node C1, since only R points to C1, R allocates 85% of its current PR value according to the outgoing edge weights. The total outgoing edge weights of R are 0.9 + 0.8 = 1.7. Therefore, the share that C1 receives from R is PR(R) * d * (weight of the edge pointing to C1 from R / total outgoing edge weights of R) = 0.33 * 0.85 * (0.9 / 1.7) ≈ 0.3 3 * 0.85 * 0.5294 ≈ 0.15, PR_new(C1) = equal share from random jump + share obtained by C1 from R = 0.05 + 0.15 = 0.2; similarly, PR_new(C2) = 0.183. By iterating continuously until the PR values ​​of the above three nodes converge, assuming the final PR(R) = 0.55, PR(C1) = 0.25, PR(C2) = 0.20, normalization is performed using the maximum value normalization method, and the normalized PR value corresponding to each node is used as the causal root cause score.

[0101] Anomaly distribution fitting score: The cosine similarity between the abnormal features of the node, such as harmonic spectrum and transient waveform, and the feature template of the corresponding typical fault mode in the preset knowledge base is calculated, and the cosine similarity value is used as the anomaly distribution fitting score. The higher the score, the better the match with the fault mode. For example, the cosine similarity between the current waveform feature vector of node R and the template vector of the "series arc" mode in the knowledge base is calculated, and the result is 0.88. This value is its anomaly distribution fitting score.

[0102] Inference confidence: This is represented by the responsibility score of each node. It should be noted that since the communication failure mode graph is formed by merging two graphs, the fault propagation directed causal graph and the fault propagation topology graph, the communication failure mode graph still retains all the attributes of the two graphs. Therefore, the responsibility score in the fault propagation directed causal graph can be directly used as the inference confidence.

[0103] Topological importance score: For each node in the dominant propagation path tree, its topological importance score is equal to the number of all its downstream leaf nodes. However, the contribution of each downstream node decays exponentially with its depth to the current node. That is, the closer a node is to the root and the more downstream nodes it can cover, the higher its topological importance score. For example, if the downstream of the root node R are leaf nodes X (depth 1), U (depth 2), and V (depth 2), then the topological importance score of R = 1*(0.8^1) + 1*(0.8^2 + 1*(0.8^2) = 0.8 + 0.64 + 0.64 = 2.08, where 0.8 is the base of the exponential decay.

[0104] Finally, the four indicators corresponding to each key node are encapsulated into a feature vector, namely [causal root cause score, outlier distribution fit score, inference confidence, topological importance score].

[0105] The feature vector corresponding to each key node is input into a pre-trained multilayer perceptron for analysis and processing. After nonlinear activation and weighted calculation in its hidden layers, an initial support is output. Subsequently, the initial support and key features such as the confidence interval of the anomaly distribution fitting score are constructed as independent evidence bodies. A basic probability assignment for the proposition "this node is the source of failure" is defined for each evidence body, with a portion of the probability assigned to "uncertainty that this node is the source of failure" to quantify the reliability of the evidence. Then, the Dempster synthesis rule is used to combine the basic probability assignments of multiple evidence bodies, and finally, specific confidence and total confidence are generated for each node. The specific confidence represents the deterministic probability of supporting that it is the source of failure, and the total confidence represents the sum of the specific confidence and the unassigned uncertainty probability. The difference between the two reflects the level of cognitive uncertainty of the current judgment.

[0106] For example, the feature vector of node N [0.92, 0.88, 0.85, 0.90] is input into a multilayer perceptron. The initial support output of the multilayer perceptron is 0.89. At the same time, the system constructs another piece of evidence based on the confidence interval of the fitted score in the feature vector of node N. After being synthesized by DS, the specific confidence of node N is 0.86 and the total confidence is 0.93. The difference of 0.07 reflects the slight uncertainty caused by the limited data.

[0107] Spatial clustering is used to divide geographically or electrically adjacent high-confidence nodes into candidate fault regions, and regional aggregate confidence is calculated. For example, the weighted average of the specific confidence of all nodes in each region is used as the regional aggregate confidence. The region with the highest regional aggregate confidence is selected, and the node with the highest specific confidence in it is determined as the most likely fault source.

[0108] It should be noted that the training data of the multilayer perceptron comes from a large number of historical cases. Each case corresponds to a confirmed fault source node and its four-dimensional feature vector [causal root cause score, anomaly distribution fitting score, inference confidence, topological importance score]. The training objective is to enable the multilayer perceptron to learn to use these features and map them into a high-accuracy support score. The loss function uses mean squared error, which aims to minimize the gap between the support score output by the multilayer perceptron and the label that the node is a true fault source.

[0109] Step S5: Obtain the actual feedback data of the communication fault location dataset, and update the system based on the actual feedback data.

[0110] Specifically, communication faults are investigated based on the communication fault location dataset, and actual feedback data is generated based on the results of the investigation. The actual feedback data includes at least fault location result labels and corresponding communication fault pattern maps and communication fault location datasets. The fault location result labels include accurate location, false alarm location, missed location, and location deviation. Based on the communication fault pattern maps and communication fault location datasets corresponding to different types of fault location result labels, corresponding incremental learning packages are generated. Based on the incremental learning packages, the corresponding modules of the collaborative identification system for power line communication signals and arc faults are incrementally updated.

[0111] In one possible implementation, for cases with accurate positioning, the key features and strategies that led to the successful diagnosis will be reinforced. For example, the threshold or parameters that trigger the alarm in the perception and early warning submodule will be fine-tuned to make it more sensitive to similar features in the future; the probability of the active perception submodule selecting the successful exploration strategy sequence under similar early warning features will be increased.

[0112] For cases of false alarms, key features and strategies that lead to false alarms in this diagnosis will be used to suppress them. For example, the threshold or parameters that trigger this alarm in the perception and early warning submodule will be adjusted to reduce the sensitivity to the interference patterns that cause this false alarm; and the probability that the active perception submodule will select this exploration strategy sequence under similar early warning features will be reduced.

[0113] For cases where the location is missed, it will be used to fill the detection blind spot. For example, the perception and early warning submodule will compensate for the missed fault type by enhancing the sensitivity of the corresponding feature extraction or adding new detection rules, and inject new strategy units designed for such faults into the exploration strategy library.

[0114] For cases of location deviation, it will be used to calibrate the inference process. For example, based on the difference between the actual fault location and the inferred location, the internal parameters of the fault location module will be adjusted to reduce such deviations.

[0115] Figure 2 This is a schematic diagram of a collaborative identification system for power line communication signals and arc faults according to the present invention.

[0116] Specifically, a collaborative identification system for power line communication signals and arc faults includes:

[0117] The structural assembly module is used to construct and train a dual-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule.

[0118] The fault perception module is used to perceive communication faults based on the dual-stream communication fault sensor and generate a communication fault perception dataset.

[0119] The pattern recognition module is used to perform fault pattern recognition on the communication fault perception dataset using a graph neural network to generate a communication fault pattern map.

[0120] The fault location module is used to locate communication faults based on the communication fault mode map and generate a communication fault location dataset.

[0121] The feedback optimization module is used to acquire the actual feedback data of the communication fault location dataset and update the system based on the actual feedback data.

[0122] It should be noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0123] The above embodiments can be implemented, in whole or in part, through software, hardware (such as circuits), firmware, or any other combination thereof.

[0124] When implemented using software, the above embodiments can be implemented in whole or in part as a computer program product, which includes one or more computer instructions or computer programs; when the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part.

[0125] It is understood that the computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device; the computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via transmission methods such as infrared, wireless, or microwave; the computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0126] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0127] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0128] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for the collaborative identification of power line communication signals and arc faults, characterized in that, Including the following steps: Construct and train a dual-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule; Based on the dual-stream communication fault sensor, communication faults are detected, and a communication fault detection dataset is generated. A graph neural network is used to perform fault pattern recognition on the communication fault perception dataset to generate a communication fault pattern map. Based on the communication failure mode map, communication failure localization is performed, and a communication failure localization dataset is generated. Obtain actual feedback data from the communication fault location dataset and update the system based on the actual feedback data.

2. The method for collaborative identification of power line communication signals and arc faults according to claim 1, characterized in that, Construct and train a two-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule, including: The perception and early warning submodule consists of two early warning channels: a communication channel and an energy channel. The communication channel is used to monitor the spectrum efficiency offset index, and the spectrum efficiency offset index is decoded and analyzed by a hidden Markov model to obtain the corresponding communication channel monitoring data. The energy channel is used to perform short-time Fourier transform on the sampled current of the power line to generate a time-frequency matrix, and simultaneously combine it with a pre-trained attention network for feature analysis to generate an attention weight matrix. Energy channel monitoring data is obtained by weighting the real-time time-spectrum matrix based on the attention weight matrix. When the monitoring data of the communication channel and the monitoring data of the energy channel reach the early warning conditions, a fault early warning is triggered, and a corresponding fault early warning vector is generated.

3. The method for collaborative identification of power line communication signals and arc faults according to claim 2, characterized in that, The method further includes: The active perception submodule consists of an exploration strategy generator and a response data collector; The exploration strategy generator takes the fault warning vector as input, uses a neural network to perform reasoning analysis on the fault warning vector, and generates a corresponding fault exploration strategy based on the analysis results. The response data collector is used to collect the response data generated when the fault exploration strategy is executed, and to generate a strategy response dataset; The strategy response dataset contains at least the response data corresponding to the communication channel and the energy channel, as well as the execution timestamps corresponding to the fault exploration strategy. The perception and early warning submodule uses normal historical data covering different loads and operating conditions to pre-train the communication channel, and combines gradient descent method to train the energy channel. The active perception submodule trains the exploration strategy generator through a simulation training process based on meta-learning, involving injecting faults, executing strategies, and collecting responses.

4. The method for collaborative identification of power line communication signals and arc faults according to claim 1, characterized in that, Based on the aforementioned dual-stream communication fault sensor, communication fault detection is performed, and a communication fault detection dataset is generated, including: Fault warning is based on the perception and early warning submodule. When the warning conditions are met, the fault warning is triggered and the corresponding fault warning vector is generated. The fault warning vector includes at least four dimensions: abnormal state identifier, deviation, duration, and abnormal confidence. The active perception submodule generates and executes the corresponding fault exploration strategy based on the fault warning vector, and obtains the strategy response dataset; A deep neural network is used to analyze the strategy response dataset to generate a corresponding communication fault perception dataset.

5. The method for collaborative identification of power line communication signals and arc faults according to claim 1, characterized in that, A graph neural network is used to perform fault mode recognition on the communication fault perception dataset to generate a communication fault mode map, including: A cross-convergence mapping and a causal discovery algorithm based on transfer entropy are used to perform causal mining on the communication fault perception dataset to generate a directed causal graph of fault transmission. The nodes of the fault propagation directed causal graph contain fault observation variables and executed fault exploration strategy actions. The directed edges represent the causal driving relationship between nodes, and the weights represent the causal strength. Construct a physical topology graph based on the real-time power grid topology, with monitoring nodes as vertices and electrical connections as connecting edges; A spatiotemporal graph convolutional network is used to perform convolutional inference on the physical topology graph to generate the corresponding spatiotemporal convolutional vector; Based on spatiotemporal convolution vectors and a preset fault propagation and diffusion mechanism, the influence diffusion intensity of each node in the physical topology graph is quantified to generate an influence diffusion intensity dataset. The physical topology map is optimized based on the dataset of influence diffusion intensity to generate a fault diffusion topology map.

6. The method for collaborative identification of power line communication signals and arc faults according to claim 5, characterized in that, The method further includes: A cross-graph attention mechanism is used to align the entity nodes of the fault propagation directed causal graph with the fault propagation topology graph, generating a corresponding set of cross-graph associated nodes; The directed causal graph of fault propagation and the topology graph of fault propagation are iteratively optimized based on a two-flow iterative reasoning mechanism that includes causal responsibility flow and spatial influence flow. The causal responsibility flow is represented by obtaining the causal contribution of each node in the directed causal graph of fault propagation to the observed fault, and optimizing the fault propagation topology graph based on the causal contribution. The spatial influence flow is represented by obtaining the likelihood strength of fault connections along the edges in the fault propagation topology graph, and optimizing the directed causal graph of fault propagation based on the likelihood strength; Based on the optimized directed cause-effect graph of fault propagation, the fault propagation topology graph, and the set of cross-graph related nodes, a communication fault mode graph is generated.

7. The method for collaborative identification of power line communication signals and arc faults according to claim 1, characterized in that, Communication fault location is performed based on a communication fault mode map, generating a communication fault location dataset, including: Feature extraction is performed on the communication failure mode graph to obtain the minimum sufficient causal subgraph and the dominant propagation path tree; The minimum sufficient causal subgraph is represented as the minimum set of nodes and edges consisting of the most likely root cause of the communication failure mode graph and the variables directly affected by it. The dominant propagation path tree represents the most likely physical propagation topology path of the fault. A set of failure mode features is generated based on the minimum sufficient causal subgraph and the dominant propagation path tree; The fault mode feature set includes at least the causal root cause score, the anomaly distribution fitting score, the inference confidence score, and the topological importance score. The fault mode feature set is quantified using multilayer perceptron and DS evidence theory to generate a communication fault location dataset.

8. The method for collaborative identification of power line communication signals and arc faults according to claim 1, characterized in that, Obtain actual feedback data from the communication fault location dataset and update the system based on the actual feedback data, including: Communication faults are investigated based on the communication fault location dataset, and actual feedback data is generated based on the results of the communication fault investigation. The actual feedback data includes at least fault location result labels and corresponding communication fault mode maps and communication fault location datasets. The fault location result labels include accurate location, false alarm location, missed location location, and location deviation. Based on the communication fault pattern map and communication fault location dataset corresponding to the labels of different types of fault location results, generate corresponding incremental learning packages. The corresponding modules of the collaborative identification system for power line communication signals and arc faults are incrementally updated based on the incremental learning package.

9. A collaborative identification system for power line communication signals and arc faults, used to implement the method described in any one of claims 1 to 8, characterized in that, include: A structural assembly module is used to construct and train a dual-stream communication fault sensor that includes a perception and early warning submodule and an active perception submodule. A fault perception module is used to perceive communication faults based on the dual-stream communication fault sensor and generate a communication fault perception dataset. The pattern recognition module is used to perform fault pattern recognition on the communication fault perception dataset using a graph neural network to generate a communication fault pattern map. A fault location module is used to locate communication faults based on a communication fault mode map and generate a communication fault location dataset. The feedback optimization module is used to acquire the actual feedback data of the communication fault location dataset and update the system based on the actual feedback data.

Images