A measurement switch control method based on intelligent fault prediction perception

Abstract

This invention discloses a measurement switch control method based on intelligent fault prediction and perception, relating to the field of measurement switches. It inputs features in parallel into instantaneous, short-term, and long-term fault prediction models and type classification models to achieve multi-timescale and multi-type fault prediction. The prediction results are fused using an improved D-S evidence theory, and a local perception summary is generated by combining model accuracy, feature quality, and time decay factor. The summary is exchanged through a local network, and implicit correlation anomalies and cascading faults are identified using graph attention networks and variational autoencoders to generate a global situational awareness. Control commands are dynamically generated and issued based on the local and global situations to achieve precise prevention and control and collaborative isolation. This invention addresses the problem of existing technologies having a single prediction scale and ignoring system-level topological correlations, resulting in an inability to comprehensively warn of multi-timescale faults and cascading risks. This invention significantly improves the reliability of measurement switches, the accuracy of distribution network fault prediction, and the intelligence level of control decision-making.

Description

Technical Field

[0001] This invention relates to the field of measurement switches, and more specifically to a measurement switch control method based on intelligent fault prediction and perception. Background Technology

[0002] With the intelligent upgrading of distribution networks and the integration of a high proportion of distributed energy resources, the operating status of measurement switches directly affects power supply reliability and system resilience. Achieving proactive prediction and coordinated control of measurement switch faults is a key technological support for improving the self-healing capability of distribution networks, avoiding cascading faults, and supporting the safe operation of new power systems. By integrating multi-source sensing data and artificial intelligence algorithms, a smart decision-making system can be constructed, moving from single-point early warning to group collaboration. This transforms the traditional passive maintenance mode into proactive preventative maintenance, significantly reducing power outage losses, optimizing the allocation of operation and maintenance resources, and providing core guarantees for building a safe, reliable, and efficient modern smart distribution network.

[0003] Currently, in the field of measurement switch status monitoring and fault prediction, existing technologies focus on extrapolating short-term trends of comprehensive health indicators, resulting in a single prediction model. They cannot distinguish or cover fault risks at different time scales; for example, they cannot simultaneously address instantaneous overcurrents occurring within seconds, overheating trends developing over several hours, or insulation aging processes lasting several months. This limits their application in broader business scenarios such as preventative maintenance and lifespan prediction. More importantly, existing solutions make closed-loop decisions based on individual measurement switches as independent units, completely ignoring the close electrical connections and physical relationships between devices in the power distribution network. Therefore, they cannot identify implicit correlation anomalies caused by adjacent device failures or abnormal network flow, nor can they provide early warnings of cascading faults that could lead to large-scale power outages, creating a system-level safety blind spot.

[0004] Therefore, a measurement switch control method based on intelligent fault prediction and perception is needed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a measurement switch control method based on intelligent fault prediction and perception. This method realizes a closed-loop system that integrates multi-source data fusion, multi-timescale parallel prediction, dynamic evidence fusion, global correlation analysis, and hierarchical collaborative control, significantly improving the reliability of measurement switches, the accuracy of power distribution network fault prediction, and the level of intelligence in control decision-making.

[0006] The present invention adopts the following technical solution:

[0007] A measurement switch control method based on intelligent fault prediction and perception includes the following steps:

[0008] S1. Real-time acquisition of multi-source raw data from measurement switches through local sensors and distributed sensor network nodes. Each measurement switch is equipped with an edge computing terminal. The edge computing terminal preprocesses the acquired multi-source raw data locally to obtain a standardized feature dataset.

[0009] S2. Input the standardized feature dataset obtained in step S1 into the instantaneous fault prediction model, short-term fault prediction model, long-term fault prediction model and fault type classification model simultaneously to achieve parallel prediction of faults at different time scales and of different types.

[0010] S3. Adaptive fusion and confidence evaluation of the parallel prediction results output in step S2 are performed using an improved DS evidence theory, and a local perceived summary is generated. The improved DS evidence theory uses the model's historical accuracy, real-time feature quality, and time decay factor as evidence weights to dynamically adjust the contribution of the parallel prediction results in the fusion process.

[0011] S4. Topologically adjacent measurement switches exchange local intelligent sensing summaries through a local communication network, and identify implicit correlation anomalies and cascaded faults based on graph attention networks and variational autoencoders to generate global situation assessment results.

[0012] S5. Based on the local situation awareness summary and the global situation assessment results, dynamically generate control commands for the measurement switches and issue them for execution.

[0013] Furthermore, the multi-source raw data includes electrical quantities, state quantities, physical quantities, and environmental quantities. The electrical quantities include at least current, voltage, zero-sequence component, and harmonic content. The state quantities include at least switch position, energy storage state, and number of operations and stroke curve. The physical quantities include at least mechanical vibration amplitude, mechanical vibration spectrum, acoustic signature signal, and partial discharge pulse. The environmental quantities include at least ambient temperature, humidity, and dust concentration.

[0014] Furthermore, in the edge computing terminal, noisy data is removed by the 3σ criterion, missing data is filled by linear interpolation, time alignment of multi-source raw data is performed based on a unified timestamp, multi-source raw data with different dimensions is mapped to a unified interval by Z-score standardization, and time-domain and frequency-domain features in multi-source raw data are extracted by wavelet transform.

[0015] Further, in step S2, the instantaneous fault prediction model performs instantaneous fault risk prediction at the second to minute level based on a hybrid structure of a one-dimensional convolutional neural network and a gated recurrent unit. The short-term fault prediction model performs short-term fault trend prediction at the minute to hour level based on a gated recurrent unit network with attention mechanism enhancement. The long-term fault prediction model performs long-term degradation trend and fault probability prediction at the day to month level based on a Transformer multi-head self-attention encoder. The fault type classification model performs high-dimensional pattern recognition on real-time features based on a deep residual network and outputs specific fault type classification results.

[0016] Furthermore, in step S2, the instantaneous fault prediction model, short-term fault prediction model, long-term fault prediction model and fault type classification model are pre-trained using the transfer learning method, and the model parameters are adjusted based on the real-time collected data.

[0017] Furthermore, in step S3, the working method of the improved DS evidence theory is as follows:

[0018] The parallel prediction results output in step S2 are defined as four independent sources of evidence. An initial confidence level is generated for each source of evidence, which is the original support level of the source of evidence.

[0019] The historical accuracy weight, real-time feature quality weight, and time decay factor weight are calculated for each evidence source. The historical accuracy weight is determined based on the prediction accuracy of the model's historical test set, the real-time feature quality weight is calculated based on the completeness and noise intensity of the preprocessed feature data, and the time decay factor weight is assigned a differentiated decay coefficient for the prediction results of the model at different time scales.

[0020] The weighted product method is used to fuse the calculated historical accuracy weights, real-time feature quality weights, and time decay factor weights of the model to obtain the comprehensive evidence weight for each evidence source.

[0021] The confidence level of each evidence source is corrected based on the comprehensive evidence weight. All evidence sources after weight correction are then merged based on the Dempster synthesis rule, and the final confidence level is calculated.

[0022] Furthermore, the local situational awareness summary includes the fault occurrence time window, fault type, risk level, and confidence level.

[0023] Further, in step S4, the specific process of identifying latent correlation anomalies and cascading faults based on graph attention network and variational autoencoder is as follows: each measurement switch is taken as a graph node, and the topological connection relationship is taken as an edge. The correlation weight between nodes is learned through graph attention network to capture the latent correlation features of different measurement switch faults; the global perception data is reconstructed using variational autoencoder, and abnormal data deviating from the normal operation mode is identified through reconstruction error; the latent correlation anomalies and cascading faults are determined by combining the correlation features and the abnormal data identification results.

[0024] Furthermore, in step S5, a hierarchical collaborative control strategy based on a preset hierarchical risk and strategy mapping matrix is ​​used to dynamically generate control commands for the measurement switch:

[0025] When the local situational awareness summary indicates that the risk level is less than the low-risk threshold and there are no globally related anomalies, an early warning command is generated and the edge computing terminal is activated for real-time monitoring.

[0026] When the local situational awareness summary indicates that the risk level is greater than or equal to the low risk threshold, less than the high risk threshold, or when there is a local correlation anomaly, a load adjustment or circuit breaker timing optimization instruction is generated.

[0027] If the local situational awareness summary indicates a risk level greater than or equal to the high-risk threshold, or if there is a risk of cascading faults, an emergency trip command is generated, and the associated measurement switches in the global situational assessment results are linked to perform coordinated isolation operations to form a fault isolation zone.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. This invention, by configuring an edge computing terminal at the measurement switch, performs noise removal, missing data completion, time alignment, standardization, and time-frequency feature extraction locally. This solves the problem of differences in dimensions, sampling rate, and spatiotemporal resolution of multi-source heterogeneous data, significantly reduces the interference of invalid and erroneous data on model inference, and enables subsequent prediction and analysis to be based on high-quality, low-latency data, thereby improving the reliability of overall intelligent analysis.

[0030] 2. This invention simultaneously inputs standardized features into instantaneous, short-term, and long-term fault prediction models and fault type classification models, which can simultaneously capture fault risks at different time scales such as second-level overcurrent, hourly-level overheating trends, and monthly-level insulation aging, and output specific fault types. This breaks through the limitation of traditional single-scale models that can only cope with limited scenarios, enabling preventive maintenance, life assessment, and emergency control to be realized under a unified framework, greatly expanding the business application scope of intelligent operation and maintenance of measurement switches.

[0031] 3. This invention adopts an improved DS evidence theory, which uses the model's historical accuracy, real-time feature quality and time decay factor as weights to dynamically weight and fuse the prediction results of different models, and generates a local perception summary based on the fusion confidence. This avoids the risk of misjudgment caused by equal weight or simple superposition fusion, making the final risk assessment more in line with the actual equipment status and operating environment, and enhancing the stability and decision accuracy of the system under complex and variable working conditions.

[0032] 4. This invention exchanges local sensing summaries through a local communication network and learns the correlation weights between nodes based on a graph attention network. Combined with the reconstruction error analysis of a variational autoencoder, it can discover hidden correlation anomalies between devices caused by electrical connections, physical environment, or power flow. It can identify cascading fault paths that may cause large-scale power outages in advance, making up for the major defect that single-point decision-making cannot discover system-level safety blind spots and significantly improving the overall resilience and anti-spread capability of the distribution network.

[0033] 5. Based on the local and global situation assessment results, this invention dynamically generates differentiated control commands such as early warning, load adjustment, circuit breaker timing optimization, or emergency circuit breaker tripping and collaborative isolation according to risk level and correlation degree. It can quickly suppress the spread of faults and avoid unnecessary network-wide power outages or load losses, achieving the best balance between safety and economy, and significantly improving the self-healing capability and power supply reliability of the distribution network. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0035] Figure 2 This is an architecture diagram of the implementation architecture in this invention. Detailed Implementation

[0036] The following will refer to the appendices in the embodiments of the present invention. Figure 1 and attached Figure 2 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] This invention discloses a measurement switch control method based on intelligent fault prediction and perception, as shown in the attached figure. Figure 1 As shown, it includes the following steps:

[0038] A measurement switch control method based on intelligent fault prediction and perception specifically includes the following steps:

[0039] S1. Multi-source data acquisition and edge preprocessing;

[0040] Edge computing terminals are configured at each measurement switch to collect real-time multi-source raw data in four categories: electrical quantities, state quantities, physical quantities, and environmental quantities via local sensors and distributed sensor network nodes. Electrical quantities include current, voltage, zero-sequence component, and harmonic content; state quantities include switch position, energy storage status, and operation count and stroke curve; physical quantities include mechanical vibration amplitude, mechanical vibration spectrum, acoustic signature signal, and partial discharge pulse; and environmental quantities include ambient temperature, humidity, and dust concentration. Within the edge computing terminals, noisy and abnormal data are first removed using the three-sigma criterion, and then missing data is completed using linear interpolation. Next, the multi-source data is time-aligned based on a unified timestamp, and Z-score standardization maps data with different dimensions to a unified numerical range. Finally, wavelet transform is applied to extract time-domain and frequency-domain features from the data, forming a standardized feature dataset.

[0041] Local sensor groups and distributed sensor network nodes are deployed around each measurement switch to achieve real-time acquisition of multi-source raw data. The specific acquisition scheme is as follows:

[0042] Electrical quantity acquisition: High-precision Hall current sensors are used to acquire three-phase current and zero-sequence current, and voltage sensors are used to acquire three-phase voltage and zero-sequence voltage. Current and voltage signals are acquired synchronously through a high-speed data acquisition card, and then the zero-sequence component and harmonic content are calculated through Fourier transform.

[0043] Status data acquisition: The switch opening and closing positions are acquired through limit switch sensors, the energy storage status is acquired through pressure sensors, the number of switch operations is recorded through encoders, and the stroke curve during the operation process is acquired synchronously.

[0044] Physical quantity acquisition: Accelerometers are used to acquire mechanical vibration amplitude and mechanical vibration spectrum; acoustic sensors are used to acquire acoustic signals during switch operation and operation; and partial discharge sensors are used to acquire partial discharge pulses.

[0045] Environmental data collection: Temperature and humidity are collected using temperature and humidity sensors, and dust concentration is collected using dust sensors.

[0046] All sensor data is transmitted in real time to the edge computing terminal configured with the corresponding measurement switch via RS485 bus.

[0047] The edge computing terminal performs preprocessing of multi-source raw data locally. The specific steps are as follows:

[0048] Noise Removal: The 3σ criterion is used to filter noise in various types of collected data. The mean of the data is set as μ and the standard deviation is σ. Abnormal noise data that exceeds the range of [μ-3σ, μ+3σ] is removed. For example, for current acquisition data, the mean of its historical 1000 sampling points is calculated as μ=500A and the standard deviation is σ=10A. Noise data below 470A and above 530A are removed.

[0049] Missing data completion: For missing data caused by sensor transmission interruption or noise removal, linear interpolation is used to complete the missing data. If the current data is missing at a certain moment, the average of the five valid sampling points before and after that moment is taken as the interpolation benchmark, and the missing data value is obtained through linear fitting.

[0050] Time alignment: Based on the system clock of the edge computing terminal (clock accuracy: 1μs), a unified timestamp is added to all collected data to align the collected data from different sensors to the same time dimension, and the time alignment error is controlled within 5ms.

[0051] Standardization: The Z-score standardization method is used to map the multi-source raw data with different dimensions to a unified interval of [0,1]. The standardization formula is: x'=(x-μ) / σ, where x is the original data, μ is the historical mean of the data, and σ is the historical standard deviation of the data. For example, voltage data (0-12kV) is standardized to dimensionless data in the interval of 0-1.

[0052] Feature extraction: Wavelet transform is used to extract time and frequency domain features from the standardized multi-source data. Time domain features include peak value, mean, variance, kurtosis, and skewness, while frequency domain features include characteristic frequency amplitude, frequency centroid, and frequency band energy. Finally, a standardized feature dataset with a dimension of 128 is obtained.

[0053] S2, Multi-model parallel fault prediction;

[0054] The standardized feature dataset obtained in step S1 is simultaneously input into four pre-trained dedicated models for parallel analysis. The instantaneous fault prediction model employs a hybrid structure of a one-dimensional convolutional neural network and a gated recurrent unit (GRU) to predict instantaneous fault risks from seconds to minutes. The short-term fault prediction model uses a GRU network enhanced with an attention mechanism to predict fault trends from minutes to hours. The long-term fault prediction model uses a Transformer multi-head self-attention encoder to predict equipment degradation trends and long-term fault probabilities from days to months. The fault type classification model is based on a deep residual network to perform high-dimensional pattern recognition on real-time features and output specific fault type classification results. All these models are pre-trained using transfer learning methods and can dynamically adjust model parameters based on real-time acquired data.

[0055] Four parallel models were constructed: an instantaneous fault prediction model, a short-term fault prediction model, a long-term fault prediction model, and a fault type classification model. All models were deployed on edge computing terminals. The specific construction scheme and parameters are as follows:

[0056] Transient fault prediction model: It is constructed based on a hybrid structure of one-dimensional convolutional neural network and gated recurrent unit. The 1D-CNN layer is set with 3 convolutional blocks, each containing 16 convolutional kernels (size 3×1), batch normalization layer and ReLU activation function. The GRU layer is set with 2 layers, each with 64 hidden units and dropout rate of 0.2. The input is a 128-dimensional normalized feature dataset, and the output is the transient fault risk probability in the range of seconds to minutes (prediction step size: 1s-5min).

[0057] Short-term fault prediction model: Based on the attention mechanism-enhanced gated recurrent unit network, the GRU layer is set with 3 layers, each with 128 hidden units. A multi-head attention mechanism (number of attention heads: 4) is added between the GRU layer and the output layer. The input is a 128-dimensional normalized feature dataset, and the output is the short-term fault trend prediction results from minute to hour (prediction step size: 5min-1h).

[0058] Long-term failure prediction model: It is built based on the Transformer multi-head self-attention encoder. The encoder contains 6 encoding layers, each encoding layer has 8 multi-head attention heads, and the FeedForward network has a hidden layer dimension of 512. The input is a standardized feature dataset of 24 consecutive hours (concatenated by time series), and the output is the long-term degradation trend and failure probability prediction results from day to month (prediction step size: 1 day to 30 days).

[0059] Fault type classification model: Based on deep residual network (ResNet-50), the fully connected layers of the original ResNet-50 are removed, and three custom fully connected layers are added (with 256, 128, and 10 hidden units respectively). The softmax activation function is used to output the classification results. The recognizable fault types include 10 specific faults such as short circuit, grounding, open circuit, mechanical jamming, insufficient energy storage, and insulation aging. The input is a 128-dimensional standardized feature dataset, and the output is the classification probability of each type of fault.

[0060] Transfer learning was used to pre-train four parallel models. The specific implementation process is as follows:

[0061] Dataset preparation: Collect historical operating data of 10kV distribution network measurement switches (including normal operation data and various fault data), with a total of 1 million data entries, which are divided into training set, validation set and test set in a ratio of 7:2:1;

[0062] Pre-training process: The training set is input into each model, and the Adam optimizer (learning rate: 0.001, β1=0.9, β2=0.999) is used for pre-training. The loss function of the instantaneous fault prediction model, short-term fault prediction model and long-term fault prediction model is the mean squared error (MSE), and the loss function of the fault type classification model is the cross-entropy loss function. The number of pre-training iterations is 200 rounds. Pre-training is stopped when the validation set loss does not decrease for 10 consecutive rounds.

[0063] Real-time parameter adjustment: After the model is deployed, the edge computing terminal collects the operating data of the measurement switch in real time, extracts 1,000 valid data points per hour as incremental training data, and uses the online gradient descent method to fine-tune the parameters of each model to ensure that the model adapts to the real-time operating status changes of the measurement switch. The fine-tuning learning rate is set to 0.0001 to avoid overfitting.

[0064] The standardized feature dataset obtained in step S1 is simultaneously input into four parallel models. Each model runs synchronously to achieve parallel prediction of different time scales and different types of faults. The prediction results are output to the fusion module in step S3 in real time.

[0065] S3. Improved evidence theory fusion and generation of localized perceived summaries;

[0066] An improved Dempster evidence theory is used to adaptively fuse and evaluate the confidence levels of the parallel prediction results from the four models in step S2. Specifically, the outputs of the four models are treated as four independent sources of evidence, each assigned an initial confidence level. Then, three dynamic weights are calculated for each source: a historical accuracy weight determined based on the accuracy of the historical test set; a real-time feature quality weight calculated based on the completeness of the preprocessed feature data and noise intensity; and a differentiated time decay factor weight assigned to the prediction results at different time scales. A weighted product method is used to fuse these three weights into a comprehensive evidence weight for each source, and the initial confidence level is adjusted accordingly. Finally, the Dempster synthesis rule is applied to fuse all the corrected evidence, calculating the final fused prediction result and its confidence level. Based on this, a local situational awareness summary is generated, including the predicted fault occurrence time window, fault type, risk level, and corresponding confidence level.

[0067] The improved DS evidence theory module is deployed on an edge computing terminal, and its specific working process is as follows:

[0068] Evidence source definition and initial confidence level generation: The output results of the four parallel models in step S2 are defined as four independent evidence sources, namely evidence source E1 instantaneous fault prediction result, E2 short-term fault prediction result, E3 long-term fault prediction result, and E4 fault type classification result; the original probability output by each model is used as the initial evidence confidence level of the corresponding evidence source. For example, if the instantaneous fault risk probability output by E1 is 0.85, then the initial confidence level of E1 is 0.85.

[0069] Weighting factor calculation:

[0070] Model historical accuracy weight calculation: Calculate the historical prediction accuracy of each model on the test set. The historical accuracy of the instantaneous fault prediction model is 92%, so its weight ω1=0.92; the historical accuracy of the short-term fault prediction model is 89%, so its weight ω2=0.89; the historical accuracy of the long-term fault prediction model is 85%, so its weight ω3=0.85; and the historical accuracy of the fault type classification model is 94%, so its weight ω4=0.94.

[0071] Real-time feature quality weight calculation: Based on the integrity and noise intensity of the feature data after preprocessing in step S1, the data integrity is defined as: data integrity = effective data volume / total collected data volume, noise intensity = noise data volume / total collected data volume, and the real-time feature quality weight ω(q) = 0.6 × integrity + 0.4 × (1 - noise intensity). If the integrity of a batch of feature data is 98% and the noise intensity is 2%, then ω(q) = 0.6 × 0.98 + 0.4 × 0.98 = 0.98.

[0072] Time decay factor weight calculation: Different decay coefficients are assigned to the prediction results of models with different time scales. For instantaneous fault prediction (second to minute level), the time decay factor α1=0.95, for short-term fault prediction (minute to hour level), α2=0.90, for long-term fault prediction (day to month level), α3=0.80, and for fault type classification models, there is no time decay, α4=1.0.

[0073] Determining the weight of comprehensive evidence: The weighted product method is used to fuse the above three types of weight factors. The weight of comprehensive evidence is ω = ω(a) × ω(q) × α, where ω(a) is the historical accuracy weight of the model; for example, the comprehensive weight of E1 is ω(E1) = 0.92 × 0.98 × 0.95 ≈ 0.857.

[0074] Evidence confidence correction and fusion: The initial confidence of each evidence source is corrected based on the comprehensive evidence weight. The correction formula is: m'(A)=ω×m(A) + (1-ω)×m0(A), where m(A) is the initial confidence and m0(A) is the global unknown confidence (set to 0.05). After correction, the fusion of the four evidence sources is completed using the improved Dempster synthesis rule (introducing the conflict coefficient correction factor).

[0075] Confidence assessment: Set the fusion conflict threshold to 0.7. If the conflict coefficient calculated during the fusion process is greater than 0.7, it is judged as an abnormal fusion. The evidence source with the largest conflict is removed and the fusion is repeated. If the conflict coefficient is less than or equal to 0.7, the final fusion result and the corresponding final confidence are output.

[0076] Based on the fusion results of the improved DS evidence theory, a local situational awareness summary is generated, which includes four core elements:

[0077] Fault occurrence time window: determined based on the prediction results of each time scale after fusion. For example, if the fusion determines that the fault may occur in 10 minutes, then the time window is "current time + 10 minutes".

[0078] Fault type: Based on the fusion result of the fault type classification model, the specific fault type is clearly defined, such as "short circuit fault";

[0079] Risk level: Risk levels are determined based on the final confidence level. A confidence level ≥ 0.9 indicates critical risk, 0.7 ≤ confidence level < 0.9 indicates severe risk, 0.5 ≤ confidence level < 0.7 indicates moderate risk, and a confidence level < 0.5 indicates minor risk.

[0080] Confidence level: Output the final confidence level after fusing the improved DS evidence theory, such as 0.88.

[0081] S4. Topology information exchange and global situation assessment;

[0082] Topologically adjacent measurement switches exchange the local intelligent sensing summaries generated in step S3 via a local communication network. These summaries are constructed into a graph structure, where each measurement switch is a node and topological connections are edges. First, a graph attention network is used to learn the association weights between nodes, thereby capturing the implicit association features between different switch fault states. Simultaneously, a variational autoencoder is used to reconstruct the global sensing data, and abnormal data deviating from the normal operating mode is identified by calculating the reconstruction error. Combining the association features mined by the graph attention network and the anomaly results identified by the variational autoencoder, potential implicit association anomalies and cascading fault risks are comprehensively analyzed and identified, ultimately generating a global situation assessment result.

[0083] Topologically adjacent measurement switches exchange local intelligent sensing summaries through a local communication network. The specific communication scheme is as follows:

[0084] The local communication network adopts LoRa industrial wireless LAN (communication frequency band: 433MHz, communication distance: 1-3km, transmission rate: 1200bps). Data exchange adopts MQTT communication protocol. The edge computing terminal of each measurement switch acts as an MQTT client, periodically (cycle: 5s) publishes local situational awareness summary and subscribes to the summary information of adjacent measurement switches in the topology to ensure real-time synchronization of summary data.

[0085] A global analysis model integrating a graph attention network and a variational autoencoder is constructed and deployed on the edge computing terminals of each measurement switch. Global analysis is achieved using a distributed collaborative computing approach. The specific implementation process is as follows:

[0086] Graph structure construction: Each measurement switch is used as a graph node, and the node features are quantitative indicators of the local situational awareness summary (fault occurrence time window, risk level, confidence level, etc.). The topological connection relationships between measurement switches (such as cable connection, association with the same feeder) are used as the edges of the graph to construct a global topology graph.

[0087] Implicit correlation feature capture: The graph attention network (GAT) is used to learn the correlation weights between nodes. The GAT has two attention layers with four attention heads in each layer and a hidden layer dimension of 64. The neighborhood features of each node are obtained by weighted summation of the attention weights, which captures the implicit correlation features of different measurement switch faults. For example, it can identify whether there is a correlation between minor faults of multiple measurement switches on the same feeder.

[0088] Global anomaly data identification: A variational autoencoder (VAE) is used to reconstruct the global perception data (all exchanged local situational awareness summaries). The VAE encoder contains two fully connected layers (hidden units: 128 and 64), and the decoder contains two fully connected layers (hidden units: 64 and 128). The reconstruction error threshold is determined by training with historical normal operation data (set to 0.1). If the reconstruction error of the perception data of a certain node is greater than 0.1, it is judged as anomaly data that deviates from the normal operation mode.

[0089] Global situation assessment results generation: Combining the correlation features captured by GAT and the abnormal data results identified by VAE, implicit correlation anomalies and cascading faults are identified, and global situation assessment results are generated. The results include the global fault propagation path (such as "measurement switch A → measurement switch B → measurement switch C"), the set of associated fault nodes, the probability of cascading fault occurrence, and the global risk coverage.

[0090] S5. Dynamic collaborative control command generation and execution;

[0091] Based on the local situational awareness summary generated in step S3 and the global situational assessment results generated in step S4, control commands for the measurement switches are dynamically generated. A preset hierarchical risk-policy mapping matrix is ​​used to guide control decisions. When the risk level indicated by the local summary is lower than the set low-risk threshold and no related anomalies are found in the global assessment, the system generates an early warning command, requiring only the edge computing terminal to strengthen real-time monitoring. When the risk level indicated by the local summary reaches or exceeds the low-risk threshold but is lower than the high-risk threshold, or when the global assessment finds local related anomalies, the system generates preventive control commands such as load adjustment or circuit breaker timing optimization. If the risk level indicated by the local summary reaches or exceeds the high-risk threshold, or the global assessment confirms the existence of cascading fault risk, the system immediately generates an emergency trip command and coordinates the related measurement switches identified in the global assessment to perform collaborative isolation operations, thereby quickly forming a fault isolation zone. All generated control commands are ultimately sent to the corresponding measurement switches for execution.

[0092] Based on a pre-defined mapping matrix of hierarchical risks and strategies, control commands for the measurement switches are dynamically generated. The specific hierarchical control scheme is as follows:

[0093] Low-risk, unrelated anomaly scenario: When the local situational awareness summary indicates a risk level of slight risk (confidence level <0.5) and the global situational assessment results show no globally related anomalies, the edge computing terminal generates an early warning command, activates the real-time monitoring enhancement mode of the edge computing terminal, increases the data acquisition frequency from 10kHz to 20kHz, and uploads the early warning information to the cloud monitoring platform.

[0094] In general risk or locally correlated anomaly scenarios: When the local situation awareness summary indicates a risk level of general risk (0.5 ≤ confidence level < 0.7) or severe risk (0.7 ≤ confidence level < 0.9), or when the global situation assessment results show locally correlated anomalies, load adjustment or circuit breaker timing optimization instructions are generated; for example, for general mechanical jamming faults, circuit breaker timing optimization instructions are generated to adjust the circuit breaker speed from 1.5m / s to 1.2m / s to reduce mechanical wear; for locally correlated anomalies, load adjustment instructions are generated to transfer the load in the fault area to the backup line;

[0095] High-risk or cascading fault scenarios: When the local situation awareness summary indicates a critical risk level (confidence ≥ 0.9), or when the global situation assessment results show a clear cascading fault risk, an emergency trip command is immediately generated. The edge computing terminal sends the command to the operating mechanism of the measurement switch via the IO interface (response time ≤ 10ms). At the same time, it coordinates with the associated measurement switches in the global situation assessment results to perform collaborative isolation operations. For example, it coordinates with all measurement switches on the fault propagation path to trip in sequence, forming a fault isolation zone to prevent the fault from escalating.

[0096] After the control command is issued and executed, the edge computing terminal collects the execution feedback data of the measuring switch in real time (such as opening and closing position, current and voltage changes, fault characteristic parameters, etc.) through sensors to evaluate the control effect:

[0097] If the feedback data shows a significant decrease in fault characteristic parameters (e.g., the peak current of a short-circuit fault returns to the normal range), the control is deemed effective, and the control strategy and its effect are recorded. If the feedback data shows that the fault has not been alleviated or the control is ineffective (e.g., the fault characteristic parameters continue to rise), the process of steps S2 to S5 is retried to adjust the model parameters and control strategy, generate optimized control commands, and continue until the fault is effectively controlled.

[0098] As attached Figure 2 As shown, the implementation architecture can be divided into four layers:

[0099] Sensing and Edge Processing Layer (S1): Intelligent terminals deployed at each measurement switch.

[0100] Local intelligent analysis layer (S2, S3): A lightweight AI model library and fusion decision module integrated on edge computing terminals or near-end devices.

[0101] Network communication layer: Local communication network connecting adjacent nodes (such as HPLC, RF Mesh, industrial Ethernet, etc.).

[0102] Regional Collaboration and Decision Layer (S4, S5): Analysis servers that can be deployed at regional control stations or in the cloud, responsible for cross-node global analysis and generating advanced control policies.

[0103] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, any equivalent modifications or substitutions made by those skilled in the art to the relevant technical features will fall within the scope of protection of the present invention.

Claims

1. A method for smart fault prediction-aware measurement switch control, the method comprising: receiving a plurality of fault prediction signals from a plurality of sensors; determining a fault prediction based on the plurality of fault prediction signals; and controlling a measurement switch based on the fault prediction. Includes the following steps: S1. Real-time acquisition of multi-source raw data from measurement switches through local sensors and distributed sensor network nodes. Each measurement switch is equipped with an edge computing terminal. The edge computing terminal preprocesses the acquired multi-source raw data locally to obtain a standardized feature dataset. S2. Input the standardized feature dataset obtained in step S1 into the instantaneous fault prediction model, short-term fault prediction model, long-term fault prediction model and fault type classification model simultaneously to achieve parallel prediction of faults at different time scales and of different types. S3. Adaptive fusion and confidence evaluation of the parallel prediction results output in step S2 are performed using an improved DS evidence theory, and a local perceived summary is generated. The improved DS evidence theory uses the model's historical accuracy, real-time feature quality, and time decay factor as evidence weights to dynamically adjust the contribution of the parallel prediction results in the fusion process. S4. Topologically adjacent measurement switches exchange local intelligent sensing summaries through a local communication network, and identify implicit correlation anomalies and cascaded faults based on graph attention networks and variational autoencoders to generate global situation assessment results. S5. Based on the local situation awareness summary and the global situation assessment results, dynamically generate control commands for the measurement switches and issue them for execution.

2. The method of claim 1, wherein the method is based on intelligent fault prediction awareness. The multi-source raw data includes electrical quantities, state quantities, physical quantities, and environmental quantities. The electrical quantities include at least current, voltage, zero-sequence component, and harmonic content. The state quantities include at least switch position, energy storage state, and number of operations and stroke curve. The physical quantities include at least mechanical vibration amplitude, mechanical vibration spectrum, acoustic signature signal, and partial discharge pulse. The environmental quantities include at least ambient temperature, humidity, and dust concentration.

3. The method of claim 1, wherein the method further comprises: In the edge computing terminal, noisy data is removed by the 3σ criterion, missing data is filled by linear interpolation, time alignment of multi-source raw data is performed based on a unified timestamp, multi-source raw data with different dimensions is mapped to a unified interval by Z-score standardization, and time-domain and frequency-domain features in multi-source raw data are extracted by wavelet transform.

4. The method of claim 1, wherein the method is based on intelligent fault prediction awareness. In step S2, the instantaneous fault prediction model performs instantaneous fault risk prediction at the second to minute level based on a hybrid structure of a one-dimensional convolutional neural network and a gated recurrent unit. The short-term fault prediction model performs short-term fault trend prediction at the minute to hour level based on a gated recurrent unit network with attention mechanism enhancement. The long-term fault prediction model performs long-term degradation trend and fault probability prediction at the day to month level based on a Transformer multi-head self-attention encoder. The fault type classification model performs high-dimensional pattern recognition on real-time features based on a deep residual network and outputs specific fault type classification results.

5. The method of claim 1, wherein the method is based on intelligent fault prediction awareness. In step S2, the instantaneous fault prediction model, short-term fault prediction model, long-term fault prediction model and fault type classification model are pre-trained using the transfer learning method, and the model parameters are adjusted based on the real-time collected data.

6. The method of claim 1, wherein the method is based on intelligent fault prediction awareness. In step S3, the working method of the improved DS evidence theory is as follows: The parallel prediction results output in step S2 are defined as four independent sources of evidence. An initial confidence level is generated for each source of evidence, which is the original support level of the source of evidence. The historical accuracy weight, real-time feature quality weight, and time decay factor weight are calculated for each evidence source. The historical accuracy weight is determined based on the prediction accuracy of the model's historical test set, the real-time feature quality weight is calculated based on the completeness and noise intensity of the preprocessed feature data, and the time decay factor weight is assigned a differentiated decay coefficient for the prediction results of the model at different time scales. The weighted product method is used to fuse the calculated historical accuracy weights, real-time feature quality weights, and time decay factor weights of the model to obtain the comprehensive evidence weight for each evidence source. The confidence level of each evidence source is corrected based on the comprehensive evidence weight. All evidence sources after weight correction are then merged based on the Dempster synthesis rule, and the final confidence level is calculated.

7. The method of claim 1, wherein the method is based on intelligent fault prediction awareness. The local situational awareness summary includes the fault occurrence time window, fault type, risk level, and confidence level.

8. The measurement switch control method based on intelligent fault prediction and perception according to claim 1, characterized in that, In step S4, the specific process of identifying latent correlation anomalies and cascade faults based on graph attention network and variational autoencoder is as follows: each measurement switch is taken as a graph node, the topological connection relationship is taken as an edge, and the correlation weight between nodes is learned through graph attention network to capture the latent correlation features of different measurement switch faults. A variational autoencoder is used to reconstruct the global perception data, and abnormal data that deviates from the normal operation mode is identified by the reconstruction error. By combining the correlation features and the results of abnormal data identification, latent correlation anomalies and cascading faults are determined.

9. The measurement switch control method based on intelligent fault prediction and perception according to claim 1, characterized in that, In step S5, a hierarchical collaborative control strategy based on a preset hierarchical risk and strategy mapping matrix is ​​used to dynamically generate control commands for the measurement switch: When the local situational awareness summary indicates that the risk level is less than the low-risk threshold and there are no globally related anomalies, an early warning command is generated and the edge computing terminal is activated for real-time monitoring. When the local situational awareness summary indicates that the risk level is greater than or equal to the low risk threshold, less than the high risk threshold, or when there is a local correlation anomaly, a load adjustment or circuit breaker timing optimization instruction is generated. If the local situational awareness summary indicates a risk level greater than or equal to the high-risk threshold, or if there is a risk of cascading faults, an emergency trip command is generated, and the associated measurement switches in the global situational assessment results are linked to perform coordinated isolation operations to form a fault isolation zone.

Images