A power distribution line abnormality monitoring and early warning method and system

Abstract

The application discloses a power distribution line abnormality monitoring and early warning method and system, and relates to the technical field of intelligent monitoring of power distribution lines. The method comprises the following steps: acquiring operation data collected by multiple types of sensors, performing time alignment, normalization and fusion processing to form a comprehensive data set; inputting the comprehensive data set into a generative denoising model based on the separability learning of noise and abnormal signal distribution, and outputting a clean data sequence through joint optimization of noise suppression and abnormal feature fidelity; identifying abnormal evolution trends with nonlinear amplification characteristics by using Lyapunov exponents and Hurst exponents, and generating risk assessment results; and constructing and dynamically adjusting a set of adaptive early warning threshold values according to the risk assessment results, and outputting abnormal early warning information when a risk index exceeds a threshold value. The application realizes parallel noise suppression and abnormal feature fidelity, has dynamic identification and self-learning capabilities, and significantly improves the accuracy of power distribution line abnormality detection and the stability of the early warning system.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power distribution lines, and in particular to a method and system for monitoring and early warning of power distribution line anomalies. Background Technology

[0002] As an important component of the power system, the operating status of power distribution lines directly affects power supply safety and equipment lifespan.

[0003] With the improvement of power grid automation, various sensors deployed on-site, such as those for temperature, current, voltage, and humidity, can collect a large amount of operational data in real time, providing fundamental support for line condition assessment and fault early warning. However, existing distribution line monitoring and early warning technologies still have the following prominent problems: During line operation, the signals collected by sensors generally contain random noise and interference components due to various factors such as environment, electromagnetic fields, and the accuracy of sampling equipment. Traditional filtering or mean smoothing methods often fail to retain weak signal features related to anomalies while removing noise, resulting in insufficient anomaly identification accuracy. Existing monitoring models are usually based on single distribution fitting or fixed threshold identification strategies. Their denoising process lacks independent modeling of the distribution characteristics of abnormal signals, which may weaken the fidelity of abnormal data while suppressing noise, making it impossible to effectively capture potential early fault features. The operating state of distribution lines has obvious non-stationary and nonlinear characteristics. Conventional feature extraction methods mainly focus on steady-state statistical indicators, which are difficult to reveal the dynamic evolution law of state changes over time, resulting in poor adaptability of risk assessment results to complex operating environments. Existing early warning mechanisms are mostly based on fixed thresholds or empirical rules. When load fluctuates or environmental conditions change, the threshold settings are often out of sync with the actual operating status, which can easily lead to missed or false alarms and make it impossible to achieve real-time dynamic adjustment.

[0004] Therefore, there is an urgent need for a power distribution line anomaly monitoring and early warning technology that can effectively distinguish between noise and abnormal signals, accurately extract non-stationary dynamic characteristics of the line, and achieve self-learning adjustment of risk thresholds, so as to improve the accuracy, real-time performance and intelligent level of anomaly detection. Summary of the Invention

[0005] To address the problems of severe noise interference, easily weakened abnormal features, difficulty in accurately depicting dynamic evolution patterns, and lack of adaptability in fixed thresholds in existing power distribution line monitoring technologies, this invention aims to provide a method and system for power distribution line anomaly monitoring and early warning. By introducing a generative denoising model and a joint optimization mechanism, the system suppresses random noise while maintaining the fidelity of abnormal signal features. Combined with dynamic feature extraction and adaptive threshold adjustment, it achieves high-precision anomaly identification and intelligent risk early warning of power distribution line operating status, thereby significantly improving the monitoring accuracy and operational stability of the system.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] On the one hand, the present invention provides a method for monitoring and early warning of power distribution line anomalies, comprising the following steps:

[0008] S1: Acquire operational data collected by multiple types of sensors in the power distribution line, and perform fusion processing on the operational data to form a comprehensive dataset;

[0009] S2: Input the comprehensive dataset into the pre-trained generative denoising model. The generative denoising model is constructed based on the separability learning of noise signal distribution and abnormal signal distribution, and trained based on historical line operation data. A joint optimization process is performed, which includes simultaneous optimization of noise suppression loss function term and abnormal feature fidelity loss function term, constraining the weight relationship between the two, and outputting a clean data sequence.

[0010] S3: Extract dynamic features characterizing non-stationary changes in the operating state from the clean data sequence, and identify abnormal evolution trends with nonlinear amplification characteristics based on the extracted dynamic features to generate risk assessment results.

[0011] S4: Determine an adaptive early warning threshold set based on the risk assessment results, and dynamically adjust the threshold set according to the changes in the distribution of risk indicators in the risk assessment results;

[0012] When the risk indicator corresponding to the risk assessment result exceeds the corresponding level threshold, an abnormal warning message is output.

[0013] Preferably, the operating data includes a temperature signal reflecting the thermal state of the equipment, a current signal reflecting the electrical load state, and a humidity signal reflecting the environmental conditions;

[0014] The fusion processing of the operational data includes:

[0015] The runtime data is then subjected to time alignment and scale normalization.

[0016] Within a preset sliding time window, the dynamic coupling degree index between various signals is calculated based on the statistical characteristics of the signals. The dynamic coupling degree index is determined by the covariance matrix or mutual information metric between the signals and is used to characterize the instantaneous correlation strength between load signals, temperature signals and environmental signals.

[0017] Preferably, the fusion processing of the operational data adaptively adjusts the fusion weights based on the changing trend of the dynamic coupling degree index and the health status factors of each sensor;

[0018] The health status factor is calculated based on sensor signal stability, historical deviation statistics, and sampling integrity.

[0019] When a sensor signal is detected to be drifting or distorted, the corresponding fusion weights are dynamically corrected based on the health status factor, and the operating data is fused based on the corrected weights to generate a comprehensive dataset that reflects the multi-dimensional coupling characteristics of the power distribution line's operating status.

[0020] Preferably, the generative denoising model includes a diffusion generation step and a feature reconstruction step;

[0021] In the diffusion generation step, multi-step back diffusion calculation is performed on the comprehensive dataset, and a generation probability model is established based on the separability learning of noise signal distribution and abnormal signal distribution. A clean data sequence is output through the collaborative optimization of noise distribution difference term, abnormal distribution difference term and residual signal constraint term.

[0022] In the feature reconstruction step, the residual signal between the intermediate state data and the original input data is calculated based on the generation probability model, and the clean data sequence is output according to the constraint of the residual signal.

[0023] Preferably, the process of extracting dynamic features from the clean data sequence includes:

[0024] Calculate the short-term rate of change, mean square fluctuation amplitude, and inter-window difference similarity of the clean data series within a preset sliding time window;

[0025] A time-series change matrix is ​​established based on the short-term change rate, mean square fluctuation amplitude, and inter-window difference similarity to characterize the degree of non-stationary change in the operating state.

[0026] By performing eigenvector decomposition and energy spectrum analysis on the time-series variation matrix, feature sequences characterizing the dynamic evolution trend of the line state are extracted;

[0027] The feature sequences serve as input data for anomaly pattern recognition and risk assessment.

[0028] Preferably, the process of identifying anomalous evolution trends with nonlinear amplification characteristics includes:

[0029] The finite-time Lyapunov and Hurst exponents are calculated for the dynamic feature sequence to quantify the nonlinear sensitivity and long-range correlation of the system state. When the calculated Lyapunov and Hurst exponents both exceed a preset threshold, nonlinear amplification characteristics are determined. The amplification factor of the abnormal signal is calculated based on the detected nonlinear amplification characteristics, and a risk assessment result is generated by combining the rate of change of the amplification factor.

[0030] Preferably, the risk assessment process includes risk clustering and risk classification steps;

[0031] In the risk clustering step, based on the risk index values ​​and amplification factor change rates in the risk assessment results, an unsupervised clustering algorithm is used to cluster historical risk samples to obtain multiple risk category centers.

[0032] In the risk classification step, the corresponding risk level boundary is calculated based on the center of each risk category, and the risk level boundary is corrected in real time by combining the time drift statistics of the risk indicator distribution, so as to generate a graded risk assessment result containing risk indicator values ​​and level information.

[0033] Preferably, the process of constructing the adaptive early warning threshold set includes:

[0034] Based on the risk indicator distribution corresponding to each risk level in the graded risk assessment results, the center value and boundary value of each risk level are calculated to construct an initial threshold set; during real-time monitoring, the drift statistics of the risk indicator distribution changes are performed; when a drift in the risk indicator distribution is detected, the threshold adjustment coefficient is calculated based on the drift amount, and the threshold set is dynamically updated to form an adaptive early warning threshold set.

[0035] Preferably, the adaptive early warning threshold construction process further includes:

[0036] Based on the updated set of early warning thresholds, risk indicators obtained from real-time monitoring are compared step by step.

[0037] When a risk indicator exceeds the corresponding threshold, an abnormal warning message is output. The warning message includes the warning time, risk level, and the identifier of the corresponding monitoring signal. Feedback data is used to correct the cluster center and threshold set, and a closed-loop self-learning mechanism for warning output and threshold update is established.

[0038] On the other hand, the present invention also provides a power distribution line anomaly monitoring and early warning system, which applies the power distribution line anomaly monitoring and early warning method described above, including:

[0039] The data acquisition module is used to acquire operational data collected by various sensors in the power distribution line. The operational data includes multi-dimensional signals reflecting the thermal state of the equipment, the electrical load state, and environmental conditions.

[0040] The data fusion module is used to perform time alignment and scale normalization on the running data, and to perform fusion processing based on the dynamic coupling relationship between signals to form a comprehensive dataset;

[0041] The generative denoising model module is used to receive the comprehensive dataset and perform generative denoising operations. The generative denoising model is constructed based on the separability learning of noise signal distribution and abnormal signal distribution, and trained based on historical line operation data. During runtime, a joint optimization process is performed. The joint optimization process includes simultaneous optimization of the noise suppression loss function term and the abnormal feature fidelity loss function term, constraining the weight relationship between the two, and outputting a clean data sequence.

[0042] The dynamic feature extraction module is used to extract dynamic features that characterize non-stationary changes in the operating state from the clean data sequence, identify abnormal evolution trends with nonlinear amplification characteristics based on the dynamic features, and generate risk assessment results.

[0043] An adaptive threshold module is used to determine an adaptive early warning threshold set based on the risk assessment results, and to dynamically adjust the threshold set according to the changes in the distribution of risk indicators in the risk assessment results.

[0044] The early warning output module is used to output abnormal early warning information when the risk indicator corresponding to the risk assessment result exceeds the threshold.

[0045] The beneficial effects of this invention are as follows: By synergistically applying key technologies such as multi-source operational data fusion, generative denoising modeling, dynamic feature extraction, and adaptive threshold optimization, this invention constructs an intelligent sensing and dynamic risk early warning system for the operating status of power distribution lines. It can achieve efficient fusion of multi-dimensional data in complex environments, improve signal consistency and data reliability through time alignment and normalization processing, and utilize a generative denoising model to suppress random noise while maintaining the authenticity of abnormal features, solving the problem of abnormal information loss during the denoising process caused by traditional filtering methods. The data after model cleaning is further analyzed in the dynamic feature extraction stage. The non-stationary evolution law of the operating status of power distribution lines is revealed through sliding time windows, rate of change, and energy spectrum characteristics. Combined with nonlinear judgment indicators such as the Lyapunov exponent and Hurst exponent, potential amplification trends of anomalies can be identified in advance, achieving trend early warning of hidden faults. In the risk assessment and early warning stage, this invention establishes an adaptive early warning threshold set based on the risk indicator distribution. Through drift detection and dynamic adjustment mechanisms, the threshold set is updated in real time, enabling the early warning judgment to be automatically corrected according to changes in the operating environment and load. When a risk indicator exceeds a threshold, the system outputs an abnormal warning message that includes the time, risk level, and monitoring signal identifier. It also uses feedback data to correct the cluster centers and threshold set, forming a closed-loop self-learning mechanism for warning output and threshold adjustment, which significantly improves the system's adaptability and long-term stability. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the method in an embodiment of the present invention; Figure 2 This is a system modular diagram in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, this is an embodiment of the present invention, which provides a method for monitoring and early warning of power distribution line anomalies, including the following steps:

[0049] S1: Acquire operational data collected by multiple types of sensors in the power distribution line, and perform fusion processing on the operational data to form a comprehensive dataset;

[0050] In this embodiment, the system can deploy various types of sensors, including temperature sensors, current sensors, and humidity sensors, at key nodes of the power distribution line. These sensors are installed at locations such as cable joints, switch cabinets, busbar trunking, or distribution ring network cabinets to collect multi-dimensional information about the line's operation in real time. The collected raw data is uploaded to the monitoring host or cloud platform for centralized processing via a communication module. In the data preprocessing stage, the system first synchronizes and formats different types of signals according to timestamps to ensure that all sensor data are comparable under a unified timeline, thus providing a consistent data foundation for subsequent fusion analysis. The purpose of fusion processing is to comprehensively reflect the overall status of the line through multi-source information, avoiding misjudgments caused by distortion of a single signal, and ultimately generating a comprehensive dataset that can be used for intelligent algorithm analysis.

[0051] The operating data includes temperature signals reflecting the thermal state of the equipment, current signals reflecting the electrical load state, and humidity signals reflecting environmental conditions.

[0052] In practical implementation, temperature signals reflect the heating status and heat dissipation capacity of equipment, and can characterize changes in contact resistance at connection points or local overload conditions; current signals reflect the electrical load level of power distribution lines, and can reflect load fluctuations and the stability of power transmission; humidity signals reflect external environmental conditions and have a significant impact on changes in insulation performance and equipment aging. These three types of signals constitute the main dimensions for describing the operating status of the line, and their comprehensive analysis results can more fully reveal the thermal, electrical, and environmental coupling characteristics of the line.

[0053] The fusion processing of the operational data includes:

[0054] The fusion processing stage unifies, structures, and weights sensor data from different physical dimensions to obtain comprehensive feature data that can be directly input into intelligent recognition models. This process can be implemented in edge computing units, centralized monitoring hosts, or cloud analysis platforms, with data preprocessing and fusion computation completed through software algorithms.

[0055] The runtime data is then subjected to time alignment and scale normalization.

[0056] Time alignment refers to unifying the timestamps of signals with different sampling periods and asynchronous timing, ensuring strict correspondence between multi-source data at the same moment during fusion. For example, interpolation resampling, delay compensation, or synchronous sampling strategies can be used to achieve time consistency. Scale normalization is used to eliminate differences in the dimensions and numerical ranges of different signals. Common methods include min-max normalization (Z-score) to ensure balanced weighting of each signal in the fusion calculation. Data that has undergone time alignment and normalization can significantly improve feature correlation and the stability of the fusion results.

[0057] Within a preset sliding time window, the dynamic coupling degree index between various signals is calculated based on the statistical characteristics of the signals. The dynamic coupling degree index is determined by the covariance matrix or mutual information metric between the signals and is used to characterize the instantaneous correlation strength between load signals, temperature signals and environmental signals.

[0058] In this embodiment, the system performs local statistical analysis on various signals using a sliding time window as the unit.

[0059] The dynamic coupling index between different signals is calculated within each time window to quantify the short-term correlation between signals.

[0060] Dynamic coupling can reflect linear correlation through the covariance matrix and nonlinear dependence through mutual information. For example, when an increase in load current leads to an increase in temperature and a slowdown in humidity change, the coupling index will exhibit characteristic changes. By continuously calculating the coupling sequence within a sliding window, the temporal synchronicity and direction of influence of different signals can be revealed. This index is the basic input for subsequent adaptive weighted fusion and can reflect the dynamic interaction strength between multidimensional physical quantities.

[0061] The fusion processing of the operational data adaptively adjusts the fusion weights based on the changing trend of the dynamic coupling degree index and the health status factors of each sensor.

[0062] During the fusion and weighting phase, the system comprehensively considers two dimensions: the dynamic coupling degree change trend and the sensor health status, assigning dynamic weights to the data collected by each sensor. When the coupling degree of a certain signal with other signals decreases, the system automatically reduces its weight in the fusion result to reduce the interference of abnormal or irrelevant data on the overall judgment. This process reflects the adaptive characteristics of the fusion algorithm, which can dynamically optimize the contribution ratio of multi-source information under different operating conditions, ensuring the representativeness and reliability of the comprehensive dataset.

[0063] The health status factor is calculated based on sensor signal stability, historical deviation statistics, and sampling integrity.

[0064] The health status factor is used to assess the data reliability of the sensor itself. Its calculation can be based on the fluctuation stability of the sensor's output signal, statistical results of deviations from historical normal values, and the integrity index of the sampled data.

[0065] For example, when a sensor experiences drift, short-term distortion, or packet loss, the health status factor will decrease accordingly.

[0066] The system can periodically assess the health of each sensor and use the results as an important reference for weighted fusion, ensuring that the fusion results are more inclined towards high-reliability signal sources.

[0067] When a sensor signal is detected to be drifting or distorted, the corresponding fusion weights are dynamically corrected based on the health status factor, and the operating data is fused based on the corrected weights to generate a comprehensive dataset that reflects the multi-dimensional coupling characteristics of the power distribution line's operating status.

[0068] During system operation, if the detection module detects drift, abnormal fluctuations, or significant deviations in a sensor signal, it will adjust the fusion weights in real time based on the sensor's health status factor. The fusion algorithm automatically reduces the influence weight of distorted signals and dynamically corrects the fusion results through reweighting calculations. The fused output data, after weight adjustment, more accurately reflects the coupling characteristics of the power distribution line in terms of heat, electricity, and environment, forming a stable and reliable comprehensive dataset. This comprehensive dataset provides high-quality input for subsequent generative denoising models and anomaly detection algorithms, significantly improving the overall monitoring accuracy and robustness of the system.

[0069] S2: Input the comprehensive dataset into the pre-trained generative denoising model. The generative denoising model is constructed based on the separability learning of noise signal distribution and abnormal signal distribution, and trained based on historical line operation data. A joint optimization process is performed, which includes simultaneous optimization of noise suppression loss function term and abnormal feature fidelity loss function term, constraining the weight relationship between the two, and outputting a clean data sequence.

[0070] In this embodiment, the generative denoising model comprises two parts: a diffusion generation step and a feature reconstruction step. Its core idea lies in achieving dual-distribution modeling of noise and anomalous signals through a separability learning mechanism, thereby suppressing random noise while maximizing the authenticity of anomalous features. This model can run on edge computing units or cloud-based intelligent analysis platforms and possesses strong self-learning and generalization capabilities.

[0071] During the diffusion generation phase, the system first performs multi-step back diffusion calculations on the input comprehensive dataset.

[0072] This step is based on the principle of the Stochastic Diffusion Process to build a generation probability model, which gradually restores the integrated data from a high-noise state to a clean signal state.

[0073] The model adopts a shared backbone network and a dual-output head architecture: the backbone network is responsible for capturing the overall evolution characteristics of the time series signal; the two output heads independently predict the noise distribution and the anomaly distribution, respectively.

[0074] The noise prediction head outputs a noise residual estimate, which is used to reconstruct the smoothing trend of the signal;

[0075] The anomaly prediction head outputs the distribution density of anomalous components to preserve weak, fault-characteristic non-stationary signal components.

[0076] During model training, through a learning mechanism that allows for the separability of noise and anomalous signal distributions, the model gradually develops an adaptive ability to recognize the boundary features of these two distributions. Specifically, the system constructs a joint optimization objective function composed of noise distribution difference terms, anomalous distribution difference terms, and residual signal constraint terms. The formula is:

[0077] ;

[0078] in: To characterize the noise suppression error, To characterize the fidelity loss of anomalous signals, The residual signal constraint term between the intermediate state and the true signal; the weight parameters of the three. The gradient balancing algorithm is used to dynamically adjust the parameters to achieve the optimal balance between noise reduction and anomaly preservation.

[0079] After multiple back-diffusion iterations, the model outputs a sequence of progressively converging intermediate states. This sequence is resampled after each iteration to improve the resolution of the generated signal in anomalous feature regions, thereby obtaining a preliminary estimate of clean data.

[0080] In the feature reconstruction stage, the system further utilizes the generation probability model formed in the diffusion generation stage to perform fine calculations on the residual signal between the intermediate state data and the original input data.

[0081] The residual signal is used to describe the difference between the clean signal and the original input, and is a key decomposition result after the mixture of abnormal information and noise.

[0082] The model performs re-constraint optimization on the clean data sequence based on the spatial distribution characteristics and temporal correlation of the residual signal.

[0083] The optimization process introduces a "Residual Consistency Constraint", which maintains the consistency of the statistical characteristics of the residual signal in two consecutive iterations to prevent abnormal features from being excessively weakened during the smoothing process.

[0084] After multiple rounds of reconstruction iterations, the model finally outputs a clean data sequence with a high signal-to-noise ratio. This sequence eliminates random noise interference while fully preserving the non-stationary features associated with potential anomalous events, providing a reliable data foundation for subsequent dynamic feature extraction and risk assessment.

[0085] After processing by the generative denoising model in this embodiment, the signal-to-noise ratio (SNR) of the power distribution line operation data is significantly improved. Compared with traditional filtering or autoencoding denoising methods, this embodiment can achieve efficient denoising without weakening abnormally sensitive features.

[0086] Generative denoising models exhibit good adaptability under multi-source data conditions, and can automatically adjust the diffusion step size and loss weights for different environmental noise intensities, thereby achieving stable and reliable clean data generation.

[0087] The model has a simple structure and highly trainable parameters, and can be embedded in edge monitoring equipment to run in real time, providing data support for anomaly identification, nonlinear evolution analysis and risk prediction of power distribution lines.

[0088] S3: Extract dynamic features characterizing non-stationary changes in the operating state from the clean data sequence, and identify abnormal evolution trends with nonlinear amplification characteristics based on the extracted dynamic features to generate risk assessment results.

[0089] In this embodiment, after denoising, the system performs dynamic feature extraction and anomaly trend identification on the resulting clean data sequence. The goal of this process is to reflect the stability of the power distribution line's operating state through the temporal variation of the signal. Unlike traditional methods that only analyze instantaneous values ​​or steady-state indicators, the dynamic analysis of this invention can detect subtle trends in line operation, thereby identifying potential faults in advance.

[0090] The process of extracting dynamic features from the clean data sequence includes:

[0091] Calculate the short-term rate of change, mean square fluctuation amplitude, and inter-window difference similarity of the clean data series within a preset sliding time window;

[0092] In its implementation, the system sets a sliding time window and calculates three statistical indicators for the data within each time period. The short-term rate of change represents the speed at which the signal rises or falls, reflecting the instantaneous fluctuations in the operating state; the mean square fluctuation amplitude characterizes the strength changes of the signal energy; and the differential similarity is used to compare the similarity of signal patterns within adjacent time windows to measure system stability. This calculation process can be completed automatically by the feature extraction module without manual intervention.

[0093] In this example, three types of statistical features are calculated simultaneously within a time window, forming a multi-dimensional dynamic characterization capability that can comprehensively capture the changing trends of signals at different time scales and significantly improve the sensitivity of anomaly identification.

[0094] A time-series change matrix is ​​established based on the short-term change rate, mean square fluctuation amplitude, and inter-window difference similarity to characterize the degree of non-stationary change in the operating state.

[0095] In this embodiment, the system integrates the above indicators to establish a time-series variation matrix, which is used to record the changing relationship of line operating characteristics over time. This matrix reflects the dynamic coupling strength between signals in different time segments and can quantify the fluctuation stability and changing trend of the line over a period of time. By storing features in matrix form, rapid modeling and analysis of complex signal relationships can be achieved.

[0096] By using matrix modeling, the originally independent time signals are transformed into analyzable structured data, enabling the system to understand the "non-stationarity" characteristics of its operating state from a higher dimension.

[0097] By performing eigenvector decomposition and energy spectrum analysis on the time-series variation matrix, feature sequences characterizing the dynamic evolution trend of the line state are extracted;

[0098] The system performs eigenvector decomposition and energy spectrum analysis on the aforementioned time-series variation matrix to identify the concentration direction and intensity of varying energy in different time periods. The analysis results are output in the form of feature sequences, which reflect the main change patterns of the line condition, such as the temperature change trend caused by load fluctuations or the impact of humidity changes on the insulation condition.

[0099] The feature sequences serve as input data for anomaly pattern recognition and risk assessment.

[0100] The generated feature sequences are input into the anomaly identification model, serving as the core basis for risk analysis. These sequences contain time-series correlation information of three types of signals: thermal, electrical, and environmental, reflecting the transition characteristics of the system from a normal to an abnormal state, and providing comprehensive data support for subsequent nonlinear trend identification.

[0101] The process of identifying anomalous evolution trends with nonlinear amplification characteristics includes:

[0102] The finite-time Lyapunov and Hurst exponents are calculated for the dynamic feature sequence to quantify the nonlinear sensitivity and long-range correlation of the system state.

[0103] In its implementation, the system calculates the finite-time Lyapunov exponent and Hurst exponent based on the feature sequence.

[0104] The finite-time Lyapunov exponent is used to evaluate the divergence rate between system trajectories. The system reconstructs the phase space, forming a trajectory set from the feature sequences with time delay and multidimensional embedding, and calculates the average growth rate of the distance between adjacent trajectories over time. When the distance between trajectories grows exponentially, the Lyapunov exponent is positive, indicating that the system is sensitive to disturbances and its operating state tends to be unstable; if it is negative, it indicates that the disturbance has decayed and the system is in a stable state.

[0105] The Hurst exponent is calculated based on the segmented fluctuation patterns of time series. The system statistically analyzes the ratio of the range to the standard deviation of the signal at different time scales and calculates the slope of the growth trend on a logarithmic coordinate system to estimate the persistence of the time series. A Hurst exponent greater than 0.5 indicates that the abnormal trend is persistent and cumulative; a value less than 0.5 indicates that the fluctuation exhibits anti-persistence characteristics.

[0106] The Lyapunov exponent reflects the system's sensitivity to small disturbances and is used to determine whether the operating state is tending to become unstable.

[0107] The Hurst index represents the persistence and correlation of a time series and is used to analyze whether anomalous behavior has a long-term cumulative effect. Combining the two can quantitatively describe whether an anomalous signal has an "amplification trend." When both the calculated Lyapunov and Hurst indices exceed preset thresholds, nonlinear amplification characteristics are identified. Based on the detected nonlinear amplification characteristics, the amplification factor of the anomalous signal is calculated, and the rate of change of the amplification factor is used to generate a risk assessment result.

[0108] Specifically, when the system detects that both of the above indicators exceed preset thresholds, it considers that the line operation exhibits nonlinear amplification characteristics. At this time, the system calculates the amplification factor of the abnormal signal to reflect the intensity of the abnormal change; simultaneously, it calculates the rate of change of the amplification factor to describe the growth trend of the risk. Combining the results of both, a risk assessment value can be generated to determine whether the line is in a normal, overloaded, or abnormal critical state.

[0109] The risk assessment process includes risk clustering and risk classification steps;

[0110] In the risk clustering step, based on the risk index values ​​and amplification factor change rates in the risk assessment results, an unsupervised clustering algorithm is used to cluster historical risk samples to obtain multiple risk category centers.

[0111] Specifically, in the risk assessment phase, the system uses risk indicator values ​​and the rate of change of amplification factors as input features, and automatically clusters historical risk samples using an unsupervised clustering algorithm. The clustering results divide the system into multiple risk category centers, representing different operational risk states. This process enables the system to automatically learn and identify characteristic patterns of different risk levels without human intervention. In the risk grading step, the corresponding risk level boundaries are calculated based on each risk category center, and the risk level boundaries are corrected in real time using the time drift statistics of the risk indicator distribution, generating a graded risk assessment result containing risk indicator values ​​and level information.

[0112] Specifically, the system calculates the risk level boundary based on the cluster centers obtained from clustering and dynamically updates the boundary by considering the drift of risk indicator distribution in real-time monitoring data. When the system detects a change in the risk level, it automatically adjusts the threshold and outputs a new risk assessment result. The output includes risk indicator values ​​and corresponding level information for visualization and early warning response on the monitoring platform.

[0113] S4: Determine an adaptive early warning threshold set based on the risk assessment results, and dynamically adjust the threshold set according to the changes in the distribution of risk indicators in the risk assessment results; when the risk indicator corresponding to the risk assessment results exceeds the threshold, output abnormal early warning information.

[0114] In this embodiment, the system constructs an adaptive early warning threshold set based on the aforementioned risk assessment results and dynamically adjusts it during operation. When the monitored risk indicator value exceeds the threshold, the system immediately generates an abnormal early warning information. Through this mechanism, the early warning threshold is no longer fixed but can be automatically corrected according to changes in the line's operating status, achieving continuous adaptation to different load conditions, environmental changes, and equipment aging.

[0115] Unlike traditional static threshold models, this solution introduces a "risk-driven self-learning threshold" mechanism, enabling the early warning system to continuously self-calibrate, significantly improving long-term stability and accuracy.

[0116] The process of constructing the adaptive early warning threshold set includes:

[0117] Based on the distribution of risk indicators corresponding to each risk level in the risk assessment results, the central value and boundary value of each risk level are calculated to construct an initial threshold set.

[0118] In the initial stage of the system, the distribution characteristics of risk indicators at different levels in the hierarchical risk assessment results are utilized.

[0119] Calculate the center value and boundary value corresponding to each level.

[0120] For example, the index intervals corresponding to "low risk", "medium risk" and "high risk" can be determined based on the clustering results, and the average value and boundary value of each interval can be used as the initial threshold setting.

[0121] The resulting initial threshold set provides a basic reference for subsequent adaptive updates.

[0122] In this embodiment, by using the output of the risk assessment model in reverse to construct the threshold, a closed-loop logic is achieved where "model judgment" feeds back into "system threshold setting," ensuring that the early warning mechanism is consistent and real-time with the risk assessment results.

[0123] During real-time monitoring, drift statistics are performed on the distribution changes of risk indicators;

[0124] The system continuously monitors the real-time changes in the distribution of risk indicators during operation. When the line operates for an extended period or external conditions change, the statistical distribution of risk indicators will gradually drift. The system uses a drift statistics module to calculate changes in parameters such as the mean and variance of the indicator distribution, using this as a basis for determining the necessity of threshold updates.

[0125] In this embodiment, a "distribution drift detection" mechanism is introduced, which enables the system to automatically identify long-term shifts in risk characteristics, thereby avoiding false alarms and missed alarms caused by environmental changes.

[0126] When a drift in the distribution of risk indicators is detected, a threshold adjustment coefficient is calculated based on the amount of drift, and the threshold set is dynamically updated to form an adaptive early warning threshold set.

[0127] Specifically, when the system identifies a significant drift in the distribution of risk indicators, it calculates a threshold adjustment coefficient based on the drift amplitude and dynamically corrects the initial threshold set. The correction method can employ linear proportional adjustment or a moving weighted average algorithm to achieve smooth threshold updates and avoid misjudgments caused by abrupt changes. The adjusted threshold set becomes the adaptive early warning threshold set, which can automatically replace the old thresholds in subsequent monitoring cycles.

[0128] The adaptive early warning threshold construction process also includes:

[0129] Based on the updated set of early warning thresholds, risk indicators obtained from real-time monitoring are compared step by step.

[0130] Specifically, during real-time operation, the system compares the monitored risk indicator values ​​sequentially with the boundary levels of each level in the adaptive threshold set. The comparison process is performed in ascending order of level. When a risk indicator value exceeds a certain level threshold, the system determines that the current line status is at the corresponding risk level. This step-by-step comparison mechanism avoids the instability of single-threshold decisions, making the risk assessment process more hierarchical and fault-tolerant.

[0131] When a risk indicator exceeds the corresponding threshold, an abnormal warning message is output. The warning message includes the warning time, risk level, and the identifier of the corresponding monitoring signal. After the warning is triggered, the abnormal intensity and duration of the risk indicator are statistically analyzed using a sliding time window. The cluster center and threshold set are corrected based on the monitoring feedback data to establish a closed-loop self-learning mechanism for warning output and threshold update.

[0132] Specifically, when the system continuously detects risk indicators exceeding limits, it statistically analyzes the intensity and duration of the anomalies within a sliding time window and feeds these results back to the risk assessment module. The risk assessment module recalculates the cluster centers and risk level distribution based on the feedback data, thereby correcting the threshold set and achieving closed-loop optimization of threshold updates and early warning outputs. As monitoring data accumulates, the system's judgment accuracy continuously improves, forming an intelligent early warning mechanism with self-correction and self-evolution capabilities. By using early warning output results to drive threshold and cluster model updates, the system achieves true "self-learning" and "closed-loop optimization," possessing self-evolutionary capabilities during long-term operation.

[0133] like Figure 2 As shown, another embodiment of the present invention provides a power distribution line anomaly monitoring and early warning system, which applies the power distribution line anomaly monitoring and early warning method described above, including:

[0134] The data acquisition module is used to acquire operational data collected by various sensors in the power distribution line. The operational data includes multi-dimensional signals reflecting the thermal state of the equipment, the electrical load state, and environmental conditions.

[0135] The data fusion module is used to perform time alignment and scale normalization on the running data, and to perform fusion processing based on the dynamic coupling relationship between signals to form a comprehensive dataset;

[0136] The generative denoising model module is used to receive the comprehensive dataset and perform generative denoising operations. The generative denoising model is constructed based on the separability learning of noise signal distribution and abnormal signal distribution, and trained based on historical line operation data. During runtime, a joint optimization process is performed. The joint optimization process includes simultaneous optimization of the noise suppression loss function term and the abnormal feature fidelity loss function term, constraining the weight relationship between the two, and outputting a clean data sequence.

[0137] The dynamic feature extraction module is used to extract dynamic features that characterize non-stationary changes in the operating state from the clean data sequence, identify abnormal evolution trends with nonlinear amplification characteristics based on the dynamic features, and generate risk assessment results.

[0138] An adaptive threshold module is used to determine an adaptive early warning threshold set based on the risk assessment results, and to dynamically adjust the threshold set according to the changes in the distribution of risk indicators in the risk assessment results.

[0139] The early warning output module is used to output abnormal early warning information when the risk indicator corresponding to the risk assessment result exceeds the threshold.

[0140] In summary, this invention constructs a sustainably evolving power distribution line anomaly monitoring and early warning technology system, realizing an intelligent closed loop from data acquisition, signal denoising, feature recognition to early warning decision-making. This method outperforms existing technologies in noise suppression accuracy, anomaly recognition sensitivity, and threshold self-learning capability, effectively improving the operational safety and predictive maintenance level of power distribution systems. This invention possesses good engineering practicality and scalability, and can be widely applied to urban power distribution networks, smart substations, and industrial power management scenarios, demonstrating significant technological promotion value and promising industrial application prospects.

[0141] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0142] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functionality involved.

[0143] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring and early warning of power distribution line anomalies, characterized in that, Includes the following steps: The system acquires operational data collected by various sensors in the power distribution line, and then fuses and processes the operational data to form a comprehensive dataset. The comprehensive dataset is input into a pre-trained generative denoising model, which is constructed based on the separability learning of noise signal distribution and abnormal signal distribution, and trained based on historical line operation data. A joint optimization process is performed, which includes simultaneous optimization of noise suppression loss function term and abnormal feature fidelity loss function term, and constrains the weight relationship between the two. The generative denoising model includes a diffusion generation step and a feature reconstruction step. In the diffusion generation step, back diffusion calculation is performed on the comprehensive dataset, and a generation probability model is established based on the separability learning of noise signal distribution and abnormal signal distribution. A clean data sequence is output through the collaborative optimization of noise distribution difference term, abnormal distribution difference term and residual signal constraint term. In the feature reconstruction step, the residual signal between the intermediate state data and the original input data is calculated based on the generation probability model, and a clean data sequence is output according to the residual signal constraint term. The model's structure employs a shared backbone network and a dual-output head architecture: the backbone network is responsible for capturing the integral of the timing signal. Evolutionary characteristics of organisms; The two output heads independently predict the noise distribution and the anomaly distribution, respectively. The noise prediction head outputs a noise residual estimate, which is used to reconstruct the smoothing trend of the signal; The anomaly prediction head outputs the distribution density of anomalous components to preserve weak, fault-characteristic non-stationary signal components. Dynamic features characterizing non-stationary changes in operating status are extracted from the clean data sequence, and abnormal evolution trends with nonlinear amplification characteristics are identified based on the extracted dynamic features to generate risk assessment results. The process of extracting dynamic features characterizing non-stationary changes in the operating state from the clean data sequence includes: Calculate the short-term rate of change, mean square fluctuation amplitude, and inter-window difference similarity of the clean data series within a preset sliding time window; A time-series change matrix is ​​established based on the short-term change rate, mean square fluctuation amplitude, and inter-window difference similarity to characterize the degree of non-stationary change in the operating state. By performing eigenvector decomposition and energy spectrum analysis on the time-series variation matrix, feature sequences characterizing the dynamic evolution trend of the line state are extracted; The feature sequences serve as input data for anomaly pattern recognition and risk assessment. The process of identifying anomalous evolution trends with nonlinear amplification characteristics includes: Finite-time Lyapunov and Hurst exponents are calculated for the feature sequences to quantify the nonlinear sensitivity and long-range correlation of the system state. When the calculated Lyapunov exponent and Hurst exponent both exceed the preset threshold, nonlinear amplification characteristics are determined to exist. The amplification factor of the abnormal signal is calculated based on the detected nonlinear amplification characteristics, and the risk assessment result is generated by combining the rate of change of the amplification factor. The process of generating the risk assessment results includes risk clustering and risk classification steps; In the risk clustering step, based on the risk index values ​​and amplification factor change rates in the risk assessment results, an unsupervised clustering algorithm is used to cluster historical risk samples to obtain multiple risk category centers. In the risk classification step, the corresponding risk level boundary is calculated based on the center of each risk category, and the risk level boundary is corrected in real time by combining the time drift statistics of the risk indicator distribution, so as to generate a graded risk assessment result containing risk indicator values ​​and risk levels. An adaptive early warning threshold set is determined based on the risk assessment results, and the threshold set is dynamically adjusted based on changes in the distribution of risk indicators in the risk assessment results. When the risk indicator corresponding to the risk assessment result exceeds the corresponding level threshold, an abnormal warning message is output.

2. The method for monitoring and early warning of power distribution line anomalies according to claim 1, characterized in that, The operating data includes temperature signals reflecting the thermal state of the equipment, current signals reflecting the electrical load state, and humidity signals reflecting environmental conditions. The fusion processing of the operational data includes: The runtime data is then subjected to time alignment and scale normalization. Within a preset sliding time window, the dynamic coupling degree index between various signals is calculated based on the statistical characteristics of the signals. The dynamic coupling degree index is determined by the covariance matrix or mutual information metric between the signals and is used to characterize the instantaneous correlation strength between load signals, temperature signals and environmental signals.

3. The method for monitoring and early warning of power distribution line anomalies according to claim 2, characterized in that, The fusion processing of the operational data adaptively adjusts the fusion weights based on the changing trend of the dynamic coupling degree index and the health status factors of each sensor. The health status factor is calculated based on sensor signal stability, historical deviation statistics, and sampling integrity. When a drift or distortion is detected in a sensor signal, the corresponding fusion weights are dynamically adjusted based on the health status factor. The operational data is then fused based on the corrected weights to generate a comprehensive dataset that reflects the multidimensional coupling characteristics of the power distribution line's operational status.

4. The method for monitoring and early warning of power distribution line anomalies according to claim 1, characterized in that, The process of constructing the adaptive early warning threshold set includes: Based on the risk indicator distribution corresponding to each risk level in the graded risk assessment results, the center value and boundary value of each risk level are calculated to construct an initial threshold set; during real-time monitoring, the drift statistics of the risk indicator distribution changes are performed; when a drift in the risk indicator distribution is detected, the threshold adjustment coefficient is calculated based on the drift amount, and the threshold set is dynamically updated to form an adaptive early warning threshold set.

5. The method for monitoring and early warning of power distribution line anomalies according to claim 4, characterized in that, The process of constructing the adaptive early warning threshold set also includes: comparing the risk indicators obtained from real-time monitoring step by step based on the updated early warning threshold set; when the risk indicator exceeds the corresponding level threshold, outputting abnormal early warning information to trigger the early warning, the abnormal early warning information including the early warning time, risk level and the identifier of the corresponding monitoring signal; after the early warning is triggered, using a sliding time window to statistically analyze the abnormal intensity and duration of the risk indicator, and correcting the cluster center and threshold set according to the monitoring feedback data, establishing a closed-loop self-learning mechanism for early warning output and threshold update.

6. A power distribution line anomaly monitoring and early warning system, employing the power distribution line anomaly monitoring and early warning method as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire operational data collected by various sensors in the power distribution line. The operational data includes multi-dimensional signals reflecting the thermal state of the equipment, the electrical load state, and environmental conditions. The data fusion module is used to perform time alignment and scale normalization on the running data, and to perform fusion processing based on the dynamic coupling relationship between signals to form a comprehensive dataset; The generative denoising model module is used to receive the comprehensive dataset and perform generative denoising operations. The generative denoising model is constructed based on the separability learning of noise signal distribution and abnormal signal distribution, and trained based on historical line operation data. During runtime, a joint optimization process is performed. The joint optimization process includes simultaneous optimization of the noise suppression loss function term and the abnormal feature fidelity loss function term, constraining the weight relationship between the two, and outputting a clean data sequence. The dynamic feature extraction module is used to extract dynamic features that characterize non-stationary changes in the operating state from the clean data sequence, identify abnormal evolution trends with nonlinear amplification characteristics based on the dynamic features, and generate risk assessment results. An adaptive threshold module is used to determine an adaptive early warning threshold set based on the risk assessment results, and to dynamically adjust the threshold set according to the changes in the distribution of risk indicators in the risk assessment results. The early warning output module is used to output abnormal early warning information when the risk indicator corresponding to the risk assessment result exceeds the corresponding level threshold.

Images