A method for full life cycle status assessment and early warning of power transmission and transformation equipment

Abstract

This invention discloses a method for full lifecycle status assessment and early warning of power transmission and transformation equipment, belonging to the field of power distribution equipment monitoring technology. It includes a four-layer structure: an edge sensing and data preprocessing layer for adaptive scheduling of edge nodes; edge node self-sensing and data reliability quantification verification; spatiotemporal alignment of multimodal data; a full lifecycle data management and evaluation layer for constructing a digital archive of the power transmission and transformation equipment throughout its entire lifecycle; multi-stage dynamic status assessment covering the entire lifecycle; equipment remaining life prediction; an intelligent early warning and decision support layer for early fault warning; fault source tracing and impact propagation analysis; and a model adaptive evolution and knowledge sharing layer for adaptive evolution of edge models; and cross-equipment fault knowledge transfer based on a multimodal large model. This enables full lifecycle status assessment and early warning processing of power transmission and transformation equipment.

Description

Technical Field

[0001] This invention belongs to the field of power distribution equipment monitoring technology, specifically relating to a method for assessing and providing early warning of the full life cycle status of power transmission and transformation equipment. Background Technology

[0002] With the deepening of the construction of new power systems and the continuous advancement of the energy transition strategy, my country's power grid scale continues to expand, the number of transmission and transformation equipment has surged, and the technical complexity has significantly increased. By the end of 2023, the total number of various status sensing components deployed in smart substations of 110kV and above nationwide had exceeded 4.7 million, with an average annual growth rate of 18.3%. At the same time, the power structure is shifting from coal-fired power, which has controllable and continuous output, to new energy sources, which have strong uncertainty and weak controllable output. The load characteristics are also shifting from rigid consumption to flexible, combining production and consumption. The randomness and volatility of power generation and load have significantly increased, posing unprecedentedly high requirements for the safe and reliable operation of transmission and transformation equipment.

[0003] Traditional power transmission and transformation equipment operation and maintenance models primarily employ a "periodic inspection and post-incident handling" approach, which has significant limitations. Manual inspections require substantial time to traverse vast areas, and are constrained by traffic and terrain conditions, making high-frequency, comprehensive equipment checks difficult. The overall operation cycle is long, response time is slow, and inspection personnel often need to work at heights or near live equipment, posing significant safety risks. Periodic testing, on the other hand, suffers from both "over-maintenance" and "under-maintenance," increasing maintenance costs and potentially affecting power supply reliability due to frequent power outages, or even causing equipment failures due to improper operation. In recent years, with the development of IoT, sensor, and AI technologies, online monitoring technology has been widely applied in the condition management of power transmission and transformation equipment. Intelligent sensing elements based on principles such as fiber Bragg gratings, piezoelectric films, and MEMS micromechanical structures have been widely used in key scenarios such as transformer oil temperature monitoring, GIS partial discharge detection, and circuit breaker mechanical characteristic analysis. In terms of data analysis, existing technologies mostly employ the analytic hierarchy process (AHP) and entropy weighting to quantify the weights of equipment status indicators, or use machine learning algorithms such as artificial neural networks, cluster analysis, and Bayesian networks to build predictive models in conjunction with historical operating data. Some advanced systems have also introduced adaptive threshold algorithms to dynamically adjust alarm thresholds based on historical equipment data and operating conditions to reduce false alarm rates.

[0004] Based on the current development trends and levels of IoT and AI, the application technology for full life cycle status assessment and early warning of power transmission and transformation equipment has revealed shortcomings that can have a substantial impact: 1) Insufficient data processing capabilities, resulting in low data quality and fusion accuracy. Edge computing resource scheduling is rigid, resulting in a high idle rate of computing power under normal operating conditions and the easy loss of critical fault data during emergencies. The lack of a reliable data measurement mechanism leads to a large number of false alarms due to data distortion caused by edge node hardware failures and electromagnetic interference. Multimodal data suffers from poor spatiotemporal alignment accuracy, fails to utilize the physical coupling characteristics of devices, and has low reliability in fusion analysis results. 2) The status assessment system is rigid and lacks dynamic adaptability throughout the entire life cycle. The assessment focuses only on the operational phase, neglecting the cumulative impact of earlier factors such as design, manufacturing, and installation. The evaluation indicators and weights are fixed and cannot be dynamically adjusted according to the equipment life cycle stage; The remaining life prediction only considers the operating load and does not take into account the coupling effect of environmental factors such as temperature, salt spray, and lightning strikes with cumulative damage. The prediction error is large under extreme conditions. 3) Weak early warning and decision-making capabilities, and low efficiency in handling faults. Relying on fixed thresholds or simple trend analysis cannot identify early, weak abnormal signals, resulting in delayed fault warnings. Fault diagnosis can only identify the type and location, lacks the ability to reason about causes and cannot trace the root cause; Without analyzing the propagation path of the fault within the equipment and the power grid, it is difficult to provide scientific support for handling decisions; 4) Lack of model iteration and knowledge sharing mechanisms leads to long-term performance degradation of the system. Once deployed, the edge model remains unchanged and cannot adapt to conceptual drift caused by equipment aging and changes in operating conditions, resulting in a gradual decline in performance over time. Data barriers exist between different types and manufacturers of equipment, making it difficult to transfer fault knowledge and causing diagnostic obstacles for new and rare equipment. Summary of the Invention

[0005] This invention provides a method for full lifecycle status assessment and early warning of power transmission and transformation equipment, comprising four layers: an edge perception and data preprocessing layer responsible for multi-source data acquisition, adaptive edge computing power scheduling, data credibility verification, and precise spatiotemporal alignment of multimodal data; a full lifecycle data management and assessment layer constructing a digital archive of the equipment's full lifecycle, enabling multi-stage dynamic status assessment and remaining life prediction considering cumulative damage and environmental coupling effects; an intelligent early warning and decision support layer realizing early fault warning based on abnormal pattern evolution and fault source tracing and impact propagation analysis based on causal reasoning; and a model adaptive evolution and knowledge sharing layer realizing online adaptive evolution of edge models and cross-equipment fault knowledge transfer based on multimodal large models. The four-layer structure forms a complete closed loop, realizing full lifecycle status assessment and early warning processing of power transmission and transformation equipment, optimizing and improving the shortcomings of current similar methods in data acquisition and processing, status assessment system, early warning decision-making, and continuous model growth.

[0006] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution: A method for full life cycle status assessment and early warning of power transmission and transformation equipment adopts an "edge-cloud" collaborative architecture, which is divided into four core layers: edge perception and data preprocessing layer, full life cycle data management and assessment layer, intelligent early warning and decision support layer, and model adaptive evolution and knowledge sharing layer. The edge sensing and data preprocessing layer completes its hierarchical functions using the following methods: S1.1: Edge node adaptive scheduling, based on sliding window data value density assessment, analyzes the rate of change, degree of anomaly and similarity with historical fault data, automatically identifies early fault symptom data and assigns it the highest processing priority; By introducing a distributed collaborative computing mechanism among edge nodes, when the computing power of a single node is insufficient, non-critical tasks are automatically offloaded to adjacent idle nodes to ensure that high-value data is processed first. Dynamically adjust the sampling frequency and storage period of different types of data to reduce the storage pressure on edge nodes while ensuring data integrity; S1.2: Edge node self-awareness and data credibility quantification verification. Construct a correlation mapping model between node working status and collected data quality. Learn the characteristic patterns of data distortion under different working states through machine learning algorithms and generate a node credibility coefficient between 0 and 1. A multi-source data fusion algorithm based on credibility weighting is proposed, which automatically reduces the weight of low-credibility node data in the fusion process and completely blocks node data with credibility below a set threshold. It enables self-diagnosis and self-isolation of edge node faults. When a serious fault is detected in a node, it automatically isolates it from the network and notifies the operation and maintenance personnel to carry out repairs. S1.3: Spatiotemporal alignment of multimodal data, based on a dual spatiotemporal alignment algorithm of cross-correlation function and physical constraints. First, the initial time delay between different sensor data is calculated by cross-correlation function, and then the initial result is corrected by using the theoretical time delay derived from the multiphysics coupling model. We designed a feature alignment network that combines dynamic temporal warping (DTW) with an attention mechanism to further eliminate local temporal biases between data. Achieve microsecond-level precise alignment of multimodal data, including at least vibration, temperature, partial discharge, and SF6 gas density, to provide a high-quality data foundation for subsequent fusion analysis; The full lifecycle data management and evaluation layer completes its hierarchical functions using the following methods: S2.1: Construct a digital archive of the entire life cycle of power transmission and transformation equipment, and establish a unified data standard for the entire life cycle of power transmission and transformation equipment; use blockchain technology to store the digital archives to ensure the immutability and traceability of the data; S2.2: Multi-stage dynamic status assessment covering the entire life cycle, with a dedicated assessment index system and weight allocation scheme designed for each stage; based on the stage transition probability dynamic assessment model, by analyzing the equipment's operating data, maintenance records, and environmental data, it automatically identifies the life cycle stage of the equipment and switches the corresponding assessment strategy in real time; comprehensively considering the equipment's current status, historical status change trends, and cumulative damage throughout the entire life cycle, it provides a comprehensive health index for the equipment and classifies the equipment status into four levels: normal, attention, abnormal, and severe. S2.3: Equipment remaining life prediction, introducing rainflow counting method and Miner's rule to calculate the cumulative fatigue damage of each key component of the equipment under alternating load; Design a remaining lifetime prediction network based on LSTM and attention mechanism to automatically learn the influence weights of different environmental factors such as temperature, humidity, salt spray, and lightning strike on the equipment damage rate. To achieve dynamic correction and early warning of the remaining lifespan of equipment under extreme environments, when it is predicted that the remaining lifespan of the equipment is lower than the set threshold, an early warning signal is issued in advance to remind maintenance personnel to arrange maintenance or replacement. The intelligent early warning and decision support layer completes its hierarchical functions using the following methods: S3.1: Early fault warning, based on the dual warning indicators of feature spatial distance and abnormal mode evolution speed, adopts an abnormal mode evolution trajectory tracking algorithm, and predicts the abnormal development trend and possible fault types by continuously monitoring changes in abnormal features; and issues different levels of warning signals according to the severity and development speed of the abnormality. S3.2: Fault source tracing and impact propagation analysis. A fault source tracing algorithm based on causal graph and Bayesian network is adopted. Starting from the observed abnormal phenomena, the root cause of the fault is deduced in reverse and the probability of each possible cause is calculated. Based on the power grid topology and electrical connections between devices, predict other devices and power grid areas that may be affected by a fault; Generate personalized fault handling suggestions and emergency plans, including fault isolation steps, maintenance personnel allocation, and spare parts preparation, to assist maintenance personnel in handling faults quickly and accurately. The model adaptive evolution and knowledge sharing layer completes its hierarchical functions using the following methods: S4.1: Edge model adaptive evolution, based on active learning and incremental learning model adaptive evolution. When concept drift is detected, the most valuable sample is automatically selected for labeling, and the model parameters are updated using incremental learning methods. During the model update process, the old and new versions of the model run simultaneously. If a problem occurs in the new version of the model, it can be rolled back to the old version immediately to ensure the continuity of system operation. S4.2: Cross-device fault knowledge transfer based on multimodal large model, which transfers fault knowledge of existing devices to new and rare devices to achieve rapid cold start of edge models; Construct a knowledge graph of equipment faults to realize the structured representation and sharing of fault knowledge, enabling equipment from different regions and manufacturers to share fault diagnosis experience.

[0007] Preferably, the time sliding window in the data value density assessment is set to sliding step size Counting sliding window sliding step size ; When the rate of change of data is detected to exceed the threshold, the window size is automatically reduced to half of its original size, and the step size is reduced to one-third of its original size; when the data in three consecutive windows is stable, the default parameters are restored. The method for evaluating the rate of change of data is as follows: 1) First-order difference rate of change First-order difference , For the sensor measurement value at the i-th sampling point, This represents the sensor measurement value at the (i-1)th sampling point; First-order difference rate of change W represents the total number of sampling points contained in the current sliding window; 2) Second-order difference acceleration Second-order difference of the i-th sampling point Second-order difference acceleration 3) Volatility Standard deviation of data within the window The arithmetic mean of the data within the window ; In-window data volatility The anomaly assessment process is as follows: Calculate the mean of the window data and standard deviation ; Abnormal scores ,when When this occurs, it is judged as a statistical anomaly; The historical fault data similarity evaluation process is as follows: 1) Build a historical fault template library Collect sensor data from the 30 minutes preceding historical equipment failures; Feature sequences are extracted for each fault type to form a fault template library. ; 2) Dynamic Time Warping (DTW) Distance Calculation Current window data sequence , Fault Template Construct the distance matrix ,in Calculate the cumulative distance matrix , DTW distance: 3) Similarity score .

[0008] Preferably, the distributed collaborative computing mechanism is implemented as follows: 1) Node status awareness and resource quantification Each edge node collects its own resource status parameters every 1 second, including: CPU utilization. Memory utilization Network uplink bandwidth Network downlink bandwidth Remaining storage space ; Node Comprehensive Load Index , , , These are the weighting coefficients, with a sum of 1; initially set to 0.4, 0.3, 0.15, and 0.15 respectively. This represents the theoretical maximum bandwidth of the node network. When the overall load index L of a node changes by more than 5%, the latest resource status information is broadcast to all adjacent edge nodes through a distributed state synchronization mechanism based on the gossip protocol. Each edge node maintains a resource status table of adjacent nodes, which records the IP address, overall load index and individual resource parameters of all adjacent nodes, and updates it regularly. 2) Task unloading based on value priority When the CPU utilization of the local node U cpu ≥80% or memory utilization U mem When the percentage is ≥70%, the task unloading decision process is triggered; The tasks to be processed are divided into four priority queues according to the data value density score V: P0 (V≥0.8), P1 (0.6≤V<0.8), P2 (0.3≤V<0.6), and P3 (V<0.3). Execute the following task to uninstall priority rules: P0 and P1 level tasks must be processed on the local node and must not be uninstalled. P2 level tasks are preferentially unloaded to adjacent free nodes, and are only processed locally if all adjacent nodes have no free resources. P3 level tasks can be directly offloaded to the cloud, or when local storage resources are insufficient, only statistical features can be retained and the original data can be discarded. For P2 level tasks that need to be unloaded, select the node with the lowest overall load index L < 50% and the lowest network latency from the adjacent node resource status table as the target unloading node. When the CPU utilization of the local node U cpu <50% and memory utilization U mem When the percentage is less than 40%, it automatically receives P2-level tasks that have been unloaded from other adjacent nodes.

[0009] Preferably, the correlation mapping model between node working status and collected data quality adopts a multi-input multi-output hierarchical fusion architecture, taking the multi-dimensional working status parameters of edge nodes and the multi-dimensional quality characteristics of collected data as inputs, and the node credibility coefficient between 0 and 1 as output, while simultaneously outputting the sub-scores of data quality in each dimension; the model structure includes: The model consists of four layers: a multi-source data input layer, which receives two types of heterogeneous input data simultaneously with a sampling period of 10 seconds; a node state feature encoding layer, which uses a stacked autoencoder (SAE) to reduce the dimensionality and encode the features of high-dimensional, nonlinear node operating state parameters, extracting low-dimensional implicit features that characterize the health of nodes; a data quality feature extraction layer, which uses a bidirectional gated recurrent unit (Bi-GRU) to extract the quality features of temporally acquired data, capturing the dynamic patterns of data quality changes over time; a cross-modal attention association layer, the core of the model, which uses a multi-head self-attention mechanism to establish a nonlinear association mapping between node operating states and data quality, dynamically learning the weights of different operating states on data quality; and a credibility quantification output layer, which uses a fully connected neural network to process the associated feature vectors, ultimately outputting a node credibility coefficient between 0 and 1. The multi-source data fusion algorithm process is as follows: 1) Preprocess heterogeneous data from different edge nodes Timestamp alignment: Linear interpolation is used to align the sampled data of all nodes to a unified time base, ensuring that the data within the same sliding window have the same timestamp; for missing sample points, the weighted average of adjacent nodes is used for temporary filling. Data standardization: The Z-score standardization method is used to map sensor data of different dimensions to a standard normal distribution; Outlier filtering: The 3σ principle is used to initially filter the standardized data, and data that exceeds the range of [−3,3] are marked as preliminary outliers and given lower weights in the subsequent fusion process; 2) Dynamic fusion weight calculation Basic credibility weight ,in Let be the credibility coefficient of the i-th node, taking a value of [0,1]; T is the credibility threshold, taking a value of 0.5; when At that time, the data of that node is completely masked and does not participate in the fusion calculation; Introducing a time decay factor, , This is the time decay coefficient, with a default value of 0.01; Let be the difference between the last update time and the current time of the credibility coefficient of the i-th node; Weight adjustment: Adjusting weights based on data quality. ,in Let i be the data missing rate within the current window of the i-th node. This represents the initial proportion of outliers; Weight normalization: The final weights of all nodes are normalized to ensure that the sum of the weights is 1. Where N is the total number of nodes participating in the fusion, if all nodes If all values ​​are 0, a system alarm will be triggered, and the most recent valid fusion result or historical data from the cloud will be used as the temporary output. 3) Weighted evidence fusion calculation ,in Let be the measured value of the i-th node; The fusion results are transformed from the normalized space back to the original dimensional space to obtain the final fused measurement value: ,in and represent the standard deviation and mean of the j-th sensor under normal operating conditions, respectively.

[0010] Preferably, the multiphysics coupling model includes: The elastic wave propagation model is used to calculate the propagation delay of vibration signals from the fault source to the vibration sensor. It is suitable for monitoring mechanical faults such as transformer winding deformation and core loosening. A three-dimensional elastic wave longitudinal wave propagation model is adopted, considering the wave impedance differences of different media including transformer oil, iron core, windings, and tank walls: ,in This is the theoretical time delay for the vibration signal to propagate to the sensor when the fault source is located at the (x,y,z) coordinates; n is the number of media layers that the vibration wave passes through along its propagation path; Let be the distance the vibration wave travels in the k-th layer of medium; Let be the longitudinal wave propagation velocity of the vibration wave in the k-th layer of medium; The heat conduction sub-model is used to calculate the propagation delay of temperature signals from the fault source to the temperature sensor, and is suitable for monitoring thermal faults such as winding overheating and core overheating. An unsteady-state heat conduction equation is adopted, considering both convective heat transfer in transformer oil and heat conduction in the solid medium: , For the density of the medium, Let T be the specific heat capacity of the medium at constant pressure, t be the temperature of the medium, t be the time, k be the thermal conductivity of the medium, and q be the intensity of the internal heat source. By solving the above partial differential equations, the temperature field distribution over time is obtained. When the heat generated by the fault source propagates to the temperature sensor location, causing the sensor reading to rise above a preset threshold (usually 0.1℃), the corresponding time is the theoretical heat conduction delay τ. temp ; A partial discharge propagation sub-model is used to calculate the propagation delay of electromagnetic waves and ultrasonic signals generated by partial discharge from the discharge source to the sensor, and is suitable for monitoring partial discharge faults. Electromagnetic wave propagation sub-model: Electromagnetic waves generated by partial discharge propagate at the speed of light inside the transformer with extremely short time delay. , For the propagation delay of electromagnetic waves, This refers to the straight-line distance from the power supply to the ultra-high frequency sensor. The speed of light in a vacuum. The relative permittivity of the propagation medium (εr≈2.2 for transformer oil); Ultrasonic propagation sub-model: Ultrasonic waves generated by partial discharge propagate in transformer oil as longitudinal waves. , For the propagation delay of ultrasound, The speed at which ultrasound propagates in transformer oil; An SF6 gas density change propagation sub-model is used to calculate the propagation delay of density changes caused by gas leakage or decomposition from the fault source to the SF6 density sensor in SF6 gas-insulated equipment. The Navier-Stokes equations for compressible fluids are used to describe the flow and density changes of SF6 gas: ,in The density of SF6 gas. For gas pressure, For gas dynamic viscosity, It is the acceleration due to gravity; By solving the above equations, the distribution of the SF6 gas density field over time is obtained; when the gas density change at the sensor location exceeds a preset threshold, the corresponding time is the theoretical SF6 gas propagation delay τ. sf6 .

[0011] Preferably, the entire lifecycle of the equipment integrates information from design to retirement, including: design and manufacturing stage, installation and commissioning stage, operation stage, maintenance stage, and retirement and scrapping stage.

[0012] Preferably, the evaluation index system adopts a general basic index library, including 6 categories and 32 core indicators: electrical performance, oil chromatography and oil quality, partial discharge, mechanical vibration, temperature and environment, and operation and maintenance. In the initial stage of operation, there are 18 indicators, with core indicators accounting for 85% and auxiliary indicators accounting for 15%. During the stable operation period, there are 24 indicators, with core indicators accounting for 70%, trend indicators accounting for 20%, and auxiliary indicators accounting for 10%. During the performance degradation period, there are 26 indicators, with core indicators accounting for 55%, trend indicators accounting for 35%, and auxiliary indicators accounting for 10%. At the end of life, there are 22 indicators, with core indicators accounting for 75%, trend indicators accounting for 20%, and auxiliary indicators accounting for 5%. The weight allocation scheme is designed as follows: Basic weighting: Calculated by combining 60% expert subjective weighting and 40% data objective weighting; Phase adjustment: The weight of core indicators is multiplied by 1.5, the weight of auxiliary indicators remains unchanged, and the weight of non-key indicators is multiplied by 0.5; Abnormal increase: When the indicator value exceeds the attention threshold, the weight is multiplied by 2; when it exceeds the abnormal threshold, the weight is multiplied by 3. All indicator weights are ultimately normalized, and the sum is 1. The dynamic evaluation model for stage transition probability includes: The multi-source data input layer accepts the following four types of data: Time-series status data: comprehensive health index, scores of each individual indicator, and rate of change of health index for three consecutive months; Event-based data: maintenance records (routine / major / emergency maintenance), fault records (general / serious faults), and technical modification records; Static attribute data: equipment model, design life, commissioning time, and cumulative operating time; Environmental and load data: average ambient temperature, number of days with extreme temperatures, cumulative overload duration, and average load rate; The dynamic transition probability calculation layer constructs a 4×4 initial basic transition matrix P0 based on historical statistical data of similar equipment throughout their entire lifecycle, representing the baseline probability of a device transitioning from stage i to stage j: Four correction coefficients that can be calculated in real time are introduced to adjust the base transition probability element by element: Stage Dwell Time Correction Factor: The longer the equipment stays in the current stage, the higher the probability of transitioning to a subsequent stage; for example, after a stable operating period of more than 15 years, the probability of transitioning to the deterioration stage increases by 20% for each additional year. Deterioration rate correction factor: When the monthly decline rate of the health index exceeds 0.02, the transition probability is multiplied by 1.5; when it exceeds 0.05, it is multiplied by 3. Event impact correction factor: After a serious failure, the probability of transitioning to the next stage is multiplied by 5; after a major overhaul is completed, the probability of transitioning back is multiplied by 0.2 (only a limited backtracking to the adjacent previous stage is allowed). Environmental load correction factor: If the annual average load rate exceeds 80% or the number of extreme temperature days exceeds 30 days, the migration probability is multiplied by 1.3; For each element of the fundamental transition matrix Multiply by the product of the corresponding correction coefficients to obtain the real-time transfer probability. Then, normalize the values ​​to ensure that the sum of probabilities for each row is 1; The stage identification and inference layer performs probabilistic inference based on the Hidden Markov Model (HMM) framework: Hidden states: 4 full lifecycle stages; Observations: current health index, rate of change of health index, and the largest outlier of a single indicator in the last 3 months; Inference algorithm: the forward algorithm is used to calculate the posterior probability of the device being in each stage at the current moment; Stage determination rules: When the probability of a certain stage is ≥0.8, the device is determined to enter that stage; when the probabilities of two adjacent stages are both between 0.5 and 0.8, it is determined to be a transition stage, and a two-stage evaluation strategy is used for weighted averaging; when the probabilities of all stages are <0.5, manual review is triggered. In the strategy switching layer, once the phase determination is completed, the corresponding phase's exclusive indicator system and weight allocation scheme are immediately loaded; in the transition phase, the evaluation strategies of the two phases are mixed according to probability weights; when switching phases, the historical evaluation data of the previous 3 months are retained for smooth transition.

[0013] Preferably, the LSTM and attention mechanism remaining lifetime prediction network includes: a data input layer for cleaning, normalizing, aligning, and sliding window segmenting the monitoring data of the device, including at least oil chromatography data, vibration signals, partial discharge, temperature, and load current; LSTM layer: As the foundation of the network, it is used to automatically learn the temporal dependencies in the process of device degradation. Through its gating mechanism, LSTM can effectively capture long-term and short-term dependency patterns in monitoring data, avoiding the limitations of traditional methods that require manual feature extraction, and directly extracting deep state features from the raw or pre-processed temporal data. Attention Mechanism Layer: Connected after the LSTM layer; this mechanism dynamically evaluates the features of all time steps of the LSTM output and assigns different weights to the features of different time steps; its core function is to enable the model to automatically focus on more indicative key time points or features in the device degradation process, suppress the interference of noise and non-critical information, and thus extract the core information representing the health status of the device more accurately.

[0014] Preferably, the fault tracing algorithm process of the causal graph and Bayesian network is as follows: 1) Data collection and preprocessing: Integrate multi-source heterogeneous data from the equipment, including online monitoring data (such as oil chromatography, partial discharge, vibration signals), inspection records, operation logs, and historical fault reports; clean, denoise, align, and extract features from the data to provide high-quality input for modeling; 2) Causal Model Construction and Validation: Generating causal structure hypotheses based on prior knowledge and data-driven methods; Statistical tests or algorithms are used to verify the existence and direction of causal relationships between variables; Complete the learning of Bayesian network parameters and construct a complete probability model; Use cross-validation or historical case backtesting methods to evaluate the accuracy and reliability of the model; 3) Root cause analysis and application: Input real-time or post-event abnormal data into the validated model for probabilistic reasoning; the algorithm can output the most likely sequence or combination of fault root causes and provide quantitative basis for maintenance decisions, such as distinguishing whether transformer oil chromatographic abnormalities are caused by partial discharge or overheating faults.

[0015] Preferably, the adaptive evolution process of the model is as follows: 1) The system detected concept drift and triggered the model evolution process; 2) From the new data of the most recent month, 100 high-value samples are selected through a three-level sampling strategy: uncertainty sampling, representative sampling, and rare sampling. The uncertainty sampling prioritizes the samples with the most uncertain model prediction results. The representative sampling selects samples from the candidate set that can represent the distribution characteristics of the new data. The rare sampling focuses on capturing novel failure modes that have never appeared before. 3) The samples are automatically pre-labeled and pushed to the cloud for expert review and confirmation; 4) After the samples are labeled, they are synchronized back to the edge nodes to construct an incremental training set; 5) The new model is trained using an incremental learning method that combines hierarchical updates with knowledge distillation; 6) After the new model passes performance testing on the historical validation set, it enters the dual-model parallel operation phase; 7) After the new model has been running stably for 72 consecutive hours and its performance is better than that of the old model, it will be officially switched to the production model.

[0016] The beneficial effects of this invention are: This invention addresses the systemic deficiencies of existing power transmission and transformation equipment condition assessment and early warning technologies by constructing a fully intelligent operation and maintenance system that integrates "edge-cloud" collaboration. It achieves comprehensive technological breakthroughs from data processing, condition assessment, early warning decision-making to system evolution, bringing significant technological value, operational benefits, and grid security gains. A highly reliable and high-value edge data processing foundation has been built. Breaking through the limitations of traditional fixed computing power allocation models, this approach enables dynamic, on-demand scheduling of edge computing resources. This significantly improves resource utilization efficiency under normal conditions and fundamentally solves the problem of data loss of critical fault symptoms under extreme conditions such as lightning strikes and sudden load changes. Through edge node self-awareness and data reliability quantification verification mechanisms, a quality assurance system is established from the data acquisition source, effectively filtering out distorted data caused by hardware failures and electromagnetic interference, greatly reducing invalid alarms and false alarms, and avoiding waste of operational resources. Simultaneously, a multimodal data spatiotemporal alignment method based on the physical coupling characteristics of devices overcomes the technical bottleneck of simply relying on timestamp alignment, achieving microsecond-level precise synchronization of multi-physical quantity data, providing a solid and reliable data foundation for subsequent multi-source fusion analysis.

[0017] A dynamic status assessment system that truly covers the entire life cycle has been established. Breaking away from the limitations of traditional assessment methods that focus solely on the operational phase of equipment, this approach constructs an immutable, full-lifecycle digital archive, bridging data barriers between design, manufacturing, installation, operation, maintenance, and decommissioning. This enables end-to-end traceability and quantification of equipment status. The dynamic assessment model, based on a five-stage equipment lifecycle division, adapts to the degradation patterns and risk characteristics of different stages, ensuring assessment results more closely reflect the actual health condition of the equipment. Furthermore, the remaining life prediction method, considering the coupling effect of operating load and environmental factors, comprehensively covers the impact of complex environments such as temperature, salt spray, and lightning strikes on equipment damage. It particularly enhances the reliability of life prediction under extreme conditions, providing a precise basis for the scientific formulation of equipment maintenance and replacement plans. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying 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.

[0019] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0021] A method for full life cycle status assessment and early warning of power transmission and transformation equipment adopts an "edge-cloud" collaborative architecture and completes the above method through four core layers: edge perception and data preprocessing layer, full life cycle data management and assessment layer, intelligent early warning and decision support layer, and model adaptive evolution and knowledge sharing layer. The edge sensing and data preprocessing layer is used to collect sensor data from various nodes of the device and preprocess the collected data; the implementation process is as follows: S1.1: Edge node adaptive scheduling, based on sliding window data value density assessment, analyzes the rate of change, degree of anomaly and similarity with historical fault data, automatically identifies early fault symptom data and assigns it the highest processing priority; Time sliding window set sliding step size Counting sliding window sliding step size When the rate of change of data is detected to exceed the threshold, the window size is automatically reduced to half of its original size, and the step size is reduced to one-third of its original size; when the data in three consecutive windows is stable, the default parameters are restored. The data change rate assessment method is as follows: 1) First-order difference rate of change First-order difference , For the sensor measurement value at the i-th sampling point, This represents the sensor measurement value at the (i-1)th sampling point; First-order difference rate of change W represents the total number of sampling points contained in the current sliding window; 2) Second-order difference acceleration Second-order difference of the i-th sampling point Second-order difference acceleration 3) Volatility Standard deviation of data within the window The arithmetic mean of the data within the window ; In-window data volatility The anomaly assessment process is as follows: Calculate the mean of the window data and standard deviation ; Abnormal scores ,when When this occurs, it is judged as a statistical anomaly; The historical fault data similarity evaluation process is as follows: 1) Build a historical fault template library Collect sensor data from the 30 minutes preceding historical equipment failures; Feature sequences are extracted for each fault type to form a fault template library. ; 2) Dynamic Time Warping (DTW) Distance Calculation Current window data sequence , Fault Template Construct the distance matrix ,in Calculate the cumulative distance matrix , DTW distance: 3) Similarity score .

[0022] By introducing a distributed collaborative computing mechanism among edge nodes, when the computing power of a single node is insufficient, non-critical tasks are automatically offloaded to adjacent idle nodes to ensure that high-value data is processed first. The implementation process of the distributed collaborative computing mechanism is as follows: 1) Node status awareness and resource quantification Each edge node collects its own resource status parameters every 1 second, including: CPU utilization. Memory utilization Network uplink bandwidth Network downlink bandwidth Remaining storage space ; Node Comprehensive Load Index , , , These are the weighting coefficients, with a sum of 1; initially set to 0.4, 0.3, 0.15, and 0.15 respectively. This represents the theoretical maximum bandwidth of the node network. When the overall load index L of a node changes by more than 5%, the latest resource status information is broadcast to all adjacent edge nodes through a distributed state synchronization mechanism based on the gossip protocol. Each edge node maintains a resource status table of adjacent nodes, which records the IP address, overall load index and individual resource parameters of all adjacent nodes, and updates it regularly. 2) Task unloading based on value priority When the CPU utilization of the local node U cpu ≥80% or memory utilization U mem When the percentage is ≥70%, the task unloading decision process is triggered; The tasks to be processed are divided into four priority queues according to the data value density score V: P0 (V≥0.8), P1 (0.6≤V<0.8), P2 (0.3≤V<0.6), and P3 (V<0.3). Execute the following task to uninstall priority rules: P0 and P1 level tasks must be processed on the local node and must not be uninstalled. P2 level tasks are preferentially unloaded to adjacent free nodes, and are only processed locally if all adjacent nodes have no free resources. P3 level tasks can be directly offloaded to the cloud, or when local storage resources are insufficient, only statistical features can be retained and the original data can be discarded. For P2 level tasks that need to be unloaded, select the node with the lowest overall load index L < 50% and the lowest network latency from the adjacent node resource status table as the target unloading node. When the CPU utilization of the local node U cpu <50% and memory utilization U mem When the percentage is less than 40%, it automatically receives P2-level tasks that have been unloaded from other adjacent nodes.

[0023] Dynamically adjust the sampling frequency and storage period of different types of data to reduce the storage pressure on edge nodes while ensuring data integrity; S1.2: Edge node self-awareness and data credibility quantification verification. Construct a correlation mapping model between node working status and collected data quality. Learn the characteristic patterns of data distortion under different working states through machine learning algorithms and generate a node credibility coefficient between 0 and 1. The correlation mapping model between node working status and collected data quality adopts a multi-input, multi-output hierarchical fusion architecture. It takes the multi-dimensional working status parameters of edge nodes and the multi-dimensional quality characteristics of collected data as inputs, and outputs the node reliability coefficient between 0 and 1, while also outputting sub-scores for data quality in each dimension. The model structure includes: The model consists of four layers: a multi-source data input layer, which receives two types of heterogeneous input data simultaneously with a sampling period of 10 seconds; a node state feature encoding layer, which uses a stacked autoencoder (SAE) to reduce the dimensionality and encode the features of high-dimensional, nonlinear node operating state parameters, extracting low-dimensional implicit features that characterize the health of nodes; a data quality feature extraction layer, which uses a bidirectional gated recurrent unit (Bi-GRU) to extract the quality features of temporally acquired data, capturing the dynamic patterns of data quality changes over time; a cross-modal attention association layer, the core of the model, which uses a multi-head self-attention mechanism to establish a nonlinear association mapping between node operating states and data quality, dynamically learning the weights of different operating states on data quality; and a credibility quantification output layer, which uses a fully connected neural network to process the associated feature vectors, ultimately outputting a node credibility coefficient between 0 and 1. A multi-source data fusion algorithm based on credibility weighting is proposed, which automatically reduces the weight of low-credibility node data in the fusion process and completely blocks node data with credibility below a set threshold. The above multi-source data fusion algorithm process is as follows: 1) Preprocess heterogeneous data from different edge nodes Timestamp alignment: Linear interpolation is used to align the sampled data of all nodes to a unified time base, ensuring that the data within the same sliding window have the same timestamp; for missing sample points, the weighted average of adjacent nodes is used for temporary filling. Data standardization: The Z-score standardization method is used to map sensor data of different dimensions to a standard normal distribution; Outlier filtering: The 3σ principle is used to initially filter the standardized data, and data that exceeds the range of [−3,3] are marked as preliminary outliers and given lower weights in the subsequent fusion process; 2) Dynamic fusion weight calculation Basic credibility weight ,in Let be the credibility coefficient of the i-th node, taking a value of [0,1]; T is the credibility threshold, taking a value of 0.5; when At that time, the data of that node is completely masked and does not participate in the fusion calculation; Introducing a time decay factor, , This is the time decay coefficient, with a default value of 0.01; Let be the difference between the last update time and the current time of the credibility coefficient of the i-th node; Weight adjustment: Adjusting weights based on data quality. ,in Let i be the data missing rate within the current window of the i-th node. This represents the initial proportion of outliers; Weight normalization: The final weights of all nodes are normalized to ensure that the sum of the weights is 1. Where N is the total number of nodes participating in the fusion, if all nodes If all values ​​are 0, a system alarm will be triggered, and the most recent valid fusion result or historical data from the cloud will be used as the temporary output. 3) Weighted evidence fusion calculation ,in Let be the measured value of the i-th node; The fusion results are transformed from the normalized space back to the original dimensional space to obtain the final fused measurement value: ,in and represent the standard deviation and mean of the j-th sensor under normal operating conditions, respectively.

[0024] It enables self-diagnosis and self-isolation of edge node faults. When a serious fault is detected in a node, it automatically isolates it from the network and notifies the operation and maintenance personnel to carry out repairs. S1.3: Spatiotemporal alignment of multimodal data, based on a dual spatiotemporal alignment algorithm of cross-correlation function and physical constraints. First, the initial time delay between different sensor data is calculated by cross-correlation function, and then the initial result is corrected by using the theoretical time delay derived from the multiphysics coupling model. The above multiphysics coupling model includes sub-models in the following dimensions: The elastic wave propagation model is used to calculate the propagation delay of vibration signals from the fault source to the vibration sensor. It is suitable for monitoring mechanical faults such as transformer winding deformation and core loosening. A three-dimensional elastic wave longitudinal wave propagation model is adopted, considering the wave impedance differences of different media including transformer oil, iron core, windings, and tank walls: ,in This is the theoretical time delay for the vibration signal to propagate to the sensor when the fault source is located at the (x,y,z) coordinates; n is the number of media layers that the vibration wave passes through along its propagation path; Let be the distance the vibration wave travels in the k-th layer of medium; Let be the longitudinal wave propagation velocity of the vibration wave in the k-th layer of medium; The heat conduction sub-model is used to calculate the propagation delay of temperature signals from the fault source to the temperature sensor, and is suitable for monitoring thermal faults such as winding overheating and core overheating. An unsteady-state heat conduction equation is adopted, considering both convective heat transfer in transformer oil and heat conduction in the solid medium: , For the density of the medium, Let T be the specific heat capacity of the medium at constant pressure, t be the temperature of the medium, t be the time, k be the thermal conductivity of the medium, and q be the intensity of the internal heat source. By solving the above partial differential equations, the temperature field distribution over time is obtained. When the heat generated by the fault source propagates to the temperature sensor location, causing the sensor reading to rise above a preset threshold (usually 0.1℃), the corresponding time is the theoretical heat conduction delay τ. temp ; A partial discharge propagation sub-model is used to calculate the propagation delay of electromagnetic waves and ultrasonic signals generated by partial discharge from the discharge source to the sensor, and is suitable for monitoring partial discharge faults. Electromagnetic wave propagation sub-model: Electromagnetic waves generated by partial discharge propagate at the speed of light inside the transformer with extremely short time delay. , For the propagation delay of electromagnetic waves, This refers to the straight-line distance from the power supply to the ultra-high frequency sensor. The speed of light in a vacuum. The relative permittivity of the propagation medium (εr≈2.2 for transformer oil); Ultrasonic propagation sub-model: Ultrasonic waves generated by partial discharge propagate in transformer oil as longitudinal waves. , For the propagation delay of ultrasound, The speed at which ultrasound propagates in transformer oil; An SF6 gas density change propagation sub-model is used to calculate the propagation delay of density changes caused by gas leakage or decomposition from the fault source to the SF6 density sensor in SF6 gas-insulated equipment. The Navier-Stokes equations for compressible fluids are used to describe the flow and density changes of SF6 gas: ,in The density of SF6 gas. For gas pressure, For gas dynamic viscosity, It is the acceleration due to gravity; By solving the above equations, the distribution of the SF6 gas density field over time is obtained; when the gas density change at the sensor location exceeds a preset threshold, the corresponding time is the theoretical SF6 gas propagation delay τ. sf6 .

[0025] The entire lifecycle of the equipment involved includes comprehensive information on the integrated equipment from design to retirement, covering the design and manufacturing stage, installation and commissioning stage, operation stage, maintenance stage, and retirement and scrapping stage. We designed a feature alignment network that combines dynamic temporal warping (DTW) with an attention mechanism to further eliminate local temporal biases between data. Achieve microsecond-level precise alignment of multimodal data, including at least vibration, temperature, partial discharge, and SF6 gas density, to provide a high-quality data foundation for subsequent fusion analysis; The full lifecycle data management and evaluation layer divides the entire lifecycle of power transmission and transformation equipment into multiple specific stages and performs dynamic evaluation of the equipment; the implementation process is as follows: S2.1: Construct a digital archive of the entire life cycle of power transmission and transformation equipment, and establish a unified data standard for the entire life cycle of power transmission and transformation equipment; use blockchain technology to store the digital archives to ensure the immutability and traceability of the data; S2.2: Multi-stage dynamic status assessment covering the entire life cycle, with a dedicated assessment indicator system and weight allocation scheme designed for each stage; The evaluation index system adopts a general basic index library, including 6 categories and 32 core indicators: electrical performance, oil chromatography and oil quality, partial discharge, mechanical vibration, temperature and environment, and operation and maintenance. The indicator system is set as follows: In the initial stage of operation, there are 18 indicators, with core indicators accounting for 85% and auxiliary indicators accounting for 15%. During the stable operation period, there are 24 indicators, with core indicators accounting for 70%, trend indicators accounting for 20%, and auxiliary indicators accounting for 10%. During the performance degradation period, there are 26 indicators, with core indicators accounting for 55%, trend indicators accounting for 35%, and auxiliary indicators accounting for 10%. At the end of life, there are 22 indicators, with core indicators accounting for 75%, trend indicators accounting for 20%, and auxiliary indicators accounting for 5%. The weight allocation scheme is designed as follows: Basic weighting: Calculated by combining 60% expert subjective weighting and 40% data objective weighting; Phase adjustment: The weight of core indicators is multiplied by 1.5, the weight of auxiliary indicators remains unchanged, and the weight of non-key indicators is multiplied by 0.5; Abnormal increase: When the indicator value exceeds the attention threshold, the weight is multiplied by 2; when it exceeds the abnormal threshold, the weight is multiplied by 3. All indicator weights are ultimately normalized, and the sum is 1. Based on the dynamic assessment model of stage transition probability, the system automatically identifies the life cycle stage of the equipment by analyzing the equipment's operating data, maintenance records, and environmental data, and switches the corresponding assessment strategy in real time. Taking into account the equipment's current status, historical status change trends, and cumulative damage throughout its entire life cycle, the system provides a comprehensive health index for the equipment and classifies the equipment status into four levels: normal, attention, abnormal, and severe. The construction and implementation process of the dynamic evaluation model for stage transition probability is as follows: The multi-source data input layer accepts the following four types of data: Time-series status data: comprehensive health index, scores of each individual indicator, and rate of change of health index for three consecutive months; Event-based data: maintenance records (routine / major / emergency maintenance), fault records (general / serious faults), and technical modification records; Static attribute data: equipment model, design life, commissioning time, and cumulative operating time; Environmental and load data: average ambient temperature, number of days with extreme temperatures, cumulative overload duration, and average load rate; The dynamic transition probability calculation layer constructs a 4×4 initial basic transition matrix P0 based on historical statistical data of similar equipment throughout their entire lifecycle, representing the baseline probability of a device transitioning from stage i to stage j: Four correction coefficients that can be calculated in real time are introduced to adjust the base transition probability element by element: Stage Dwell Time Correction Factor: The longer the equipment stays in the current stage, the higher the probability of transitioning to a subsequent stage; for example, after a stable operating period of more than 15 years, the probability of transitioning to the deterioration stage increases by 20% for each additional year. Deterioration rate correction factor: When the monthly decline rate of the health index exceeds 0.02, the transition probability is multiplied by 1.5; when it exceeds 0.05, it is multiplied by 3. Event impact correction factor: After a serious failure, the probability of transitioning to the next stage is multiplied by 5; after a major overhaul is completed, the probability of transitioning back is multiplied by 0.2 (only a limited backtracking to the adjacent previous stage is allowed). Environmental load correction factor: If the annual average load rate exceeds 80% or the number of days with extreme temperatures exceeds 30, the migration probability is multiplied by 1.3; For each element of the fundamental transition matrix Multiply by the product of the corresponding correction coefficients to obtain the real-time transfer probability. Then, normalize the values ​​to ensure that the sum of probabilities for each row is 1; The stage identification and inference layer performs probabilistic inference based on the Hidden Markov Model (HMM) framework: Hidden states: 4 full lifecycle stages; Observations: current health index, rate of change of health index, and the largest outlier of a single indicator in the last 3 months; Inference algorithm: the forward algorithm is used to calculate the posterior probability of the device being in each stage at the current moment; Stage determination rules: When the probability of a certain stage is ≥0.8, the device is determined to enter that stage; when the probabilities of two adjacent stages are both between 0.5 and 0.8, it is determined to be a transition stage, and a two-stage evaluation strategy is used for weighted averaging; when the probabilities of all stages are <0.5, manual review is triggered. In the strategy switching layer, once the phase determination is completed, the corresponding phase's exclusive indicator system and weight allocation scheme are immediately loaded; in the transition phase, the evaluation strategies of the two phases are mixed according to probability weights; when switching phases, the historical evaluation data of the previous 3 months are retained for smooth transition.

[0026] S2.3: Equipment remaining life prediction, introducing rainflow counting method and Miner's rule to calculate the cumulative fatigue damage of each key component of the equipment under alternating load; Design a remaining lifetime prediction network based on LSTM and attention mechanism to automatically learn the influence weights of different environmental factors such as temperature, humidity, salt spray, and lightning strike on the equipment damage rate. The LSTM and attention mechanism residual lifetime prediction network includes: a data input layer for cleaning, normalizing, aligning, and sliding window segmenting the monitoring data of the device, including at least oil chromatography data, vibration signals, partial discharge, temperature, and load current; LSTM layer: As the foundation of the network, it is used to automatically learn the temporal dependencies in the process of device degradation. Through its gating mechanism, LSTM can effectively capture long-term and short-term dependency patterns in monitoring data, avoiding the limitations of traditional methods that require manual feature extraction, and directly extracting deep state features from the raw or pre-processed temporal data. Attention Mechanism Layer: Connected after the LSTM layer; this mechanism dynamically evaluates the features of all time steps of the LSTM output and assigns different weights to the features of different time steps; its core function is to enable the model to automatically focus on more indicative key time points or features in the device degradation process, suppress the interference of noise and non-critical information, and thus extract the core information representing the health status of the device more accurately.

[0027] To achieve dynamic correction and early warning of the remaining lifespan of equipment under extreme environments, when it is predicted that the remaining lifespan of the equipment is lower than the set threshold, an early warning signal is issued in advance to remind maintenance personnel to arrange maintenance or replacement. The intelligent early warning and decision support layer provides early warnings and generates corresponding decision support based on the assessment results of power transmission and transformation equipment; the implementation process is as follows: S3.1: Early fault warning, based on the dual warning indicators of feature spatial distance and abnormal mode evolution speed, adopts an abnormal mode evolution trajectory tracking algorithm, and predicts the abnormal development trend and possible fault types by continuously monitoring changes in abnormal features; and issues different levels of warning signals according to the severity and development speed of the abnormality. S3.2: Fault source tracing and impact propagation analysis. A fault source tracing algorithm based on causal graph and Bayesian network is adopted. Starting from the observed abnormal phenomena, the root cause of the fault is deduced in reverse and the probability of each possible cause is calculated. The fault source tracing algorithm using cause-effect graphs and Bayesian networks is as follows: 1) Data collection and preprocessing: Integrate multi-source heterogeneous data from the equipment, including online monitoring data such as oil chromatography, partial discharge, and vibration signals, inspection records, operation logs, and historical fault reports; clean, denoise, align, and extract features from the data to provide high-quality input for modeling; 2) Causal Model Construction and Validation: Generating causal structure hypotheses based on prior knowledge and data-driven methods; Statistical tests or algorithms are used to verify the existence and direction of causal relationships between variables; Complete the learning of Bayesian network parameters and construct a complete probability model; The accuracy and reliability of the model are evaluated using methods such as cross-validation and backtesting with historical cases. 3) Root cause analysis and application: Input real-time or post-event abnormal data into the validated model for probabilistic reasoning; the algorithm can output the most likely sequence or combination of fault root causes and provide quantitative basis for maintenance decisions, such as distinguishing whether transformer oil chromatographic abnormalities are caused by partial discharge or overheating faults.

[0028] Based on the power grid topology and electrical connections between devices, predict other devices and power grid areas that may be affected by a fault; Generate personalized fault handling suggestions and emergency plans, including fault isolation steps, maintenance personnel allocation, and spare parts preparation, to assist maintenance personnel in handling faults quickly and accurately. The model adaptive evolution and knowledge sharing layer is used to acquire high-value samples and actively optimize the model using these samples; the process is as follows: S4.1: Edge model adaptive evolution, based on active learning and incremental learning model adaptive evolution. When concept drift is detected, the most valuable sample is automatically selected for labeling, and the model parameters are updated using incremental learning methods. The adaptive evolution process of the model is as follows: 1) The system detected concept drift and triggered the model evolution process; 2) From the new data of the most recent month, 100 high-value samples are selected through a three-level sampling strategy: uncertainty sampling, representative sampling, and rare sampling. The uncertainty sampling prioritizes the samples with the most uncertain model prediction results. The representative sampling selects samples from the candidate set that can represent the distribution characteristics of the new data. The rare sampling focuses on capturing novel failure modes that have never appeared before. 3) The samples are automatically pre-labeled and pushed to the cloud for expert review and confirmation; 4) After the samples are labeled, they are synchronized back to the edge nodes to construct an incremental training set; 5) The new model is trained using an incremental learning method that combines hierarchical updates with knowledge distillation; 6) After the new model passes performance testing on the historical validation set, it enters the dual-model parallel operation phase; 7) After the new model has been running stably for 72 consecutive hours and its performance is better than that of the old model, it will be officially switched to the production model.

[0029] During the model update process, the old and new versions of the model run simultaneously. If a problem occurs in the new version of the model, it can be rolled back to the old version immediately to ensure the continuity of system operation. S4.2: Cross-device fault knowledge transfer based on multimodal large model, which transfers fault knowledge of existing devices to new and rare devices to achieve rapid cold start of edge models; Construct a knowledge graph of equipment faults to realize the structured representation and sharing of fault knowledge, enabling equipment from different regions and manufacturers to share fault diagnosis experience.

Claims

1. A method for full life-cycle status assessment and early warning of power transmission and transformation equipment, characterized in that, It is implemented based on a four-layer structure, namely: edge perception and data preprocessing layer, full life cycle data management and evaluation layer, intelligent early warning and decision support layer, and model adaptive evolution and knowledge sharing layer. The edge sensing and data preprocessing layer completes its hierarchical functions using the following methods: S1.1: Edge node adaptive scheduling, based on sliding window data value density assessment, analyzes the rate of change, degree of anomaly and similarity with historical fault data, automatically identifies early fault symptom data and assigns it the highest processing priority; By introducing a distributed collaborative computing mechanism among edge nodes, when the computing power of a single node is insufficient, non-critical tasks are automatically offloaded to adjacent idle nodes to ensure that high-value data is processed first. Dynamically adjust the sampling frequency and storage period of different types of data to reduce the storage pressure on edge nodes while ensuring data integrity; S1.2: Edge node self-awareness and data credibility quantification verification. Construct a correlation mapping model between node working status and collected data quality. Learn the characteristic patterns of data distortion under different working states through machine learning algorithms and generate a node credibility coefficient between 0 and 1. A multi-source data fusion algorithm based on credibility weighting is proposed, which automatically reduces the weight of low-credibility node data in the fusion process and completely blocks node data with credibility below a set threshold. It enables self-diagnosis and self-isolation of edge node faults. When a serious fault is detected in a node, it automatically isolates it from the network and notifies the operation and maintenance personnel to carry out repairs. S1.3: Spatiotemporal alignment of multimodal data, based on a dual spatiotemporal alignment algorithm of cross-correlation function and physical constraints. First, the initial time delay between different sensor data is calculated by cross-correlation function, and then the initial result is corrected by using the theoretical time delay derived from the multiphysics coupling model. Design a feature alignment network that combines DTW with an attention mechanism to further eliminate local temporal biases between data; Achieve microsecond-level precise alignment of multimodal data; The full lifecycle data management and evaluation layer completes its hierarchical functions using the following methods: S2.1: Construct a digital archive of the entire life cycle of power transmission and transformation equipment, and establish a unified data standard for the entire life cycle of power transmission and transformation equipment; use blockchain technology to store the digital archives to ensure the immutability and traceability of the data; S2.2: Multi-stage dynamic status assessment covering the entire life cycle, with a dedicated assessment index system and weight allocation scheme designed for each stage; based on the stage transition probability dynamic assessment model, it automatically identifies the life cycle stage of the equipment and switches the corresponding assessment strategy in real time; taking into account the current status of the equipment, historical status change trends and cumulative damage throughout the entire life cycle, it provides a comprehensive health index for the equipment and classifies the equipment status into four levels: normal, attention, abnormal and severe. S2.3: Equipment remaining life prediction, introducing rainflow counting method and Miner's rule to calculate the cumulative fatigue damage of each key component of the equipment under alternating load; Design a remaining lifetime prediction network based on LSTM and attention mechanism to automatically learn the influence weights of different environmental factors on the device damage rate; To achieve dynamic correction and early warning of the remaining lifespan of equipment under extreme environments, when it is predicted that the remaining lifespan of the equipment is lower than the set threshold, an early warning signal is issued in advance to remind maintenance personnel to arrange maintenance or replacement. The intelligent early warning and decision support layer completes its hierarchical functions using the following methods: S3.1: Early fault warning, based on the dual warning indicators of feature spatial distance and abnormal mode evolution speed, adopts an abnormal mode evolution trajectory tracking algorithm, and predicts the abnormal development trend and possible fault types by continuously monitoring changes in abnormal features; and issues different levels of warning signals according to the severity and development speed of the abnormality. S3.2: Fault source tracing and impact propagation analysis. A fault source tracing algorithm based on causal graph and Bayesian network is adopted. Starting from the observed abnormal phenomena, the root cause of the fault is deduced in reverse and the probability of each possible cause is calculated. Based on the power grid topology and electrical connections between devices, predict other devices and power grid areas that may be affected by a fault; Generate personalized fault handling suggestions and emergency plans; assist maintenance personnel in quickly and accurately handling faults; The model adaptive evolution and knowledge sharing layer completes its hierarchical functions using the following methods: S4.1: Edge model adaptive evolution, based on active learning and incremental learning model adaptive evolution. When concept drift is detected, the most valuable sample is automatically selected for labeling, and the model parameters are updated using incremental learning methods. During the model update process, the old and new versions of the model run simultaneously. If a problem occurs in the new version of the model, it can be rolled back to the old version immediately to ensure the continuity of system operation. S4.2: Cross-device fault knowledge transfer based on multimodal large model, which transfers fault knowledge of existing devices to new and rare devices to achieve rapid cold start of edge models; Construct a knowledge graph of equipment faults to realize the structured representation and sharing of fault knowledge, enabling equipment from different regions and manufacturers to share fault diagnosis experience.

2. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The time sliding window in the data value density assessment is set as follows: Sliding step size Counting sliding window Sliding step size ; When the rate of change of data is detected to exceed the threshold, the window size is automatically reduced to half of its original size, and the step size is reduced to one-third of its original size; when the data in three consecutive windows is stable, the default parameters are restored. The method for evaluating the rate of change of data is as follows: 1) First-order difference rate of change First-order difference , For the sensor measurement value at the i-th sampling point, This represents the sensor measurement value at the (i-1)th sampling point; First-order difference rate of change W represents the total number of sampling points contained in the current sliding window; 2) Second-order difference acceleration Second-order difference of the i-th sampling point Second-order difference acceleration 3) Volatility Standard deviation of data within the window The arithmetic mean of the data within the window ; In-window data volatility The anomaly assessment process is as follows: Calculate the mean of the window data and standard deviation ; Abnormal scores ,when When this occurs, it is judged as a statistical anomaly; The historical fault data similarity evaluation process is as follows: 1) Build a historical fault template library Collect sensor data from the 30 minutes preceding historical equipment failures; Feature sequences are extracted for each fault type to form a fault template library. ； 2) DTW distance calculation Current window data sequence , Fault Template Construct the distance matrix ,in Calculate the cumulative distance matrix , DTW distance: 3) Similarity score .

3. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The implementation process of the distributed collaborative computing mechanism is as follows: 1) Node status awareness and resource quantification Each edge node collects its own resource status parameters every 1 second, including: CPU utilization. Memory utilization Network uplink bandwidth Network downlink bandwidth Remaining storage space ; Node Comprehensive Load Index , , , These are the weighting coefficients, with a sum of 1; initially set to 0.4, 0.3, 0.15, and 0.15 respectively. This represents the theoretical maximum bandwidth of the node network. When the overall load index L of a node changes by more than 5%, the latest resource status information is broadcast to all adjacent edge nodes through a distributed state synchronization mechanism based on the gossip protocol. Each edge node maintains a resource status table of adjacent nodes, recording the IP address, overall load index and individual resource parameters of all adjacent nodes, and updates it periodically. 2) Task unloading based on value priority When the CPU utilization of the local node U cpu ≥80% or memory utilization U mem When the percentage is ≥70%, the task unloading decision process is triggered; The tasks to be processed are divided into four priority queues according to the data value density score V: P0 (V≥0.8), P1 (0.6≤V<0.8), P2 (0.3≤V<0.6), and P3 (V<0.3). Execute the following task to uninstall priority rules: P0 and P1 level tasks must be processed on the local node and must not be uninstalled. P2 level tasks are preferentially unloaded to adjacent free nodes, and are only processed locally if all adjacent nodes have no free resources. P3 level tasks can be directly offloaded to the cloud, or when local storage resources are insufficient, only statistical features can be retained and the original data can be discarded. For P2 level tasks that need to be unloaded, select the node with the lowest overall load index L < 50% and the lowest network latency from the adjacent node resource status table as the target unloading node. When the CPU utilization of the local node U cpu <50% and memory utilization U mem When the percentage is less than 40%, it automatically receives P2-level tasks that have been unloaded from other adjacent nodes.

4. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The correlation mapping model between node working status and collected data quality adopts a multi-input multi-output hierarchical fusion architecture. It takes the multi-dimensional working status parameters of edge nodes and the multi-dimensional quality characteristics of collected data as inputs, and the node credibility coefficient between 0 and 1 as outputs. At the same time, it outputs the sub-scores of data quality in each dimension. The model structure includes: The system comprises the following layers: a multi-source data input layer, which simultaneously receives two types of heterogeneous input data with a sampling period of 10 seconds; a node state feature encoding layer, which uses a stacked autoencoder to reduce the dimensionality and encode the features of high-dimensional, nonlinear node operating state parameters, extracting low-dimensional implicit features that characterize the health of nodes; a data quality feature extraction layer, which uses a bidirectional gated recurrent unit to extract the quality features of temporally acquired data, capturing the dynamic patterns of data quality changes over time; a cross-modal attention association layer, which uses a multi-head self-attention mechanism to establish a nonlinear association mapping between node operating states and data quality, dynamically learning the influence weights of different operating states on data quality; and a credibility quantification output layer, which uses a fully connected neural network to process the associated feature vectors, ultimately outputting a node credibility coefficient between 0 and 1. The multi-source data fusion algorithm process is as follows: 1) Preprocess heterogeneous data from different edge nodes Timestamp alignment: Linear interpolation is used to align the sampled data of all nodes to a unified time base, ensuring that the data within the same sliding window have the same timestamp; for missing sample points, the weighted average of adjacent nodes is used for temporary filling. Data standardization: The Z-score standardization method is used to map sensor data of different dimensions to a standard normal distribution; Outlier filtering: The 3σ principle is used to initially filter the standardized data, and data that exceeds the range of [−3,3] are marked as preliminary outliers and given lower weights in the subsequent fusion process; 2) Dynamic fusion weight calculation Basic credibility weight ,in Let be the credibility coefficient of the i-th node, taking a value of [0,1]; T is the credibility threshold, taking a value of 0.5; when At that time, the data of that node is completely masked and does not participate in the fusion calculation; Introducing a time decay factor, , This is the time decay coefficient, with a default value of 0.01; Let be the difference between the last update time and the current time of the credibility coefficient of the i-th node; Weight adjustment: Adjusting weights based on data quality. ,in Let i be the data missing rate within the current window of the i-th node. This represents the initial proportion of outliers; Weight normalization: The final weights of all nodes are normalized to ensure that the sum of the weights is 1. Where N is the total number of nodes participating in the fusion, if all nodes If all values ​​are 0, a system alarm will be triggered, and the most recent valid fusion result or historical data from the cloud will be used as the temporary output. 3) Weighted evidence fusion calculation ,in Let be the measured value of the i-th node; The fusion results are transformed from the normalized space back to the original dimensional space to obtain the final fused measurement value: ,in and represent the standard deviation and mean of the j-th sensor under normal operating conditions, respectively.

5. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The multiphysics coupling model includes: The elastic wave propagation model is used to calculate the propagation delay of vibration signals from the fault source to the vibration sensor. It is suitable for monitoring mechanical faults such as transformer winding deformation and core loosening. A three-dimensional elastic wave longitudinal wave propagation model is adopted, considering the wave impedance differences of different media including transformer oil, iron core, windings, and tank walls: ,in This is the theoretical time delay for the vibration signal to propagate to the sensor when the fault source is located at the (x,y,z) coordinates; n is the number of media layers that the vibration wave passes through along its propagation path; Let be the distance the vibration wave travels in the k-th layer of medium; Let be the longitudinal wave propagation velocity of the vibration wave in the k-th layer of medium; The heat conduction sub-model is used to calculate the propagation delay of temperature signals from the fault source to the temperature sensor, and is suitable for monitoring thermal faults such as winding overheating and core overheating. An unsteady-state heat conduction equation is adopted, considering both convective heat transfer in transformer oil and heat conduction in the solid medium: , For the density of the medium, Let T be the specific heat capacity of the medium at constant pressure, t be the temperature of the medium, t be the time, k be the thermal conductivity of the medium, and q be the intensity of the internal heat source. By solving the above partial differential equations, the temperature field distribution over time is obtained. When the heat generated by the fault source propagates to the temperature sensor location, causing the sensor reading to rise above a preset threshold, the corresponding time is the theoretical heat conduction delay τ. temp ; A partial discharge propagation sub-model is used to calculate the propagation delay of electromagnetic waves and ultrasonic signals generated by partial discharge from the discharge source to the sensor, and is suitable for monitoring partial discharge faults. Electromagnetic wave propagation sub-model: Electromagnetic waves generated by partial discharge propagate at the speed of light inside the transformer with extremely short time delay. , For the propagation delay of electromagnetic waves, This refers to the straight-line distance from the power supply to the ultra-high frequency sensor. The speed of light in a vacuum. The relative permittivity of the propagation medium; Ultrasonic propagation sub-model: Ultrasonic waves generated by partial discharge propagate in transformer oil as longitudinal waves. , For the propagation delay of ultrasound, The speed at which ultrasound propagates in transformer oil; An SF6 gas density change propagation sub-model is used to calculate the propagation delay of density changes caused by gas leakage or decomposition from the fault source to the SF6 density sensor in SF6 gas-insulated equipment. The Navier-Stokes equations for compressible fluids are used to describe the flow and density changes of SF6 gas: ,in The density of SF6 gas. For gas pressure, For gas dynamic viscosity, It is the acceleration due to gravity; By solving the above equations, the distribution of the SF6 gas density field over time is obtained; when the gas density change at the sensor location exceeds a preset threshold, the corresponding time is the theoretical SF6 gas propagation delay τ. sf6 .

6. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The entire lifecycle of the equipment integrates information from design to retirement, including: design and manufacturing stage, installation and commissioning stage, operation stage, maintenance stage, and retirement and scrapping stage.

7. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The evaluation index system adopts a general basic index library, including 6 categories and 32 core indicators: electrical performance, oil chromatography and oil quality, partial discharge, mechanical vibration, temperature and environment, and operation and maintenance. In the initial stage of operation, there are 18 indicators, with core indicators accounting for 85% and auxiliary indicators accounting for 15%. During the stable operation period, there are 24 indicators, with core indicators accounting for 70%, trend indicators accounting for 20%, and auxiliary indicators accounting for 10%. During the performance degradation period, there are 26 indicators, with core indicators accounting for 55%, trend indicators accounting for 35%, and auxiliary indicators accounting for 10%. At the end of life, there are 22 indicators, with core indicators accounting for 75%, trend indicators accounting for 20%, and auxiliary indicators accounting for 5%. The weight allocation scheme is designed as follows: Basic weighting: Calculated by combining 60% expert subjective weighting and 40% data objective weighting; Phase adjustment: The weight of core indicators is multiplied by 1.5, the weight of auxiliary indicators remains unchanged, and the weight of non-key indicators is multiplied by 0.5; Abnormal increase: When the indicator value exceeds the attention threshold, the weight is multiplied by 2; when it exceeds the abnormal threshold, the weight is multiplied by 3. All indicator weights are ultimately normalized, and the sum is 1. The dynamic evaluation model for stage transition probability includes: The multi-source data input layer accepts the following four types of data: Time-series status data: comprehensive health index, scores of each individual indicator, and rate of change of health index for three consecutive months; Event-based data: maintenance records, fault records, and technical modification records; Static attribute data: equipment model, design life, commissioning time, and cumulative operating time; Environmental and load data: average ambient temperature, number of days with extreme temperatures, cumulative overload duration, and average load rate; The dynamic transition probability calculation layer constructs a 4×4 initial basic transition matrix P0 based on historical statistical data of the entire life cycle of similar equipment, representing the baseline probability of equipment transitioning from stage i to stage j. Four real-time calculable correction coefficients are introduced to adjust the basic transition probability element by element; Stage dwell time correction factor: The longer the equipment stays in the current stage, the higher the probability of it moving to the next stage; Deterioration rate correction factor: When the monthly decline rate of the health index exceeds 0.02, the transition probability is multiplied by 1.5; when it exceeds 0.05, it is multiplied by 3. Event impact correction factor: After a serious failure, the probability of transitioning to subsequent stages is multiplied by 5; after a major overhaul is completed, the probability of transitioning back is multiplied by 0.

2. Environmental load correction factor: If the annual average load rate exceeds 80% or the number of extreme temperature days exceeds 30 days, the migration probability is multiplied by 1.3; For each element of the fundamental transition matrix Multiply by the product of the corresponding correction coefficients to obtain the real-time transfer probability. Then, normalize the values ​​to ensure that the sum of probabilities for each row is 1; The stage-based identification and inference layer performs probabilistic inference based on a Hidden Markov Model framework. Hidden states: 4 full lifecycle stages; Observations: current health index, rate of change of health index, and the largest outlier of a single indicator in the last 3 months; Inference algorithm: the forward algorithm is used to calculate the posterior probability of the device being in each stage at the current moment; Stage determination rules: When the probability of a certain stage is ≥0.8, the device is determined to enter that stage; when the probabilities of two adjacent stages are both between 0.5 and 0.8, it is determined to be a transition stage, and a two-stage evaluation strategy is used for weighted averaging; when the probabilities of all stages are <0.5, manual review is triggered. In the strategy switching layer, once the phase determination is completed, the corresponding phase's exclusive indicator system and weight allocation scheme are immediately loaded; in the transition phase, the evaluation strategies of the two phases are mixed according to probability weights; when switching phases, the historical evaluation data of the previous 3 months are retained for smooth transition.

8. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The remaining lifetime prediction network of the LSTM and attention mechanism includes: a data input layer, used to clean, normalize, align and slide window segment the monitoring data of the device, including at least oil chromatography data, vibration signals, partial discharge, temperature and load current; LSTM layer: As the foundation of the network, it is used to automatically learn the temporal dependencies in the process of device degradation. Through its gating mechanism, LSTM can effectively capture long-term and short-term dependency patterns in monitoring data, avoiding the limitations of traditional methods that require manual feature extraction, and directly extracting deep state features from the raw or pre-processed temporal data. Attention Mechanism Layer: Connected after the LSTM layer; this mechanism dynamically evaluates the features of all time steps of the LSTM output and assigns different weights to the features of different time steps; its core function is to enable the model to automatically focus on more indicative key time points or features in the device degradation process, suppress the interference of noise and non-critical information, and thus extract the core information representing the health status of the device more accurately.

9. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The fault source tracing algorithm process using cause-effect graphs and Bayesian networks is as follows: 1) Data collection and preprocessing: Integrate multi-source heterogeneous data from equipment, including online monitoring data, inspection records, operation logs, and historical fault reports; clean, denoise, align, and extract features from the data to provide high-quality input for modeling; 2) Causal Model Construction and Validation: Generating causal structure hypotheses based on prior knowledge and data-driven methods; Statistical tests or algorithms are used to verify the existence and direction of causal relationships between variables; Complete the learning of Bayesian network parameters and construct a complete probability model; Use cross-validation or historical case backtesting methods to evaluate the accuracy and reliability of the model; 3) Root cause analysis and application: Input real-time or post-event abnormal data into the validated model for probabilistic reasoning; the algorithm can output the most likely sequence or combination of fault root causes and provide quantitative basis for maintenance decisions.

10. The method for full life-cycle status assessment and early warning of power transmission and transformation equipment according to claim 1, characterized in that, The adaptive evolution process of the model is as follows: 1) The system detected concept drift and triggered the model evolution process; 2) From the new data of the most recent month, 100 high-value samples are selected through a three-level sampling strategy of uncertainty sampling, representativeness sampling, and rarity sampling; uncertainty sampling prioritizes the samples with the most uncertain model prediction results; representativeness sampling selects samples from the candidate set that can represent the distribution characteristics of the new data. The rarity sampling focuses on capturing novel failure modes that have never appeared before. 3) The samples are automatically pre-labeled and pushed to the cloud for expert review and confirmation; 4) After the samples are labeled, they are synchronized back to the edge nodes to construct an incremental training set; 5) The new model is trained using an incremental learning method that combines hierarchical updates with knowledge distillation; 6) After the new model passes performance testing on the historical validation set, it enters the dual-model parallel operation phase; 7) After the new model has been running stably for 72 consecutive hours and its performance is better than that of the old model, it will be officially switched to the production model.

Images