A transformer heat failure early warning method based on edge computing

Abstract

The present application relates to the technical field of transformer fault early warning, and discloses a transformer heating fault early warning method based on edge computing. The method comprises collecting multi-source temperature signals during transformer operation, decomposing the signals into regional temperature signals of different heating areas; then constructing a dynamic thermodynamic model of the transformer heating process based on the regional temperature signals, linearizing the model to establish a linear thermal dynamic model, decomposing the model into a longitudinal heat conduction equation set and a transverse heat diffusion equation set, combining the longitudinal heat conduction equation set and the regional temperature signals, and expressing the dynamic thermodynamic model as a linear heating system; finally, setting an adaptive early warning rule containing a conventional early warning rule and a thermal disturbance compensation rule (the thermal disturbance compensation rule covers time-varying and non-time-varying thermal disturbances) for the linear heating system, and executing the rule during transformer operation, which can comprehensively capture the regional temperature state, accurately depict the thermal dynamic characteristics, and compensate the influence of thermal disturbance.

Description

Technical Field

[0001] This invention relates to the field of transformer fault early warning technology, specifically to a transformer overheating fault early warning method based on edge computing. Background Technology

[0002] A transformer is a device that uses the principle of electromagnetic induction to change alternating current voltage. Its main components are the primary coil, secondary coil, and iron core (magnetic core). A transformer is a device that uses electromagnetic mutual induction to transform voltage, current, and impedance. The components of a transformer include the transformer body (iron core, windings, insulation, leads), transformer oil, oil tank and cooling device, voltage regulating device, protection device (dehumidifier, safety vent, gas relay, oil conservator and temperature measuring device, etc.) and outgoing bushings. Among them, the iron core is the main magnetic circuit part of the transformer, and the windings are the electrical circuit part of the transformer, which are made of double-insulated flat wire or enameled round wire.

[0003] As a core piece of equipment in the power system, the transformer's operating status directly affects the stability and security of the entire power network. During long-term operation, transformers are prone to localized overheating due to factors such as load changes, aging internal components, and insulation damage. If this is not detected and addressed in time, it can gradually develop into a serious fault, leading to equipment shutdown or even large-scale power outages, causing significant economic losses and social impact. Therefore, effective early warning of transformer overheating faults has become an important research direction in the field of power equipment operation and maintenance. Currently, early warning methods for transformer overheating faults mainly rely on single-point or multi-point temperature data collected by temperature sensors, and determine the existence of fault risk by setting a fixed threshold. However, this type of method has obvious limitations: the internal structure of a transformer is complex, and the heating characteristics of different components (such as windings, core, and tank) vary significantly. Single-point or multi-point temperature data cannot fully reflect the true temperature status of each heating area, and misjudgments or omissions are prone to occur. During transformer operation, it is affected by a variety of time-varying and non-time-varying thermal disturbances, such as changes in ambient temperature, load fluctuations, and changes in cooling system efficiency. Fixed threshold early warning methods cannot effectively compensate for these disturbances, resulting in the early warning accuracy being greatly affected by the external environment. With the application of edge computing technology in industrial equipment monitoring, some studies have attempted to combine edge computing with temperature monitoring to achieve real-time processing of transformer temperature data. However, most existing edge computing-based early warning methods remain at the level of data transmission and simple threshold comparison, failing to construct specialized thermodynamic models for the dynamic characteristics of transformer heating processes. This makes it impossible to fundamentally reveal the inherent laws governing temperature changes in different heating regions, and also hinders adaptive adjustments to thermal disturbances. Furthermore, existing methods lack a systematic approach in the regional temperature signal decomposition stage, failing to accurately distinguish the temperature contributions of different heating regions, further affecting the reliability of early warning results. These problems mean that current transformer overheating fault early warning methods are insufficient in terms of practicality, accuracy, and adaptability to meet the high requirements of power systems for equipment operation and maintenance. Therefore, an early warning technology capable of accurate analysis of multi-source temperature signals, dynamic thermal characteristic modeling, and adaptive disturbance compensation is needed. Summary of the Invention

[0004] The purpose of this invention is to provide a transformer overheating fault early warning method based on edge computing to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a transformer overheating fault early warning method based on edge computing, the method comprising: Collect multi-source temperature signals during transformer operation and decompose the multi-source temperature signals into regional temperature signals of different heating areas; Based on the temperature signal of the region, a dynamic thermodynamic model of the transformer heating process is constructed. The dynamic thermodynamic model is linearized to establish a linear thermal dynamic model. The linear thermal dynamic model is decomposed into a longitudinal heat conduction equation set and a transverse heat diffusion equation set. Combining the longitudinal heat conduction equation set and the temperature signal of the region, the dynamic thermodynamic model is expressed as a linear heating system. For the linear heating system, an adaptive early warning rule for transformer overheating faults is set. The adaptive early warning rule includes a conventional early warning rule and a thermal disturbance compensation rule. The thermal disturbance compensation rule includes compensation for time-varying thermal disturbances and time-invariant thermal disturbances. The adaptive early warning rule is executed during transformer operation.

[0006] Preferably, the multi-source temperature signal is acquired in the following way: Temperature data of the transformer core, windings and radiator surface are collected at fixed time intervals using a group of temperature sensors in the edge computing node network. The collected temperature data is uploaded to the edge computing gateway for signal preprocessing to generate a standardized temperature signal stream.

[0007] Preferably, the decomposition process of the regional temperature signal includes: The edge computing gateway receives the standardized temperature signal stream and divides the signal into core temperature zone signal, winding temperature zone signal and surface temperature zone signal according to the physical structural characteristics of the transformer. Each temperature zone signal is processed independently using wavelet transform to extract the characteristic temperature fluctuation patterns of each zone.

[0008] Preferably, the process of establishing the linear thermal dynamics model includes: Based on the characteristic temperature fluctuation pattern, determine the thermal coupling coefficient and thermal conduction parameters for each temperature range; Based on the aforementioned thermal coupling coefficient and thermal conduction parameters, a multi-temperature zone coordinated thermal state equation set is constructed. The frequency domain heat transfer function matrix is ​​obtained by performing a Laplace transform on the thermal state equations.

[0009] Preferably, the adaptive early warning rule is generated in the following way: The edge computing gateway periodically retrieves a historical fault temperature pattern library from the cloud server; The current characteristic temperature fluctuation pattern is matched and calculated with the historical fault temperature pattern library; The threshold parameters of the conventional early warning rule and the compensation intensity of the thermal disturbance compensation rule are dynamically adjusted based on the matching results.

[0010] Preferably, the method further includes an early warning verification step: When the adaptive warning rule is triggered, the edge computing gateway broadcasts a warning verification request to adjacent edge computing nodes; Receive verification temperature data returned by adjacent edge computing nodes and perform multi-node temperature data consistency verification; The warning information will only be uploaded to the cloud monitoring platform when the verification is successful.

[0011] Preferably, the specific process of the multi-node temperature data consistency verification includes: Calculate the Euclidean distance and correlation coefficient between the temperature data of this node and the temperature data of neighboring nodes; The verification is deemed successful when the Euclidean distance is less than a set threshold and the correlation coefficient is greater than a set threshold. Otherwise, start the re-acquisition program to reacquire the verification temperature data.

[0012] Preferably, the method further includes an early warning optimization step: The cloud-based monitoring platform receives early warning information uploaded by multiple edge computing gateways; Spatiotemporal correlation analysis is performed on the aforementioned early warning information to generate a regional thermal fault risk map; The regional thermal failure risk map is distributed to each edge computing gateway to optimize the adaptive early warning rules.

[0013] Preferably, the process of generating the regional thermal fault risk map includes: Extract temperature anomaly patterns, occurrence times, and geographical location information from early warning information; Density clustering algorithm is used to aggregate and analyze early warning information with similar characteristics; Based on the aggregation results, a distribution map of thermal failure risk levels and a time evolution trend map were drawn.

[0014] Preferably, the method further includes an early warning feedback step: After receiving the regional thermal fault risk map, the edge computing gateway adjusts the local temperature sampling frequency and signal processing parameters. The adjusted operational status data is fed back to the cloud monitoring platform to form a closed-loop optimization system.

[0015] Compared with the prior art, the beneficial effects of the present invention are: At the temperature information processing level, this method collects multi-source temperature signals from the transformer during operation and decomposes them into regional temperature signals for different heating areas. This comprehensively covers the temperature status of key heating components inside the transformer (such as windings, core, and tank), avoiding misjudgments and omissions caused by incomplete information in traditional single-point or multi-point temperature monitoring. Compared to the limitations of existing methods that rely solely on local temperature data, this method, through regional temperature signal decomposition, can accurately capture the temperature change differences in each heating area, providing more comprehensive basic information for subsequent fault location and risk assessment. This allows the early warning results to more accurately reflect the actual heating situation of various parts of the transformer, improving the efficiency and accuracy of temperature information utilization. At the thermal characteristic modeling level, this method constructs a dynamic thermodynamic model of the transformer heating process based on regional temperature signals, and establishes a linear thermodynamic model through linearization. This model is further decomposed into longitudinal heat conduction equations and transverse heat diffusion equations, ultimately representing the dynamic thermodynamic model as a linear heating system. This modeling approach reveals the dynamic laws of temperature changes in each heating region of the transformer from a physical perspective, clearly depicting the conduction and diffusion characteristics of heat between different regions. Compared to existing methods that lack dynamic thermal characteristic analysis, this method more accurately reflects the temporal evolution characteristics of the transformer heating process, providing theoretical support based on thermophysical mechanisms for fault early warning. Simultaneously, the construction of a linear heating system reduces the computational complexity of the model. Combined with the real-time data processing capabilities of edge computing, it enables rapid analysis of temperature change patterns, meeting the requirements of transformer operation and maintenance for timely early warning. At the level of early warning rule design, this method sets adaptive early warning rules for linear heating systems, including conventional early warning rules and thermal disturbance compensation rules. The thermal disturbance compensation rules can simultaneously cover time-varying thermal disturbances (such as ambient temperature fluctuations and load changes) and non-time-varying thermal disturbances (such as inherent losses in the cooling system and changes in fixed thermal resistance due to component aging). This design effectively counteracts the interference of external disturbances on the early warning results, avoiding the problem of decreased early warning accuracy caused by the inability of traditional fixed-threshold early warning methods to adapt to disturbances. During actual transformer operation, regardless of whether it is affected by time-varying factors such as sudden changes in ambient temperature and temporary load adjustments, or non-time-varying factors such as cooling system efficiency degradation, this method can adjust the early warning judgment criteria in real time through the thermal disturbance compensation rules, ensuring that the early warning results always maintain high stability and accuracy, significantly improving the adaptability of the early warning method to complex operating environments. This method deeply integrates edge computing technology with dynamic thermal modeling and adaptive rules, enabling rapid processing, model calculation, and early warning judgment of multi-source temperature signals locally on the transformer. It eliminates the need for remote cloud data transmission and processing, reducing data transmission latency and network dependence, making it more suitable for power system equipment maintenance scenarios with high real-time requirements. Furthermore, the entire solution requires no additional complex hardware and can be upgraded based on existing temperature monitoring systems and edge computing nodes, reducing the cost and difficulty of implementation. This facilitates its widespread application in transformer maintenance across various voltage levels, providing stronger protection for the safe and stable operation of power system equipment. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the working principle of the transformer overheating fault early warning method based on edge computing described in this invention.

[0017] Figure 2 A flowchart for establishing a linear thermal dynamics model.

[0018] Figure 3 A flowchart for generating adaptive early warning rules. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1This invention provides a transformer overheating fault early warning method based on edge computing. The method includes: acquiring multi-source temperature signals from temperature sensors at different parts of the transformer during operation; decomposing the multi-source temperature signals into regional temperature signals of different heating areas, with signals from the core temperature area, winding temperature area, and surface temperature area processed independently; constructing a dynamic thermodynamic model of the transformer heating process based on the regional temperature signals, which describes the heat transfer and diffusion process inside the transformer; linearizing the dynamic thermodynamic model to simplify calculations and improve real-time performance, establishing a linear thermal dynamic model; and decomposing the linear thermal dynamic model into a longitudinal heat conduction equation set and a transverse heat diffusion equation set, with the longitudinal equation set processed... Heat is conducted vertically along the transformer, while the horizontal equations handle heat diffusion in the horizontal direction. Combining the vertical heat conduction equations and regional temperature signals, the dynamic thermodynamic model is expressed as a linear heating system, which can simulate the thermal behavior of the transformer during operation. For the linear heating system, adaptive early warning rules for transformer overheating faults are set. These rules include conventional early warning rules and thermal disturbance compensation rules. Conventional early warning rules are triggered based on temperature thresholds, while thermal disturbance compensation rules include compensation for time-varying and time-invariant thermal disturbances. Time-varying thermal disturbances include changes in ambient temperature, while time-invariant thermal disturbances include equipment aging. During transformer operation, edge computing devices continuously execute the adaptive early warning rules to achieve early warning of faults.

[0021] Example 1: Temperature sensor arrays, serving as the fundamental data acquisition unit, are carefully positioned at key heat-generating areas of the transformer, including the core area, winding coil area, and radiator outer surface area. These sensors employ high-precision digital temperature probes, capable of capturing minute temperature changes in real time. The sensors acquire temperature data at fixed time intervals, typically once per second, but this interval can be dynamically adjusted based on the transformer's actual operating load and environmental conditions. For example, under high-temperature or high-load conditions, the system automatically shortens the acquisition interval to 0.5 seconds to obtain more densely packed data points. The acquired raw temperature data is transmitted to the edge computing gateway via wired or wireless communication protocols (such as RS-485 or ZigBee), employing a data packet verification mechanism to ensure integrity during transmission. The edge computing gateway, acting as the local processing core, immediately performs signal preprocessing upon receiving the raw data streams from multiple sensors. The preprocessing stage comprises three main steps: data cleaning, noise filtering, and signal normalization. The data cleaning stage identifies and removes outliers caused by momentary sensor malfunctions or communication interference, for example, by detecting abrupt changes exceeding physically reasonable ranges using a sliding window algorithm. Noise filtering employs digital filters (such as low-pass FIR filters) to suppress high-frequency interference while retaining low-frequency components that reflect true thermal changes. Signal standardization converts the raw voltage or resistance values ​​collected by different sensors into a unified temperature unit (degrees Celsius) and aligns the timestamps, ultimately generating a continuous and consistent standardized temperature signal stream. This signal stream is stored in the gateway's memory using a ring buffer structure. The standardized temperature signal stream then enters the region decomposition stage. The edge computing gateway divides the transformer into three independent temperature zones based on its physical structural characteristics: core temperature zone signal, winding temperature zone signal, and surface temperature zone signal. The partitioning logic is based on sensor IDs and a pre-defined temperature zone mapping table; each sensor is labeled with its temperature zone category upon deployment. The core temperature zone signal corresponds to sensor data from the core and adjacent areas, the winding temperature zone signal covers all coil measurement points, and the surface temperature zone signal comes from sensors on the casing and radiator surface. After partitioning, each temperature zone signal is processed independently to extract its unique temperature fluctuation characteristics.

[0022] Wavelet transform is the core technique for feature extraction, performed separately for each temperature zone signal. The processing uses Daubechies wavelet basis functions (typically db4), with a decomposition level of 3, striking a balance between computational complexity and feature resolution. The transform process decomposes the original temperature signal into approximation coefficients and detail coefficients. Approximation coefficients capture the macroscopic trend components of the signal, while detail coefficients reflect high-frequency fluctuations. For the core temperature zone signal, wavelet transform focuses on extracting its slowly changing temperature rise trend due to the large thermal inertia of the iron core; for the winding temperature zone signal, it focuses on its potential rapid fluctuation patterns, stemming from Joule heating caused by current changes; and for the surface temperature zone signal, it analyzes its periodic fluctuations related to ambient temperature. These characteristic temperature fluctuation patterns are thus quantified into a set of wavelet coefficients, which characterize the behavior of each temperature zone in the time-frequency domain. The wavelet-transformed temperature zone signals are stored in the local database of the edge computing gateway, with each signal accompanied by a timestamp, temperature zone identifier, and feature coefficient vector. The storage format uses structured tables for easy retrieval and use by the subsequent model building module. The entire signal acquisition and decomposition process is completed in a closed loop at the edge, without cloud intervention, thus significantly reducing data transmission latency and bandwidth consumption, and achieving real-time response. Simultaneously, the system is designed with fault tolerance mechanisms; for example, when a sensor fails, the gateway generates a replacement signal based on historical data interpolation, ensuring the continuity of the decomposition process.

[0023] A 220kV oil-immersed power transformer in a certain region was put into operation during the spring load increase period. This transformer is equipped with a complete edge computing temperature monitoring system. The temperature sensor group uses PT100 platinum resistance temperature detectors. Three sensors are arranged in the core area to monitor the upper, middle, and lower parts of the iron core, respectively. Nine sensors are evenly distributed at key points of the three-phase windings in the winding area. Six sensors are arranged on the radiator surface to monitor the oil pipe inlet and outlet and the surface temperature of the radiator fins. All sensors are connected to the edge computing gateway via shielded twisted-pair cables, with a sampling interval of 1 second and a measurement accuracy of ±0.5℃. At 10:00 AM on a certain weekday, the transformer load rate gradually increased to 85%, and the temperature sensors began collecting operating data. The sensors in the core area recorded a slow temperature increase from 45℃, the sensors in the winding area showed a temperature increase starting from 65℃, and the sensors on the radiator surface monitored a gradual increase in temperature from 38℃. The collected raw temperature data contained a small amount of noise interference, mainly from electromagnetic field interference and thermal noise from the sampling circuit. Data is transmitted to the edge computing gateway via the Modbus RTU protocol at a rate of 9600 bps. Each data packet contains the current readings and timestamps of 16 sensors. Upon receiving the data, the edge computing gateway immediately initiates a preprocessing procedure. The data cleaning stage employs a sliding median filter algorithm with a window width of 5 sampling points to effectively eliminate sudden interference pulses. The noise filtering stage uses a second-order Butterworth low-pass filter with a cutoff frequency of 0.1 Hz to retain low-frequency signals related to the transformer's thermal process. In the signal standardization stage, the original resistance values ​​are converted to temperature values, linear interpolation is performed using the PT100 calibration table, and all sensor data is aligned to a unified time base, ultimately generating a continuous standardized temperature signal stream. The gateway then performs region segmentation processing, dividing the standardized signal into three independent data streams according to a preset temperature zone mapping table. The core temperature zone signal contains data from 3 sensors, corresponding to the temperature changes in the core area; the winding temperature zone signal contains data from 9 sensors, reflecting the winding's heating status; and the surface temperature zone signal contains data from 6 sensors, monitoring the operating status of the cooling system. Each temperature zone signal is allocated an independent data buffer, using a first-in-first-out queue structure to store data from the most recent 1200 seconds. Wavelet transform processing is performed independently for each temperature zone. The core temperature zone signal is decomposed into three levels using the db4 wavelet basis to extract low-frequency approximation coefficients reflecting the slow heating of the core; the winding temperature zone signal is decomposed into four levels using the db6 wavelet basis to obtain detailed coefficients characterizing the dynamic heating of the winding; and the surface temperature zone signal is decomposed into three levels using the db2 wavelet basis to separate the influence of ambient temperature and heat dissipation fluctuations. The Mallat algorithm is used for fast calculation during the transform process, generating a set of feature coefficient vectors for each temperature zone. These vectors capture the unique temperature fluctuation patterns of each region. The processed feature data is stored in the gateway's local database, which is organized in a time-series structure. Each record contains a timestamp, temperature zone identifier, feature coefficient array, and quality flag.The system simultaneously performs data integrity checks. When missing data from a sensor is detected, it interpolates by using a weighted average of data from adjacent sensors to ensure data continuity. The entire signal acquisition and processing process is completed at the edge, with a processing time controlled within 50 milliseconds, meeting the requirements of real-time monitoring. Through this implementation, the multi-source temperature signals during transformer operation are effectively converted into characteristic signals with clear physical meaning, providing a high-quality data foundation for subsequent thermal model construction and fault early warning. The system design considers various interference factors in actual operation, ensuring data reliability and accuracy through multiple processing steps, enabling subsequent analysis to be based on real operating conditions.

[0024] Example 2: See Figure 2 The edge computing gateway retrieves previously extracted wavelet coefficients for each temperature zone from the local database. These coefficients reflect the dynamic thermal behavior characteristics of the core, winding, and surface temperature zones. Based on these characteristic patterns, the system first determines the thermal coupling coefficient between each temperature zone. The thermal coupling coefficient quantifies the intensity of heat transfer between different temperature zones; for example, the thermal influence of the core temperature zone on the winding temperature zone is calculated using the temperature covariance from historical operating data. Simultaneously, thermal conduction parameters, including the material's thermal conductivity and specific heat capacity, are obtained partly from the transformer's design specifications and partly through offline experimental measurements, such as thermal characteristic tests on the transformer winding material to obtain accurate values. After determining the thermal coupling coefficient and thermal conduction parameters, the system constructs a multi-temperature zone coordinated thermal state equation set. This equation set uses partial differential equations to describe the heat transfer process in spatial and temporal dimensions. The energy balance equation for the core temperature zone considers the balance between heat generated by core losses and heat conducted outwards. The thermal diffusion equation for the winding temperature zone describes the diffusion process of Joule heat generated by current along the coil direction, and the heat dissipation equation for the surface temperature zone establishes the mathematical relationship between surface temperature and ambient cooling conditions. Each equation in the system contains time and space derivative terms, accurately characterizing the dynamic changes in temperature. The parameters of the equation system are continuously updated based on real-time temperature data, ensuring that the model remains consistent with the actual operating state of the transformer.

[0025] After constructing the thermal state equations, the system undergoes a Laplace transform to convert them to the frequency domain for analysis. The transformation assumes zero initial conditions, simplifying computational complexity while maintaining the integrity of the physical meaning. Through the Laplace transform, the partial differential equations in the time domain are converted into a set of algebraic equations in the frequency domain, resulting in the system's frequency-domain heat transfer function matrix. This matrix is ​​a set of multiple-input, multiple-output transfer functions; rows correspond to output temperature zones, columns to input temperature zones, and each element is a transfer function describing the thermal response characteristics of one temperature zone to another. For example, an element in the matrix might represent the frequency response characteristics of the winding temperature zone to changes in the core temperature zone. The frequency-domain heat transfer function matrix provides the foundation for subsequent linearization. The system approximates the transfer function matrix using rational fractions, transforming it into a linear time-invariant system expression. This linearization process preserves the main dynamic characteristics of the original system while significantly reducing computational complexity. The final linear thermal dynamic model is stored in the memory of the edge computing gateway in state-space or transfer function form. The model parameters are updated periodically based on real-time operating data to ensure that it accurately reflects the current thermal behavior characteristics of the transformer. This model serves as the core computational foundation for subsequent early warning analysis, providing a theoretical basis and computational framework for the execution of adaptive early warning rules.

[0026] Taking a 110kV oil-immersed power transformer operating during the high temperatures of summer in a certain region as an example, maintenance personnel discovered abnormal temperature fluctuations in the winding temperature zone through an edge monitoring system. The transformer is equipped with a complete temperature sensor network: two sensors in the core area, six sensors in the winding area, and four sensors on the radiator surface. At 2:30 PM one afternoon, the edge computing gateway detected simultaneous temperature anomalies from multiple sensors in the winding area, with the highest reading reaching 87℃, 12℃ higher than the normal operating temperature. The gateway immediately initiated the model building process of Example 2, first calling up historical temperature data from the past 24 hours to extract characteristic temperature fluctuation patterns for each temperature zone. Analysis revealed that the core temperature zone remained stable at 65±2℃, the surface temperature zone fluctuated normally with the ambient temperature, while the winding temperature zone showed a continuous upward trend with increased fluctuation amplitude. Based on these characteristic patterns, the system began calculating the thermal coupling coefficient. By analyzing the covariance matrix of the temperature data from the core and winding areas, the thermal coupling coefficient was determined to be 0.76, indicating that the core heat has a strong influence on the winding. Simultaneously, the thermal coupling coefficients between the winding phases were calculated, revealing that the coupling coefficient between phase B and phase C reached 0.88, indicating significant thermal interaction between the two winding phases. The thermal conduction parameters were determined based on material parameters from the transformer's technical specifications: the thermal conductivity of the winding copper conductor was 401 W / (m·K), the thermal conductivity of the insulating paper was 0.8 W / (m·K), and the thermal conductivity of the transformer oil was 0.12 W / (m·K). The system dynamically corrected these parameters using real-time temperature data; for example, the thermal conductivity of the transformer oil was adjusted to 0.125 W / (m·K) based on oil temperature changes to more accurately reflect the current operating state. Based on these parameters, the system constructed a multi-temperature zone coordinated thermal state equation set. The equation set includes three core equations: the core temperature zone energy balance equation describes the relationship between core loss heat generation and heat transfer to the windings; the winding temperature zone thermal diffusion equation establishes the mathematical relationship between current heating effects and axial heat transfer; and the surface temperature zone heat dissipation equation quantifies the heat exchange between the radiator and the environment. Each equation contains time and space differential terms; for example, the winding temperature zone equation considers the temperature gradient distribution along the winding axis and its rate of change over time. After constructing the equation set, the system performs a Laplace transform. The transform process converts the partial differential equations into an algebraic equation set in the complex frequency domain by introducing a complex variable s to replace the time differential operator. The transformed equation set forms a frequency domain heat transfer function matrix, which is a 3×3 structure, with rows corresponding to the output temperature zone and columns corresponding to the input temperature zone. The matrix elements show that the transfer function of the core temperature zone to the winding temperature zone has a large gain, confirming the previous thermal coupling analysis results. The final linear thermal dynamic model is stored in the gateway memory in state-space form, containing the system matrix A, input matrix B, output matrix C, and direct transfer matrix D. This model can simulate the thermal dynamic response of the transformer under conditions such as current disturbance and environmental changes, providing a mathematical model basis for subsequent early warning analysis.The system uses this model to simulate the current abnormal temperature and predict the temperature change trend in the next 30 minutes. The results show that if no intervention measures are taken, the winding temperature may rise to above 95°C, triggering a thermal fault warning.

[0027] Example 3: See Figure 3 The edge computing gateway, according to a preset time schedule (typically set to midnight UTC daily), initiates a data request to the cloud server to obtain the latest historical fault temperature pattern library. This pattern library is a structured database containing temperature change patterns under various typical fault scenarios, such as temperature time-series data and corresponding fault tags for cases like localized overheating, insulation aging, and cooling system failure. The gateway downloads incremental updates to the pattern library via secure communication protocols (such as HTTPS or MQTT with TLS) to reduce network bandwidth consumption. After downloading, it performs data verification and decompression locally to ensure data integrity and availability.

[0028] After acquiring the historical fault temperature pattern library, the edge computing gateway initiates a pattern matching calculation process, performing a similarity analysis between the currently processed characteristic temperature fluctuation pattern and the records in the pattern library. The current characteristic temperature fluctuation pattern originates from the coefficient vector extracted by wavelet transform in Example 1, representing a set of multi-dimensional features. The matching calculation employs a multi-dimensional similarity measurement method, combining Euclidean distance and cosine similarity for comprehensive evaluation. The core of the matching algorithm can be described as follows: in: This represents the overall similarity score, with a value range of [0,1]. This represents the Euclidean distance between the current feature vector and the historical pattern vector; This represents the angle between two vectors; This is a weighting coefficient, with a value range of (0,1), used to balance the contributions of distance similarity and orientation similarity; This is a scale adjustment parameter used to control the sensitivity to distance effects. The symbols in this formula are not repeated in the formulas of Examples 2 and 4. , This is specifically used for similarity calculation.

[0029] Based on the matching calculation results, the system dynamically adjusts the parameters of the early warning rules. The threshold parameters for conventional early warning rules include absolute temperature thresholds and rate of change thresholds. When similarity analysis shows a high match between the current pattern and a certain historical fault pattern, the system correspondingly lowers the temperature threshold or increases the rate of change sensitivity. For example, if a local overheating pattern is matched, the upper temperature threshold may be lowered from the default 90℃ to 85℃; if a progressive aging pattern is matched, the monitoring window for the rate of temperature change will be shortened from 10 minutes to 5 minutes. The adjustment of thermal disturbance compensation rules is handled separately for time-varying and time-invariant thermal disturbances. For time-varying thermal disturbances (such as diurnal temperature differences or seasonal changes), the compensation rule is implemented by introducing a time-dependent compensation function. This function fits the trend of ambient temperature changes based on recent temperature data and cancels it out in the calculation. For time-invariant thermal disturbances (such as sensor calibration deviations or slow changes in material properties), the compensation rule is implemented by updating the bias parameters in the model. These parameters are calibrated based on steady-state differences in long-term operating data. The compensation intensity is automatically adjusted based on the similarity score; the higher the matching degree, the greater the compensation magnitude. The rule adjustments take effect immediately upon completion. The new parameters are written to the edge computing gateway's configuration register, and the real-time alert module uses these parameters for judgment. Simultaneously, adjustment logs are recorded and periodically uploaded to the cloud for iterative optimization of the pattern library. The entire adjustment process forms a closed loop, ensuring the alert system can adapt to transformer aging, environmental changes, and operational status transitions, maintaining high alert accuracy.

[0030] After several days of continuous high temperatures, the edge computing gateway of a 35kV dry-type transformer at a substation detected abnormal temperature data in the winding temperature zone. Around 3 PM, the winding temperature rose from 78℃ to 86℃ within half an hour, exceeding the initial threshold of 82℃ set by the standard warning rules, triggering the warning assessment process. The gateway immediately initiated data interaction with the cloud server, requesting the download of the latest historical fault temperature pattern library via an encrypted link. The pattern library contains various fault cases recorded for this type of transformer over the past three years, such as overheating caused by insulation aging, temperature rise caused by cooling fan failure, and localized overheating caused by loose connections. Each pattern includes multi-dimensional data features such as temperature change curves, environmental conditions, and load rate. The incremental update package received by the gateway contains newly confirmed fault patterns from the past week, especially two cases of winding overheating that also occurred during high-temperature weather. Pattern matching calculations were then performed, and the gateway extracted the characteristic temperature fluctuation pattern of the current winding temperature zone—this pattern shows a continuous step-like temperature increase, with an hourly increase of approximately 8℃, accompanied by small, high-frequency fluctuations. The system performs a similarity analysis between the current pattern and records in the historical database, calculating the comprehensive similarity score between the current feature vector and each historical pattern vector. The analysis revealed a high degree of similarity between the current pattern and a recorded "cooling efficiency decline" fault pattern, with a similarity score of 0.87, and a similarity score of 0.72 with the "partial insulation damage" pattern. Based on the matching results, the system begins dynamically adjusting the warning rule parameters. Regarding the conventional warning rules, since the matching shows a high correlation between the fault characteristics and cooling system problems, the system lowers the upper limit threshold for winding temperature from 82℃ to 79℃ and shortens the monitoring window for temperature change rate from 30 minutes to 15 minutes to improve sensitivity to the temperature rise rate. The adjustment of the thermal disturbance compensation rules addresses both time-varying and time-invariant thermal disturbances separately. For time-varying thermal disturbances, the system detected that the ambient temperature had reached 38℃ at noon and was continuing to rise; therefore, a temperature compensation based on a time function is introduced to offset the impact of the rising ambient temperature in the calculation. For time-invariant thermal disturbances, the system, based on long-term operating data, found that the temperature reading of phase B winding of the transformer was consistently about 2°C higher than that of other phases. This inherent deviation was incorporated into the compensation model and corrected in the calculations. All adjusted parameters took effect immediately, and the new early warning rules were implemented. The system continuously monitored the winding temperature changes. In the following hour, the temperature fluctuated around the 79°C threshold but did not consistently exceed the new threshold. The adjustment log was recorded in detail, including the matched pattern ID, similarity score, and parameter values ​​before and after the adjustment. This data will be periodically uploaded to the cloud for optimization and updates to the pattern library. The entire adaptive adjustment process was completed independently at the edge, without manual intervention, demonstrating the system's real-time response capability to changes in the operating environment.Through continuous pattern matching and rule optimization, the early warning system can gradually adapt to the individual characteristics and operating environment of transformers, thereby improving the accuracy and timeliness of fault identification.

[0031] Example 4: When the gateway determines that a temperature zone is abnormal based on a linear thermal dynamics model and adaptive rules, such as a winding temperature zone rising above a set threshold in a short period of time, the system does not immediately report an early warning. Instead, it first initiates a multi-node verification procedure to eliminate false alarms. The gateway generates an early warning verification request message, which includes the identifier of the abnormal temperature zone, the current temperature reading, a timestamp, and a geolocation code. This request is sent to physically adjacent nodes via the broadcast channel of the edge computing node network. The range of adjacent nodes is typically defined as other transformer monitoring nodes deployed in the same substation or distribution area. Upon receiving the verification request, the adjacent edge computing nodes first check the timeliness of the message, discarding requests exceeding the valid time window (e.g., 5 seconds) to prevent processing expired data. Subsequently, each node retrieves temperature records from its own database at the same time, which are from monitoring data of similar transformers. The node organizes a verification response data packet, including its own node ID, the monitored transformer equipment ID, the temperature value of the corresponding temperature zone, and the acquisition timestamp, and returns it to the requesting gateway via a dedicated reply channel. The entire request-response process uses a lightweight communication protocol to ensure real-time performance, typically completing within hundreds of milliseconds.

[0032] After collecting response data from all neighboring nodes, the gateway initiating the verification performs a multi-node temperature data consistency check. The verification process is based on spatial correlation and temporal synchronization principles. First, it aligns the timestamps of the data from each node, using interpolation to synchronize data within small time deviations (e.g., ±500 milliseconds). Then, it calculates the Euclidean distance and correlation coefficient between the local node's temperature data and the temperature data of each neighboring node. The Euclidean distance measures the absolute difference in temperature, while the correlation coefficient assesses the consistency of the changing trend. The gateway's preset judgment thresholds are: the Euclidean distance must not exceed 3 degrees Celsius, and the Pearson correlation coefficient must be greater than 0.75. When the data from all neighboring nodes meets both conditions, the system determines that the verification has passed. If the verification fails, the gateway initiates a re-acquisition procedure, which includes three steps: first, it sends a data retransmission request to neighboring nodes, requesting the resending of temperature data at a specific time point; second, the local node rereads the sensor data to eliminate possible instantaneous acquisition errors; and finally, it performs a second verification on the re-acquisition data. Re-acquisition is repeated a maximum of three times. If it still fails, the warning verification is abandoned, and the event is recorded as a suspected false alarm. The entire verification process ensures spatial consistency of early warning information, avoiding false warnings caused by single-node sensor failures or localized interference. See Table 1 for a specific verification request and response data.

[0033] Table 1: Warning Verification Request and Response Data Based on the data shown in the table, the gateway calculated the Euclidean distances to each node to be 2.4, 2.7, and 1.3 degrees Celsius, with correlation coefficients of 0.82, 0.79, and 0.88, respectively. All values ​​met the preset threshold requirements, therefore the verification passed. The gateway then generated a formal early warning message, including the abnormal temperature value, the list of verified nodes, verification statistics, and other complete evidence chains, and uploaded it to the cloud monitoring platform. This multi-node verification mechanism significantly improves the reliability of early warning information and reduces false alarms caused by local factors.

[0034] Example 5: The cloud monitoring platform, as the core processing unit, continuously receives verified early warning information uploaded from the distributed edge computing gateway. This early warning information is transmitted using a standardized data format, including fields such as: early warning ID, trigger timestamp, edge gateway device ID, substation number, abnormal temperature zone type, abnormal temperature value, duration, geographic coordinates, and a list of verification nodes. The platform receiving module performs a data integrity check, discarding records with missing key fields and storing valid data in a spatiotemporal database for later use. The spatiotemporal correlation analysis module performs multi-dimensional analysis on the stored early warning information. The time dimension analysis uses a sliding window mechanism, with a 1-hour window as the basic unit, statistically analyzing the frequency and duration distribution of similar early warnings within the window. The spatial dimension analysis is based on geographic coordinates, using a gridded processing method, dividing the monitoring area into 500m × 500m grid units, and calculating the density value of early warning events within each grid. The module identifies the spatiotemporal clustering patterns of early warning events. For example, if winding temperature zone early warnings occur in the same grid within three consecutive time windows, it is marked as a potential hotspot area; or if multiple adjacent grids experience surface temperature zone early warnings within a similar time period, it is identified as a regional heat dissipation anomaly.

[0035] Based on the spatiotemporal correlation analysis results, the system initiates the process of generating a regional thermal fault risk map. Map generation first extracts core features from the early warning information: temperature anomaly patterns are categorized into three types: instantaneous spikes, sustained exceedances, and gradient increases; occurrence time is converted into categorical variables such as weekday / holiday and day / night periods; and geographical location information is mapped to a predefined power grid topology. The density clustering algorithm (DBSCAN) performs joint analysis on these features, with the following parameter settings: neighborhood radius Epsilon of 1 km, minimum sample size MinPts of 5, and spatiotemporal weight ratio of 0.7:0.3. The clustering results form multiple early warning event clusters, each representing a potential risk area. When drawing the thermal fault risk level distribution map, a red-yellow-green three-color early warning system is used: red represents high risk (number of events within the cluster > 10 and high degree of temperature exceedance), yellow represents medium risk (5 < number of events within the cluster ≤ 10), and green represents low risk (number of events within the cluster ≤ 5). The distribution map is overlaid on the power grid geographic information system base map, marking the scope, level, and main anomaly patterns of the risk areas. The time evolution trend chart uses a line graph format to show the change curve of the number of areas at each risk level in the past 24 hours, and marks key inflection point events.

[0036] The cloud platform distributes the generated risk map to each edge computing gateway via a secure downlink channel, using the MQTT with QoS 1 protocol to ensure at least one delivery. After receiving the map, the edge gateway parses the content and first matches the risk level based on its geographical location. If it is in a high-risk area, it automatically adjusts local monitoring parameters: the temperature sampling frequency is increased from the baseline 1Hz to 2Hz; the wavelet transform level is increased from 3 to 4; and the Euclidean distance threshold for early warning verification is tightened from 3℃ to 2℃. If it is in a low-risk area, the parameters are appropriately relaxed to save computing resources.

[0037] The adjusted operational status data is encapsulated into feedback messages, including the gateway ID, adjustment time, original parameter value, new parameter value, and adjustment reason code, and is periodically reported to the cloud monitoring platform. Upon receiving the feedback data, the platform updates the device profiles of each gateway and assesses the impact of parameter adjustments on the accuracy of early warnings. This data is further used to optimize the update strategy of the historical fault temperature pattern library, such as increasing the sample weight of fault patterns in high-frequency adjustment areas. The entire process forms a closed-loop optimization system from edge perception to cloud analysis and then to underlying execution, enabling early warning capabilities to dynamically evolve with the operating environment. The system is designed with an anomaly handling mechanism. When the cloud platform detects multiple gateways reporting parameter adjustment failures, it automatically triggers a rollback procedure to restore the previous stable configuration and generates a diagnostic report for operations and maintenance personnel to analyze. All data exchange processes are recorded in audit logs to ensure the traceability of the optimization process.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A transformer overheating fault early warning method based on edge computing, characterized in that, Includes the following steps: Collect multi-source temperature signals during transformer operation and decompose the multi-source temperature signals into regional temperature signals of different heating areas; Based on the temperature signal of the region, a dynamic thermodynamic model of the transformer heating process is constructed. The dynamic thermodynamic model is linearized to establish a linear thermal dynamic model. The linear thermal dynamic model is decomposed into a longitudinal heat conduction equation set and a transverse heat diffusion equation set. Combining the longitudinal heat conduction equation set and the temperature signal of the region, the dynamic thermodynamic model is expressed as a linear heating system. For the linear heating system, an adaptive early warning rule for transformer overheating faults is set. The adaptive early warning rule includes a conventional early warning rule and a thermal disturbance compensation rule. The thermal disturbance compensation rule includes compensation for time-varying thermal disturbances and time-invariant thermal disturbances. The adaptive early warning rule is executed during transformer operation.

2. The transformer overheating fault early warning method based on edge computing according to claim 1, characterized in that, The method for acquiring the multi-source temperature signal is as follows: Temperature data of the transformer core, windings and radiator surface are collected at fixed time intervals using a group of temperature sensors in the edge computing node network. The collected temperature data is uploaded to the edge computing gateway for signal preprocessing to generate a standardized temperature signal stream.

3. The transformer overheating fault early warning method based on edge computing according to claim 2, characterized in that, The decomposition process of the regional temperature signal includes: The edge computing gateway receives the standardized temperature signal stream and divides the signal into core temperature zone signal, winding temperature zone signal and surface temperature zone signal according to the physical structural characteristics of the transformer. Each temperature zone signal is processed independently using wavelet transform to extract the characteristic temperature fluctuation patterns of each zone.

4. The transformer overheating fault early warning method based on edge computing according to claim 3, characterized in that, The process of establishing the linear thermal dynamics model includes: Based on the characteristic temperature fluctuation pattern, determine the thermal coupling coefficient and thermal conduction parameters for each temperature range; Based on the aforementioned thermal coupling coefficient and thermal conduction parameters, a multi-temperature zone coordinated thermal state equation set is constructed. The frequency domain heat transfer function matrix is ​​obtained by performing a Laplace transform on the thermal state equations.

5. The transformer overheating fault early warning method based on edge computing according to claim 4, characterized in that, The adaptive early warning rule is generated as follows: The edge computing gateway periodically retrieves a historical fault temperature pattern library from the cloud server; The current characteristic temperature fluctuation pattern is matched and calculated with the historical fault temperature pattern library; The threshold parameters of the conventional early warning rule and the compensation intensity of the thermal disturbance compensation rule are dynamically adjusted based on the matching results.

6. The transformer overheating fault early warning method based on edge computing according to claim 5, characterized in that, It also includes an early warning verification step: When the adaptive warning rule is triggered, the edge computing gateway broadcasts a warning verification request to adjacent edge computing nodes; Receive verification temperature data returned by adjacent edge computing nodes and perform multi-node temperature data consistency verification; The warning information will only be uploaded to the cloud monitoring platform when the verification is successful.

7. The transformer overheating fault early warning method based on edge computing according to claim 6, characterized in that, The specific process of the multi-node temperature data consistency verification includes: Calculate the Euclidean distance and correlation coefficient between the temperature data of this node and the temperature data of neighboring nodes; The verification is deemed successful when the Euclidean distance is less than a set threshold and the correlation coefficient is greater than a set threshold. Otherwise, start the re-acquisition program to reacquire the verification temperature data.

8. The transformer overheating fault early warning method based on edge computing according to claim 7, characterized in that, It also includes early warning optimization steps: The cloud-based monitoring platform receives early warning information uploaded by multiple edge computing gateways; Spatiotemporal correlation analysis is performed on the aforementioned early warning information to generate a regional thermal fault risk map; The regional thermal failure risk map is distributed to each edge computing gateway to optimize the adaptive early warning rules.

9. A transformer overheating fault early warning method based on edge computing according to claim 8, characterized in that, The process of generating the regional thermal failure risk map includes: Extract temperature anomaly patterns, occurrence times, and geographical location information from early warning information; Density clustering algorithm is used to aggregate and analyze early warning information with similar characteristics; Based on the aggregation results, a distribution map of thermal failure risk levels and a time evolution trend map were drawn.

10. A transformer overheating fault early warning method based on edge computing according to claim 9, characterized in that, It also includes an early warning feedback step: After receiving the regional thermal fault risk map, the edge computing gateway adjusts the local temperature sampling frequency and signal processing parameters. The adjusted operational status data is fed back to the cloud monitoring platform to form a closed-loop optimization system.

Images