An oil-immersed transformer equipment fire identification method and linkage control system based on operation monitoring data

Abstract

The present application relates to the technical field of data processing, in particular to an oil-immersed transformer equipment fire identification method and a linkage control system based on operation monitoring data; heterogeneous physical quantities and environmental load data of a variable transformer equipment are synchronously collected based on an edge node, and multi-dimensional time sequence data are constructed; measured thermal data are extracted from the multi-dimensional time sequence data, the measured thermal data are compared with an adaptive thermal behavior benchmark model, and a thermal anomaly residual sequence excluding environmental interference is generated; the thermal anomaly residual sequence and non-thermal heterogeneous time sequence data are synchronously input into a multi-head attention fusion network to extract cross-modal derived features and confirm an abnormal node; a fire cause knowledge graph containing phenomenon nodes and mechanism nodes is constructed, node features of the abnormal node are input into a graph neural network, mapping reasoning is performed under prior topological constraints of the knowledge graph, and a fire cause evolution link and a confidence under a multi-abnormal interweaving are output; a structured early warning report is generated according to the fire cause evolution link to search and match a disposal strategy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and linkage control system for fire identification of oil-immersed transformer equipment based on operational monitoring data. Background Technology

[0002] As a core hub of the new power system, the operating status of transformer equipment directly affects the safety of the power grid. With the popularization of intelligent sensing technology, transformer equipment monitoring systems have accumulated massive amounts of multimodal time-series data, including temperature, partial discharge, and vibration.

[0003] However, existing data-driven pattern recognition and processing methods face significant bottlenecks when dealing with such complex industrial data. First, in the feature extraction stage, existing models often struggle to separate normal thermal fluctuations caused by complex meteorological environments and power grid loads from actual physical heat generation, resulting in extracted feature sequences containing a large amount of environmental artifact noise, which easily triggers algorithmic misjudgments. Second, in the multi-source data fusion stage, existing fusion mechanisms mostly employ static weighting. When faced with partial missing sensor data or intermodal logical conflicts caused by strong electromagnetic interference, they lack dynamic weight suppression and conflict resolution capabilities, resulting in poor model robustness. Finally, traditional deep learning classification models can only output black-box risk probabilities, failing to map discrete anomalous pattern nodes to underlying physical decay mechanisms, making it difficult to infer evolution paths under multimodal anomaly interweaving.

[0004] In summary, the core problem that urgently needs to be solved in this field is how to achieve dynamic and reliable fusion of multi-source cross-modal features in a noisy environment, and to complete causal link pattern matching and inference decision with strong interpretability.

[0005] To address this, a fire detection method and linkage control system for oil-immersed transformer equipment based on operational monitoring data are proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method and linkage control system for fire identification of oil-immersed transformer equipment based on operation monitoring data. Under the prior topological constraints of the knowledge graph, the method performs mapping reasoning and outputs the fire cause evolution link and confidence level under multiple anomalies. Based on the fire cause evolution link, the method retrieves and matches the disposal strategy to generate a structured early warning report.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for fire detection in oil-immersed transformer equipment based on operational monitoring data includes: Data on heterogeneous physical quantities and environmental loads of transformer equipment are collected synchronously based on edge nodes, and multidimensional time-series data are constructed. Historical long-term operating data is extracted and seasonal trend smoothing decomposition and multivariate nonlinear regression are performed to construct an adaptive thermal behavior benchmark model; measured thermal data are extracted from the multidimensional time series data, and the measured thermal data are compared with the adaptive thermal behavior benchmark model to generate a thermal anomaly residual sequence that removes environmental interference. The thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from multidimensional time series data are synchronously input into a multi-head attention fusion network to extract cross-modal derived features. The feature vector is transformed into an evidence body and conflict determination and dynamic comprehensive risk fusion are performed using evidence theory to confirm the anomaly node. A fire cause knowledge graph containing phenomenon nodes and mechanism nodes is constructed. The node features of abnormal nodes are input into a graph neural network. Under the prior topological constraints of the knowledge graph, mapping reasoning is performed, and the fire cause evolution link and confidence level under multiple anomalies are output. Based on the fire cause evolution link, a matching and treatment strategy is retrieved to generate a structured early warning report.

[0008] Preferably, the process of collecting and constructing multidimensional time-series data includes: An edge node sensing array is deployed at the physical monitoring points of the transformer equipment to acquire raw heterogeneous signals; wherein, the heterogeneous physical quantities include top oil temperature, winding temperature, partial discharge high-frequency pulse, and enclosure mechanical vibration; the environmental load data includes ambient temperature and humidity and grid load current; By using a time-sensitive network edge gateway, the collected heterogeneous physical quantities and environmental load data are timestamped with a unified clock cycle to obtain discrete time-series data. To address the uneven sampling frequency caused by differences in the physical properties of different sensors, alignment and reconstruction are performed. The aligned and reconstructed data are orthogonally stitched together according to the time dimension and the feature dimension to output multidimensional time-series data with temporal consistency.

[0009] Preferably, the process of constructing an adaptive thermal behavior benchmark model includes: Extract historical data from multiple complete operating cycles as historical long-cycle operating data; The local weighted scatter smoothing algorithm is used to perform time-series decoupling on the thermal parameters in the historical long-cycle operation data, and to separate the slow-change trend component reflecting the evolution of equipment insulation aging and the periodic seasonal component reflecting the seasonal change and diurnal temperature difference. To address the thermal inertia hysteresis effect caused by transformer volume and insulating oil, a multivariate nonlinear prediction unit is established. This unit uses historical grid load current and historical ambient temperature and humidity as input features, and the superposition of the gradually changing trend component and the periodic seasonal component as the benchmark prediction target. The multivariate nonlinear prediction unit introduces a dynamic delay factor for heat fluid conduction and a cumulative compensation coefficient for solar radiation. The backpropagation algorithm is used to optimize the parameters, and the resulting adaptive thermal behavior benchmark model can adaptively track fluctuations in grid load and micro-meteorological conditions.

[0010] Preferably, the process of obtaining the thermal anomaly residual sequence includes: The measured thermal data and current environmental load data are extracted from multidimensional time series data. The current environmental load data is input into the adaptive thermal behavior benchmark model, and the model is solved by combining the feedback parameters of the current start-stop status of the transformer's cooling equipment. The theoretically expected normal temperature benchmark value that should appear under the current actual operating conditions is output. The measured thermal data specifically includes the top oil temperature and winding temperature. To address the interference of localized sudden drops in transformer surface temperature caused by micro-meteorological fluctuations, a dynamic differential unit is constructed between the expected normal temperature baseline value and the measured thermodynamic data. The dynamic differential unit smooths negative temperature abrupt changes caused by rapid external cooling and dynamically compensates for the deviation of external cold sources in the thermal balance calculation. The smoothed continuous differential results are output as a thermal anomaly residual sequence.

[0011] Preferably, the process for confirming the abnormal node includes: The thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from the multidimensional time series data are synchronously input into the multi-head attention fusion network; the non-thermal heterogeneous time series data includes partial discharge high-frequency pulses and box mechanical vibrations; To address the issue of data packet loss from single sensors in the complex and strong electromagnetic environment of transformer substations, an adaptive channel masking mechanism is introduced at the network input to dynamically suppress missing and abrupt data. Inside the network, cross-modal attention weights are used to calculate the degree of physical coupling response between thermal abrupt changes and mechanical vibration of the enclosure, and cross-modal derived feature vectors are extracted. These cross-modal derived feature vectors are then mapped to multiple basic probability assignment functions containing independent fire risk assessment probabilities as evidence. Calculate the logical conflict coefficient between each piece of evidence. If the conflict coefficient is lower than the safety limit, use the Dempster synthesis rule to perform orthogonal sum operation to complete dynamic comprehensive risk fusion, identify abnormal nodes and output abnormal node characteristics.

[0012] Preferably, the process of obtaining the fire cause evolution chain includes: A directed knowledge graph of fire causes is pre-constructed, in which phenomenon nodes include sensor distortion feature parameters such as temperature exceeding limits and sudden increase in discharge, mechanism nodes include hidden physical defect parameters such as winding loosening and deformation and insulation medium aging and peeling, and fire cause nodes are topologically connected through directed edges with historical real occurrence conditional probability weights. The node features of the abnormal nodes are received and transformed into initial activation sources in the graph structure, and then input into the multi-layer graph neural network. Under the rigid prior topological path constraints of the fire cause knowledge graph, message passing between neighboring nodes and high-dimensional feature aggregation reasoning are performed. The connected subgraph that maximizes the joint activation probability is found through global search optimization, and the most complete fire cause evolution link from the current multi-dimensional physical anomaly starting point to the final local deflagration endpoint is output, and the corresponding comprehensive confidence value is output simultaneously.

[0013] Preferably, the process of generating a structured early warning report based on the fire cause and evolution chain retrieval matching and response strategy includes: The system receives the fire cause evolution chain and the underlying mechanism and inducing node included, performs a high-dimensional vector space similarity search in the pre-established transformer safety operation and maintenance strategy database, retrieves historical handling experience based on the highest matching degree, and generates a corresponding risk reduction and handling strategy; the risk reduction and handling strategy includes physical cooling, load shedding, and fire extinguishing device standby; Combining the risk transmission rate parameters in the fire cause evolution chain, the remaining rescue time window from the current deterioration state to the irreversible fire outbreak is calculated; the multi-source anomaly characteristic description, the inferred cause evolution conclusion, the remaining rescue time window, and the risk reduction and disposal strategy are assembled and recombined into a structured early warning report.

[0014] A transformer fire detection system based on operational monitoring data includes: The data acquisition module synchronously collects heterogeneous physical quantities and environmental load data of transformer equipment based on edge nodes, and constructs multi-dimensional time-series data; The data processing module extracts historical long-cycle operating data, performs seasonal trend smoothing decomposition and multivariate nonlinear regression, and constructs an adaptive thermal behavior benchmark model. It extracts measured thermal data from the multidimensional time series data, compares the measured thermal data with the adaptive thermal behavior benchmark model, and generates a thermal anomaly residual sequence that eliminates environmental interference. The anomaly detection module synchronously inputs the thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from multidimensional time series data into a multi-head attention fusion network to extract cross-modal derived features, transforms the feature vector into evidence, and uses evidence theory to perform conflict detection and dynamic comprehensive risk fusion to confirm the anomaly node. The anomaly early warning module constructs a fire cause knowledge graph containing phenomenon nodes and mechanism nodes. The node features of the anomaly nodes are input into the graph neural network, and mapping reasoning is performed under the prior topological constraints of the knowledge graph. The module outputs the fire cause evolution link and confidence level under multiple anomalies. Based on the fire cause evolution link, a matching and disposal strategy is retrieved to generate a structured early warning report.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts long-term historical data, decomposes it through time-series smoothing to reveal insulation aging trends and cyclical seasonal patterns, and integrates heat transfer delay factors and solar radiation compensation to construct an adaptive thermal behavior benchmark model. Based on this, by constructing dynamic differential units, it can intelligently smooth negative abrupt changes caused by rapid external cooling and dynamically compensate for the bias of external cold sources in thermal balance calculations; providing high-signal-to-noise ratio, clean feature inputs for subsequent pattern recognition networks, reducing false alarms caused by environmental artifacts, and improving recognition accuracy.

[0016] 2. The multi-head attention fusion network of this invention introduces an adaptive channel masking mechanism at its input end, which can keenly capture data missing states and dynamically implement weight suppression, blocking the interference of dirty data on the global feature space from the source. Simultaneously, it utilizes a cross-modal attention mechanism to deeply mine the cooperative response patterns between heterogeneous signals such as thermal abrupt changes and mechanical vibrations; and maps the extracted derived features into evidence bodies, performing rigorous logical conflict judgment before fusion; effectively resolving identification discrepancies between multi-source sensors, ensuring that high-confidence anomalous pattern nodes can still be output even in harsh industrial noise environments.

[0017] 3. This invention constructs a knowledge graph of fire causes that includes phenomenon nodes and mechanism nodes, transforming prior knowledge into rigid graph topological constraint paths. When the confirmed abnormal pattern nodes are input into the multi-layer graph neural network as initial activation sources, high-dimensional feature aggregation and global search optimization are performed along the directed edges with historical true conditional probability weights. This pattern matching reasoning based on graph topological constraints clearly outlines the causal evolution chain under multiple anomalies, and based on this, a structured early warning report containing the remaining rescue time window is generated by retrieving historical databases, providing action guidelines for emergency rescue. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for fire identification in oil-immersed transformer equipment based on operational monitoring data, according to the present invention. Figure 2 This is a data logic diagram of a fire identification method for oil-immersed transformer equipment based on operational monitoring data according to the present invention. Figure 3 This is a schematic diagram of the structure of a transformer fire identification system based on operation monitoring data according to the present invention. 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] Example 1: This invention proposes a method for fire detection in oil-immersed transformer equipment based on operational monitoring data. The process of the method is as follows: Figure 1 As shown, the data logic of the method is as follows: Figure 2 As shown, it includes: Data on heterogeneous physical quantities and environmental loads of transformer equipment are collected synchronously based on edge nodes, and multidimensional time-series data are constructed. Historical long-term operating data is extracted and seasonal trend smoothing decomposition and multivariate nonlinear regression are performed to construct an adaptive thermal behavior benchmark model; measured thermal data are extracted from the multidimensional time series data, and the measured thermal data are compared with the adaptive thermal behavior benchmark model to generate a thermal anomaly residual sequence that removes environmental interference. The thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from multidimensional time series data are synchronously input into a multi-head attention fusion network to extract cross-modal derived features. The feature vector is transformed into an evidence body and conflict determination and dynamic comprehensive risk fusion are performed using evidence theory to confirm the anomaly node. A fire cause knowledge graph containing phenomenon nodes and mechanism nodes is constructed. The node features of abnormal nodes are input into a graph neural network. Under the prior topological constraints of the knowledge graph, mapping reasoning is performed, and the fire cause evolution link and confidence level under multiple anomalies are output. Based on the fire cause evolution link, a matching and treatment strategy is retrieved to generate a structured early warning report.

[0021] Preferably, the process of collecting and constructing multidimensional time-series data includes: An edge node sensing array is deployed at the physical monitoring points of the transformer equipment to acquire raw heterogeneous signals; wherein, the heterogeneous physical quantities include top oil temperature, winding temperature, partial discharge high-frequency pulse, and enclosure mechanical vibration; the environmental load data includes ambient temperature and humidity and grid load current; By using a time-sensitive network edge gateway, the collected heterogeneous physical quantities and environmental load data are timestamped with a unified clock cycle to obtain discrete time-series data. To address the uneven sampling frequency caused by differences in the physical properties of different sensors, alignment and reconstruction are performed. The aligned and reconstructed data are orthogonally stitched together according to the time dimension and the feature dimension to output multidimensional time-series data with temporal consistency.

[0022] Edge node sensing arrays are deployed at the physical monitoring points of the transformer equipment. These arrays consist of multiple types of sensors, each responsible for collecting raw signals of different physical properties. A top-level oil temperature sensor, using a platinum resistance thermometer, is placed at the top of the transformer's oil tank flange. It continuously records the real-time temperature of the top-level transformer oil at a sampling frequency of once per second, reflecting the overall thermal load status of the transformer. A winding temperature sensor, employing a fiber optic grating method, is directly embedded between the layers of the high-voltage winding conductors. It acquires the real-time temperature of the winding body at a sampling frequency of once per second. This location is a high-risk area where localized overheating first occurs.

[0023] The partial discharge high-frequency pulse sensor uses an ultra-high frequency antenna attached to the outside of the transformer tank wall. It acquires ultra-high frequency pulse signals caused by internal insulation degradation and air gap breakdown at a high sampling frequency of 100 to 1000 times per second. The occurrence of partial discharge indicates the presence of electrical defects inside the insulating medium. The tank mechanical vibration sensor uses a piezoelectric accelerometer attached to the outer wall of the transformer tank. It continuously acquires the vibration acceleration sequence of the core and windings under electromagnetic force at a sampling frequency of 200 to 500 times per second. Abnormal vibration signals can reflect mechanical defects such as winding loosening and deformation.

[0024] The ambient temperature and humidity sensor is installed in the shaded area outside the transformer's protective fence to eliminate interference from direct sunlight, acquiring ambient air temperature and relative humidity values ​​once per minute. The power grid load current sensor uses a through-hole current transformer connected to the high-voltage side of the transformer, recording the effective value of the actual current passing through the transformer equipment once per second. This value is the core input parameter for assessing the equipment's thermal load.

[0025] Due to their different physical measurement principles, the aforementioned sensors have significantly different inherent sampling frequencies. Directly splicing these sensors without processing would result in numerous gaps in the data matrix, making subsequent multivariate joint analysis impossible. To address this issue, edge nodes use a time-sensitive network (TSN) edge gateway to timestamp the raw signals collected by all sensors with a unified clock cycle. TSN is a deterministic Ethernet protocol based on the IEEE 802.1 standard. Its core capability lies in controlling the clock error of all nodes within the network to the sub-microsecond level through a precise clock synchronization mechanism. This ensures the global uniformity of the timestamps accompanying the data reported by different sensors, thereby eliminating time base deviations caused by transmission delays or clock drift, and obtaining discrete time-series data from each sensor.

[0026] After obtaining discrete time-series data with a unified timestamp, alignment and reconstruction are performed to address the uneven sampling frequency caused by differences in the physical properties of different sensors. The goal of alignment and reconstruction is to unify all sensor data onto the same time resolution grid. Specifically, using the sensor with the lowest sampling frequency in the system (i.e., the environmental temperature and humidity sensor, once per minute) as the baseline target resolution, downsampling is performed on the high-frequency sensor data. This involves averaging the high-frequency data within each target time window, compressing multiple sampled values ​​within the window into a single representative value. For low-frequency sensor data that has already been sampled within the baseline time window, it is directly referenced. For moments without new samples within the window, nearest neighbor interpolation is used to fill the gap, i.e., the most recent valid sampled value is used for forward propagation. This alignment and reconstruction operation ensures that the sampling points of all physical quantities on the time axis are completely corresponding, eliminating spurious correlation interference caused by time misalignment.

[0027] After alignment and reconstruction, the one-dimensional time-series data of each physical quantity are orthogonally concatenated along the time and feature dimensions to form a multi-dimensional time-series data matrix. The rows of this matrix represent the time series index, meaning each row corresponds to a unified sampling time; the columns represent the feature dimension index, meaning each column corresponds to a physical or environmental quantity. The final output multi-dimensional time-series data matrix contains seven feature dimensions: top-layer oil temperature, winding temperature, statistical characteristics of high-frequency pulses of partial discharge, statistical characteristics of mechanical vibration of the enclosure, ambient temperature, ambient humidity, and grid load current. It exhibits complete time-series consistency and can be directly used as the data source for subsequent steps.

[0028] This invention achieves comprehensive acquisition of signals from multiple physical domains, including thermal, electrical, mechanical, and environmental domains, by deploying multi-type sensors to form a sensing array at key physical monitoring points of transformer equipment. The introduction of a time-sensitive network edge gateway for sub-microsecond clock synchronization eliminates inconsistencies in the time reference of multi-sensor data, avoiding spurious feature correlations introduced by time deviations. A specifically designed alignment and reconstruction strategy unifies heterogeneous data of different frequencies to the same time resolution grid, preserving the statistical characteristics of high-frequency signals while avoiding the computational resource consumption of directly using high-frequency data. The final output multi-dimensional time-series data matrix possesses temporal consistency and feature integrity, providing a high-quality, structured data foundation for subsequent thermal behavior benchmark modeling, anomaly feature extraction, and multimodal fusion analysis, significantly improving the comprehensiveness and reliability of transformer equipment operating status perception.

[0029] Preferably, the process of constructing an adaptive thermal behavior benchmark model includes: Extract historical data from multiple complete operating cycles as historical long-cycle operating data; The local weighted scatter smoothing algorithm is used to perform time-series decoupling on the thermal parameters in the historical long-cycle operation data, and to separate the slow-change trend component reflecting the evolution of equipment insulation aging and the periodic seasonal component reflecting the seasonal change and diurnal temperature difference. To address the thermal inertia hysteresis effect caused by transformer volume and insulating oil, a multivariate nonlinear prediction unit is established. This unit uses historical grid load current and historical ambient temperature and humidity as input features, and the superposition of the gradually changing trend component and the periodic seasonal component as the benchmark prediction target. The multivariate nonlinear prediction unit introduces a dynamic delay factor for heat fluid conduction and a cumulative compensation coefficient for solar radiation. The backpropagation algorithm is used to optimize the parameters, and the resulting adaptive thermal behavior benchmark model can adaptively track fluctuations in grid load and micro-meteorological conditions.

[0030] First, historical data from multiple complete operating cycles are extracted from the historical repository of multidimensional time-series data to form historical long-cycle operating data. A complete operating cycle is defined as continuous time-series data covering at least one full calendar year to ensure that the data contains complete seasonal temperature variation patterns and complete diurnal temperature range cycles. The historical data includes all characteristic dimensions such as historical top-layer oil temperature, historical winding temperature, historical grid load current, and historical ambient temperature and humidity.

[0031] To decouple the thermal parameters (i.e., historical top-layer oil temperature and historical winding temperature) in historical long-term operating data, a locally weighted scatter smoothing algorithm is used. This algorithm is a non-parametric regression method. Its core idea is to select several neighboring data points within the time neighborhood of each data point in the target time series, assigning weighting coefficients based on distance from the target point (closer points have greater weights; typically, a cubic polynomial kernel function is used to calculate the weights). Then, a low-order polynomial (usually a first- or second-order polynomial) is fitted to the weighted local dataset, and the function value of this polynomial at the target point is used as the smoothed output value. This process is repeated point-by-point to complete the smoothing of the entire time series.

[0032] The advantage of the local weighted scatter smoothing algorithm lies in its good adaptability to local nonlinear structures and the fact that it does not require a pre-defined global function form. It is suitable for handling the complex nonlinear time series patterns of transformer thermal parameters caused by the combined effects of equipment aging, load fluctuations, and seasonal changes.

[0033] After time series smoothing, two types of components are further extracted from the smoothing results. The first type is the gradually changing trend component, which reflects the slow rise of the thermal parameter baseline caused by irreversible aging of transformer insulation materials under long-term thermal and electrical stress, with a time scale of several months to several years. The extraction method is to perform secondary smoothing with a larger smoothing window (e.g., 30 days) on the basis of local weighted scatter smoothing, filtering out seasonal fluctuations and retaining only the ultra-long-term trend. The second type is the periodic seasonal component, which reflects the periodic modulation effect of environmental temperature changes caused by seasonal alternation and diurnal temperature range on transformer temperature. The extraction method is to subtract the gradually changing trend component from the smoothed time series, retaining the annual seasonal component with a period of 1 year and the daily periodic component with a period of 1 day. The superposition of the two is the periodic seasonal component.

[0034] After extracting the gradually varying trend component and the periodic seasonal component, the two are superimposed as the baseline prediction target to establish a multivariate nonlinear prediction unit. The input features of the multivariate nonlinear prediction unit are historical power grid load current and historical ambient temperature and humidity, and the output target is the superimposed value of the aforementioned gradually varying trend component and periodic seasonal component. This unit adopts a multi-layer feedforward neural network architecture, which, for example, includes one input layer, two to three hidden layers, and one output layer. The activation function uses a linear rectified function to introduce nonlinear mapping capability.

[0035] In the model design of the multivariate nonlinear prediction unit, two key correction parameters are introduced to improve the model's adaptability to the physical characteristics of transformers. The first is the thermal fluid conduction dynamic delay factor, used to model the thermal inertia hysteresis effect of the insulating oil inside the transformer as a heat conduction medium. This means that after a change in load current, the heat generated takes a certain amount of time to be conducted to the top oil temperature sensor via oil convection. This delay time is related to the viscosity, density, and circulation path of the insulating oil, and is typically several minutes to tens of minutes. In the model, this delay effect is quantified by introducing a time lag term (i.e., using historical current values ​​from several minutes prior to the current moment as additional input features) to the input load current characteristics. The second is the solar radiation cumulative compensation coefficient, used to compensate for the additional heat absorption effect of the transformer tank caused by solar radiation. This effect cannot be directly reflected by current and ambient temperature characteristics. In the model, this is compensated by introducing an estimated theoretical solar radiation intensity calculated based on geographical location and calendar time as an additional input feature. The backpropagation algorithm is used to iteratively optimize all parameters of the multivariate nonlinear prediction unit. The mean square error between the prediction output on the training set and the benchmark target is used as the loss function. An adaptive moment estimation optimizer is used to perform gradient descent updates until the loss function converges. Finally, an adaptive thermal behavior benchmark model that can adaptively track the fluctuations of power grid load and micro-meteorological conditions is output.

[0036] Furthermore, the specific process for obtaining the dynamic delay factor of the heat fluid conduction includes: When extracting historical long-cycle operating data, a sliding time window is used to filter historical data segments where the grid load current undergoes a step change. The load current time-series variation curve and the top oil temperature time-series variation curve are extracted from the historical data segments respectively. The sliding cross covariance calculation is performed on the load current time-series variation curve and the top oil temperature time-series variation curve in the time axis direction, and the time offset corresponding to the highest peak value of the cross covariance is extracted. The time offset is used as the heat conduction lag time of the transformer under seasonal and load base conditions. A time delay matrix is ​​constructed using the heat conduction lag time, and the time delay matrix is ​​used as the dynamic delay factor of heat fluid conduction and fused into the input layer network weight of the multivariate nonlinear prediction unit.

[0037] In actual substation operation, transformers contain a large amount of insulating oil. When the grid load suddenly increases significantly, the heat generated in the windings rises instantaneously. However, this heat needs to be slowly transferred to the oil temperature sensor at the top of the transformer through the convection circulation of the insulating oil. First, we identified typical periods in historical data where the current increased by more than 30% of the rated value within a short period. Then, we extracted the current data curve and the top oil temperature data curve for this period separately. Using a cross-covariance algorithm in mathematics, we gradually slid the current curve backward on the time axis, continuously calculating the correlation score between the two curves. When we found that the two curves had the highest overlap after sliding backward for five minutes, it indicated that the actual heat transfer lag time of this transformer under this operating condition was five minutes. This was used as a dynamic delay factor for heat fluid conduction and input into the neural network.

[0038] This invention accurately replicates the unique thermal inertia physical properties of large solid-liquid mixing equipment by precisely calculating the time offset corresponding to the peak value of the cross-covariance. This delay factor, dynamically calculated based on real historical data, gives the benchmark model a strong sense of physical reality, enabling it to cope with drastic fluctuations in grid load and avoid prediction distortion and false fire alarms caused by the time difference in heat conduction.

[0039] Furthermore, the specific process for obtaining the cumulative solar radiation compensation coefficient includes: Obtain the specific geographical latitude and longitude coordinates of the transformer equipment's deployment location; combine the current calendar date and daily time, and calculate the theoretical solar altitude angle and theoretical clear-sky solar radiation intensity of the transformer equipment's location based on an astronomical trajectory algorithm; simultaneously retrieve the current cloud cover ratio and weather status data released by the meteorological department, and use the cloud cover ratio to perform physical attenuation conversion on the theoretical clear-sky solar radiation intensity to obtain the actual transient solar radiation intensity reaching the transformer tank surface; perform discrete integral calculation on the time axis on the transient solar radiation intensity over several consecutive hours to obtain the cumulative solar heat absorbed by the tank surface; map this solar heat value to a preset compensation activation function and output the corresponding cumulative solar radiation compensation coefficient.

[0040] Outdoor transformers are exposed to direct sunlight for extended periods during the summer daytime, causing their black metal casings to absorb heat. Before operation, the substation's precise longitude and latitude information is recorded. During daily operation, based on the current precise time, astronomical algorithms accurately calculate the sun's altitude and azimuth angles, determining the solar radiation intensity without cloud cover. Subsequently, real-time cloud thickness data is obtained via a meteorological interface. If the current weather is cloudy, for example, with a 50% cloud cover, the theoretical radiation intensity is halved to obtain the actual transient radiation intensity. Since heat absorption by the casing is a cumulative process, the system adds up the actual radiation intensities from past time windows to calculate how much solar heat the transformer casing has absorbed during that period. Finally, the accumulated heat is input into a nonlinear activation function to calculate a specific compensation coefficient, used to adjust the expected surface temperature increase accordingly.

[0041] This invention simulates the actual physical heat absorption and accumulation process of a metal enclosure by integrating geographical latitude and longitude, astronomical solar altitude angle, and real-time meteorological cloud cover data, and employing discrete integral calculations within a time window. The obtained compensation coefficients can accurately isolate environmental interference heat from the external sun, ensuring that the extracted residual sequence only reflects the internal fault heating of the transformer, greatly improving the purity of fire identification under complex meteorological conditions.

[0042] This invention achieves refined differentiation of thermal driving factors at different time scales by performing local weighted scatter smoothing time-series decoupling on historical long-term operating data. This is achieved by separating the transformer temperature time series into a slowly varying trend component reflecting the long-term health status of the equipment and a seasonal periodic component reflecting the influence of periodic environmental factors. In the design of the multivariate nonlinear prediction unit, a dynamic delay factor for heat fluid conduction is introduced to explicitly model the thermal inertia hysteresis of the insulating oil. Furthermore, a cumulative compensation coefficient for solar radiation is used to compensate for the often-neglected additional heat source, solar radiation. This allows the benchmark model to output high-precision expected normal temperatures under different seasons, load conditions, and solar radiation environments, effectively eliminating the interference of environmental load factors on thermal anomaly judgment. This provides a highly reliable benchmark reference for subsequent thermal anomaly residual extraction and significantly reduces the false alarm rate caused by normal load fluctuations.

[0043] Preferably, the process of obtaining the thermal anomaly residual sequence includes: The measured thermal data and current environmental load data are extracted from multidimensional time series data. The current environmental load data is input into the adaptive thermal behavior benchmark model, and the model is solved by combining the feedback parameters of the current start-stop status of the transformer's cooling equipment. The theoretically expected normal temperature benchmark value that should appear under the current actual operating conditions is output. The measured thermal data specifically includes the top oil temperature and winding temperature. To address the interference of localized sudden drops in transformer surface temperature caused by micro-meteorological fluctuations, a dynamic differential unit is constructed between the expected normal temperature baseline value and the measured thermodynamic data. The dynamic differential unit smooths negative temperature abrupt changes caused by rapid external cooling and dynamically compensates for the deviation of external cold sources in the thermal balance calculation. The smoothed continuous differential results are output as a thermal anomaly residual sequence.

[0044] The measured thermal data at the current moment is extracted from the real-time updated multi-dimensional time series data. Specifically, it includes the current top oil temperature measurement value and the current winding temperature measurement value, as well as the current environmental load data, including the current ambient temperature, the current ambient humidity and the current grid load current.

[0045] The aforementioned current environmental load data is input into the trained adaptive thermal behavior baseline model to perform extrapolation calculations. During the extrapolation calculation process, it is also necessary to combine the feedback parameters of the current start-stop status of the transformer's cooling equipment for joint calculation.

[0046] The start / stop status feedback parameter of the heat dissipation and air cooling equipment refers to the current operating status of the transformer's external cooling devices (including cooling fan groups and oil pumps). The operation of the cooling devices will significantly reduce the transformer tank temperature. If it is not included in the calculation, it will lead to an overestimation of the baseline prediction value, resulting in a systematic negative bias in the residual sequence. A binary code of the start / stop status of the cooling devices (1 for running, 0 for stopping) is added as an auxiliary input feature to the input of the adaptive thermal behavior baseline model. After integrating all input features, the model outputs the theoretically expected normal temperature baseline value that should appear under the current actual operating conditions. This value is output independently for the top oil temperature and winding temperature, forming the corresponding theoretical expected temperature baseline vector.

[0047] After obtaining the measured thermodynamic data and the expected normal temperature baseline value, theoretically, the thermal anomaly residual can be obtained by directly subtracting them. However, in actual engineering environments, a special interference phenomenon exists: micro-meteorological fluctuations (such as sudden rainfall, cold wind gusts, etc.) can cause the transformer surface temperature to drop sharply by several degrees within minutes, resulting in the measured temperature being briefly lower than the baseline prediction value, forming a large negative differential abrupt change. If left untreated, such negative temperature abrupt changes will appear as obvious negative pulses in the residual sequence, which may interfere with the threshold judgment of subsequent anomaly detection, or even mask the true thermal anomaly signal. To address this problem, a dynamic differential unit is constructed between the expected normal temperature baseline value and the measured thermodynamic data.

[0048] The working mechanism of the dynamic differential unit is as follows: First, the original difference between the measured temperature and the baseline predicted temperature at the current moment is calculated. Then, a smoothing filter based on an exponentially weighted moving average is introduced into this differential time series. Its time constant is set to match the transformer's thermal time constant (typically on the order of 10 to 30 minutes). This smooths and suppresses sharp negative abrupt changes in the differential time series. That is, when the differential value experiences a large negative jump in a short period, the memory property of the exponentially weighted moving average ensures that the smoothed residual does not immediately jump to a negative value of the same magnitude, but rather decreases slowly at an exponential decay rate, effectively filtering out brief negative disturbances caused by rapid external cooling. Simultaneously, the rate of decrease of the differential value is analyzed to determine whether the current negative change exceeds the physical rate limit of the transformer's natural cooling. If it does, the deviation of the excess portion on the thermal balance calculation is dynamically compensated, bringing the residual sequence back to a reasonable physical range. The physical rate limit of the transformer's natural cooling is calculated using Newton's law of cooling combined with the transformer's specific heat capacity. The smoothed continuous differential result sequence is output as the thermal anomaly residual sequence for subsequent multimodal fusion analysis.

[0049] Furthermore, auxiliary input features of the start-stop status of the cooling device are added to the input of the adaptive thermal behavior benchmark model. The specific process includes: real-time acquisition of the electrical contact closure status of the external cooling fan and forced oil circulation pump of the transformer equipment through the substation monitoring system; conversion of the acquired closure status into binary codes, where the running state is encoded as the Arabic numeral 1 and the stopped state is encoded as the Arabic numeral 0, and the binary code is concatenated as an independent auxiliary input feature vector to the input layer of the adaptive thermal behavior benchmark model; the adaptive thermal behavior benchmark model integrates historical load, micro-meteorological parameters and the auxiliary input features, and performs nonlinear calculation through a hidden layer network, and sets two independent neuron nodes at the output end; the two neuron nodes independently output the expected normal top oil temperature benchmark value and the expected normal winding temperature benchmark value that should theoretically appear under the current real operating conditions, and combine the two to form the corresponding theoretical expected temperature benchmark vector.

[0050] Large transformers are typically equipped with multiple sets of forced air cooling or oil-water cooling devices. The operational status of these fans and oil pumps is determined by directly reading the contact signals of the industrial relays controlling them. A fan running at high speed is assigned a value of 1; a fan stopped is assigned a value of 0. This sequence of 1s and 0s, along with ambient temperature and load current, is input into a neural network. As the neural network processes this data, its internal multilayer perceptron structure learns that encountering the Arabic numeral 1 signifies strong external cooling. In the final layer of the network, two completely separate output ports are designed: one port calculates the optimal oil temperature at the top of the transformer, and the other calculates the optimal temperature of the innermost winding copper wire. Finally, these two temperature values ​​are packaged into a vector, serving as the current normal temperature reference standard.

[0051] This invention introduces concise binary encoding, enabling the model to recognize artificial physical intervention in cooling. Simultaneously, it employs an architecture with independent dual-node outputs for top-layer oil temperature and winding temperature because the winding is close to the heat-generating core and cools extremely rapidly, while the top-layer oil accumulates at the top and cools slowly. This independent dual-output completely decouples these two monitoring points with drastically different thermodynamic properties, providing a precise benchmark for subsequently extracting thermal anomaly residuals from different locations.

[0052] This invention addresses the engineering challenge of false negative anomalies in outdoor transformer equipment under complex weather conditions. It constructs a dynamic differential unit between measured thermodynamic data and a baseline predicted temperature, and introduces an exponentially weighted moving average mechanism to smooth and suppress negative temperature abrupt changes caused by micro-meteorological disturbances. This is achieved by smoothing and suppressing negative temperature jumps caused by micro-meteorological disturbances. By incorporating the start-stop status feedback parameters of the cooling system into the baseline model's derivation input, the systematic bias in thermal balance calculations caused by the switching actions of the cooling device is avoided. The final output thermal anomaly residual sequence eliminates the influence of environmental load fluctuations, seasonal variations, cooling device status changes, and micro-meteorological disturbances, accurately reflecting the deviation of the transformer equipment's internal thermal state from the normal operating baseline. This provides a high signal-to-noise ratio thermal feature input for subsequent multi-modal anomaly fusion and judgment.

[0053] Preferably, the process for confirming the abnormal node includes: The thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from the multidimensional time series data are synchronously input into the multi-head attention fusion network; the non-thermal heterogeneous time series data includes partial discharge high-frequency pulses and box mechanical vibrations; To address the issue of data packet loss from single sensors in the complex and strong electromagnetic environment of transformer substations, an adaptive channel masking mechanism is introduced at the network input to dynamically suppress missing and abrupt data. Inside the network, cross-modal attention weights are used to calculate the degree of physical coupling response between thermal abrupt changes and mechanical vibration of the enclosure, and cross-modal derived feature vectors are extracted. These cross-modal derived feature vectors are then mapped to multiple basic probability assignment functions containing independent fire risk assessment probabilities as evidence. Calculate the logical conflict coefficient between each piece of evidence. If the conflict coefficient is lower than the safety limit, use the Dempster synthesis rule to perform orthogonal sum operation to complete dynamic comprehensive risk fusion, identify abnormal nodes and output abnormal node characteristics.

[0054] The thermal anomaly residual sequence and non-thermal heterogeneous time-series data extracted from multidimensional time-series data are synchronously input into a multi-head attention fusion network. The non-thermal heterogeneous time-series data includes partial discharge high-frequency pulse time series and enclosure mechanical vibration time series. The partial discharge high-frequency pulse time series, after alignment and reconstruction, is represented by three statistical features: peak count, mean peak amplitude, and maximum peak amplitude within its window. The enclosure mechanical vibration time series is represented by three statistical features: root mean square value of vibration acceleration within its window, amplitude of dominant frequency component, and harmonic distortion. These features, together with the thermal anomaly residual sequence, constitute the input feature tensor.

[0055] Since the thermal anomaly residual sequence is a one-dimensional scalar time series, while the partial discharge and mechanical vibration features are multi-dimensional vectors, a dimension alignment module is set up before calculating the dot product similarity: a single-layer fully connected neural network is used to increase the dimension of the one-dimensional thermal anomaly residual sequence, and then project it together with the partial discharge feature sequence and the vibration feature sequence into a latent space of the same dimension. Subsequently, the projected thermal anomaly residual sequence is used as the query vector, and the projected partial discharge feature sequence and vibration feature sequence are used as the key vector and value vector, respectively, to calculate the degree of cross-modal physical coupling response between thermal mutations and non-thermal signals.

[0056] Considering the strong electromagnetic interference environment at the transformer substation, devices such as partial discharge sensors are prone to data packet loss during data transmission. This means that sampled values ​​at certain times cannot be effectively uploaded, resulting in missing values ​​or abrupt changes to zero in the input data. If such missing data is directly input into the network, it will cause serious deviations in the network's judgment at that moment. To address this, an adaptive channel masking mechanism is introduced at the input of the multi-head attention fusion network: by monitoring the data integrity rate (i.e., the proportion of effective sampling points to all time steps) of each feature channel within the sliding window, the confidence weight of that channel is dynamically calculated. When the integrity rate is lower than a preset threshold (e.g., below 80%), the confidence weight of that channel is dynamically reduced to near zero. In subsequent attention calculations, the network will assign a very low weight to that channel, thereby achieving dynamic weight suppression of missing and abrupt data and preventing the impact of missing data from a single sensor from affecting the overall judgment.

[0057] The architecture of the multi-head attention fusion network is designed as follows: the network contains multiple parallel attention heads, each independently learning the correlation weights between different modes or different time steps. In the cross-modal attention calculation mechanism, the thermal anomaly residual sequence is used as the query vector, and the partial discharge feature sequence and vibration feature sequence are used as the key vector and value vector, respectively, to calculate the degree of cross-modal physical coupling response between thermal abrupt changes and non-thermal signals. The magnitude of the cross-modal attention weights reflects the degree to which the discharge signal and vibration signal respond in time to specific changes in thermal parameters. That is, if the winding overheating is accompanied by a sudden increase in discharge and an enhancement of vibration at a specific frequency, the attention weights among the three will increase significantly, indicating that the joint response mode has a high correlation with the fire situation. After concatenation and linear transformation of the calculation results from multiple attention heads, cross-modal derived feature vectors containing the spatiotemporal collaborative modes of thermal anomalies and non-thermal anomalies are extracted.

[0058] Specifically, the multi-head attention fusion network includes an input layer, three cascaded attention computation layers, a fully connected fusion layer, and an output layer connected in sequence. The attention computation layer has 4 to 8 parallel attention heads, each of which independently learns the correlation weights between different modalities. The input layer has a dimension alignment module, which is used to upscale the one-dimensional thermal anomaly residual sequence to the same latent space dimension as the partial discharge feature sequence and vibration feature sequence through a single-layer fully connected neural network. The computation results of each attention head are integrated through a splicing operation, and then processed by linear transformation and nonlinear activation function to output a cross-modal derived feature vector.

[0059] After obtaining the cross-modal derived feature vectors, they are mapped to multiple basic probability assignment functions through a fully connected layer, serving as evidence bodies within the evidence theory framework. Each evidence body contains an independent probability assessment of the existence of a fire risk at the current moment, a probability assessment of the absence of a fire risk, and a probability assignment for uncertainty in the current judgment. These multiple evidence bodies originate from independent inferences from different attention heads within the network, forming a mutually corroborating multi-evidence system. Before synthesizing multiple evidence bodies, the logical conflict coefficient between each evidence body is first calculated. The logical conflict coefficient reflects the inconsistency in the degree of support two evidence bodies have for the same proposition; a higher value indicates a more contradictory judgment between the two evidence bodies. If the conflict coefficient is below a preset safety limit (usually set to 0.6), it indicates good consistency in judgments among the modal sensors. In this case, the Dempster synthesis rule is used to perform an orthogonal sum operation on all evidence bodies, synthesizing the basic probability assignments of each evidence body into a unified dynamic comprehensive risk value. If the conflict coefficient exceeds the safety limit, direct synthesis is abandoned, and a weighted average method is used to synthesize the evidence bodies, with a warning message indicating low judgment credibility in the output.

[0060] Specifically, the mapping process of the basic probability assignment function includes: the cross-modal derived feature vectors are mapped through a fully connected layer to output a three-dimensional vector, which represents the probability values ​​of three mutually exclusive propositions: "there is a fire risk", "there is no fire risk" and "the judgment is uncertain"; each attention head outputs an independent evidence body; the conflict coefficient is quantified by calculating the cosine distance between the evidence body vectors; when the conflict coefficient exceeds 0.6, a weighted average method is used to replace the Dempster rule for evidence fusion.

[0061] Finally, based on the comparison between the comprehensive risk value and the preset anomaly judgment threshold, it is confirmed whether there is an abnormal node at the current moment, and the node feature vector of the abnormal node is output for subsequent graph neural network inference.

[0062] This invention achieves cross-modal deep fusion of thermal anomaly residuals and non-thermal signals such as partial discharge and mechanical vibration through a multi-head attention fusion network, overcoming the limitation of single physical domain signals being susceptible to environmental interference and misjudgment in fire identification. An adaptive channel masking mechanism is introduced to effectively address data loss issues in strong electromagnetic environments, ensuring the network's robust judgment capability under conditions of incomplete sensor data. An evidence theory framework is employed to perform conflict determination and dynamic synthesis of the inference results from multiple independent attention heads, achieving explicit quantitative evaluation of the consistency of multimodal evidence. This avoids judgment distortion caused by forced synthesis of conflicting evidence, significantly improving the accuracy and noise robustness of anomaly node confirmation.

[0063] Preferably, the process of obtaining the fire cause evolution chain includes: A directed knowledge graph of fire causes is pre-constructed, in which phenomenon nodes include sensor distortion feature parameters such as temperature exceeding limits and sudden increase in discharge, mechanism nodes include hidden physical defect parameters such as winding loosening and deformation and insulation medium aging and peeling, and fire cause nodes are topologically connected through directed edges with historical real occurrence conditional probability weights. The node features of the abnormal nodes are received and transformed into initial activation sources in the graph structure, and then input into the multi-layer graph neural network. Under the rigid prior topological path constraints of the fire cause knowledge graph, message passing between neighboring nodes and high-dimensional feature aggregation reasoning are performed. The connected subgraph that maximizes the joint activation probability is found through global search optimization, and the most complete fire cause evolution link from the current multi-dimensional physical anomaly starting point to the final local deflagration endpoint is output, and the corresponding comprehensive confidence value is output simultaneously.

[0064] A directed knowledge graph of fire causes is pre-constructed. This knowledge graph consists of two types of nodes and one type of directed edges. The first type of nodes are phenomenon nodes, which contain abnormal characterization parameters observable by sensors, such as the top oil temperature exceeding the limit node (parameters are the magnitude and duration of the oil temperature exceeding the rated value), the winding temperature exceeding the limit node (parameters are the abnormal winding temperature gradient), the partial discharge increase node (parameters are the rate of increase of the discharge pulse count per unit time), and the vibration abnormal node (parameters are the amplitude abnormalities of specific frequency components), etc. The parameters of these phenomenon nodes directly correspond to the abnormal node features output by the aforementioned multi-head attention fusion network. The second type of nodes are mechanism nodes, which contain descriptive parameters of hidden physical defects inside the equipment, such as the winding loosening and deformation node (parameters are the estimated axial displacement of the winding), the insulation medium aging and peeling node (parameters are the estimated degree of polymerization of cellulose in the insulation paper), and the oil passage partial blockage node (parameters are the abnormal decrease in the thermal conductivity of the oil flow), etc. The parameters of these mechanism nodes are derived from domain knowledge of the physical mechanisms of transformer faults, reflecting the essential causes behind various sensor anomalies. Fire causal nodes (including phenomenon nodes and mechanism nodes) are topologically connected through directed edges with conditional probability weights based on historical occurrences. The conditional probability weight of each directed edge is calculated using the maximum likelihood estimation method. Specifically, the frequency of simultaneous occurrence of a preceding fault node (such as winding temperature exceeding the limit) and a subsequent fault node (such as winding loosening and deformation) in the historical transformer fault case database is statistically analyzed. This frequency is divided by the total frequency of preceding fault nodes, and the calculated ratio is used as the prior probability weight of the directed edge. For example, the weight of the directed edge from winding temperature exceeding the limit to winding loosening and deformation represents the conditional probability of simultaneous winding loosening and deformation in historical cases under the condition of winding temperature exceeding the limit. Finally, the local deflagration node is used as the convergence endpoint node of the knowledge graph.

[0065] The abnormal node features confirmed in the previous step are received and transformed into the initial activation values ​​of the corresponding phenomenon nodes in the fire cause knowledge graph, forming the initial activation source set in the graph structure, which is then input into the multi-layer graph neural network.

[0066] The architecture of a multi-layer graph neural network is as follows: the network consists of multiple stacked graph convolutional layers. Each graph convolutional layer performs one round of neighbor node message passing and feature aggregation operations on all nodes in the graph. The message passing rules are as follows: each node sends its current feature vector (including activation probability and physical defect level estimate) to its neighbor nodes along the directed edges. The neighbor nodes sum the messages from all upstream nodes according to the conditional probability weights of the directed edges, and then update their own feature vectors through learnable linear transformations and nonlinear activation functions. This process is executed under the rigid prior topological path constraints of the knowledge graph, that is, messages can only propagate along the predefined directed edge directions in the knowledge graph, and are not allowed to propagate backward or across node pairs without directed edges. This strictly limits the feature propagation path of the graph neural network to the topological space that conforms to the physical mechanism of transformer faults, avoiding the physically unreasonable inferences that may occur in purely data-driven methods.

[0067] After the multi-layer graph convolutional layer is completed, a global search is performed on the final feature vectors of all nodes in the graph to find the connected subgraph path with the highest activation probability from the initial activation source node (i.e., the anomaly node) to the local deflagration endpoint node. This path maximizes the joint activation probability of the fire causal evolution chain. This chain completely describes the complete causal inference chain starting from the currently observable multidimensional physical anomaly, through the progressive deterioration evolution of each intermediate mechanism node, and finally reaching the local deflagration final state. The comprehensive confidence value of this chain is output simultaneously. This value is calculated by multiplying the conditional probability weights of all directed edges on the chain and then normalizing it, reflecting the degree of consistency between the inferred causal evolution path and historical failure patterns.

[0068] For example, the graph neural network adopts a graph convolutional network (GCN) architecture, which includes three graph convolutional layers. Each graph convolutional layer performs neighbor node message passing and feature aggregation operations. The message propagates along the predefined directed edge direction, and the aggregation is performed by weighted summation according to the directed edge conditional probability weights. After feature aggregation is completed, the node features are updated through learnable linear transformations and nonlinear activation functions. After multiple convolutions, the shortest path search from the initial activation source to the local deflagration endpoint is performed on the final feature vectors of all nodes in the graph, and the connected subgraph with the highest joint activation probability is taken as the fire cause evolution link.

[0069] This invention pre-constructs a directed fire causation knowledge graph containing phenomenon nodes and mechanism nodes, explicitly encoding the domain physical mechanism knowledge of transformer fire evolution in the form of a graph topology. This provides rigid prior constraints for feature propagation in graph neural networks, effectively preventing physically unreasonable inferences from purely data-driven methods when training samples are insufficient. The directed edges of the knowledge graph are quantified using historical real-occurrence conditional probability weights, making the reasoning conclusions probabilistically interpretable. The message passing and high-dimensional feature aggregation mechanisms of the multi-layer graph neural network can effectively handle complex scenarios where multiple abnormal nodes are activated simultaneously and multiple fault links are intertwined. The output fire causation evolution link not only locates the most dangerous causal path but also provides quantifiable comprehensive confidence, offering physically interpretable reasoning basis for maintenance personnel's decision-making.

[0070] Preferably, the process of generating a structured early warning report based on the fire cause and evolution chain retrieval matching and response strategy includes: The system receives the fire cause evolution chain and the underlying mechanism and inducing node included, performs a high-dimensional vector space similarity search in the pre-established transformer safety operation and maintenance strategy database, retrieves historical handling experience based on the highest matching degree, and generates a corresponding risk reduction and handling strategy; the risk reduction and handling strategy includes physical cooling, load shedding, and fire extinguishing device standby; Combining the risk transmission rate parameters in the fire cause evolution chain, the remaining rescue time window from the current deterioration state to the irreversible fire outbreak is calculated; the multi-source anomaly characteristic description, the inferred cause evolution conclusion, the remaining rescue time window, and the risk reduction and disposal strategy are assembled and recombined into a structured early warning report.

[0071] Receive the fire cause evolution link and the underlying mechanism causal nodes contained therein; the underlying mechanism causal nodes refer to the core mechanism nodes that are located near the starting point of the fire cause evolution link and directly drive the subsequent deterioration process, such as insulation medium aging and peeling nodes or oil channel partial blockage nodes. These nodes represent the fundamental physical causes of the current fire risk.

[0072] A high-dimensional vector space similarity retrieval is performed in a pre-established transformer safety operation and maintenance strategy database. This database contains a large number of historical transformer fault cases and their corresponding handling experiences. Each historical handling record is encoded as a high-dimensional feature vector, with each dimension component corresponding to a quantized encoding value of attributes such as different fault mechanism types, different anomaly degrees, and different equipment model parameters. The feature vectors of all nodes in the current fire cause evolution chain are processed through a feature compression module. This module introduces an average pooling operation to reduce the dimensionality of the link node feature vectors of varying lengths by averaging them over time or node sequence dimensions, compressing them into a fixed, fully connected vector with a standardized dimension, which serves as the retrieval query vector. The cosine similarity between the query vector and all historical handling record vectors is calculated in the high-dimensional vector space of the database. The vectors are sorted from highest to lowest similarity, and the highest similarity historical handling experiences are retrieved as references. Based on the highest matching degree, a corresponding risk mitigation strategy is generated.

[0073] Specifically, the historical handling records in the transformer safety operation and maintenance strategy database are encoded as 128-dimensional feature vectors, with each dimension corresponding to different fault mechanism types, abnormality levels, and quantitative encoding values ​​of equipment parameters; similarity retrieval uses the cosine similarity algorithm to calculate the cosine value of the angle between the query vector and the historical record vector, and selects the top 5 historical records with the highest similarity as reference.

[0074] The risk mitigation and response strategy includes the following three types of operational instructions: physical cooling operations, including forcibly activating all cooling fans and oil pumps to maximize heat dissipation capacity; load shedding operations, including reducing the grid load current flowing through the transformer to a specified percentage of the rated value through the dispatch system; and fire extinguishing device standby operations, including activating the fixed fire extinguishing devices on-site to a quasi-activated standby state and sending an early warning notification to the fire duty personnel.

[0075] The remaining rescue time window is calculated by combining the risk propagation rate parameters of each directed edge in the fire cause evolution chain. The risk propagation rate parameters reflect the speed at which the current fault state develops into the next level of deterioration along the evolution chain. The historical evolution trajectory of the comprehensive confidence level within continuous time steps is extracted in real time, and the instantaneous evolution velocity and evolution acceleration are calculated. A nonlinear exponential degradation model incorporating instantaneous evolution velocity and evolution acceleration parameters is used to curve-fit the remaining life evolution trajectory of the equipment. On the generated nonlinear fitting curve, the x-axis intersection point where the curve spatially intersects with the preset fire outbreak critical confidence threshold on the time axis is determined. The absolute time difference between the current moment and this x-axis intersection point is calculated, and this time difference is used as the remaining rescue time window from the current deterioration state to an irreversible fire outbreak, output in hours or minutes, providing maintenance personnel with a clear reference for action time limits.

[0076] The final structured early warning report is generated by assembling and reassembling the data. The report contains the following four structured modules: The first module is a description of the characteristics of multi-source anomalies, describing the anomalies detected by each sensor, their degree of anomaly, and their duration in natural language; the second module is the inferred causal evolution conclusions, showing the complete causal evolution chain from the current anomaly to the local deflagration endpoint and the comprehensive confidence level of each node; the third module is the remaining rescue time window, displaying the calculated remaining response time limit; the fourth module is the risk mitigation strategy, listing specific operational instructions and execution parameters such as physical cooling, load shedding, and fire extinguishing equipment standby in priority, and outputting a complete structured early warning report.

[0077] This invention utilizes a high-dimensional vector space similarity retrieval mechanism to accurately match historical handling experience with the current causes of faults, avoiding the incomplete coverage problem of static handling strategies based on rule tables when facing complex and intertwined fault modes. It introduces a risk transmission rate parameter to quantify the remaining rescue time window, transforming abstract risk confidence into action time limits with clear temporal significance, significantly enhancing the practical guidance value of early warning information for maintenance personnel. The structured early warning report standardizes and assembles four types of information—abnormal phenomenon description, causal reasoning conclusions, time limit reminders, and handling instructions—forming a clear and comprehensive report format. This not only meets the needs of maintenance personnel to quickly understand the current risk situation but also provides directly executable operational instructions, effectively shortening the response time from early warning issuance to intervention action and reducing the probability of fire and explosion in transformer equipment.

[0078] Example 2: This invention proposes a transformer fire detection and linkage control system based on operational monitoring data. The structure of the system is as follows: Figure 3 As shown, it includes: a data acquisition module, a data processing module, an anomaly detection module, and an anomaly warning module; The data acquisition module synchronously collects heterogeneous physical quantities and environmental load data of the transformer equipment based on edge nodes, and constructs multi-dimensional time-series data. Furthermore, after obtaining discrete time-series data with a unified timestamp, to address the issue of uneven sampling frequencies caused by differences in the physical properties of different sensors, it also includes downsampling by extracting high-frequency signal envelope features. Using the sensor with the lowest sampling frequency as the baseline target resolution, downsampling processing based on envelope extraction is performed on the high-frequency sensor data. This involves extracting the statistical feature envelope of the high-frequency data within each target time window (e.g., extracting the peak value, mean peak amplitude, and root mean square value within the window), compressing the high-frequency waveform features within the window into a low-frequency feature vector representing the energy state of that frequency band. For low-frequency sensor data that has already been sampled within the baseline time window, it is directly used; for moments within the window without new sampling, nearest neighbor interpolation is used to fill the gap. This operation ensures that all physical quantities have completely corresponding sampling points on the time axis, preserving the transient physical characteristics of the high-frequency signal while eliminating spurious correlation interference caused by time misalignment.

[0079] The data processing module extracts historical long-cycle operating data, performs seasonal trend smoothing decomposition and multivariate nonlinear regression, and constructs an adaptive thermal behavior benchmark model. It extracts measured thermal data from the multidimensional time series data, compares the measured thermal data with the adaptive thermal behavior benchmark model, and generates a thermal anomaly residual sequence that eliminates environmental interference. The process of obtaining the thermal anomaly residual sequence includes: directionally extracting the measured thermal data and the current environmental load data from the multidimensional time series data; inputting the current environmental load data into the adaptive thermal behavior benchmark model, and performing calculations in conjunction with the feedback parameters of the current start-stop status of the transformer's cooling and air-cooling equipment, and outputting the expected normal temperature benchmark value that should theoretically occur under the current actual operating conditions; the measured thermal data specifically includes the top oil temperature and winding temperature; To address the interference of localized sudden drops in transformer surface temperature caused by micro-meteorological fluctuations, a dynamic differential unit is constructed between the expected normal temperature baseline value and the measured thermodynamic data. The dynamic differential unit smooths negative temperature abrupt changes caused by rapid external cooling and dynamically compensates for the deviation of external cold sources in the thermal balance calculation. The smoothed continuous differential results are output as a thermal anomaly residual sequence.

[0080] To address the imbalance of spatial temperature gradients in transformers caused by severe convective weather, an anisotropic cooling compensation mechanism is introduced when constructing the dynamic differential unit. The specific process includes: deploying independent surface temperature sensors on the four exterior facades (east, south, west, and north) of the transformer enclosure to acquire multi-dimensional measured surface thermal data; calculating the spatial temperature range between the windward and leeward temperature sensors in real time; triggering the anisotropic cooling compensation module when the spatial temperature range exceeds a preset thermal balance tear threshold; assigning asymmetrical external cold source deviation compensation coefficients to the temperature measurement points on the four facades based on wind direction and speed sensor data; independently smoothing and compensating the differential results for each facade using these asymmetrical external cold source deviation compensation coefficients; and finally, spatially weighting and fusing the compensated values ​​from each facade to output a global thermal anomaly residual sequence.

[0081] In coastal areas or open spaces, transformers frequently encounter extreme convective weather such as typhoons and torrential rains. In these situations, extreme weather often only violently impacts the windward side of the transformer, causing a sharp drop in temperature on the windward side, while the temperature on the leeward side decreases slowly. This severe imbalance in the spatial temperature gradient not only leads to mechanical noise caused by the contraction and cooling of one side of the transformer housing, but also renders the data from a single-point temperature sensor unrepresentative globally, resulting in significant calculation errors by the dynamic differential unit when compensating for external cold sources.

[0082] For example, temperature probes were installed on all four sides of the transformer's massive enclosure. When a sudden downpour accompanied by strong winds struck, the temperature of the probe on the windward north side dropped by 15 degrees Celsius within three minutes, while the temperature of the probe on the leeward south side dropped by only 2 degrees Celsius. At this point, the calculated temperature difference between the north and south sides reached 13 degrees Celsius, exceeding the preset thermal equilibrium tear threshold. Anemometer data was then retrieved, confirming a strong northerly wind. The anisotropy compensation module was activated, assigning a higher cold source compensation coefficient to the north probe to offset the scouring effect of the downpour, while assigning a lower compensation coefficient to the south probe. After independent dynamic compensation calculations, the data from both the north and south sides were restored to the true internal heating level, eliminating the influence of weather. Finally, the restored data from these four sides were weighted and averaged to obtain a unique residual sequence representing the overall true thermal state of the transformer.

[0083] Specifically, the current rainfall per unit time and the decrease in ambient temperature are retrieved, and the preset convective heat transfer mapping relationship is substituted to calculate the global basic compensation base representing the ideal cooling capacity of the current weather. Retrieve instantaneous wind direction angle data and calculate the spatial angle between the real-time wind direction vector and the normal vectors of the four exterior facades of the transformer. Using the cosine mapping relationship of this angle, calculate the independent windward projection factor of each of the four exterior facades (the smaller the angle, the larger the factor; the leeward side approaches zero). Extract the instantaneous wind speed value, substitute it into the empirical curve of wind-cooled convection heat transfer, and obtain the dynamic wind speed amplification factor of nonlinear surge to quantify the error interference of strong winds and rain forcibly carrying away heat. For each independent facade, the global basic compensation base, the windward projection factor corresponding to that facade, and the dynamic wind speed amplification factor are multiplied together to accurately output the dynamic cold source compensation coefficients for the four facades (east, south, west, and north) where the values ​​are not equivalent. The convective heat transfer basic mapping relationship and the wind-cooled convective heat transfer empirical curve described in this invention do not use fixed, purely theoretical physical formulas. Instead, they are constructed by pre-collecting massive amounts of historical micro-meteorological and temperature fluctuation data from the site where the equipment is located, and using machine learning algorithms for multivariate nonlinear regression training, thereby building a data-driven empirical mapping surface that fits the individual physical characteristics of the equipment.

[0084] Traditional single-point monitoring or full-scale average compensation methods, when facing typhoons and heavy rain, mistakenly equate rapid local cooling with overall temperature drop, leading to over-compensation or under-compensation. This solution, by deploying multi-dimensional spatial probes and introducing anisotropic compensation coefficients based on wind direction and speed, reconstructs the complex three-dimensional heat dissipation process of the transformer under the impact of strong winds and heavy rain. This enables the linkage control system to accurately identify early thermal runaway fires inside the transformer during extremely severe convective weather.

[0085] The anomaly detection module synchronously inputs the thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from multidimensional time series data into a multi-head attention fusion network to extract cross-modal derived features, transforms the feature vector into evidence, and uses evidence theory to perform conflict detection and dynamic comprehensive risk fusion to confirm the anomaly node. The anomaly early warning module constructs a fire cause knowledge graph containing phenomenon nodes and mechanism nodes. The node features of the anomaly nodes are input into the graph neural network, and mapping reasoning is performed under the prior topological constraints of the knowledge graph. The module outputs the fire cause evolution link and confidence level under multiple anomalies. Based on the fire cause evolution link, a matching and disposal strategy is retrieved to generate a structured early warning report.

[0086] 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 method for fire identification in oil-immersed transformer equipment based on operational monitoring data, characterized in that, include: Data on heterogeneous physical quantities and environmental loads of transformer equipment are collected synchronously based on edge nodes, and multidimensional time-series data are constructed. Historical long-term operating data is extracted and seasonal trend smoothing decomposition and multivariate nonlinear regression are performed to construct an adaptive thermal behavior benchmark model; measured thermal data are extracted from the multidimensional time series data, and the measured thermal data are compared with the adaptive thermal behavior benchmark model to generate a thermal anomaly residual sequence that removes environmental interference. The thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from multidimensional time series data are synchronously input into a multi-head attention fusion network to extract cross-modal derived features. The feature vector is transformed into an evidence body and conflict determination and dynamic comprehensive risk fusion are performed using evidence theory to confirm the anomaly node. A fire cause knowledge graph containing phenomenon nodes and mechanism nodes is constructed. The node features of abnormal nodes are input into a graph neural network. Under the prior topological constraints of the knowledge graph, mapping reasoning is performed, and the fire cause evolution link and confidence level under multiple anomalies are output. Based on the fire cause evolution link, a matching and treatment strategy is retrieved to generate a structured early warning report.

2. The method for fire identification of oil-immersed transformer equipment based on operation monitoring data according to claim 1, characterized in that: The process of collecting and constructing multidimensional time series data includes: An edge node sensing array is deployed at the physical monitoring points of the transformer equipment to acquire raw heterogeneous signals; wherein, the heterogeneous physical quantities include top oil temperature, winding temperature, partial discharge high-frequency pulse, and enclosure mechanical vibration; the environmental load data includes ambient temperature and humidity and grid load current; By using a time-sensitive network edge gateway, the collected heterogeneous physical quantities and environmental load data are timestamped with a unified clock cycle to obtain discrete time-series data. To address the uneven sampling frequency caused by differences in the physical properties of different sensors, alignment and reconstruction are performed. The aligned and reconstructed data are orthogonally stitched together according to the time dimension and the feature dimension to output multidimensional time-series data with temporal consistency.

3. The method for fire identification of oil-immersed transformer equipment based on operation monitoring data according to claim 1, characterized in that: The process of constructing an adaptive thermal behavior benchmark model includes: Extract historical data from multiple complete operating cycles as historical long-cycle operating data; The local weighted scatter smoothing algorithm is used to perform time-series decoupling on the thermal parameters in the historical long-cycle operation data, and to separate the slow-change trend component reflecting the evolution of equipment insulation aging and the periodic seasonal component reflecting the seasonal change and diurnal temperature difference. To address the thermal inertia hysteresis effect caused by transformer volume and insulating oil, a multivariate nonlinear prediction unit is established. This unit uses historical grid load current and historical ambient temperature and humidity as input features, and the superposition of the gradually changing trend component and the periodic seasonal component as the benchmark prediction target. The multivariate nonlinear prediction unit introduces a dynamic delay factor for heat fluid conduction and a cumulative compensation coefficient for solar radiation. The backpropagation algorithm is used to optimize the parameters, and the resulting adaptive thermal behavior benchmark model can adaptively track fluctuations in grid load and micro-meteorological conditions.

4. The method for fire identification of oil-immersed transformer equipment based on operation monitoring data according to claim 1, characterized in that: The process of obtaining the thermal anomaly residual sequence includes: The measured thermal data and current environmental load data are extracted from multidimensional time series data. The current environmental load data is input into the adaptive thermal behavior benchmark model, and the model is solved by combining the feedback parameters of the current start-stop status of the transformer's cooling equipment. The theoretically expected normal temperature benchmark value that should appear under the current actual operating conditions is output. The measured thermal data specifically includes the top oil temperature and winding temperature. To address the interference of localized sudden drops in transformer surface temperature caused by micro-meteorological fluctuations, a dynamic differential unit is constructed between the expected normal temperature baseline value and the measured thermodynamic data. The dynamic differential unit smooths negative temperature abrupt changes caused by rapid external cooling and dynamically compensates for the deviation of external cold sources in the thermal balance calculation. The smoothed continuous differential results are output as a thermal anomaly residual sequence.

5. The method for fire identification of oil-immersed transformer equipment based on operation monitoring data according to claim 1, characterized in that: The process for confirming the abnormal node includes: The thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from the multidimensional time series data are synchronously input into the multi-head attention fusion network; the non-thermal heterogeneous time series data includes partial discharge high-frequency pulses and box mechanical vibrations; To address the issue of data packet loss from single sensors in the complex and strong electromagnetic environment of transformer substations, an adaptive channel masking mechanism is introduced at the network input to dynamically suppress missing and abrupt data. Inside the network, cross-modal attention weights are used to calculate the degree of physical coupling response between thermal abrupt changes and mechanical vibration of the enclosure, and cross-modal derived feature vectors are extracted. These cross-modal derived feature vectors are then mapped to multiple basic probability assignment functions containing independent fire risk assessment probabilities as evidence. Calculate the logical conflict coefficient between each piece of evidence. If the conflict coefficient is lower than the safety limit, use the Dempster synthesis rule to perform orthogonal sum operation to complete dynamic comprehensive risk fusion, identify abnormal nodes and output abnormal node characteristics.

6. The method for fire identification of oil-immersed transformer equipment based on operation monitoring data according to claim 1, characterized in that: The process of obtaining the fire cause evolution chain includes: A directed knowledge graph of fire causes is pre-constructed, in which phenomenon nodes include sensor distortion feature parameters such as temperature exceeding limits and sudden increase in discharge, mechanism nodes include hidden physical defect parameters such as winding loosening and deformation and insulation medium aging and peeling, and fire cause nodes are topologically connected through directed edges with historical real occurrence conditional probability weights. The node features of the abnormal nodes are received and transformed into initial activation sources in the graph structure, and then input into the multi-layer graph neural network. Under the rigid prior topological path constraints of the fire cause knowledge graph, message passing between neighboring nodes and high-dimensional feature aggregation reasoning are performed. The connected subgraph that maximizes the joint activation probability is found through global search optimization, and the most complete fire cause evolution link from the current multi-dimensional physical anomaly starting point to the final local deflagration endpoint is output, and the corresponding comprehensive confidence value is output simultaneously.

7. The method for fire identification of oil-immersed transformer equipment based on operation monitoring data according to claim 1, characterized in that: The process of generating a structured early warning report based on the fire cause and evolution chain retrieval and matching strategy includes: The system receives the fire cause evolution chain and the underlying mechanism and inducing node included, performs a high-dimensional vector space similarity search in the pre-established transformer safety operation and maintenance strategy database, retrieves historical handling experience based on the highest matching degree, and generates a corresponding risk reduction and handling strategy; the risk reduction and handling strategy includes physical cooling, load shedding, and fire extinguishing device standby; Combining the risk transmission rate parameters in the fire cause evolution chain, the remaining rescue time window from the current deterioration state to the outbreak of the fire is calculated; the multi-source anomaly characteristic description, the inferred cause evolution conclusion, the remaining rescue time window, and the risk reduction and disposal strategy are assembled and recombined into a structured early warning report.

8. A fire detection and linkage control system for oil-immersed transformer equipment based on operational monitoring data, characterized in that, include: The data acquisition module synchronously collects heterogeneous physical quantities and environmental load data of transformer equipment based on edge nodes, and constructs multi-dimensional time-series data; The data processing module extracts historical long-cycle operating data, performs seasonal trend smoothing decomposition and multivariate nonlinear regression, and constructs an adaptive thermal behavior benchmark model. It extracts measured thermal data from the multidimensional time series data, compares the measured thermal data with the adaptive thermal behavior benchmark model, and generates a thermal anomaly residual sequence that eliminates environmental interference. The anomaly detection module synchronously inputs the thermal anomaly residual sequence and the non-thermal heterogeneous time series data extracted from multidimensional time series data into a multi-head attention fusion network to extract cross-modal derived features, transforms the feature vector into evidence, and uses evidence theory to perform conflict detection and dynamic comprehensive risk fusion to confirm the anomaly node. The anomaly early warning module constructs a fire cause knowledge graph containing phenomenon nodes and mechanism nodes. The node features of the anomaly nodes are input into the graph neural network, and mapping reasoning is performed under the prior topological constraints of the knowledge graph. The module outputs the fire cause evolution link and confidence level under multiple anomalies. Based on the fire cause evolution link, a matching and disposal strategy is retrieved to generate a structured early warning report.

Images