Methods, devices, electronic equipment and storage media for monitoring temperature of electromagnetic transformers

Abstract

This invention discloses a method, device, electronic equipment, and storage medium for monitoring the temperature of electromagnetic transformers. The method includes: S1, initialization and multimodal calibration; S2, collaborative sampling and preprocessing; S3, three-dimensional temperature field reconstruction and digital twin prediction, projecting spatiotemporally registered multimodal temperature samples onto the device's observation grid, generating temperature volume data and local confidence levels using distance- and confidence-based weighted interpolation or radial basis function interpolation, and simultaneously constructing a heat-flow-electric coupling digital twin model and outputting short-term thermal field predictions and prior uncertainties for 10-30 minutes; S4, data assimilation, hotspot diagnosis, and forward-looking early warning. This invention achieves a closed loop consisting of multimodal sensing, three-dimensional reconstruction, digital twin, and data assimilation, enabling high-confidence estimation of the global temperature field under conditions of incomplete measurements, and improving early warning lead time and robustness through adaptive encrypted sampling based on uncertainty feedback.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and intelligent diagnostics technology for power equipment, and more specifically, to a method, device, electronic equipment, and storage medium for monitoring the temperature of an electromagnetic transformer. Background Technology

[0002] With the continuous increase in the scale and voltage level of power systems, the demand for safe and stable operation of oil-immersed power equipment such as electromagnetic transformers has significantly increased. The temperature field distribution inside in-service equipment is directly related to the evolution trend of insulation aging rate and thermal runaway risk. Traditional shell temperature measurement or a few fixed-point temperature measurements cannot reflect the true three-dimensional temperature distribution inside the equipment, and it is even more difficult to identify the formation, growth, and migration process of hot spots. However, under heavy load and environmental fluctuation conditions, local overheating often exhibits rapid evolution characteristics. If it cannot be identified and warned in time, it can easily induce faults and outage risks.

[0003] To obtain more comprehensive internal temperature information, common solutions attempt to increase the number of measurement points or adopt a single new sensing channel. However, due to limited point deployment, insufficient coverage, and signal-to-noise ratio issues, the accuracy and timeliness of three-dimensional thermal field reconstruction are still limited.

[0004] In existing technologies, fixed sensor arrays and simple interpolation are commonly used to estimate local areas in two or quasi-three dimensions. However, if strict registration with the device's geometric model, scanning pose, and time is lacking, and if model prediction and data assimilation mechanisms are not introduced, significant uncertainty will arise in the neighborhood of hotspots, leading to false alarms, missed alarms, or delayed warnings, thereby reducing the reliability and availability of early warnings. Summary of the Invention

[0005] The present invention provides at least one method, device, electronic device and storage medium for monitoring temperature of electromagnetic transformers, realizing a closed loop consisting of multimodal sensing, three-dimensional reconstruction, digital twin and data assimilation, achieving high confidence estimation of the global temperature field under incomplete measurement conditions, and improving early warning lead time and robustness through uncertainty feedback adaptive encrypted sampling.

[0006] In a first aspect, embodiments of the present invention provide a method for monitoring the temperature of an electromagnetic transformer, comprising: S1, initialization and multimodal calibration, performing temperature scaling, time and geometric registration on a miniature thermocouple array, a mechanical multi-point scanning mechanism and an optional distributed fiber optic temperature sensor or infrared miniature array, establishing a channel confidence model and setting initial fusion weights, scanning path, step size and dwell time; S2, collaborative sampling and preprocessing, continuously sampling with a fixed array and combining mechanical scanning with point-by-point sampling along the path, performing spatiotemporal alignment based on a unified timestamp and pose encoding, performing zero-point / temperature drift compensation, sliding median and low-pass filtering, outlier removal or interpolation on the collected data and completing geometric mapping; S3, three-dimensional temperature field reconstruction and digital twin prediction, projecting the spatiotemporally registered multimodal temperature samples onto the device observation grid, and using weighted interpolation or radial basis function based on distance and confidence. Function interpolation generates temperature volume data and local confidence scores, while simultaneously constructing a heat-flow-electricity coupled digital twin model and outputting short-term thermal field predictions and prior uncertainties for 10-30 minutes; S4, data assimilation, hotspot diagnosis, and forward-looking early warning: Bayesian or Kalman assimilation is used to fuse the reconstructed field with the twin prediction field to obtain the posterior temperature field and uncertainty, identifying the hotspot centroid, intensity, area or volume, and migration vector. Based on absolute temperature, temperature rise rate, hotspot migration speed, volume expansion rate, and temporal anomaly detection and small model risk scoring, multi-condition joint early warning is performed, providing a 50-120 minute advance warning and generating disposal suggestions; wherein, the interaction mechanism of data assimilation and adaptive sampling includes: issuing sampling encryption requests for areas where uncertainty exceeds a threshold; adjusting the scanning step size, dwell time, and local path based on online learning or reinforcement learning.

[0007] In this embodiment of the invention, a closed loop of "perception—assimilation—diagnosis—decision—execution—feedback—learning" is formed by fusing multimodal temperature sensing, 3D reconstruction, and digital twin prediction, along with assimilation-driven uncertainty feedback. This technical solution can significantly reduce hotspot neighborhood reconstruction errors while ensuring temperature resolution, stably achieve forward-looking early warnings of 50-120 minutes, and improve operational efficiency and security through a closed loop of human-computer interactive handling suggestions.

[0008] According to the first aspect, in one possible implementation, the channel confidence model established in S1 is initialized based on channel noise level, coverage density and historical stability, and is adaptively updated during operation based on reconstruction residual and confidence interval width to dynamically adjust multimodal fusion weights.

[0009] According to the first aspect, in one possible implementation, S2 includes synchronizing temperature samples from the fixed array and the scanning channel at the millisecond level with a unified clock and binding pose encoding, performing zero-point / temperature drift compensation, sliding median filtering and low-pass filtering, performing rule and statistical joint detection on outliers and compensating them with neighborhood interpolation.

[0010] In this embodiment of the invention, the accuracy of spatiotemporal registration is ensured by time synchronization and pose constraints, and the preprocessing link significantly suppresses pulse interference and high-frequency noise, thereby improving the quality of the reconstruction input data.

[0011] According to the first aspect, in one possible implementation, S3 includes constructing a voxel or unstructured observation grid based on a simplified geometric model of a device layered cylinder or rectangle, employing distance and measurement point confidence weighted interpolation and radial basis function interpolation, and assigning high weights to scanned measurement points in hotspot neighborhoods to reduce local reconstruction errors.

[0012] According to the first aspect, in one possible implementation, S4 includes using Bayesian or Kalman assimilation to fuse the reconstructed field and the twin prediction field to obtain the posterior temperature field and uncertainty. When the width of the 95% confidence interval of the hotspot neighborhood exceeds a preset threshold, an encrypted sampling request is triggered to reduce the scan step size and extend the dwell time.

[0013] In this invention, encrypted sampling triggered by an uncertainty threshold is used to form an adaptive measurement density scheduling for hotspot neighborhoods, which balances timeliness and accuracy.

[0014] According to the first aspect, in one possible implementation, the execution of the encrypted sampling request includes: an online learning model updating a scanning strategy based on a mapping from "hotspot location, growth rate, uncertainty distribution" to "step size, dwell time, path shape"; and fine-tuning the local strategy by reinforcement learning under security constraints to minimize a weighted objective of "reconstruction error, early warning delay, and scanning energy consumption".

[0015] According to the first aspect, in one possible implementation, the early warning adopts a joint classification strategy of rule threshold, time-series anomaly detection and small model risk scoring to output risk level and lead time, and presents a three-dimensional heat map, hot spot trajectory, confidence distribution and future prediction heat map on the human-machine interface, generating a handling suggestion matrix including load reduction, cooling enhancement, oil circulation adjustment and maintenance scheduling, and supporting suggestion-confirmation-execution closed loop.

[0016] Secondly, embodiments of the present invention also provide an electromagnetic transformer temperature monitoring device, comprising: a multi-modal acquisition module for acquiring temperature data from a miniature thermocouple array, a mechanical scanning channel, and an optional distributed optical fiber or infrared microarray, and performing time synchronization and geometric registration; a preprocessing and reconstruction module for performing zero-point / temperature drift compensation, filtering, and outlier processing on the acquired data, and generating three-dimensional temperature volume data and local confidence on the equipment observation grid based on distance and confidence weighting or radial basis function interpolation; a digital twin and assimilation module for constructing a thermal-fluid-electric coupling twin model to output short-term thermal field prediction, and fusing it with the three-dimensional reconstructed field using Bayesian or Kalman assimilation to obtain the posterior temperature field and uncertainty; and a diagnosis and linkage module for identifying hotspot characteristics, performing multi-condition joint early warning, outputting advance and handling suggestions, and interacting with the human-machine interface and the station control system; wherein, the diagnosis and linkage module generates a sampling encryption request for areas where the uncertainty exceeds a threshold and sends it to the multi-modal acquisition module and the preprocessing and reconstruction module to achieve adaptive scanning and reconstruction encryption.

[0017] According to the second aspect, in one possible implementation, the multimodal acquisition module and the preprocessing and reconstruction module are decoupled and connected through a standard data contract, which includes at least a timestamp, pose index, geometric coordinate index, channel identifier, and confidence weight field.

[0018] According to the second aspect, in one possible implementation, the diagnostic and linkage module interacts bidirectionally with the station's SCADA / EMS system via a communication bus so that it can automatically issue load reduction and cooling scheduling commands when an early warning is triggered, or be executed by the operator after confirmation through a human-machine interface.

[0019] According to the second aspect, in one possible implementation, the digital twin and assimilation module supports online parameter calibration and grayscale strategy rollback. When the model update effect does not meet expectations, it automatically reverts to the previous stable version, ensuring operational security and continuity.

[0020] Thirdly, embodiments of the present invention also provide an electronic device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the electromagnetic transformer temperature monitoring method described in the first aspect above, or in any possible implementation of the first aspect, is executed.

[0021] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the electromagnetic transformer temperature monitoring method described in the first aspect or any possible implementation of the first aspect. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to the present invention and, together with the specification, serve to explain the technical solutions of the present invention. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart of an electromagnetic transformer temperature monitoring method provided by an embodiment of the present invention is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] The internal temperature field of in-service electromagnetic transformers is highly non-uniform and time-varying, with hot spots potentially forming and migrating along the winding ends and oil passages. Traditional external casing or limited point measurements are insufficient to characterize the internal three-dimensional distribution, leading to delayed early warnings and the risk of misjudgments.

[0026] The closed loop formed by multimodal sensing, 3D reconstruction, digital twin, and data assimilation described in this invention achieves high-confidence estimation of the global temperature field under conditions of incomplete measurement, and improves early warning lead time and robustness through adaptive encrypted sampling based on uncertainty feedback.

[0027] like Figure 1 As shown:

[0028] I. Overall Architecture

[0029] Sensing layer: Miniature thermocouple array for high-efficiency monitoring at fixed points; mechanical multi-point scanning for flexible coverage and local encryption; optional DTS / infrared microarray.

[0030] Processing layer: unified time and pose synchronization, preprocessing and spatiotemporal registration, 3D reconstruction, digital twin prediction, Bayesian / Kalman assimilation.

[0031] Diagnosis and Decision-Making Layer: Hotspot identification and trajectory extraction, joint early warning (threshold + time series anomaly + small model scoring), handling suggestion matrix, human-computer interaction and execution closed loop.

[0032] Learning and self-healing layer: reinforcement learning optimizes scanning strategies, self-healing of channel anomalies and mechanical jams, in-site collaboration and edge-cloud incremental training.

[0033] II. Detailed Implementation Method

[0034] Step S1: Initialization and Multimodal Calibration

[0035] Temperature calibration: Zero point and sensitivity calibration of each channel are performed using a standard temperature source.

[0036] Time and geometry registration: A unified clock source (preferably with an accuracy of ≤5ms) is used to establish a mapping between the sensor points and the device's geometric model through a pose encoder and calibration fixture.

[0037] Confidence model: Initial values ​​are set based on channel noise, coverage density, and historical stability, which serve as weights for subsequent fusion and assimilation.

[0038] Scan configuration: Set the circumferential × axial grid path, step size (10-30mm), and dwell time (≥3-5 times the sensor response constant).

[0039] Step S2: Co-sampling and Preprocessing

[0040] Synchronous acquisition: The fixed array continuously samples; the scanning mechanism samples point by point along the path and records the pose and timestamp.

[0041] Preprocessing: Perform zero-point / temperature drift compensation; use sliding median filtering to suppress pulse interference;

[0042] Low-pass filtering smooths high-frequency noise; outlier detection is performed using a combination of rule-based and statistical methods, with neighborhood interpolation compensation.

[0043] Spatiotemporal registration: accurately mapping the sample to the device's geometric grid coordinate system.

[0044] Step S3: Three-dimensional temperature field reconstruction and digital twin prediction

[0045] Geometric modeling and meshing: Construct voxels or unstructured meshes based on simplified models of layered cylinders or rectangles.

[0046] Interpolation fusion: Distance and channel confidence weighting and radial basis function interpolation are used; the weight of scan points is increased in the hotspot neighborhood to reduce local errors.

[0047] Twin prediction: Construct a thermal-fluid-electric coupling model to output a short-term prediction field and prior uncertainty for 10-30 minutes; calibrate thermal conductivity and convection parameters online based on historical and current observations.

[0048] Step S4: Data assimilation, Hotspot Diagnosis, and Forward-Looking Early Warning

[0049] Data assimilation: Bayesian or Kalman assimilation is used to fuse the reconstructed field and the twin prediction field to obtain the posterior temperature field and uncertainty distribution.

[0050] Uncertainty-driven: If the width of the 95% confidence interval of the hot spot neighborhood exceeds the threshold, an encrypted sampling request is triggered to reduce the step size and extend the dwell time.

[0051] Hotspot diagnosis: Extract features such as hotspot centroid, peak / mean value, area / volume, migration vector and temperature rise rate, and expansion rate.

[0052] Joint early warning: Based on absolute temperature, temperature rise rate, migration speed, volume expansion rate, and time-series anomaly detection and small model risk score, the system classifies and judges the alarm, calculates a 50-120 minute lead time, and generates a matrix of handling suggestions (load reduction, cooling enhancement, oil circulation adjustment, and maintenance scheduling).

[0053] Human-computer interaction and execution closed loop:

[0054] HMI presents a 3D heatmap, hotspot trajectory, confidence distribution, and predicted heatmap, providing an operational process and drill mode for suggestion-confirmation-execution.

[0055] It integrates with the on-site SCADA / EMS system, supports automatic or semi-automatic execution of control policies, and retains audit logs.

[0056] Learning and self-healing:

[0057] Online learning / reinforcement learning: The scanning strategy and fusion weights are adaptively updated based on "hotspot location, growth rate and uncertainty distribution"; optimization is performed with "reconstruction error, early warning delay and energy consumption / latency" as weighted objectives.

[0058] Self-healing mechanisms include: abnormal shielding of sensor channels and neighborhood interpolation, zeroing of scan card lag and path detour, online small-amplitude excitation / self-heating recalibration; and reporting to the cloud to improve fault prior knowledge.

[0059] Through the above steps, multimodal fusion and assimilation are achieved to reduce the reconstruction error of hotspot neighborhoods, the temperature resolution reaches ±0.03℃, and the uncertainty can be quantified.

[0060] Early warning lead time: By combining trend and model judgment, a stable early warning of 50-120 minutes can be achieved, which is significantly better than the lag of single-point temperature measurement.

[0061] Timeliness and closed-loop: Uncertainty-driven adaptive encryption and online learning ensure timeliness and resource efficiency under complex operating conditions; closed-loop handling improves operational feasibility and compliance auditability.

[0062] Integrability: Modular method packages and standard data contracts facilitate integration with on-site systems and subsequent upgrades.

[0063] The present invention also provides an electronic device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the above-described electromagnetic transformer temperature monitoring method is performed.

[0064] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described electromagnetic transformer temperature monitoring method.

[0065] This invention also provides an electromagnetic transformer temperature monitoring device, comprising: a multi-modal acquisition module for acquiring temperature data from a miniature thermocouple array, a mechanical scanning channel, and an optional distributed optical fiber or infrared microarray, and performing time synchronization and geometric registration; a preprocessing and reconstruction module for performing zero-point / temperature drift compensation, filtering, and outlier processing on the acquired data, and generating three-dimensional temperature volume data and local confidence scores on the device observation grid based on distance and confidence weighting or radial basis function interpolation; a digital twin and assimilation module for constructing a thermal-fluid-electric coupled twin model to output short-term thermal field predictions, and fusing the predicted three-dimensional temperature field and uncertainty with Bayesian or Kalman assimilation; and a diagnosis and linkage module for identifying hotspot characteristics, performing multi-condition joint early warning, outputting lead time and handling suggestions, and interacting with the human-machine interface and the station control system; wherein, the diagnosis and linkage module generates sampling encryption requests for areas where the uncertainty exceeds a threshold and sends them to the multi-modal acquisition module and the preprocessing and reconstruction module to achieve adaptive scanning and reconstruction encryption.

[0066] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0067] This invention also provides a verification example of another method for monitoring the temperature of an electromagnetic transformer:

[0068] Step 201: Build the device and data environment, connect to the multimodal sensing channel and complete initialization and multimodal calibration.

[0069] In this embodiment, a high-capacity electromagnetic transformer with a rated voltage of 220kV is selected as the object under test. A 24-point miniature thermocouple array is arranged at key locations such as the winding end, clamps, and oil passages. A circumferential × axial two-dimensional guide rail mechanical scanning mechanism (adjustable step distance 10–30mm) is configured, and one distributed fiber optic temperature sensing (DTS) channel can be optionally connected. A unified clock source (PTP, synchronization accuracy ≤5ms) is used to synchronize the time of each channel. Geometric registration is completed through a calibration fixture and a position encoder. Zero-point and sensitivity calibration are performed using a constant temperature bath and a standard heat source. A channel confidence model is established, and the initial fusion weights, scanning path, step distance (initial value 20mm), and dwell time (initial value 2.5s, corresponding to a sensing response constant of approximately 0.6s) are set.

[0070] Step 202: Implement collaborative sampling and preprocessing to achieve data spatiotemporal alignment and quality improvement.

[0071] A fixed array continuously samples at a frequency of 2 Hz; a mechanical scanner samples point-by-point along a preset path and records the pose; a DTS outputs the temperature curve along the fiber optic path at a resolution of 1 minute. Zero-point / temperature drift compensation, sliding median filtering (windows 3–5), and low-pass filtering (cutoff 0.3 Hz) are performed on the multi-channel data. Outliers are identified using a combination of a regular threshold (3σ) and statistical analysis (ESD), and compensation is performed using neighborhood interpolation. All samples are mapped to the device's observation grid coordinate system using a unified timestamp and pose index.

[0072] Step 203: Perform 3D temperature field reconstruction and digital twin prediction to generate current and short-term prior thermal fields. Based on the equipment's layered cylindrical simplified geometry, construct a voxel mesh (typical resolution 10×10×10mm). Use distance and channel confidence-weighted interpolation and radial basis function (RBF) interpolation to fuse fixed-point, scan-point, and DTS data, outputting 3D temperature volume data and local confidence scores. Concurrently construct a heat-fluid-electric coupling digital twin model, inputting load current, ambient temperature, and oil pump status. Online calibration of thermal conductivity and convective heat transfer coefficient is performed, outputting a 10–30 minute short-term predicted thermal field and its prior uncertainty.

[0073] Step 204: Conduct data assimilation, hotspot diagnosis, and forward-looking early warning to generate warning conclusions and handling recommendations. Using Bayesian / Kalman assimilation, the reconstructed field from Step 203 is fused with the twin prediction field to obtain the posterior temperature field and uncertainty distribution. Based on the posterior field, hotspot centroids, peak / mean values, area / volume, migration vectors, temperature rise rates, and volume expansion rates are extracted. A joint strategy of "rule-based thresholding + time-series anomaly detection (CUSUM / ESD) + small model risk scoring" is employed for graded warning assessment, calculating the early warning lead time and generating handling recommendations (load reduction, enhanced cooling, oil circulation adjustment, and maintenance scheduling).

[0074] Step 205: Trigger uncertainty-driven adaptive sampling and online strategy learning to optimize measurement density and timeliness in a closed loop.

[0075] When the width of the 95% confidence interval of the hotspot neighborhood exceeds the threshold (e.g., ≥0.12℃), the system automatically issues an encrypted sampling request: reducing the scanning step size from 20mm to 10mm and extending the dwell time from 2.5s to 3.0s; the online learning model updates the scanning strategy mapping according to "hotspot location, growth rate, and uncertainty distribution"; under safety constraints, reinforcement learning fine-tunes the local path to minimize the weighted objective of "reconstruction error + early warning delay + scanning energy consumption" (0.5 / 0.3 / 0.2).

[0076] Step 206: Implement coordinated action and human-computer interaction, record audit results, and retest and verify the effectiveness.

[0077] The human-machine interface (HMI) displays a 3D heat map, hotspot trajectory, confidence distribution, and predicted heat map. It outputs a treatment suggestion matrix, which is then "suggested-confirmed-executed" by the on-duty personnel. Simultaneously, it coordinates with SCADA / EMS to implement a 10% phased load reduction and forced oil circulation. Within 5 minutes after treatment, it performs resampling and assimilation updates, compares the changes in hotspot peak value and migration speed to evaluate the effectiveness of the treatment, and leaves a complete record of the entire process.

[0078] Step 207: Setting up and defining evaluation indicators for the traditional technology control experiment.

[0079] To verify the improvement effect of this invention compared with the traditional point temperature measurement / simple interpolation scheme, a control group was set up: using only fixed point temperature measurement (24 points) and two-dimensional linear interpolation reconstruction, without pose synchronization correction, without twin prediction and assimilation, and without adaptive encryption and learning mechanisms, while maintaining the same equipment and operating conditions.

[0080] Evaluation metrics include: mean square error of temperature reconstruction (MSE, based on built-in reference point / calibrated heat source), hot spot peak error (absolute value), hot spot migration speed error, early warning time, false alarm rate / false negative rate, reconstruction refresh cycle, sampling energy consumption, and system availability (percentage of continuous operation time under abnormal conditions).

[0081] Step 208: Experimental conditions and parameter settings.

[0082] Operating Condition 1: Peak daily load of 1.2 pu during summer high temperature, ambient temperature 34±2℃, oil pump ON.

[0083] Operating Condition 2: Daily load fluctuation of 0.8–1.1 pu, ambient temperature of 28±3℃, oil pump AUTO.

[0084] Operating Condition 3: Emergency Overload Test 1.3pu for 30 minutes, ambient temperature 31±1℃, oil pump ON.

[0085] The invention group and the control group alternated in operation for 1 hour at different times within the same day, and the average was taken for 3 days. The warning threshold was set based on the baseline and compared using a unified standard.

[0086] Step 209: Experimental Results and Comparison.

[0087] The average values ​​of key indicators were calculated for each of the three operating conditions and then combined and displayed in the table below.

[0088] Comparison Dimensions Traditional technical data Technical data of this invention Improvement effect Technical Advantages Explanation Temperature reconstruction MSE (°C²) 0.022 0.006 Reduced by 72.7% Assimilation and fusion with adaptive multimodal weights significantly reduce hotspot neighborhood errors. Hotspot peak error (°C) 0.45 0.12 Reduced by 73.3% Hotspot neighborhood encryption scanning and RBF interpolation improve local accuracy Migration speed error (mm / min) 4.8 1.2 Reduced by 75.0% Spatiotemporal alignment and trajectory extraction are based on the posterior field, reducing trajectory jitter. Early warning time (min) 18–35 58–115 An increase of approximately 3.2× Combining twin prediction, temporal anomalies, and scoring to improve forward-looking capabilities. False alarm rate (%) 9.6 3.1 Reduced by 67.7% Uncertainty gating and multi-index joint suppression of false alarms Missed report rate (%) 7.3 2.4 Reduced by 67.1% Combining assimilation posterior with trend enhances sensitivity. Reconstruction refresh cycle (min) 3.8 1.2 (Warning status 0.8) Accelerate by approximately 3.2× Adaptive encryption and edge optimization improve timeliness Sampling energy consumption (relative unit) 1.00 0.74 26% lower Learn to optimize step size / dwell time and path to reduce invalid scans. System availability (%) 93.2 98.6 An increase of 5.4 percentage points Multimodal redundancy and self-healing strategies improve continuous operation capability

[0089] Step 210: Result Analysis and Problem Identification.

[0090] 1. This invention fuses multimodal data by distance + confidence weighting and RBF interpolation, and then obtains the posterior field through assimilation, which significantly reduces the MSE of hotspot neighborhoods and improves the accuracy of temperature reconstruction; the control group lacks pose correction and assimilation, and the error is concentrated in the sparse area of ​​the measurement points.

[0091] 2. Enhanced early warning capability: Digital twin short-term prediction provides trend priors, and by combining time series anomalies with small model scores, the early warning lead time is increased from an average of about 26 minutes to about 86 minutes; the control group is based only on absolute thresholds and has weak trend capture capability.

[0092] 3. Uncertainty gating triggers encrypted scanning to ensure information density in risk areas; multimodal redundancy and self-healing strategies maintain 98.6% availability under occasional sensor channel anomalies and scanning lag, achieving system stability and robustness.

[0093] 4. Timeliness and energy efficiency: Edge computing shortens the reconstruction cycle to 1.2 minutes (0.8 minutes in the early warning state); online / reinforcement learning reduces invalid scans, resulting in a 26% reduction in relative sampling energy consumption.

[0094] 5. Alarm quality improvement: Joint alarm judgment reduces false alarms / missed alarms; HMI interpretability and response drills improve response accuracy; post-execution retesting shows that hotspot peak temperature drops by >0.8℃ and migration speed decreases by >40%.

[0095] 6. Under extreme overload (≥1.4 pu) short-term impact, twin parameters need to be quickly recalibrated; subsequently, finer-grained online oil flow field identification and adaptive mesh refinement can be introduced to further reduce hotspot errors.

[0096] Step 211: Reproduce the scenario and verify the parameters in the example.

[0097] In the emergency overload test under operating condition 3, after the encrypted scan was triggered in step 205, the scan step distance was reduced from 20mm to 10mm, the dwell time was increased from 2.5s to 3.0s, and the reconstruction cycle was reduced from 2.0min to 0.9min. After the load was reduced by 10% and the oil circulation was strengthened in step 206, the hot spot peak value was reduced from 96.2℃ to 95.3℃ after 5 minutes of retesting, and then reduced to 94.8℃ after 10 minutes. The warning level was reduced from high to medium.

[0098] The control group, due to the inability to adaptively encrypt during the same period, experienced a delay of approximately 7–10 minutes in estimating hotspot trajectories, resulting in a delayed response. The peak temperature dropped to 95.6℃ only after 15 minutes, and the migration speed had a large error.

[0099] Step 212, Conclusion.

[0100] The above embodiments demonstrate that the present invention is significantly superior to traditional technologies under multiple operating conditions, achieving quantitative improvements in key dimensions such as reconstruction accuracy, early warning, stability, timeliness, and energy efficiency. This verifies that the technical route of the present invention—multimodal perception—digital twin—data assimilation—adaptive scanning—joint early warning—closed-loop handling—online learning and self-healing—has good engineering applicability and broad application prospects.

Claims

1. A method for monitoring the temperature of an electromagnetic transformer, characterized in that, include: S1, Initialization and multimodal calibration, performing temperature scale, time and geometric registration on micro thermocouple arrays, mechanical multi-point scanning mechanisms and distributed fiber optic temperature sensors or infrared micro arrays, establishing channel confidence models and setting initial fusion weights, scanning paths, step sizes and dwell times. S2, Cooperative sampling and preprocessing, fixed array continuous sampling combined with mechanical scanning to sample point by point along the path, spatiotemporal alignment based on unified timestamp and pose coding, zero point / temperature drift compensation, sliding median and low-pass filtering, outlier removal or interpolation, and geometric mapping are performed on the collected data. S3, 3D temperature field reconstruction and digital twin prediction, projects spatiotemporally registered multimodal temperature samples onto the equipment observation grid, and generates temperature volume data and local confidence using distance- and confidence-based weighted interpolation or radial basis function interpolation. At the same time, it constructs a heat-flow-electric coupling digital twin model and outputs short-term thermal field predictions and prior uncertainties for 10-30 minutes. S4, Data Assimilation, Hotspot Diagnosis and Proactive Early Warning: Bayesian or Kalman assimilation is used to fuse the three-dimensional temperature reconstruction field with the short-term thermal field prediction output by the digital twin model to obtain the posterior temperature field and uncertainty. The centroid, intensity, area or volume and migration vector of hotspots are identified. Based on absolute temperature, temperature rise rate, hotspot migration speed, volume expansion rate and time anomaly detection and small model risk score, multi-condition joint early warning is performed, and a warning lead of 50-120 minutes is given and disposal suggestions are generated. The interaction mechanism of data assimilation and adaptive sampling includes: when the width of the 95% confidence interval of the hot spot neighborhood exceeds a preset threshold, triggering an encrypted sampling request, reducing the mechanical scanning step size and extending the dwell time; establishing a mapping of hot spot location, growth rate and uncertainty distribution to step size, dwell time and path shape through an online learning model to update the scanning strategy, and fine-tuning the local strategy by reinforcement learning under security constraints to minimize the weighted objective of reconstruction error, early warning delay and scanning energy consumption.

2. The method according to claim 1, characterized in that, The channel confidence model established in S1 is initialized based on channel noise level, coverage density and historical stability, and is adaptively updated during operation according to reconstruction residual and confidence interval width to dynamically adjust multimodal fusion weights.

3. The method according to claim 2, characterized in that, The S2, collaborative sampling and preprocessing, includes: synchronizing temperature samples from the fixed array and scanning channel at millisecond-level time according to a unified clock and binding pose encoding; performing zero-point / temperature drift compensation, sliding median filtering to suppress pulse interference, and low-pass filtering to smooth high-frequency noise; performing rule-based and statistical joint detection on outliers and using neighborhood interpolation for compensation.

4. The method according to claim 3, characterized in that, The S3, three-dimensional temperature field reconstruction includes: constructing voxel or unstructured observation grids based on a simplified geometric model of the device's layered cylinders or rectangles; using distance and measurement point confidence weighted interpolation; using radial basis function interpolation; and assigning high weights to scanned measurement points in the hotspot neighborhood to reduce local reconstruction errors.

5. The method according to claim 1, characterized in that, The early warning system employs a combined grading strategy of rule-based thresholds, time-series anomaly detection, and small-model risk scoring to output risk levels and lead times. It also presents a 3D heatmap, hotspot trajectory, confidence distribution, and future prediction heatmap on the human-machine interface, generating a matrix of handling suggestions including load reduction, cooling enhancement, oil circulation adjustment, and maintenance scheduling, and supporting a suggestion-confirmation-execution closed loop.

6. A temperature monitoring device for an electromagnetic transformer, characterized in that, The method described in any one of claims 1-5 comprises: a multimodal acquisition module for acquiring temperature data from a micro thermocouple array, a mechanical scanning channel, and an optional distributed optical fiber or infrared microarray, and performing time synchronization and geometric registration; a preprocessing and reconstruction module for performing zero-point / temperature drift compensation, filtering, and outlier processing on the acquired data, and generating three-dimensional temperature volume data and local confidence scores on the equipment observation grid based on distance and confidence weighting or radial basis function interpolation; a digital twin and assimilation module for constructing a heat-fluid-electric coupling twin model to output short-term thermal field predictions, and fusing the predicted three-dimensional temperature field with Bayesian or Kalman assimilation to obtain a posterior temperature field and uncertainty; and a diagnosis and linkage module for identifying hotspot characteristics, performing multi-condition joint early warning, outputting lead time and handling suggestions, and interacting with the human-machine interface and the station control system; wherein the diagnosis and linkage module generates sampling encryption requests for areas where the uncertainty exceeds a threshold and sends them to the multimodal acquisition module and the preprocessing and reconstruction module to achieve adaptive scanning and reconstruction encryption.

7. An electronic device, characterized in that, include: The system includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the electromagnetic transformer temperature monitoring method as described in any one of claims 1 to 5 is performed.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, performs the electromagnetic transformer temperature monitoring method as described in any one of claims 1 to 5.

Images