Method for measuring temperature of switch cabinet in real time by infrared matrix based on multi-modal sensing

Abstract

The invention relates to the technical field of switch cabinet temperature measurement, and discloses a method for measuring the temperature of a switch cabinet in real time through an infrared matrix based on multi-modal sensing. The method comprises the following steps: synchronously acquiring an infrared image sequence, an acoustic vibration signal and current waveform data of the switch cabinet through a multi-mode sensor array, and capturing equipment operation and temperature associated information; constructing a three-dimensional temperature field reconstruction mechanism model based on a heat conduction physical equation, outputting a theoretical temperature distribution value fitting a physical rule, processing multi-source data by using a spatial-temporal feature extraction network, and outputting a temperature prediction value reflecting a real-time working condition; inputting the two into a dynamic fusion module based on a Markov decision process to generate a fusion temperature distribution value, and generating a temperature anomaly positioning instruction through a rolling time domain optimization strategy according to the fusion temperature distribution value; when the deviation between the actual temperature sampling value and the fused temperature distribution value exceeds a preset threshold value, calibrating a network output layer parameter; and triggering a linkage control protocol of the switch cabinet protection system according to the abnormal positioning instruction.

Description

Technical Field

[0001] This invention relates to the field of switchgear temperature measurement technology, specifically a method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix. Background Technology

[0002] During the operation of a power system, switchgear, as the core equipment for power distribution and control, directly affects the stable operation of the power system due to its internal temperature. With the continuous increase in power load and the aging of switchgear, critical components such as busbar joints and circuit breaker contacts are prone to abnormal heating due to increased contact resistance. Failure to monitor and address these temperature anomalies in a timely manner can lead to serious accidents such as insulation aging, equipment burnout, or even large-scale power outages, posing a significant threat to the safety and reliability of the power system. Currently, the industry primarily uses traditional single-point temperature measurement methods for switchgear temperature measurement, such as thermocouples and infrared thermometers. These methods can only obtain temperature data at specific points within the switchgear, failing to provide a comprehensive understanding of the overall temperature distribution within the cabinet and easily overlooking potential temperature anomalies in non-measured areas. Furthermore, single-point temperature measurement is susceptible to interference from environmental factors, such as electromagnetic radiation and airflow fluctuations within the cabinet, making it difficult to guarantee the accuracy and stability of the measurement data. With the development of infrared imaging technology, some solutions have begun to use infrared thermal imagers to monitor the temperature of switchgear, acquiring two-dimensional temperature distribution images within the cabinet. However, these methods rely solely on infrared images as a single data source, neglecting other key physical parameters during switchgear operation. For example, the acoustic vibration signals generated by equipment during operation are closely related to equipment fault conditions, while current waveform data directly reflects the equipment load. Changes in these parameters all affect the temperature distribution. Relying solely on infrared images for temperature judgment makes it difficult to accurately distinguish whether temperature anomalies are caused by equipment malfunctions or normal load fluctuations or environmental changes, easily leading to misjudgments or missed diagnoses. Existing temperature measurement methods have shortcomings in data processing and fusion. Traditional methods typically use simple threshold comparisons to determine whether a temperature is abnormal, lacking the ability to deeply fuse and dynamically analyze multi-source data. When switchgear is under complex operating conditions, such as rapid load changes and large fluctuations in ambient temperature, fixed threshold judgment methods cannot adapt to changes in operating conditions, leading to a decrease in the timeliness and accuracy of anomaly identification. Furthermore, most existing methods lack adaptive calibration capabilities; as equipment operating time increases, sensor performance may drift, causing a gradual decrease in measurement accuracy and failing to meet long-term stable temperature monitoring requirements. Summary of the Invention

[0003] The purpose of this invention is to provide a method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix, the method comprising: The infrared image sequence, acoustic vibration signal and current waveform data of the switchgear are acquired synchronously through a multimodal sensor array. A three-dimensional temperature field reconstruction mechanism model is constructed based on the physical equation of heat conduction, and the theoretical temperature distribution value is output. The infrared image sequence, acoustic vibration signal and current waveform data are processed by a spatiotemporal feature extraction network to output real-time temperature prediction values. The theoretical temperature distribution value and the real-time temperature prediction value are input into a dynamic fusion module based on Markov decision process to generate a fused temperature distribution value. Based on the fused temperature distribution value, a temperature anomaly location instruction is generated using a rolling time-domain optimization strategy. When the deviation between the actual temperature sample value of the switch cabinet and the fused temperature distribution value exceeds a preset threshold, the output layer parameters of the spatiotemporal feature extraction network are adaptively calibrated. The linkage control protocol of the switchgear protection system is triggered according to the temperature anomaly location command.

[0005] Preferably, the step of constructing a three-dimensional temperature field reconstruction mechanism model based on the physical equation of heat conduction and outputting theoretical temperature distribution values ​​includes: Based on the thermal conductivity of the switchgear material and the heat dissipation boundary conditions, an initial temperature gradient constraint is set for the heat conduction equation. Based on the initial temperature gradient constraint, the theoretical temperature distribution value is calculated iteratively using the finite volume method.

[0006] Preferably, the step of processing the infrared image sequence, acoustic vibration signal, and current waveform data through a spatiotemporal feature extraction network to output a real-time temperature prediction value includes: The infrared image sequence is input into a three-dimensional convolution module to extract spatial thermal radiation features; The acoustic vibration signal and current waveform data are input into a bidirectional time encoder to extract time correlation features. By integrating the spatial thermal radiation characteristics and time correlation characteristics, a real-time temperature prediction value is output through a multilayer sensor.

[0007] Preferably, the step of inputting the theoretical temperature distribution value and the real-time temperature prediction value into a dynamic fusion module based on a Markov decision process to generate a fused temperature distribution value includes: Define a thermal distribution state space, which includes the theoretical temperature distribution values ​​and real-time temperature prediction values ​​of each area of ​​the switchgear. The confidence weights of the theoretical temperature distribution values ​​and the confidence weights of the real-time temperature prediction values ​​are calculated using a policy network. The theoretical temperature distribution value and the real-time temperature prediction value are weighted and fused based on the confidence weight to generate a fused temperature distribution value.

[0008] Preferably, the step of generating a temperature anomaly location instruction based on the fused temperature distribution value using a rolling time-domain optimization strategy includes: Construct an objective function for temperature gradient change within a sliding time window; The minimum anomaly energy point of the objective function is found by gradient descent. The physical coordinates corresponding to the minimum abnormal energy point are used as the temperature anomaly location command.

[0009] Preferably, when the deviation between the actual temperature sample value of the switchgear and the fused temperature distribution value exceeds a preset threshold, adaptive calibration of the output layer parameters of the spatiotemporal feature extraction network is performed, including: The actual temperature sampling values ​​of key nodes in the switch cabinet are collected by contact thermocouples. The mean absolute error between the actual temperature sampling values ​​and the fused temperature distribution values ​​is calculated. When the mean absolute error exceeds the set tolerance, adaptive calibration is triggered. The adaptive calibration includes: The convolutional layer and encoder parameters of the spatiotemporal feature extraction network are fixed; Based on the error backpropagation gradient between the actual temperature sample value and the fused temperature distribution value; Only update the fully connected weight parameters of the multilayer perceptron until the mean absolute error is lower than the set tolerance.

[0010] Preferably, the step of inputting the infrared image sequence into a three-dimensional convolution module to extract spatial thermal radiation features includes: Non-uniformity correction is performed on the infrared image sequence to generate standardized thermal imaging data; The standardized thermal imaging data is scanned using deformable convolutional kernels to extract multi-scale hot spot topology. Spatial thermal radiation characteristics are generated based on the spatial correlation of hot spot topology.

[0011] Preferably, the step of synchronously acquiring infrared image sequences, acoustic vibration signals, and current waveform data of the switchgear via a multimodal sensor array includes: The detection angle of the infrared focal plane array is deployed according to the switch cabinet partition topology diagram, the sampling frequency of the vibration sensor is configured based on the acoustic signal propagation model, and the data acquisition timing of the multi-modal sensor array is triggered according to the current phase synchronization.

[0012] Preferably, after using the physical coordinates corresponding to the minimum abnormal energy point as the temperature anomaly location command, the process includes: Map the physical coordinates to the actual physical space coordinate system of the switchgear; Based on the equipment safety margin threshold, a graded alarm code is generated and written into the register queue of the switchgear protection system.

[0013] Preferably, after writing the graded alarm code into the register queue of the switchgear protection system, the process includes: Based on the device type code corresponding to the physical coordinates, a preset protection action response time constraint is matched; Based on the abnormal gradient change rate of the fused temperature distribution value, the priority weight coefficient of the protection system action is calculated; The circuit breaker tripping sequence command is generated based on the priority weighting coefficient and response time constraint.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By simultaneously acquiring infrared image sequences, acoustic vibration signals, and current waveform data using a multimodal sensor array, this method overcomes the limitations of traditional single-source measurements. Traditional methods often rely on only a single type of temperature-related data, failing to comprehensively reflect the complex operating states and temperature-related factors within the switchgear. This method, however, simultaneously acquires multi-dimensional data, capturing switchgear operating information from different physical levels. This provides richer and more comprehensive basic data support for subsequent temperature analysis, enabling the judgment of temperature status to move beyond a single indicator and instead combine the equipment's thermal radiation, vibration characteristics, and current load conditions, better aligning with the objective laws governing the interaction of multiple factors during actual switchgear operation. In terms of temperature field construction and prediction, this method constructs a three-dimensional temperature field reconstruction mechanism model based on the physical equations of heat conduction. Simultaneously, it processes multi-source collected data through a spatiotemporal feature extraction network to generate both theoretical temperature distribution values ​​and real-time temperature prediction values. The mechanism model, starting from the physical essence, derives the temperature distribution based on the fundamental laws of heat conduction, ensuring the theoretical rationality of the temperature analysis. The spatiotemporal feature extraction network can uncover the spatiotemporal correlation features in the multi-source data, capturing the dynamic patterns of data changes over time and spatial distribution differences, thus achieving accurate prediction of real-time temperature. The combination of these two methods avoids the difficulty of handling parameter changes under complex operating conditions when relying solely on the mechanism model, and also compensates for the shortcomings of relying solely on data-driven models, which may deviate from the physical essence. This makes the temperature analysis both theoretically grounded and supported by actual data. The dynamic fusion module based on Markov decision processes (MDF) fuses theoretical temperature distribution values ​​with real-time temperature prediction values ​​to generate a fused temperature distribution value, effectively improving the reliability and accuracy of the temperature distribution results. MDF possesses the ability to analyze and make decisions for dynamic systems, dynamically adjusting the weights of different data sources in the fusion process based on real-time changes in multi-source data, fully leveraging the advantages of both the theoretical and data models. When the reliability of a data source decreases due to interference, the fusion module can automatically reduce its weight to avoid excessive impact on the overall temperature distribution results; conversely, when the data quality of a data source is high, its weight can be appropriately increased to enhance the accuracy of the fusion results, thus enabling the final fused temperature distribution value to more realistically reflect the actual temperature conditions inside the switchgear. The application of a rolling time-domain optimization strategy makes the generation of temperature anomaly location commands more timely and targeted. This strategy can continuously perform dynamic analysis and optimization of temperature distribution within a certain time window based on current and recent multi-source data, tracking temperature change trends in real time and promptly detecting the initial stages of temperature anomalies. Compared to traditional anomaly judgment methods with fixed periods or fixed thresholds, the rolling time-domain optimization strategy can flexibly adjust the analysis period and judgment logic according to the actual temperature changes, enabling faster location of temperature anomaly areas, saving valuable time for subsequent fault handling, and effectively reducing the risk of fault escalation. The introduction of an adaptive calibration mechanism ensures the measurement accuracy and stability of the method over long-term operation. During actual operation of the switchgear, sensor performance may fluctuate due to environmental factors and equipment aging, leading to measurement data deviations. This method compares the actual temperature sampling value of the switchgear with the fused temperature distribution value. When the deviation exceeds a preset threshold, it automatically calibrates the output layer parameters of the spatiotemporal feature extraction network. This dynamic calibration method can promptly correct sensor performance drift or deviations generated during data processing, ensuring that temperature prediction and fusion results maintain high accuracy, avoiding measurement accuracy degradation due to long-term operation, and extending the effective lifespan of the method. This method triggers the linkage control protocol of the switchgear protection system based on temperature anomaly location commands, achieving seamless integration of temperature monitoring and protection control. Traditional temperature measurement methods mostly only have data acquisition and display functions; after an anomaly is detected, manual intervention is required for judgment and control operations, resulting in a certain response delay. In contrast, this method, after accurately locating the temperature anomaly, can directly trigger the linkage control of the protection system, quickly initiating corresponding protective measures without manual intervention, such as cutting off the fault circuit and activating the cooling device. This significantly improves the response speed of switchgear safety protection, further reducing the possibility of equipment damage and accident escalation, and providing more comprehensive protection for the safe and stable operation of the switchgear. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the real-time temperature measurement method for switchgear based on multimodal sensing using an infrared matrix, as described in this invention. Figure 2 A flowchart for a spatiotemporal feature extraction network processing multimodal data; Figure 3 This is a flowchart of dynamic fusion based on Markov decision processes; Figure 4 A flowchart for an adaptive calibration spatiotemporal feature extraction network; Figure 5 This is a flowchart for extracting the characteristics of space thermal radiation. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 This invention provides a method for real-time measurement of switchgear temperature using an infrared matrix based on multimodal sensing, the method comprising: Infrared image sequences, acoustic vibration signals, and current waveform data of the switchgear are synchronously acquired using a multimodal sensor array. A three-dimensional temperature field reconstruction mechanism model is constructed based on the physical equations of heat conduction, which outputs a theoretical temperature distribution value. Simultaneously, a spatiotemporal feature extraction network processes the acquired multimodal sensor data to output a real-time temperature prediction value. The theoretical temperature distribution value and the real-time temperature prediction value are input into a dynamic fusion module based on a Markov decision process, which generates a fused temperature distribution value. Based on this fused value, a rolling time-domain optimization strategy is used to generate a temperature anomaly location command. When the deviation between the actual temperature sample value and the fused temperature distribution value at a key node of the switchgear exceeds a preset threshold, the output layer parameters of the spatiotemporal feature extraction network are adaptively calibrated. Finally, based on the temperature anomaly location command, the linkage control protocol of the switchgear protection system is triggered.

[0018] Example 1: See Figure 2In constructing a three-dimensional temperature field reconstruction mechanism model based on the physical equations of heat conduction, the first step is to establish an accurate mathematical model based on the physical properties of the switchgear's constituent materials and its operating environment. Switchgear typically consists of metallic conductors, insulating materials, and a housing. The thermal conductivity of each material varies significantly, requiring precise values ​​for each. Simultaneously, the heat exchange boundary conditions between the cabinet and its surrounding environment, including convective heat transfer coefficients and radiative heat dissipation coefficients, need to be obtained through experimental measurements or by consulting engineering manuals. These parameters collectively constitute the initial temperature gradient constraint of the heat conduction equations. This constraint defines the initial distribution characteristics and rate of change boundary of the temperature field in the spatial dimension, forming the basis for solving the temperature field.

[0019] The heat conduction process follows Fourier's law and the law of conservation of energy, and its governing equations can be expressed in partial differential form. For three-dimensional entities like switchgear, which contain multiple materials and have complex structures, directly solving these equations analytically is extremely difficult; therefore, numerical methods are used for approximate calculations. The finite volume method is chosen as the discretization method due to its inherent conservation properties. This method systematically divides the entire three-dimensional computational domain of the switchgear into a large number of non-overlapping discrete control volume elements, each of which follows the principle of energy conservation. During discretization, the time domain is also discretized into a series of time steps. For each control volume element, within each time step, the rate of change of its internal energy is equal to the net heat flow into that element plus any heat source terms that may exist within the element. The net heat flow includes the heat conducted through the interfaces of the element, and its calculation depends on the temperature gradient between adjacent elements and the thermal conductivity of the materials. The internal heat source terms mainly consist of Joule heat generated by the current-carrying conductors within the switchgear, the intensity of which is proportional to the square of the operating current. Thus, discretized algebraic equations can be established for each element, and the equations of all elements are combined to form a large sparse linear system of equations.

[0020] Solving this system of equations requires an iterative algorithm. Given the symmetric positive definiteness of its coefficient matrix, the conjugate gradient method or its preprocessed variants are frequently used due to their high convergence efficiency. The iterative process begins with a set initial temperature field and gradually updates the temperature value of each element until the temperature change between two adjacent iterations is less than the set convergence tolerance. At this point, the solution is considered to have reached a stable state, and the output is the theoretical temperature distribution value inside the switchgear. This distribution value is a numerical representation of a spatially continuous function covering the entire computational domain at discrete grid points.

[0021] In the parallel path of processing multimodal sensing data through a spatiotemporal feature extraction network to output real-time temperature predictions, the infrared image sequence is first input into a specially designed 3D convolutional module. The input data for this module is a four-dimensional tensor, with dimensions corresponding to the image batch size, time frame number, image height, image width, and number of infrared channels, respectively. The 3D convolutional kernel performs sliding computations on this data volume, operating across both spatial and temporal dimensions, thus effectively capturing the spatial distribution patterns of thermal radiation signals and their dynamic characteristics evolving over time. Multiple 3D convolutional layers and 3D pooling layers are stacked alternately to gradually abstract and refine higher-level spatial thermal radiation features. These features can characterize the spatial topology of local hot spots, temperature gradient distributions, and anomalous heating regions.

[0022] Acoustic vibration signals and current waveform data, serving as time-series auxiliary information closely related to thermal phenomena, are input into a bidirectional temporal encoder for processing. This encoder is typically built upon gated recurrent units or long short-term memory (LSTM) networks, its core being the introduction of a gating mechanism that effectively learns dependencies in long time series and mitigates the vanishing gradient problem. The bidirectional structure means that data is processed in both forward and backward directions. The forward network processes data from the beginning to the end of the sequence, capturing historical information; the backward network processes data from the end to the beginning, capturing future contextual information. For each time step, the hidden states from the forward and backward directions are fused to obtain a temporally relevant feature representation rich in contextual information. This feature reflects dynamic events related to temperature fluctuations, such as mechanical vibration states, arc discharge signs, and current load changes. The spatial thermal radiation feature tensor output from the 3D convolutional module and the temporally relevant feature vector output from the bidirectional temporal encoder need to be effectively fused. Since the two types of features may exist in different feature spaces, the spatial feature tensor is usually first converted into a feature vector through flattening or global pooling operations. This vector is then concatenated or weighted and added with the temporal feature vector to form a joint feature representation. This joint feature is fed into a multilayer perceptron for further nonlinear transformation and dimensionality reduction. The perceptron typically consists of several fully connected layers, with nonlinear activation functions, such as the ReLU function, between layers. The last layer is the output layer, with the number of neurons corresponding to the number of temperature points to be predicted. The output value is the real-time temperature prediction for each area of ​​the switch cabinet. The parameters of the entire spatiotemporal feature extraction network are obtained through supervised learning training on a large amount of historical data, with the goal of making the predicted temperature as close as possible to the actual temperature measurement.

[0023] Example 2: See Figure 3When inputting theoretical temperature distribution values ​​and real-time temperature prediction values ​​into a dynamic fusion module based on Markov decision processes, the primary task is to construct a state space that comprehensively describes the thermal state of the switchgear. This thermal distribution state space is a high-dimensional mathematical construct, with its number of dimensions strictly corresponding to the number of physical regions into which the switchgear is divided. Each point in the state space, i.e., each specific thermal state, is represented by a multi-dimensional vector. Each component of this vector contains two key data points for the corresponding region: one is the theoretical temperature distribution value calculated by the heat conduction mechanism model, which represents the theoretical expectation based on physical laws; the other is the real-time temperature prediction value output by the spatiotemporal feature extraction network, which reflects the current state presented by multimodal sensing data. This joint state vector constitutes the foundational data for all subsequent decisions and calculations performed by the dynamic fusion module.

[0024] Markov decision processes provide a formal framework for this dynamic fusion problem, where the probabilities of state transitions and the payoffs (or rewards) for taking different actions need to be defined. A specially trained deep neural network acts as the policy network, which maps the observed current thermal state to an optimal action policy. The number of neurons in the input layer of this policy network is exactly the same as the dimension of the state vector, ensuring that information from all regions is fully received. The network typically contains multiple hidden layers, which introduce complex transformation capabilities through nonlinear activation functions, enabling the network to learn a highly nonlinear mapping between states and optimal actions. The output of the policy network is an action vector with the same dimension as the state vector, but each component no longer represents temperature, but rather a pair of confidence weight coefficients. Specifically, for each region, the policy network outputs two weight values: a weight value... Another weighting value is used to weight the theoretical values ​​of the mechanistic model for this region. The neural network predictions used to weight this region are calculated dynamically, and their magnitude depends on the uncertainty, reliability, and historical performance of the various information reflected in the current state.

[0025] Training the policy network is an offline process that requires a large amount of labeled historical running data. The training objective is to maximize a cumulative reward function, the core of which is to penalize the deviation between the fused temperature and the true reference temperature. A common reward function design is a negative mean squared error, meaning the closer the fused temperature is to the true value, the higher the reward (or the smaller the penalty). Through backpropagation algorithms and policy gradient methods, such as the REINFORCE algorithm or the Actor-Critic method, the internal parameters of the policy network are continuously adjusted, ultimately enabling it to learn when to rely more on the theoretical deductions of the mechanistic model and when to rely more on the real-time perception of the neural network. Once the confidence weight vector for the current state is obtained from the policy network, the calculation of the final fused temperature distribution becomes a deterministic weighted fusion operation. For each area in the switch cabinet... Its fusion temperature value The theoretical values ​​from the mechanistic model of this region and neural network predictions The result is obtained by multiplying each component by its corresponding dynamic weight and then summing the results. This calculation process can be uniformly expressed by the following mathematical formula: In this formula: Representative area The final estimated fusion temperature is the core basis for the system to make anomaly judgments and control. Represents the region calculated based on the physical equation of heat conduction. The theoretical temperature distribution value. This represents the region output after processing multimodal data through a spatiotemporal feature extraction network. The real-time temperature prediction value. The representative is dynamically calculated by the policy network for the region. The confidence weight of the theoretical temperature value reflects the reliability of the mechanistic model under the current conditions. The representative is dynamically calculated by the policy network for the region. The confidence weights for the real-time predicted values ​​reflect the quality of the sensor data and the performance of the neural network in the current state. These weights typically need to satisfy normalization constraints, such as for each region. ,have This is to ensure the numerical stability of the fusion process.

[0026] After obtaining the fused temperature distribution values ​​for the entire switchgear, the system initiates a rolling time-domain optimization strategy to generate precise temperature anomaly location commands. This strategy operates within a fixed-length sliding time window, which continuously moves forward, always containing all fused temperature data from the current moment back to a past period. On this spatiotemporal data block, an objective function needs to be constructed to quantify the degree of temperature anomalies. This function is designed to highlight spatial locations where temperature change patterns significantly deviate from expected normal heat dissipation or uniform heating. An effective construction method is to calculate the second-order gradient (Laplace operator) of the temperature field at each point in space, or to calculate the sum of squared residuals between it and a reference temperature distribution model built based on historical normal data.

[0027] After the objective function is constructed, the next task is to find the point in the spatial domain of the switchgear where the function value reaches its minimum. These minimum points correspond to the locations with the lowest "abnormal energy," which are usually the areas with the most abnormal temperature distribution and the most noteworthy. Since the objective function may be very complex and non-convex, numerical optimization algorithms are usually required to solve it. Gradient descent calculates the gradient of the function and iteratively updates the search position along the inverse direction of the gradient, eventually approximating a local minimum point. For large-scale problems, quasi-Newton methods (such as the L-BFGS algorithm) often achieve faster convergence by constructing and utilizing an approximation matrix of the objective function's Hessian matrix. The result of the optimization solution is one or more spatial coordinates located in discrete computational grids or image pixel coordinate systems. The final step is to map these grid coordinates to the actual, continuous physical spatial coordinate system of the switchgear using a pre-calibrated coordinate transformation relationship. The resulting physical coordinates accurately pinpoint the actual location of potential temperature anomalies, forming the core of the temperature anomaly location command sent to the switchgear protection system.

[0028] Example 3: See Figure 4In the operation of a switchgear temperature monitoring system, adaptive calibration of the output layer parameters of the spatiotemporal feature extraction network is a crucial maintenance step aimed at maintaining the long-term accuracy and reliability of real-time temperature predictions. This process is triggered by actual temperature samples collected by contact thermocouple sensors installed at key nodes inside the switchgear. These key nodes typically include busbar connections, circuit breaker moving contacts, cable terminations, and other known heat-prone areas. The system periodically or after specific events reads the actual temperature measurements at these discrete locations and compares them one by one with the temperature estimates for the corresponding locations in the fused temperature distribution, calculating the absolute error at each point. The arithmetic mean of the absolute errors of all these sampling points is calculated and compared to a pre-set tolerance threshold in the system parameters. When the calculated mean absolute error consistently and significantly exceeds this set tolerance, the system logic determines that the current real-time temperature prediction has a systematic deviation, automatically triggering an adaptive calibration procedure for the neural network output layer parameters to ensure the accuracy of the sensing system output.

[0029] The adaptive calibration process follows the fundamental principles of transfer learning and incremental learning. Its core principle is to fine-tune only the output layer responsible for the final regression prediction without compromising the network's existing feature extraction capabilities. Therefore, during calibration, all parameters of the layers used for feature extraction in the spatiotemporal feature extraction network are frozen. This includes all convolutional and pooling layers and their associated parameters within the 3D convolutional module processing infrared image sequences, as well as all recurrent units, gating parameters, and connection weights in the bidirectional time encoder processing acoustic vibration and current waveform data. These fixed parameters represent the network's ability to extract effective spatiotemporal features from raw multimodal signals, learned from a large amount of historical data. These capabilities have universal applicability and should not be altered by short-term deviations in local measurements. The parameter freezing operation ensures that gradient backpropagation during calibration does not update any weights of these layers, thus preserving the learned feature representations.

[0030] The core calculations in the calibration process rely on the error backpropagation algorithm to adjust the unfrozen parameters. The system calculates the actual temperature sampling values ​​at key nodes. The real-time temperature prediction value currently output by the neural network The error between these values ​​is the driving force behind the parameter updates of the network output layer. For each calibration sample... Its error This is defined as the difference between the actual and predicted values. The errors of all calibration samples are summed and used to calculate an overall loss function, which is typically expressed as mean squared error, and its expression is: In this formula: The overall loss function value represents the current batch of calibration data. Its magnitude directly reflects the average deviation between network predictions and actual measurements, and is the target that needs to be minimized in the optimization process. This represents the total number of critical nodes (thermocouple sampling points) participating in the current calibration calculation. The averaging operation ensures that the loss value is independent of the number of sampling points and is comparable. Representative at the The actual temperature sample value obtained by directly measuring the key node locations using contact thermocouples is regarded as the true reference value for the current calibration process. The spatiotemporal feature extraction network represents the first The real-time temperature prediction values ​​output at key node locations are model outputs that need to be adjusted to approximate the actual values.

[0031] The calculated loss function value The gradients of the trainable parameters of the neural network are calculated using the backpropagation algorithm. Since the parameters of the convolutional and encoder layers are frozen, gradient calculations are only backtracked to the multilayer perceptron (MLP), the network's output module. This MLP typically consists of several fully connected layers, whose parameters include the connection weight matrix and bias vectors between layers. Optimization algorithms, such as stochastic gradient descent or variants like the Adam optimizer, iteratively update these unfrozen fully connected weights and bias parameters using the calculated gradient information. Each iteration adjusts the network's output in a direction that reduces the difference between the predicted and actual measurements. This iterative update process continues until the mean absolute error calculated on independent validation datasets decreases and stabilizes below a previously set tolerance threshold, or a preset maximum number of iterations is reached. In this way, the adaptive calibration process effectively corrects the systematic bias of the neural network's output layer, realigning its predictions with the actual temperature measurements at key points, thereby restoring the accuracy of the temperature monitoring system without requiring retraining the entire network.

[0032] Example 4: See Figure 5When implementing infrared image sequence processing and multimodal sensor synchronous acquisition schemes, the raw thermal imaging data acquired by the infrared focal plane array inherently suffers from non-uniformity. This non-uniformity manifests as objects at the same temperature exhibiting different grayscale values ​​at different locations in the image, primarily caused by inconsistencies in the response characteristics of individual pixels within the detector array. Actual measurement data for a certain type of switchgear shows that, under constant temperature testing conditions, the grayscale value fluctuation of the same uniform thermal target in different regions of the infrared image can reach over 15%. The table below shows the typical data distribution of uncorrected infrared detector pixel response differences when performing multi-point temperature measurements on a standard blackbody radiation source under controlled laboratory conditions. The data in this table clearly demonstrates the necessity of non-uniformity correction and provides a basis for setting the parameters of the correction algorithm. See Table 1.

[0033] Table 1: Test data on non-uniformity of pixel response in infrared focal plane array When processing raw infrared image sequences using a scene-based statistical non-uniformity correction algorithm, it is first necessary to establish the gain coefficient and offset parameters for each pixel. These parameters are obtained by analyzing the gray-level statistical characteristics of the same pixel in multiple consecutive frames of images, effectively compensating for differences in the response characteristics of each pixel in the detector. In the corrected image sequence, the gray-level performance of objects at the same temperature in different spatial locations tends to be consistent, providing standardized thermal imaging data for feature extraction. A field test case of a substation switchgear shows that after correction processing, the gray-level fluctuation of the same busbar connection point in different areas of the image is reduced to within 3%, significantly improving the spatial consistency of temperature measurement.

[0034] Deformable convolutional kernels exhibit unique advantages in extracting multi-scale hotspot topology. Taking an overheating fault in the contacts of a 10kV switchgear circuit breaker as an example, the hotspot region in the infrared image sequence has an irregular shape and changes over time. Traditional fixed-shape convolutional kernels struggle to effectively capture these dynamically changing boundary features, while deformable convolution, by introducing learnable offset parameters, allows the sampling grid to adaptively fit the actual shape of the hotspot. In this case, the deformable convolutional kernel automatically adjusts the sampling point distribution in the localized overheating area caused by poor contact, using dense sampling in the hotspot center and sparse sampling in areas with gentle temperature gradients, thus extracting the geometric features of the hotspot more accurately.

[0035] Synchronous acquisition of multimodal sensor arrays requires precise timing control design. In the implementation scheme of a smart substation, the system uses a GPS clock synchronization signal as a unified time reference to coordinate the sampling times of infrared cameras, vibration sensors, and current transformers. Specific configuration parameters include: the infrared focal plane array acquires thermal images at a frame rate of 30Hz, the vibration sensor records acoustic signals at a sampling rate of 25.6kHz, and the current transformer acquires waveform data at a sampling rate of 10kHz. The sampling start time of all sensors is aligned to the rising edge of the PPS (pulses per second) signal, with time synchronization accuracy controlled within 100 microseconds. This precise synchronization mechanism ensures strict alignment of different modal data in the time dimension, creating conditions for cross-modal feature correlation analysis.

[0036] The establishment of an acoustic signal propagation model needs to consider the complexity of the internal structure of the switchgear. Acoustic characteristic tests on a certain type of metal-enclosed switchgear revealed a significant multipath effect in the propagation path of vibration signals within the cabinet. The signals received by vibration sensors installed near the circuit breaker operating mechanism contain multiple components, including direct waves, reflected waves from the cabinet wall, and structural resonance. The propagation model established based on measured data shows significant differences in the frequency response of vibration sensors at different locations, with amplitude differences exceeding 20 dB at certain specific frequency points within the 1kHz to 8kHz frequency band. These characteristics were used to optimize the placement of vibration sensors to ensure that the vibration characteristics of critical mechanical components can be effectively captured.

[0037] The design of a current phase synchronization triggering mechanism needs to adapt to the dynamic characteristics of the power system. In the implementation scheme of a wind farm collector line switchgear, the system tracks the fundamental phase of the current in real time through a phase-locked loop circuit, generating a trigger pulse signal at the current zero-crossing point. This pulse signal is simultaneously sent to the infrared camera and vibration sensor, triggering the start of the acquisition of a new frame of data. Since the current signal itself may have harmonic distortion and noise interference, the trigger circuit is designed with appropriate filtering and hysteresis comparison functions to ensure stable triggering accuracy under the condition that the current waveform distortion rate does not exceed 15%. This design ensures that the acquisition of multi-modal data maintains strict phase synchronization with the actual operating state of the power system, which is beneficial for analyzing the dynamic correlation between electrical load changes and equipment temperature rise.

[0038] Example 5: After obtaining the physical coordinates corresponding to the temperature anomaly location command, they must be accurately mapped to the actual physical space coordinate system of the switchgear. This mapping process relies on a pre-established and rigorously calibrated coordinate transformation model. This model comprehensively considers the installation position of the infrared thermal imager, optical lens parameters, spatial orientation angle, and installation offset relative to the switchgear reference point. Through a coordinate transformation algorithm, the two-dimensional coordinates of the anomaly point in the infrared image pixel coordinate system are first transformed to a three-dimensional camera coordinate system centered on the thermal imager, and then further transformed to a world coordinate system with a fixed corner point or grounding bolt of the switchgear as the origin. The final output is a data structure containing three-dimensional spatial coordinates (X, Y, Z), which clearly indicates the actual physical location of the abnormal heat point inside the switchgear, such as a precise description of "1.2 meters from the front door panel, 0.8 meters from the left side panel, and 1.5 meters from the bottom surface." After mapping, the system needs to generate a graded alarm code based on the equipment type corresponding to the physical location. The system maintains an internal equipment information database, which stores the physical spatial range, equipment type identifier, rated operating parameters, and safety margin thresholds of all important components within the switchgear. By querying this database, the system determines which equipment's spatial envelope the coordinates of an anomaly point fall within, thus identifying the specific equipment involved, such as "A-phase busbar connection bar," "vacuum circuit breaker moving contact," or "outgoing cable bushing." Each equipment type is associated with a set of predefined multi-level temperature thresholds, typically based on the equipment's insulation material heat resistance rating, temperature rise limits under rated current, and historical operating experience. The system compares the temperature at that point in the current integrated temperature distribution with these thresholds level by level, generating a graded alarm code. This code includes not only an anomaly status identifier but also the severity level of the alarm; for example, "01" represents the observation level (temperature exceeds ambient temperature by a certain range but is below the warning value), "02" represents the warning level (temperature exceeds the warning value but is below the alarm value), and "03" represents the emergency alarm level (temperature exceeds the alarm value and approaches the danger value).

[0039] The generated tiered alarm codes are written to a specific register queue within the switchgear protection system. This register queue is a shared memory area within the protection system, specifically used to receive instructions and data from the upper-level monitoring system. The write operation follows a strict communication protocol to ensure data integrity and timing correctness. The protection system's central processing unit periodically polls this register queue; once a new alarm code is detected, it immediately triggers the corresponding interrupt service routine to begin parsing and executing the subsequent logic associated with that alarm code.

[0040] After parsing the alarm code, the protection system needs to match the preset protection action response time constraints from the pre-set plan library based on the type code of the abnormal equipment. Different types of equipment have significantly different protection response time requirements due to differences in their material properties, insulation strength, and the severity of the fault consequences. For example, for epoxy resin insulated current transformers overheating, the insulation material may age rapidly or even crack due to sustained high temperatures, thus requiring a shorter response time, possibly within a few seconds. However, for busbar joints with metal connections, the heat capacity is larger, and short-term overheating may not immediately lead to catastrophic consequences; the allowable response time window may be longer, reaching tens of seconds or even minutes. The system queries the pre-set plan library to obtain the maximum allowable response delay time bound to the specific equipment type code.

[0041] The system analyzes the merged temperature distribution values ​​of the anomaly point and its surrounding area in real time, calculating its temperature gradient change rate. This rate is obtained by comparing the temperature values ​​at the current moment with those at several previous moments using a numerical differentiation method; its magnitude directly reflects the rate of deterioration of the temperature anomaly. A rapidly increasing temperature gradient change rate indicates that the fault is developing rapidly, requiring more urgent handling measures. Based on the calculated gradient change rate, the system calculates a dynamic protection system action priority weight coefficient through a predefined functional relationship. This coefficient is a value between 0 and 1; a higher value indicates a higher degree of urgency and requires priority handling. The protection system's decision logic integrates the priority weight coefficient with the response time constraints obtained from the contingency plan library to generate the final circuit breaker tripping sequence command. This command is not a simple binary "yes" or "no" decision, but a precise timing command containing specific execution times. For example, for high-priority faults requiring rapid response, the instruction might require the relevant circuit breaker to immediately trip at the next current zero-crossing point; for faults with slightly lower priority and more lenient response time constraints, the instruction might be set to delay tripping for a certain period, or issue a warning signal first for operator intervention. This tripping sequence instruction includes the circuit breaker number requiring action, the specified tripping time or delay time, and the expected action result, forming a complete, clear, and executable control command, ultimately completing closed-loop control from temperature sensing to protection action.

[0042] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0043] 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 real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix, characterized in that, include: The infrared image sequence, acoustic vibration signal and current waveform data of the switchgear are acquired synchronously through a multimodal sensor array. A three-dimensional temperature field reconstruction mechanism model is constructed based on the physical equation of heat conduction, and the theoretical temperature distribution value is output. The infrared image sequence, acoustic vibration signal and current waveform data are processed by a spatiotemporal feature extraction network to output real-time temperature prediction values. The theoretical temperature distribution value and the real-time temperature prediction value are input into a dynamic fusion module based on Markov decision process to generate a fused temperature distribution value. Based on the fused temperature distribution value, a temperature anomaly location instruction is generated using a rolling time-domain optimization strategy. When the deviation between the actual temperature sample value of the switch cabinet and the fused temperature distribution value exceeds a preset threshold, the output layer parameters of the spatiotemporal feature extraction network are adaptively calibrated. The linkage control protocol of the switchgear protection system is triggered according to the temperature anomaly location command.

2. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 1, characterized in that, The three-dimensional temperature field reconstruction mechanism model constructed based on the physical equation of heat conduction outputs theoretical temperature distribution values, including: Based on the thermal conductivity of the switchgear material and the heat dissipation boundary conditions, an initial temperature gradient constraint is set for the heat conduction equation. Based on the initial temperature gradient constraint, the theoretical temperature distribution value is calculated iteratively using the finite volume method.

3. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 1, characterized in that, The process of processing the infrared image sequence, acoustic vibration signal, and current waveform data through a spatiotemporal feature extraction network to output a real-time temperature prediction value includes: The infrared image sequence is input into a three-dimensional convolution module to extract spatial thermal radiation features; The acoustic vibration signal and current waveform data are input into a bidirectional time encoder to extract time correlation features. By integrating the spatial thermal radiation characteristics and time correlation characteristics, a real-time temperature prediction value is output through a multilayer sensor.

4. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 1, characterized in that, The step of inputting the theoretical temperature distribution value and the real-time temperature prediction value into a dynamic fusion module based on a Markov decision process to generate a fused temperature distribution value includes: Define a thermal distribution state space, which includes the theoretical temperature distribution values ​​and real-time temperature prediction values ​​of each area of ​​the switchgear. The confidence weights of the theoretical temperature distribution values ​​and the confidence weights of the real-time temperature prediction values ​​are calculated using a policy network. The theoretical temperature distribution value and the real-time temperature prediction value are weighted and fused based on the confidence weight to generate a fused temperature distribution value.

5. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 1, characterized in that, The step of generating temperature anomaly location instructions based on the fused temperature distribution values ​​using a rolling time-domain optimization strategy includes: Construct an objective function for temperature gradient change within a sliding time window; The minimum anomaly energy point of the objective function is found by gradient descent. The physical coordinates corresponding to the minimum abnormal energy point are used as the temperature anomaly location command.

6. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 1, characterized in that, When the deviation between the actual temperature sample value of the switchgear and the fused temperature distribution value exceeds a preset threshold, adaptive calibration is performed on the output layer parameters of the spatiotemporal feature extraction network, including: The actual temperature sampling values ​​of key nodes in the switch cabinet are collected by contact thermocouples. The mean absolute error between the actual temperature sampling values ​​and the fused temperature distribution values ​​is calculated. When the mean absolute error exceeds the set tolerance, adaptive calibration is triggered. The adaptive calibration includes: The convolutional layer and encoder parameters of the spatiotemporal feature extraction network are fixed; Based on the error backpropagation gradient between the actual temperature sample value and the fused temperature distribution value; Only update the fully connected weight parameters of the multilayer perceptron until the mean absolute error is lower than the set tolerance.

7. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 3, characterized in that, The step of inputting the infrared image sequence into a three-dimensional convolution module to extract spatial thermal radiation features includes: Non-uniformity correction is performed on the infrared image sequence to generate standardized thermal imaging data; The standardized thermal imaging data is scanned using deformable convolutional kernels to extract multi-scale hot spot topology. Spatial thermal radiation characteristics are generated based on the spatial correlation of hot spot topology.

8. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 1, characterized in that, The method of synchronously acquiring infrared image sequences, acoustic vibration signals, and current waveform data of the switchgear via a multimodal sensor array includes: The detection angle of the infrared focal plane array is deployed according to the switch cabinet partition topology diagram, the sampling frequency of the vibration sensor is configured based on the acoustic signal propagation model, and the data acquisition timing of the multi-modal sensor array is triggered according to the current phase synchronization.

9. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 5, characterized in that, After using the physical coordinates corresponding to the minimum abnormal energy point as the temperature anomaly location command, the following steps are included: Map the physical coordinates to the actual physical space coordinate system of the switchgear; Based on the equipment safety margin threshold, a graded alarm code is generated and written into the register queue of the switchgear protection system.

10. The method for real-time measurement of switchgear temperature based on multimodal sensing using an infrared matrix according to claim 9, characterized in that, After writing the graded alarm code into the register queue of the switchgear protection system, the following steps are included: Based on the device type code corresponding to the physical coordinates, a preset protection action response time constraint is matched; Based on the abnormal gradient change rate of the fused temperature distribution value, the priority weight coefficient of the protection system action is calculated; The circuit breaker tripping sequence command is generated based on the priority weighting coefficient and response time constraint.

Images