A Method and System for Diagnosing Temperature Anomalies in Power Equipment Based on Infrared Images

Abstract

This invention discloses a method and system for diagnosing temperature anomalies in power equipment based on infrared images, belonging to the fields of photoelectric detection and radiation thermometry. It includes acquiring binocular infrared images, depth maps, and operating parameters to generate a multimodal dataset; denoising and 3D structured fusion of the dataset; generating a prediction benchmark for future moments based on the fusion results and operating parameters using a prediction model; generating a residual map by differencing the future actual image with the prediction benchmark; and analyzing the residual map to output a diagnostic report. This invention uses binocular stereo vision to acquire the 3D geometry and infrared radiation information of the equipment and constructs a prediction model for thermal state prognostic analysis. It can accurately decouple the radiation field of the target from the environment in complex backgrounds and predict changes in radiation characteristics caused by internal defects in advance, significantly improving the accuracy, robustness, and foresight of radiation thermometry diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of photoelectric detection and radiation temperature measurement technology, and in particular to a method and system for diagnosing temperature anomalies in power equipment based on infrared images. Background Technology

[0002] During operation, the thermal state of power equipment is a key indicator for assessing operational safety. Radiation thermometry, particularly infrared thermal imaging, is a non-contact detection method widely used in power system inspection. This method captures the infrared radiation energy emitted from the equipment surface and converts it into a visualized thermal distribution map, enabling remote, uninterrupted safety monitoring of operating equipment and providing crucial technical support for ensuring the stable operation of the power system.

[0003] However, accurately acquiring thermal radiation information of specific components in complex on-site environments remains a challenge. In existing technologies, target devices are often interfered with by foreground or background heat sources, making target identification difficult. Furthermore, traditional two-dimensional thermal imaging lacks spatial depth information, making it difficult to locate abnormal heat sources in three dimensions. Simultaneously, diagnostic conclusions largely rely on static threshold judgments and human experience, resulting in low diagnostic efficiency and unstable results. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method and system for diagnosing temperature anomalies in power equipment based on infrared images. It employs binocular stereo vision to acquire three-dimensional geometric and infrared radiation information of the equipment and constructs a predictive model for thermal state prognosis analysis. This method can accurately decouple the radiation field of the target from the environment in complex backgrounds and predict changes in radiation characteristics caused by internal defects in advance, significantly improving the accuracy, robustness, and foresight of radiation-based temperature diagnosis.

[0005] The above objectives can be achieved through the following approach:

[0006] The method for diagnosing temperature anomalies in power equipment based on infrared images includes: acquiring binocular infrared thermal image sequences, pixel-level depth map sequences, real-time operating parameters of the equipment, and environmental parameters to generate a multimodal raw dataset; performing dual-domain joint denoising processing on the multimodal raw dataset to extract temporal smoothing features and restore frequency domain details, outputting a denoised infrared image dataset and a denoised depth map dataset; segmenting the foreground thermal image based on the denoised depth map dataset, and combining it with the denoised infrared image dataset to generate a structured thermal depth fusion map through image spatial registration and feature fusion; inputting the structured thermal depth fusion map, the real-time operating parameters of the multimodal raw dataset, and environmental parameters into a preset thermal state prediction model to perform time-series extrapolation and generate a predicted thermal distribution benchmark for a preset future time point; acquiring binocular infrared thermal image sequences and pixel-level depth map sequences at the future time point, and performing differential operations with the predicted thermal distribution benchmark to generate a spatiotemporal prediction residual map; extracting spatiotemporal evolution features from the spatiotemporal prediction residual map, classifying and interpreting them, and outputting an infrared image anomaly diagnosis report.

[0007] Optionally, generating the multimodal raw dataset includes: acquiring binocular infrared thermal image sequences, pixel-level depth map sequences, real-time device operating parameters, and environmental parameters; identifying and locking natural calibration objects in the pixel-level depth map sequences; acquiring the apparent infrared radiation temperature of the natural calibration objects in real time; calculating the influence of environmental radiation and atmospheric attenuation on the temperature to obtain a dynamic correction factor; using the dynamic correction factor to perform frame-by-frame correction on the binocular infrared thermal image sequences; and fusing the corrected binocular infrared thermal image sequences with the real-time operating parameters and the environmental parameters to generate the multimodal raw dataset.

[0008] Optionally, the output denoised infrared image dataset and denoised depth map dataset include: analyzing the ideal thermal equilibrium state of the equipment based on the real-time operating parameters and environmental parameters of the equipment in the multimodal original dataset, and constructing a theoretical thermal distribution map; performing differential processing on the infrared image sequence in the multimodal original dataset and the theoretical thermal distribution map to generate a hybrid residual signal map; performing spatiotemporal correlation analysis on the hybrid residual signal map to identify and suppress random uncorrelated noise components and retain locally correlated signal components; superimposing the locally correlated signal components back onto the theoretical thermal distribution map to reconstruct the output denoised infrared image dataset, and simultaneously processing the pixel-level depth map sequence in the multimodal original dataset to output a denoised depth map dataset.

[0009] Optionally, generating the structured thermal depth fusion map includes: using the three-dimensional spatial position of the natural calibration object as a reference anchor point, performing initial rigid spatial registration between the denoised depth map dataset and the denoised infrared image dataset to obtain an initial registration dataset; using the theoretical thermal distribution map as a physical prior, fine-tuning the initial registration dataset to align the thermal field gradient with the physical geometric boundary to obtain a precisely aligned dataset; and performing semantic consistency segmentation on the precisely aligned dataset, using thermal properties as constraints, identifying and associating electrical components and thermophysical parameters to generate a structured thermal depth fusion map.

[0010] Optionally, the method further includes: evaluating the volatility of the dynamic correction factor and statistically analyzing the energy distribution of the mixed residual signal graph; calculating a model reliability index based on the volatility and the energy distribution to determine the degree of mismatch between the thermal state prediction model and the physical entity; dynamically evaluating the model reliability index and triggering online parameter fine-tuning of the thermal state prediction model based on the evaluation results.

[0011] Optionally, the step of generating a predictive thermal distribution benchmark for a preset time point through time-series extrapolation includes: based on the component categories and spatial topological relationships in the structured thermal deep fusion map, abstracting electrical components as nodes and heat conduction paths between components as edges to construct a component thermodynamic map; inputting the component thermodynamic map, real-time operating parameters and environmental parameters of the multimodal original dataset into the thermal state prediction model for iterative solution of heat flow conduction, simulating the thermal state evolution from the current time to the future time point, and generating a prognostic thermodynamic map; mapping the final thermal state of each node in the prognostic thermodynamic map back to the image space to generate a predictive thermal distribution benchmark.

[0012] Optionally, generating the spatiotemporal prediction residual map includes: acquiring a pixel-level depth map sequence at the future time point and performing three-dimensional registration with the geometric structure of the component's thermodynamic spectrum to calculate a non-rigid deformation field; using the non-rigid deformation field to perform a geometric spatial transformation on the predictive thermal distribution benchmark to generate a prediction benchmark; acquiring a binocular infrared thermal image sequence at the future time point and performing a difference operation with the prediction benchmark to generate a preliminary hybrid prediction residual map; performing residual component attribution separation on the preliminary hybrid prediction residual map, and decoupling the radiation artifact component and the real thermodynamic anomaly component by analyzing the spatial frequency and gradient distribution of the residuals to construct the spatiotemporal prediction residual map.

[0013] Optionally, the output infrared image anomaly diagnosis report includes: attributing the residual energy in the spatiotemporal prediction residual map to the corresponding node of the component thermodynamic spectrum, and tracing the causal propagation path of the residual energy in the spectrum topology; fusing the causal propagation path, the non-rigid deformation field, and the model credibility index to construct a multi-dimensional fault feature vector; performing probabilistic reasoning to determine the cause of the fault based on the multi-dimensional fault feature vector, and generating an infrared image anomaly diagnosis report in conjunction with the model credibility index.

[0014] Optionally, the method further includes: correlating the risk level in the infrared image anomaly diagnosis report with the spatial distribution of the non-rigid deformation field, identifying key monitoring areas of the thermo-mechanical coupling effect, and generating attention weights for the key monitoring areas; and based on the attention weights, feeding back and dynamically adjusting the acquisition field of view and spatial resolution of the binocular infrared thermal image sequence.

[0015] Based on the same inventive concept, this invention also provides a power equipment temperature anomaly diagnosis system based on infrared images. The system includes: a data acquisition module for acquiring binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters, and environmental parameters to generate a multimodal raw dataset; a dual-domain denoising processing module for performing dual-domain joint denoising processing on the multimodal raw dataset, extracting temporal smoothing features and restoring frequency domain details, and outputting a denoised infrared image dataset and a denoised depth map dataset; and an image registration and feature fusion module for segmenting the foreground thermal image based on the denoised depth map dataset and combining it with the denoised infrared image dataset through image spatial registration. The system integrates features to generate a structured thermal depth fusion map; a thermal state prediction module inputs the structured thermal depth fusion map, real-time operating parameters and environmental parameters of the multimodal original dataset into a preset thermal state prediction model to perform time-series extrapolation and generate a preset predictive thermal distribution benchmark for future time points; a spatiotemporal residual generation module acquires binocular infrared thermal image sequences and pixel-level depth map sequences at the future time points, performs differential operations with the predictive thermal distribution benchmark, and generates a spatiotemporal prediction residual map; an anomaly diagnosis and analysis module extracts spatiotemporal evolution features from the spatiotemporal prediction residual map, classifies and interprets them, and outputs an infrared image anomaly diagnosis report.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. This invention utilizes binocular stereo vision technology to acquire the three-dimensional geometric information of the device, enabling precise segmentation and locking of the target radiation source under complex backgrounds and obstruction conditions. Combined with an on-site dynamic calibration method, it can compensate for environmental and atmospheric interference on radiation transmission in real time, ensuring the authenticity and accuracy of infrared radiation intensity measurement, and fundamentally improving the reliability of radiation thermometry in variable field environments;

[0018] 2. This invention constructs a predictive model that integrates physical constraints, overcoming the limitation of traditional radiation thermometry, which can only perform static diagnosis. This model can extrapolate the future trends of equipment's infrared radiation characteristics based on current operating conditions. By analyzing the discrepancies between predictions and reality, it can detect early signs of faults, achieving a leap from "fault diagnosis" to "risk prognosis," and providing a forward-looking decision-making basis for preventative maintenance.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the method for diagnosing abnormal temperatures in power equipment based on infrared images, according to an embodiment of the present invention.

[0022] Figure 2 This is a cross-sectional evolution diagram of the spatiotemporal prediction residual map according to an embodiment of the present invention.

[0023] Figure 3 This is a rain cloud diagram showing the residual energy distribution of components according to an embodiment of the present invention.

[0024] Figure 4 This is a fault attribution and causal relationship chord diagram according to an embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of the power equipment temperature anomaly diagnosis system based on infrared images according to an embodiment of the present invention. Detailed Implementation

[0026] 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, 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.

[0027] Reference Figure 1 One embodiment of the present invention proposes a method for diagnosing temperature anomalies in power equipment based on infrared images. It uses binocular stereo vision to acquire the three-dimensional geometry and infrared radiation information of the equipment, and constructs a prediction model for thermal state prognosis analysis. It can accurately decouple the radiation field of the target and the environment in complex backgrounds, and predict changes in radiation characteristics caused by internal defects in advance, which significantly improves the accuracy, robustness and foresight of radiation temperature measurement diagnosis.

[0028] The method described in this embodiment specifically includes:

[0029] Collect binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters and environmental parameters to generate a multimodal raw dataset;

[0030] Perform dual-domain joint denoising processing on the original multimodal dataset, extract temporal smoothing features and restore frequency domain details, and output denoised infrared image dataset and denoised depth map dataset;

[0031] Based on the denoised depth map dataset, the foreground thermal image is segmented, and combined with the denoised infrared image dataset, a structured thermal depth fusion map is generated through image spatial registration and feature fusion.

[0032] The structured thermal deep fusion map, the real-time operating parameters and environmental parameters of the multimodal original dataset are input into a preset thermal state prediction model to perform time-series extrapolation and generate a preset predictive thermal distribution benchmark for future time points.

[0033] A sequence of binocular infrared thermal images and a sequence of pixel-level depth maps are acquired at the future time point, and a difference operation is performed with the predictive thermal distribution benchmark to generate a spatiotemporal prediction residual map.

[0034] Spatiotemporal evolution features are extracted from the spatiotemporal prediction residual map, and then classified and interpreted to output an infrared image anomaly diagnosis report.

[0035] By employing binocular stereo vision to acquire three-dimensional geometric and infrared radiation information of the equipment and constructing a predictive model for thermal state prognosis analysis, it can accurately decouple the radiation field of the target and the environment in complex backgrounds and predict changes in radiation characteristics caused by internal defects in advance, significantly improving the accuracy, robustness and foresight of radiation temperature measurement diagnosis.

[0036] Optionally, generating the multimodal raw dataset includes:

[0037] Collect binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters and environmental parameters, and identify and lock natural markers in the pixel-level depth map sequences;

[0038] Specifically, a Natural Calibration Object (NCO) here refers to an object whose three-dimensional geometric features and thermophysical properties are relatively stable and known within the imaging field of view. A natural calibration object feature library is pre-built, generated by inputting the three-dimensional models and material property parameters of fixed objects in common power scenarios. For example, a fixed cement utility pole, a metal bracket with a uniform surface material, or the wall of a distant building can all serve as a natural calibration object. During processing, a processing device reconstructs the three-dimensional geometric contour of the scene using the acquired pixel-level depth map sequence and matches it with the models in the natural calibration object feature library to identify available natural calibration objects in the current scene, and continuously locks their spatial positions using tracking algorithms such as Kalman filters.

[0039] The apparent infrared radiation temperature of the natural calibrator is acquired in real time, and the effects of environmental radiation and atmospheric attenuation on the temperature are calculated to obtain a dynamic correction factor.

[0040] Specifically, the processing device first acquires the apparent infrared radiation temperature of the locked natural calibrator in real time through an infrared imaging unit; this is an uncorrected apparent measurement. Simultaneously, the processing device retrieves the pre-stored true emissivity and reference temperature of the calibrator from a natural calibrator feature library. Here, the reference temperature, for example, for a non-heat-generating calibrator, can be directly taken as the ambient temperature measured in real time by an environmental parameter sensor. Based on the above inputs, a dynamic correction factor can be calculated using a thermal radiation transfer inverse solution model. An exemplary physical model can be represented by the following formula:

[0041] ,

[0042] in, The calculated dynamic correction factor; The true reference temperature for natural calibration objects; The apparent infrared radiation temperature of a natural calibrator acquired in real time; The surface emissivity of a natural calibrator; The background ambient temperature is obtained through an environmental parameter sensor; This is the integral of Planck's blackbody radiation law over a specific wavelength band, used to characterize the nonlinear relationship between temperature and radiation intensity. It is its inverse function. This dynamic correction factor quantifies in real time the combined error caused by all external factors in the current environment to radiation thermometry.

[0043] The binocular infrared thermal image sequence is corrected frame by frame using the dynamic correction factor, and the corrected binocular infrared thermal image sequence is fused with the real-time operating parameters and the environmental parameters to generate a multimodal raw dataset.

[0044] Specifically, the processing device performs frame-by-frame correction on the binocular infrared thermal image sequence. This means that for each pixel in each frame of the sequence, its original apparent infrared radiation temperature value is corrected according to a dynamic correction factor, resulting in a corrected binocular infrared thermal image sequence that eliminates environmental interference and reflects the true thermal state of the equipment. Finally, the processing device rigorously aligns and structurally encapsulates this corrected sequence with synchronously acquired pixel-level depth map sequences, real-time equipment operating parameters, and environmental parameters, based on a unified high-precision timestamp, generating a multimodal raw dataset that provides high-quality data input for all subsequent diagnostic steps.

[0045] Optionally, the output denoised infrared image dataset and denoised depth map dataset include:

[0046] Based on the real-time operating parameters and environmental parameters of the equipment in the multimodal raw dataset, the ideal thermal equilibrium state of the equipment is analyzed, and a theoretical heat distribution map is constructed.

[0047] Specifically, the theoretical heat distribution map here is not an actual photograph, but rather generated through a multiphysics coupled simulation model for this power equipment. This simulation model is based on the equipment's precise 3D design drawings, material thermophysical properties, and is established using finite element analysis (FEA). During real-time diagnostics, the processing unit extracts real-time operating parameters and environmental parameters from the multimodal raw dataset as real-time inputs to drive the simulation model to solve thermodynamic partial differential equations, thereby calculating the ideal steady-state temperature at each point on the equipment surface under that specific operating condition, ultimately generating the theoretical heat distribution map for that moment.

[0048] The infrared image sequence in the multimodal raw dataset is differentially processed with the theoretical thermal distribution map to generate a hybrid residual signal map;

[0049] Specifically, the processing device performs pixel-level differential operations on the dynamically corrected infrared image sequence from the multimodal raw dataset and the theoretical thermal distribution map generated in the previous step, which is perfectly aligned with the timestamps. The result is a mixed residual signal map, in which the value of each pixel represents the deviation between the actual measured temperature and the theoretical healthy temperature at that location. Ideally, this residual map should contain only two types of information: random, irregular, and uncorrelated noise components introduced by the imaging sensor and signal transmission, and locally correlated signal components with physical evolution characteristics caused by early potential defects in the equipment.

[0050] Spatiotemporal correlation analysis was performed on the hybrid residual signal map to identify and suppress random uncorrelated noise components while retaining locally correlated signal components;

[0051] Specifically, the processing device performs spatiotemporal correlation analysis on the mixed residual signal map. The physical basis for this analysis is that random noise is isolated and uncorrelated in both spatial and temporal dimensions, while a real physical thermal anomaly, even if very weak, will have its energy propagated spatially to neighboring regions and persist in consecutive frames temporally, exhibiting strong spatiotemporal local correlation. An exemplary calculation for quantifying this correlation can be expressed by the following formula:

[0052] ,

[0053] in, For pixels exist The final spatiotemporal correlation score at any given moment; For the mixed residual signal map at the pixel point and Intensity value at any given moment; For pixels The set of pixels in the spatial neighborhood; for A set of time steps preceding a given moment; and These are the weighting coefficients for spatial and temporal correlation, respectively; the rest of the formula is calculated using the standard Pearson correlation coefficient. The mean, The standard deviation is denoted as . The processing device calculates the spatiotemporal correlation score of each pixel in the image. Pixels with scores below a predetermined correlation threshold are identified as random uncorrelated noise components and suppressed, while those with scores above the threshold are identified as locally correlated signal components and are fully preserved.

[0054] The locally correlated signal components are superimposed back onto the theoretical heat distribution map to reconstruct and output a denoised infrared image dataset. Simultaneously, the pixel-level depth map sequence in the multimodal original dataset is processed to output a denoised depth map dataset.

[0055] Specifically, this reconstruction process not only eliminates random noise but also preserves early fault characteristics crucial for diagnosis, ultimately outputting a denoised infrared image dataset. Simultaneously, to ensure modal consistency of the dataset, the processing unit also applies standard denoising algorithms, such as median filtering, to the pixel-level depth map sequences in the original multimodal dataset to eliminate sensor noise, outputting a denoised depth map dataset.

[0056] Optionally, the generation of the structured thermal depth fusion map includes:

[0057] Using the three-dimensional spatial position of the natural calibration object as the reference anchor point, the denoised depth map dataset and the denoised infrared image dataset are initially rigidly registered to obtain the initial registration dataset.

[0058] Specifically, the processing unit first invokes the identified and locked natural landmarks. Since the 3D spatial positions of these natural landmarks are stable and precisely known within the scene, they can serve as highly reliable reference anchor points. Subsequently, a processing unit employs, for example, an Iterative Closest Point (ICP) algorithm or a feature-point-based registration method, using these anchor points as constraints, to perform a comprehensive, deformation-free, rigid spatial transformation on the denoised depth map dataset and the denoised infrared image dataset, thereby aligning them to a unified coordinate system and generating an initial registration dataset.

[0059] Using the theoretical heat distribution map as a physical prior, the initial registration dataset is fine-tuned to align the thermal field gradient with the physical geometric boundary, resulting in a precisely aligned dataset.

[0060] Specifically, the physical prior here refers to the constructed theoretical thermal distribution map, which contains ideal temperature gradient information that conforms to physical laws. For example, there should be a significant temperature jump at the interface between components made of different materials. The processing device employs a deformable registration algorithm. During fine-tuning, it not only matches the features of the image itself but also, guided by the theoretical thermal distribution map, maximizes the overlap between the actual thermal field gradient of the thermal image and the physical geometric boundary of the depth map in the initial registration dataset. Through this physical prior constraint, the physical correctness of the registration result is ensured, ultimately resulting in a precisely aligned dataset.

[0061] On the precisely aligned dataset, semantic consistency segmentation is performed with thermal properties as constraints to identify and associate electrical components and thermophysical parameters, generating a structured thermal depth fusion map.

[0062] Specifically, the semantically consistent segmentation performed by the processing device is an intelligent segmentation method that incorporates prior knowledge. It is constrained by a component knowledge base containing various electrical components and their thermal properties. During segmentation, the algorithm not only utilizes the color and geometric information of the image but also determines whether the segmented region conforms to the thermal properties of the identified component. For example, a region identified as a "copper connector" must have a temperature distribution that conforms to the thermal characteristics of a conductor. In this way, the processing device can accurately identify each electrical component and correlate its position, shape, real-time temperature field, and pre-stored thermophysical parameters in three-dimensional space to generate a structured thermal depth fusion map—a highly integrated digital snapshot of the device that can be used for subsequent advanced analysis.

[0063] Optionally, the method further includes:

[0064] The volatility of the dynamic correction factor is evaluated, and the energy distribution of the mixed residual signal plot is statistically analyzed.

[0065] Specifically, the processing device periodically performs two-dimensional evaluations within a preset time window. First, it evaluates the volatility of the generated dynamic correction factor, for example, by calculating its standard deviation or coefficient of variation. This volatility directly reflects the stability of the measurement environment; drastic fluctuations indicate a decrease in the reliability of the data source input to the prediction model. The overall energy distribution of the mixed residual signal map and the spatiotemporal prediction residual map is calculated, for example, by calculating the L2 norm of the intensity of all pixels within the map. The energy of the mixed residual signal map represents the degree of deviation of the model from its interpretation of the normal thermal behavior of the equipment, while the energy of the spatiotemporal prediction residual map represents the accuracy of the model's prediction of future states.

[0066] Based on the volatility and energy distribution, calculate the model credibility index of the degree of mismatch between the thermal state prediction model and the physical entity;

[0067] Specifically, the model credibility index is a comprehensive score used to characterize the degree of match between the prediction results of the thermal state prediction model at the current moment and the actual physical state of the equipment. An exemplary model for calculating this index can be represented by the following formula:

[0068] ,

[0069] in, The final calculated model credibility index has a value range between 0 and 1, with the value closer to 1 indicating higher credibility. The standard deviation of the dynamic correction factor within a time window characterizes its volatility; The weighted sum of the energy distributions of the hybrid residual signal map and the spatiotemporal prediction residual map characterizes the overall prediction error of the model. and These are weighting coefficients used to balance the impact of input data quality and model fitting accuracy on overall reliability. This is the natural exponential function, used to map the error term to the confidence score. The calculation of this index provides an objective and quantitative basis for deciding whether subsequent model adjustments are necessary.

[0070] The reliability index of the model is dynamically evaluated, and the online parameter fine-tuning of the thermal state prediction model is triggered based on the evaluation results.

[0071] Specifically, the processing unit continuously monitors the changing trend of the model's reliability index. When this index continues to decline over a period of time, or when a single evaluation result falls below a preset dynamic reliability threshold, it is determined that the thermal state prediction model can no longer accurately match the current physical state of the equipment, and an online parameter fine-tuning mechanism is automatically triggered. This fine-tuning process does not completely retrain the model, but rather uses recently collected and validated high-quality data to perform small-scale iterative corrections to the key physical parameters or network weights within the model, generating an updated thermal state prediction model. This updated model will be applied to subsequent diagnostic cycles, thereby achieving closed-loop adaptation and continuous evolution of the entire diagnostic system.

[0072] Optionally, the step of generating a predictive heat distribution benchmark at a preset time point through time-series simulation includes:

[0073] Based on the component categories and spatial topology relationships in the structured thermal fusion graph, electrical components are abstracted as nodes and heat conduction paths between components are abstracted as edges to construct a component thermodynamic graph.

[0074] Specifically, the processing device first parses the generated structured thermal depth fusion map. Each identified electrical component in the map, such as a circuit breaker contact or a section of busbar, is abstracted as a node in the map. The attributes of this node are assigned the component's comprehensive thermophysical parameters, including its current average temperature, material, geometry, and heat capacity and thermal conductivity retrieved from the component knowledge base. Simultaneously, based on the spatial topology provided by the structured thermal depth fusion map, an edge is established between the nodes corresponding to two physically contacting components or those with significant thermal radiation relationships. The weight of the edge is assigned a value representing the thermal conductivity or thermal resistance between the two components, thereby constructing a component thermodynamic map that can fully describe the thermodynamic relationships of the equipment.

[0075] The component thermodynamic map, the real-time operating parameters and environmental parameters of the multimodal raw dataset are input into the thermal state prediction model to perform heat flow conduction iterative solution, simulate the thermal state evolution from the current time to the future time point, and generate a prognostic thermodynamic map;

[0076] Specifically, the thermal state prediction model here is implemented as a numerical simulation solver that iteratively solves the heat conduction equation on the component's thermodynamic map. In each iteration, the model updates the state of all nodes in the graph. This update process couples three major thermophysical effects: First, internal heat source generation, where the model calculates the heat generated inside nodes due to current based on real-time operating parameters from the multimodal raw dataset, such as using Joule's law; second, inter-component heat conduction, where the model calculates and updates the heat transfer between connected nodes along the graph edges according to Fourier's law; and third, heat exchange with the environment, where the model calculates the heat lost by each node to the surrounding environment through convection and radiation based on environmental parameters, using Newton's law of cooling and the Stefan-Boltzmann law. By continuously executing an equivalent number of iterations to a preset future time point, the model fully simulates the dynamic evolution of thermal energy throughout the entire device topology. Its final output is a prognostic thermodynamic map containing the predicted temperatures of all nodes at future times.

[0077] The final thermal state of each node in the prognostic thermodynamic map is mapped back to the image space to generate a predictive thermal distribution baseline.

[0078] Specifically, the processing device retrieves the final predicted temperature of each node in the pre-thermodynamic map. Then, based on the precise pixel regions of each component identified in the structured thermal depth fusion map within the image space, the predicted temperature value of the corresponding node is "rendered" or "drawn" back to that region. For the temperature distribution within the components, methods such as gradient interpolation can be used to ensure that the generated temperature field is visually smooth and conforms to physical laws. By mapping all components, a predictive thermal distribution reference image characterizing the device's ideal future thermal state is finally generated in the same coordinate system as the original infrared image.

[0079] Optionally, generating the spatiotemporal prediction residual map includes:

[0080] At the future time point, a pixel-level depth map sequence is acquired and three-dimensionally registered with the geometric structure of the component's thermodynamic map to calculate the non-rigid deformation field.

[0081] Specifically, equipment may experience minute deformations or displacements due to thermal expansion and contraction or mechanical stress. Directly comparing two thermal images from different times would introduce significant geometric errors. Therefore, the processing device first acquires a new sequence of pixel-level depth maps at a future time point to obtain the equipment's true 3D geometry at that moment. Then, using a non-rigid 3D registration algorithm, this new geometry is compared with the initial geometry implied in the constructed thermodynamic map of the component, thereby calculating a non-rigid deformation field (NDF). This deformation field is a three-dimensional vector field that precisely describes the spatial displacement of each point on the equipment surface from its initial position to its current position.

[0082] The predictive thermal distribution benchmark is generated by performing a geometric spatial transformation on the non-rigid deformation field.

[0083] Specifically, the processing device uses the non-rigid deformation field as a spatial transformation function, applying it to the predictive thermal distribution baseline. This transformation process involves pixel-level resampling and interpolation of the predictive baseline map, "distorting" or "mapping" it from its initial geometry to its current deformed geometry. This step generates a pose-corrected predictive baseline. This baseline map describes the temperature distribution that the device should theoretically exhibit in its current deformed physical state, assuming no thermal anomalies.

[0084] At the future time point, a sequence of binocular infrared thermal images is acquired and differentially analyzed with the prediction benchmark to generate a preliminary hybrid prediction residual map.

[0085] Specifically, the processing device will perform pixel-level differential calculations between the binocular infrared thermal image sequence acquired synchronously at future time points and the pose-corrected prediction benchmark generated in the previous step. Since the geometry is already perfectly aligned, the result of this differential calculation can more purely reflect the deviation at the thermodynamic level. Its output is a preliminary mixed prediction residual map, in which the non-zero regions theoretically contain only two types of deviations: radiation artifacts caused by changes in the emissivity of the device surface, and temperature anomalies caused by actual internal fault sources.

[0086] The residual components of the preliminary mixed prediction residual map are attributed and separated. By analyzing the spatial frequency and gradient distribution of the residuals, the radiation artifact component and the real thermodynamic anomaly component are decoupled, and a spatiotemporal prediction residual map is constructed.

[0087] Specifically, the processing device performs residual component attribution separation on the preliminary mixed prediction residual map. The physical principle behind this method is that radiation artifacts are typically caused by changes in surface state, and are characterized by low spatial frequency, gentle gradients, and patchy distribution on the residual map; while true thermodynamic anomalies are caused by concentrated internal heat sources, and are typically characterized by high spatial frequency, sharp temperature gradients, and localization. The processing device can use frequency domain analysis tools such as wavelet transform or Fourier transform, combined with gradient operators, to analyze the residual map, thereby successfully decoupling these two types of components. Finally, the processing device extracts and reconstructs only the true thermodynamic anomaly components to construct the final, clean spatiotemporal prediction residual map, ensuring extremely high signal-to-noise ratio and reliability for subsequent diagnostic analysis. Figure 2 As shown in the figure, the X-axis represents the spatial location of the device surface, the Y-axis represents the time evolution, and the Z-axis represents the residual energy intensity. By stacking and filling one-dimensional residual profile curves at different times, this figure comprehensively demonstrates how the residual signal of an early fault emerges from the background noise and how its amplitude and range increase over time.

[0088] Optionally, the output infrared image anomaly diagnostic report includes:

[0089] The residual energy in the spatiotemporal prediction residual map is attributed to the corresponding node of the component thermodynamic spectrum, and the causal propagation path of the residual energy in the spectrum topology is traced.

[0090] Specifically, the processing device first spatially correlates the generated spatiotemporal prediction residual map with the constructed component thermodynamic map. By calculating the residual energy integral within the pixel region covered by each component, the residual energy is attributed to the corresponding node in the map. Subsequently, the processing device analyzes the sequence of residual energy changes at each node within consecutive time frames. Through gradient analysis or dynamic time warping algorithms, it tracks the sequence and direction of energy peak propagation between edges and nodes in the map, thereby determining the causal propagation path. For example, if the residual energy of node A increases first, followed by an increase in the energy of its neighboring node B, a causal propagation path from A to B is determined. This strongly suggests that the root cause of the fault is located in or near component A. Figure 3 As shown, this is a rain cloud map of energy distribution after attributing the spatiotemporal prediction residual map to different electrical components. This map compositely displays the probability density distribution, key statistical quantiles, and original data points of the residual energy for each component, allowing for intuitive comparison and identification of core abnormal components from a statistical perspective.

[0091] By integrating the causal propagation path, the non-rigid deformation field, and the model credibility index, a multi-dimensional fault feature vector is constructed.

[0092] Specifically, the processing device fuses multiple key pieces of information from different technical dimensions. This fusion process combines and normalizes the causal propagation path obtained in the previous step, the key statistics of the calculated non-rigid deformation field, and the calculated model credibility index, ultimately constructing a multi-dimensional fault feature vector. This vector comprehensively describes the thermodynamic evolution characteristics, mechanical correlation effects, and credibility of this observation of an anomalous event.

[0093] Based on the multi-dimensional fault feature vector, probabilistic reasoning is performed to determine the cause of the fault, and an infrared image anomaly diagnosis report is generated by combining the model credibility index.

[0094] Specifically, the processing device employs a probabilistic reasoning method based on Bayesian networks. The relational structure of this network is defined by a pre-defined fault reasoning knowledge base, which is constructed from the experience of power industry experts and a large amount of historical fault data. This knowledge base includes conditional probability relationships between different root causes of faults and various feature components in a multi-dimensional fault feature vector. The processing device inputs the real-time constructed feature vectors as observational evidence into the Bayesian network and calculates the posterior probability of various potential fault causes using Bayes' theorem. An exemplary calculation can be expressed by the following formula:

[0095] ,

[0096] in, In order to observe the feature vector Under the condition that the k-th type of failure occurs Posterior probability; Stored in the knowledge base, when a fault occurs Observed eigenvectors The likelihood probability; For fault The prior probability is calculated. The processing device calculates and selects the fault type with the highest posterior probability as the most likely root cause. The final infrared image anomaly diagnosis report not only includes a description of the fault cause, but also a diagnostic confidence score calculated by weighting the posterior probability and the model confidence index. Based on the severity of the fault and the confidence score, it matches and outputs the corresponding risk level and maintenance priority from a preset risk matrix, such as... Figure 4 As shown in the diagram, the outer nodes represent the key components and core fault characteristics of the equipment, and the chords between the nodes demonstrate the strength of their correlation. For example, the chord emanating from the "top connector" node clearly quantifies the salience of this component in characteristics such as "residual energy" and "causal propagation source," while the chord connecting the "top connector" and the "main insulating sleeve" indicates a strong causal propagation relationship between the two.

[0097] Optionally, the method further includes:

[0098] Correlation analysis is performed between the risk level in the infrared image anomaly diagnosis report and the spatial distribution of the non-rigid deformation field to identify key monitoring areas of the thermo-mechanical coupling effect, and attention weights are generated for the key monitoring areas.

[0099] Specifically, the processing device first retrieves the generated infrared image anomaly diagnostic report and extracts risk level information for each component. Simultaneously, it retrieves the calculated non-rigid deformation field. Subsequently, the processing device performs spatial correlation analysis on these two datasets to identify areas with high risk levels and significant physical deformation. These areas are identified as key monitoring areas for thermo-mechanical coupling effects, as the coexistence of thermal and mechanical stresses typically indicates more severe or urgent failures. Finally, based on the risk level and the magnitude of deformation, the processing device generates a quantified, high-value attention weight for these key monitoring areas, while assigning lower weights to other normal areas, thus forming an "attention map" to guide future perception.

[0100] Based on the attention weight, the acquisition field of view and spatial resolution of the binocular infrared thermal image sequence are fed back and dynamically adjusted.

[0101] Specifically, the processing device uses the attention weight map generated in the previous step as a control command and feeds it back to the front-end binocular infrared thermal image acquisition unit. This feedback triggers the acquisition unit to dynamically adjust its subsequent acquisition strategy: First, in terms of the acquisition field of view, the acquisition unit prioritizes aligning the camera's field of view (FOV) with and covering key monitoring areas with high attention weights; second, in terms of spatial resolution, the acquisition unit can increase the imaging pixel density in key monitoring areas through optical zoom or digital image cropping. Through this dynamic adjustment based on diagnostic result feedback, the method of this invention can intelligently focus limited sensing resources on the weakest links most likely to fail, realizing a shift from a passive, fixed scanning mode to an active, focused enhanced perception and key monitoring, thereby greatly improving the ability to capture early signs of major faults and diagnostic efficiency.

[0102] To verify the feasibility and effectiveness of this invention in practice, it was applied to a 220kV hub substation to conduct continuous online diagnostic monitoring of a critical SF6 high-voltage circuit breaker for six months. This substation has a complex environment with small equipment spacing, numerous background interference sources, and variable weather conditions, posing a significant challenge to traditional infrared detection technology. Traditional periodic manual inspection methods struggle to accurately quantify the thermal state of equipment under different operating conditions and environments, and cannot provide early warnings for progressive faults. This substation aims to use the method of this invention to achieve accurate and proactive diagnosis of the circuit breaker's condition, preventing potential problems before they occur.

[0103] In this embodiment, the processing device is deployed at the monitoring master station, and the front end uses the binocular infrared imaging device of this invention to continuously observe the target circuit breaker. The collected data is processed through our constructed complete diagnostic process. First, on-site dynamic calibration technology ensures the accuracy of the measurement data; then, a physical constraint intelligent denoising method is used to extract meaningful weak thermal signals from complex background noise; next, structured fusion is used to construct a digital snapshot of the circuit breaker; based on this, the core thermal state prediction model performs prognostic analysis on the future thermal trend of the circuit breaker, and through a series of advanced analysis methods, a reliable, interpretable, and action-guided diagnostic report is finally output.

[0104] To verify the beneficial effects of this invention, we recorded several key events and corresponding data during actual operation. On the afternoon of August 5, 2025, with strong sunlight and gusts of wind, a fixed steel structure support next to the circuit breaker was used as a natural calibration object. A dynamic correction factor as high as +3.5℃ was calculated in real time, successfully offsetting the measurement errors caused by solar radiation reflection and wind cooling effects, and the corrected true temperature data was obtained, as shown in Table 1. On the night of August 10, by comparing the theoretical thermal distribution map with the actual measured values, spatiotemporal correlation analysis was used to successfully separate a slight temperature rise of only 0.8℃ at a connection terminal from the background sensor noise from the mixed residual signal map, whereas traditional filtering methods would smooth out this signal.

[0105] The most representative event occurred on September 15th. After a normal high-load operation, the structured thermal depth fusion map of the circuit breaker showed everything was normal. However, the predictive thermal distribution benchmark output by the thermal state prediction model, based on the component's thermodynamic map, indicated that the temperature of the bushing connection seat on the A-phase outgoing side would experience a nonlinear abnormal temperature rise after 15 minutes. After 15 minutes, new actual images were acquired. By performing three-dimensional registration with the component's thermodynamic map, a non-rigid deformation field of 0.2 mm was calculated at this location. The spatiotemporal prediction residual map generated after geometric correction of the prediction benchmark clearly showed a real thermodynamic anomaly component that highly matched the prediction, as shown in Table 2.

[0106] Finally, in the diagnostic phase, the residual energy was attributed to the component's thermodynamic spectrum, tracing a causal propagation path from the inside of the connector to the external lead. By integrating the characteristics of this path, the non-rigid deformation field, and the then-high model confidence index of 0.95, a multi-dimensional fault feature vector was constructed. Through Bayesian probabilistic inference, the output infrared image anomaly diagnostic report clearly indicated that the root cause of the fault was "loose bolts on the conductive rod inside the bushing connector," with a diagnostic confidence level of 92%, a risk level of "high," and a maintenance priority of "urgent." Simultaneously, based on this high-risk diagnostic result, a high-attention weight was generated and fed back to the front-end acquisition device, automatically adjusting the acquisition strategy for the next cycle, and performing intensive observation and high-definition imaging of the fault location, as shown in Table 3.

[0107] Table 1. Data on Dynamic Environmental Correction and Intelligent Noise Reduction Effects

[0108]

[0109] Table 2 Comparison of Predictive Diagnosis and Actual Temperature Rise

[0110]

[0111] Table 3 Multimodal Fault Reasoning and Response Processing Data Table

[0112]

[0113] As can be seen from the data in Tables 1-3 above, the method of this invention exhibits superior performance in practical applications. Table 1 shows that under strong environmental interference, dynamic correction controls the measurement error to a very small range, and intelligent denoising successfully preserves weak early fault signals. The data in Table 2 strongly demonstrates the core predictive capability of this invention, successfully providing an early warning of a significant abnormal temperature rise 15 minutes in advance, with the predicted value highly consistent with the actual value. Table 3 showcases the highest level of intelligence of this invention, which not only provides the superficial indication of "where it is hot," but also accurately determines the root cause of "why it is hot" through multimodal information fusion and causal reasoning, providing decision-making criteria including confidence level and risk level. Finally, it can optimize its monitoring behavior through feedback loop, and its comprehensive effect far exceeds that of existing technologies.

[0114] Based on the same inventive concept, this invention also provides a power equipment temperature anomaly diagnosis system based on infrared images, such as... Figure 5 As shown, the system includes:

[0115] The data acquisition module is used to acquire binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters and environmental parameters, and generate multimodal raw datasets;

[0116] The dual-domain denoising module is used to perform dual-domain joint denoising processing on the multimodal raw dataset, extract temporal smoothing features and restore frequency domain details, and output denoised infrared image dataset and denoised depth map dataset.

[0117] The image registration and feature fusion module is used to segment the foreground thermal image based on the denoised depth map dataset, and generate a structured thermal depth fusion map by combining the denoised infrared image dataset through image spatial registration and feature fusion.

[0118] The thermal state prediction module is used to input the structured thermal deep fusion map, the real-time operating parameters and environmental parameters of the multimodal original dataset into the preset thermal state prediction model, and perform time-series extrapolation to generate a preset predictive thermal distribution benchmark for future time points;

[0119] The spatiotemporal residual generation module is used to acquire binocular infrared thermal image sequences and pixel-level depth map sequences at the future time point, and perform differential operations with the predictive thermal distribution benchmark to generate a spatiotemporal prediction residual map.

[0120] The anomaly diagnosis and analysis module is used to extract spatiotemporal evolution features from the spatiotemporal prediction residual map, classify and interpret them, and output an infrared image anomaly diagnosis report.

[0121] It should be noted that the functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between them represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for diagnosing temperature anomalies in power equipment based on infrared images, characterized in that, The method includes: Collect binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters and environmental parameters to generate a multimodal raw dataset; Perform dual-domain joint denoising processing on the original multimodal dataset, extract temporal smoothing features and restore frequency domain details, and output denoised infrared image dataset and denoised depth map dataset; Based on the denoised depth map dataset, the foreground thermal image is segmented, and combined with the denoised infrared image dataset, a structured thermal depth fusion map is generated through image spatial registration and feature fusion. The structured thermal fusion map, real-time operating parameters, and environmental parameters of the multimodal original dataset are input into a preset thermal state prediction model to generate a predictive thermal distribution benchmark for a preset future time point through time-series extrapolation. The generation of the predictive thermal distribution benchmark for the preset future time point through time-series extrapolation includes: based on the component categories and spatial topology relationships in the structured thermal fusion map, abstracting electrical components as nodes and heat conduction paths between components as edges to construct a component thermodynamic map; inputting the component thermodynamic map, real-time operating parameters, and environmental parameters of the multimodal original dataset into the thermal state prediction model for iterative heat flow conduction solution, simulating the thermal state evolution from the current time to the future time point, and generating a prognostic thermodynamic map; mapping the final thermal state of each node in the prognostic thermodynamic map back to the image space to generate a predictive thermal distribution benchmark. At the future time point, a sequence of binocular infrared thermal images and a sequence of pixel-level depth maps are acquired, and differential operations are performed with the predictive thermal distribution benchmark to generate a spatiotemporal prediction residual map. Generating the spatiotemporal prediction residual map includes: at the future time point, acquiring a sequence of pixel-level depth maps and performing three-dimensional registration with the geometric structure of the component's thermodynamic spectrum to calculate a non-rigid deformation field; using the non-rigid deformation field to perform a geometric spatial transformation on the predictive thermal distribution benchmark to generate a prediction benchmark; at the future time point, acquiring a sequence of binocular infrared thermal images and performing differential operations with the prediction benchmark to generate a preliminary hybrid prediction residual map; performing residual component attribution separation on the preliminary hybrid prediction residual map, and decoupling the radiation artifact component and the true thermodynamic anomaly component by analyzing the spatial frequency and gradient distribution of the residuals to construct the spatiotemporal prediction residual map. Spatiotemporal evolution features are extracted from the spatiotemporal prediction residual map, and then classified and interpreted to output an infrared image anomaly diagnosis report.

2. The method for diagnosing temperature anomalies in power equipment based on infrared images according to claim 1, characterized in that, The generated multimodal raw dataset includes: Collect binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters and environmental parameters, and identify and lock natural markers in the pixel-level depth map sequences; The apparent infrared radiation temperature of the natural calibrator is acquired in real time, and the effects of environmental radiation and atmospheric attenuation on the temperature are calculated to obtain a dynamic correction factor. The binocular infrared thermal image sequence is corrected frame by frame using the dynamic correction factor, and the corrected binocular infrared thermal image sequence is fused with the real-time operating parameters and the environmental parameters to generate a multimodal raw dataset.

3. The method for diagnosing temperature anomalies in power equipment based on infrared images according to claim 2, characterized in that, The output denoised infrared image dataset and the denoised depth map dataset include: Based on the real-time operating parameters and environmental parameters of the equipment in the multimodal raw dataset, the ideal thermal equilibrium state of the equipment is analyzed, and a theoretical heat distribution map is constructed. The infrared image sequence in the multimodal raw dataset is differentially processed with the theoretical thermal distribution map to generate a hybrid residual signal map; Spatiotemporal correlation analysis was performed on the hybrid residual signal map to identify and suppress random uncorrelated noise components while retaining locally correlated signal components; The locally correlated signal components are superimposed back onto the theoretical heat distribution map to reconstruct and output a denoised infrared image dataset. Simultaneously, the pixel-level depth map sequence in the multimodal original dataset is processed to output a denoised depth map dataset.

4. The method for diagnosing temperature anomalies in power equipment based on infrared images according to claim 3, characterized in that, The generated structured thermal depth fusion map includes: Using the three-dimensional spatial position of the natural calibration object as the reference anchor point, the denoised depth map dataset and the denoised infrared image dataset are initially rigidly registered to obtain the initial registration dataset. Using the theoretical heat distribution map as a physical prior, the initial registration dataset is fine-tuned to align the thermal field gradient with the physical geometric boundary, resulting in a precisely aligned dataset. On the precisely aligned dataset, semantic consistency segmentation is performed with thermal properties as constraints to identify and associate electrical components and thermophysical parameters, generating a structured thermal depth fusion map.

5. The method for diagnosing temperature anomalies in power equipment based on infrared images according to claim 3, characterized in that, The method further includes: The volatility of the dynamic correction factor is evaluated, and the energy distribution of the mixed residual signal plot is statistically analyzed. Based on the volatility and energy distribution, calculate the model credibility index of the degree of mismatch between the thermal state prediction model and the physical entity; The reliability index of the model is dynamically evaluated, and the online parameter fine-tuning of the thermal state prediction model is triggered based on the evaluation results.

6. The method for diagnosing temperature anomalies in power equipment based on infrared images according to claim 5, characterized in that, The abnormality diagnosis report of the output infrared image includes: The residual energy in the spatiotemporal prediction residual map is attributed to the corresponding node of the component thermodynamic spectrum, and the causal propagation path of the residual energy in the spectrum topology is traced. By integrating the causal propagation path, the non-rigid deformation field, and the model credibility index, a multi-dimensional fault feature vector is constructed. Based on the multi-dimensional fault feature vector, probabilistic reasoning is performed to determine the cause of the fault, and an infrared image anomaly diagnosis report is generated by combining the model credibility index.

7. The method for diagnosing temperature anomalies in power equipment based on infrared images according to claim 1, characterized in that, The method further includes: Correlation analysis is performed between the risk level in the infrared image anomaly diagnosis report and the spatial distribution of the non-rigid deformation field to identify key monitoring areas of the thermo-mechanical coupling effect, and attention weights are generated for the key monitoring areas. Based on the attention weight, the acquisition field of view and spatial resolution of the binocular infrared thermal image sequence are fed back and dynamically adjusted.

8. A power equipment temperature anomaly diagnosis system based on infrared images, applied to the power equipment temperature anomaly diagnosis method based on infrared images as described in any one of claims 1-7, characterized in that, The system includes: The data acquisition module is used to acquire binocular infrared thermal image sequences, pixel-level depth map sequences, real-time equipment operating parameters and environmental parameters, and generate multimodal raw datasets; The dual-domain denoising module is used to perform dual-domain joint denoising processing on the multimodal raw dataset, extract temporal smoothing features and restore frequency domain details, and output denoised infrared image dataset and denoised depth map dataset. The image registration and feature fusion module is used to segment the foreground thermal image based on the denoised depth map dataset, and generate a structured thermal depth fusion map by combining the denoised infrared image dataset through image spatial registration and feature fusion. The thermal state prediction module is used to input the structured thermal deep fusion map, the real-time operating parameters and environmental parameters of the multimodal original dataset into the preset thermal state prediction model, and perform time-series extrapolation to generate a preset predictive thermal distribution benchmark for future time points; The spatiotemporal residual generation module is used to acquire binocular infrared thermal image sequences and pixel-level depth map sequences at the future time point, and perform differential operations with the predictive thermal distribution benchmark to generate a spatiotemporal prediction residual map. The anomaly diagnosis and analysis module is used to extract spatiotemporal evolution features from the spatiotemporal prediction residual map, classify and interpret them, and output an infrared image anomaly diagnosis report.

Images