1380 results about "Learning network" patented technology

A computer system for adaptive operation of deep learning networks through hierarchical supervision, meta-level pattern tracking, cross-network signal coordination, and selective activation prioritization. The system operates a layered neural network monitored by a hierarchical supervisory system that collects activation data, identifies operational patterns, implements architectural modifications, detects network sparsity, coordinates pruning decisions, and manages resource redistribution. A meta-supervisory system tracks supervisory behavior, stores successful pruning and modification patterns, and extracts generalizable optimization principles. The system manages signal transmission pathways that enable direct communication between non-adjacent network regions, with signal modification and temporal coordination. A greedy neural system selectively processes activation patterns based on utility metrics and includes a competitive bidding manager to allocate limited computational resources to high-value signals. This architecture enables real-time optimization of network behavior and resource usage while maintaining operational stability and responsiveness across diverse applications.

The invention discloses a physical information neural network-driven ocean island typhoon scene generation method and system. The method comprises the steps of collecting multi-source heterogeneous meteorological data and performing space-time alignment preprocessing; constructing a coarse-scale space-time probability prediction model, capturing space correlation of meteorological elements by using a graph topology learning network, efficiently processing long-time-sequence dependence of typhoon evolution by integrating a state space model with linear complexity, and generating a probabilistic typhoon scene with coarse resolution through a multivariable joint distribution probability model; further constructing a physical downscaling model, taking a coarse-scale prediction result as condition input, and performing physical consistency downscaling on a coarse-scale scene by embedding an atmospheric fluid mechanics equation in a loss function as a physical hard constraint; and finally, outputting a high-resolution typhoon scene with probability reliability and physical authenticity.

The invention provides a medical RFQ accelerator low-level signal phase stability control method and system. The method comprises the following steps: constructing a time-frequency energy spectrum feature vector based on wavelet packet transformation; extracting a second disturbance feature based on a lightweight convolutional neural network and an attention mechanism; constructing a phase dynamic trend prediction module based on a long short-term memory network, and obtaining a first prediction phase error; constructing a phase compensation module based on a residual control network to obtain a second phase compensation amount; and outputting a real-time driving control signal based on the extended Kalman filter. According to the method, the time-frequency energy spectrum feature vector based on wavelet packet transformation is constructed, accurate characterization of the multi-scale disturbance features of the low-level signals is achieved, a medical RFQ accelerator phase dynamic compensation system is established in combination with a deep learning network and an extended Kalman filtering algorithm, the control precision and the anti-interference capability of signal phase stability are remarkably improved, and the method is suitable for popularization and application. The method is suitable for a high-precision medical particle accelerator control system.

The invention relates to the technical field of engineering cost control, and discloses an engineering cost control method and system based on big data. The method comprises the steps of collecting engineering project full-cycle cost data streams, and generating a standardized cost data set through data cleaning; constructing a dynamic cost feature library, and extracting multi-dimensional features such as time sequence fluctuation, resource allocation discretization and supplier association; inputting the feature library into a cost anomaly detection model constructed by a pre-training deep learning network, outputting a cost deviation index, and triggering a correction instruction if the cost deviation index exceeds a threshold value; matching historical cases to generate an optimization strategy set containing material replacement, construction period adjustment and supplier replacement; and virtual deduction is carried out on the optimization strategy through an engineering digital twin system, and the strategy with the predicted cost curve closest to the target value is screened as a final execution scheme. The method depends on big data and a deep learning technology, full-cycle dynamic management and control of the engineering cost are achieved, and the precision and feasibility of cost control are improved.

The invention discloses an intelligent agent digital image interaction generation method based on multi-modal perception, which comprises the following steps: collecting multi-modal input data of a user, and respectively carrying out preprocessing and feature extraction on the multi-modal input data; inputting to an improved efficient modal cross learning network, and carrying out multi-modal feature fusion processing; constructing a semantic intention map, introducing a time index edge weight and an emotion driving edge weight, and encoding the map by using a structure perception map neural network; a modal style vector is extracted through a cross-modal style contrast learning mechanism, and a personalized style coding vector is generated through a hierarchical nested structure; inputting a personalized regulation and control gating mechanism, and regulating and controlling the middle layer representation in the interaction strategy generation process by adopting a feature channel linear modulation method; inputting the representation vector into a behavior strategy generation module to generate a multi-modal behavior output sequence; and the sequence is output to drive the digital image to perform synchronous response, and natural response generation in the user interaction process is completed.

The invention discloses a traffic scheduling method and an electronic device, and relates to the technical field of traffic scheduling, and the method comprises the steps: determining the priority weight of a micro-service, and predicting a target traffic according to the historical traffic information of a network device; constructing a graph model according to the topological information of the network equipment and the dependency relationship of the micro-service, and performing embedded learning on nodes in the graph model to generate a state vector representing a network state; the priority weight, the state vector and the target traffic of the micro-service serve as input of a reinforcement learning model, and a traffic scheduling strategy of the network equipment is obtained; performing iterative search according to iterative particles formed by encoding the strategy network parameters of the reinforcement learning model and the feature learning network parameters of the graph model to determine reinforcement learning model parameters; and issuing the traffic scheduling strategy to the network equipment and executing the traffic scheduling strategy so as to solve the technical problem that a traffic scheduling method in related technologies is difficult to adapt to a dynamic and complex network environment and service requirements under a micro-service architecture, and the reliability of traffic scheduling is improved.

The present invention relates to the technical field of path planning, and provides a deep reinforcement learning-based path exploration parameter optimization system. The system comprises: a variable parameter path planning module, configured to perform node exploration based on a deep reinforcement learning network, conduct collision detection on child nodes in a child node set, calculate cost values for all child nodes, and finally generate a loading and parking path using a Reeds-Shepp curve; an environmental state space modeling module, configured to perform regional division of obstacles around a current node and conduct environmental state space modeling; and a deep learning parameter optimization module, configured to construct a deep learning network to compute an optimal step size and an optimal steering angle, build a reward function to optimize the deep learning network, and simultaneously execute a training process of the deep learning network.

The invention discloses a command and control system resource trend prediction method based on fusion of long and short time sequence characteristics. The method comprises the following steps: acquiring a public power load or similar time sequence monitoring data set, and preprocessing the data in the data set; a deep learning network model based on a TCN-Transformer hybrid model is constructed, a TCN model and a Transformer model are adopted for parallel computing to achieve feature extraction, the TCN model extracts short-term information, the Transformer model extracts long-term features, then fusion features are obtained through a cross attention mechanism and multi-layer perceptron (MLP) weighting, and finally prediction output is generated through full connection layer mapping. Taking data in the training set as input, training the constructed TCN-Transform hybrid model, and continuously optimizing the model until convergence meets a set requirement; and performing prediction by using the trained network model. According to the method, the TCN-Transform hybrid model is constructed, so that local fine-grained features are reserved, the global time trend is effectively captured, and the accuracy of command decision making is improved.

The invention relates to the technical field of battery management, in particular to an intelligent monitoring control method and system for a lithium battery BMS. According to the application, through the multi-mode sensing array integrated on the surface of the inactive correlation structure of the single cell, the limitation of traditional voltage and current monitoring is broken through, and microcosmic parameter signals such as high-frequency impedance, microstrain and characteristic steam spectrum are synchronously acquired; an empirical mode decomposition algorithm is innovatively adopted to separate electrochemical and mechanical process intrinsic modes, and fusion with spectral signal time domain features is carried out, so that deep fusion of multi-physical field information is realized; a channel coupling deep learning network is constructed, a process fingerprint spectrum is constructed through a dynamic updating algorithm, and full-life-cycle state tracking is achieved; based on failure mode recognition of dynamic time warping distance and similarity matching, a target control strategy is generated in combination with multi-physics coupling simulation; according to the application, multi-dimensional microscopic parameter fusion monitoring is realized, and the fault early warning precision and the battery safety protection capability are remarkably improved.

The invention discloses an operation optimization method, system and equipment for cooperative carbon reduction of an intelligent rail line network, and relates to the technical field of energy-saving scheduling and low-carbon operation of urban public transport. A refined section-level energy consumption and carbon emission quantitative model is constructed based on dynamic characteristics of intelligent rail vehicles and carbon emission factors of a power grid; a deep learning network is used for fusing multi-source data to carry out cross-line short-time passenger flow prediction, real-time passenger flow demands, energy consumption and carbon emission indexes and intersection signal phase time window constraints are uniformly incorporated into a multi-target collaborative optimization framework, and solving is carried out through mixed integer linear programming in a rolling vision field. According to the method, combined optimization and dynamic closed-loop control of departure intervals, section operation speeds and road right strategies are realized, so that the total operation energy consumption and carbon emission of the intelligent rail line network are remarkably reduced on the premise of ensuring passenger flow transportation requirements and service quality, and the line network level energy-saving and carbon-reducing operation targets are achieved.

The invention relates to an RGB-T saliency target detection method based on a query guide specific learning network, and the method comprises the steps: constructing a network model of an encoder-decoder architecture, and setting a visible light image saliency detection module, an infrared image saliency detection module and a bimodal saliency detection module according to the network model, visual features of the visible light RGB image are obtained through an encoder of the visible light image saliency detection module, and infrared features of the infrared thermal imaging image are obtained through the infrared image saliency detection module; in a bimodal information fusion module in the bimodal saliency detection module, generating an RGB modal feature, an infrared modal feature and a cross-modal fusion feature; and the three-feature cross block outputs the enhanced multi-modal fusion features. Compared with the prior art, the method has the advantages of high accuracy, high modal adaptability, high efficiency and the like.

The invention discloses a remote sensing image change detection method based on deep learning and a GIS, and relates to the technical field of image processing and geographic information system application. Comprising the steps of 1, carrying out data preparation and preprocessing, 2, constructing a deep learning network model, carrying out feature extraction on preprocessed remote sensing image data by adopting a multi-layer convolution layer and a pooling layer, and aiming at obtained GIS data, carrying out feature extraction on the remote sensing image data by adopting different feature extraction methods according to data types, 3, performing feature fusion: fusing the remote sensing image feature vector and the GIS feature vector in a deep learning network model by using a feature fusion layer to obtain a fused feature vector; and step 5, performing change detection and result output: applying the trained deep learning network model to the preprocessed to-be-detected remote sensing image data and GIS data, outputting a change probability value of each pixel point through forward propagation calculation of the model to obtain a change probability graph, and performing post-processing operation on an obtained binary change detection result to obtain a change probability graph. And finally, outputting a clear and accurate remote sensing image data change detection result, and displaying the position, range and type information of a change area in the form of image or vector data.

The invention is applicable to the technical field of geophysical detection, and provides an underwater target detection method and device based on acoustoelectric joint and width learning, and the method comprises the steps: carrying out the signal collection through an acoustoelectric joint detection network which is disposed in an underwater monitoring region in advance, the method comprises the following steps: acquiring an acoustic signal and an electromagnetic signal synchronously acquired by each detection node in an acoustic-electric joint detection network, performing feature extraction on the acoustic signal and the electromagnetic signal based on a data sliding window mechanism to obtain an acoustic-electric fusion feature matrix, and calculating the acoustic-electric fusion feature matrix according to the acoustic-electric fusion feature matrix. Whether a target object capable of autonomously radiating an acoustic signal and an electromagnetic signal underwater exists in an underwater monitoring area or not is determined through a pre-trained width learning network, so that false alarm interference caused by background noise in an underwater environment is effectively reduced, and the underwater target recognition capability under a weak signal condition is remarkably improved; and the accuracy of underwater target detection is improved.

The invention discloses a cloth surface flaw detection method and device based on texture perception and anomaly detection. The method comprises the following steps: acquiring a surface image of detected cloth; inputting the surface image of the detected cloth into a deep learning network model; a multi-scale feature map is extracted through the backbone network; processing the feature map through the texture perception feature extraction module so as to fuse cross-channel and cross-space texture information; integrating anomaly detection branches through the check network, and outputting feature maps of different scales; performing frequency domain enhancement and spatial domain feature extraction operation and fusion on the features through the adaptive frequency domain convolution module; and outputting a detection result through the YoloHead detection head so as to judge whether the cloth has flaws or not. According to the method, the texture feature information in the cloth image can be effectively utilized, and the detection accuracy and robustness of the cloth surface flaws and the recognition capability of unknown flaws are improved.

The invention discloses a three-dimensional geological model dynamic updating system and method based on multi-modal data fusion, and belongs to the technical field of geological engineering and information. According to the method, multi-source data of geological radar, earthquake, drilling, remote sensing and the like are integrated, and after unified formatting and quality optimization preprocessing, geological features of a stratum interface, a fault zone and the like are automatically extracted by using a deep learning network; multi-source feature data fusion is realized through space-time registration and attention; constructing a three-dimensional geologic model by adopting a TIN and Octree mixed structure based on a fusion result; dynamic updating of the model is realized through incremental detection and a local reconstruction mechanism, and the boundary of the model is optimized by adopting a radial basis function interpolation. According to the method, the problems of single data, updating lag and the like of a traditional method are solved, the modeling precision and efficiency are remarkably improved, and reliable support is provided for geological engineering decision making.

The invention discloses a high-voltage switch cabinet partial discharge fault identification method and a storage medium. The method comprises the following steps: synchronously acquiring acoustic, electric, magnetic and thermal signals through a multi-physical field sensor array; performing preprocessing and feature extraction on the signals; using a cascade deep learning network to fuse multi-modal features to identify fault types and levels; realizing accurate positioning of the discharge source in the digital twinborn model by adopting a dynamic simulation coherent positioning method; and risk assessment and early warning are carried out by integrating diagnosis and positioning results. According to the method, deep learning and physical mechanism simulation are deeply fused through a data and model dual-drive normal form, so that the defects of shallow fusion level, low positioning precision, lack of interpretability and the like in the prior art are overcome, and intelligent, precise and prospective identification and early warning of the partial discharge fault of the high-voltage switch cabinet are realized.

The invention discloses a bridge service intelligent evaluation and early warning method and system based on dynamic and static load-multi-modal data fusion, and the method comprises the steps: S1, installing a vibration acceleration sensor and a strain gauge at a bridge control part, and collecting data; locally preprocessing data by adopting an edge computing framework, extracting feature values, encrypting and transmitting the feature values to a cloud; identifying and collecting diseases, and combining a deep learning network model to realize pixel-level crack segmentation; s2, establishing a space-time-measuring point-disease three-dimensional correlation model; s3, a three-level early warning threshold system is set, different early warning levels have different requirements for vibration amplitude and crack length indexes, and when a monitoring index exceeds a threshold value, early warning information is pushed to a terminal in real time through 5G; S4, a decision engine is constructed based on a Q-Learning algorithm or a deep reinforcement learning model, early warning levels, residual life and maintenance cost parameters are input, and the early warning level, the residual life and the maintenance cost parameters are calculated. And outputting the optimal maintenance scheme.

The invention discloses a spine image key point detection algorithm based on multi-task learning. The algorithm comprises the following steps: constructing a multi-task deep learning network model comprising a segmentation branch and a key point positioning branch; a feature aggregation module based on multi-scale cavity convolution is embedded in the segmented branches, and the multi-scale cone feature extraction capability of the model is enhanced through splicing fusion of multiple cavity rate convolution branches and global pooling branches; a cross-task attention fusion module is introduced between the two branches, and bidirectional dynamic interaction and complementation between segmentation features and key point features are realized by generating and fusing first-order and second-order context attention maps; semantic alignment loss is designed in a training stage, collaborative optimization of two tasks is promoted by constraining the consistency of segmentation masks and key point heat maps in a high-level feature space, global information of segmentation and local information of key point detection are fully utilized, and the accuracy and stability of spine centrum key point positioning and the accuracy of a segmentation result are improved.

A system for learning optimal driving behavior for autonomous vehicles comprises a deep neural network, a first stage training module, and a second stage training module. The deep neural network comprises a feature learning network configured to receive sensor data from a vehicle as input and output spatial temporal feature embeddings and a decision action network configured to receive the spatial temporal feature embeddings as input and output an optimal driving policy for the vehicle. The first training stage module is configured to, during a first training stage, train the feature learning network using object detection loss. The second stage training module is configured to, during a second training stage, train the decision action network using reinforcement learning.

The invention relates to a meteorological deduction method and device fusing physical constraints and a neural network, and the method comprises the steps: obtaining multi-source meteorological data, and constructing a spatial-temporal feature input tensor; the spatio-temporal feature input tensor is subjected to standardization processing and then input into a deep learning network model, and a future weather prediction result is obtained; the model extracts time sequence evolution features and space attention features through a neural network module and a space attention module respectively, and integrates the time sequence evolution features and the space attention features in a splicing form; for a forecast task of a future gamma day, a deep learning network model and a physical mode are adopted for prediction respectively, and a splicing time point is determined according to an error minimum principle, so that splicing of prediction results is carried out; when the physical mode is used for prediction, the improved regional numerical weather prediction model is used as a basis, atmospheric basic equation sets are integrated, and weather prediction at future moments is carried out. Compared with the prior art, the method has the advantages that the atmospheric physical law and data driving advantages are fused, and the extreme weather prediction precision and stability are improved.

The invention particularly relates to a multi-task motion recognition method based on a physical constraint guide condition diffusion model, which comprises the following steps: acquiring original observation data of multi-task motion based on an inertial measurement unit and preprocessing to obtain preprocessed original observation data; constructing a conditional diffusion model, and training the conditional diffusion model by using the preprocessed original observation data based on a preset loss function to obtain a trained conditional diffusion model; performing data enhancement on a key motion state category determined based on the motion state category label corresponding to the original observation data according to the trained conditional diffusion model to obtain a key motion state category sample, and constructing a sample balance data set based on the key motion state category sample; and based on the sample balance data set and a preset multi-task learning network loss function, constructing a multi-task learning pedestrian motion recognition model and carrying out multi-task recognition. Therefore, the problems of few key state samples, unreasonable data and the like in motion recognition are solved.

The invention discloses a mechanical arm autonomous grabbing system and method based on visual guidance and visual touch fusion. The mechanical arm autonomous grabbing system at least comprises a visual semantic guidance and grabbing pose generation module, a smooth grabbing control module based on visual touch fusion and a dynamic planning module based on model prediction path integration. The visual module efficiently generates a six-degree-of-freedom grabbing pose from the point cloud data by using a deep learning network; the smooth grabbing control module monitors object slippage in real time and adaptively adjusts clamping force by fusing visual and tactile information, and stable grabbing of fragile or slippery objects is achieved. And the dynamic planning module adopts a model prediction path integral algorithm MPPI to realize collision-free path planning of the mechanical arm in a complex dynamic environment. The problems that an existing mechanical arm is insufficient in sensing capacity, poor in grabbing robustness and low in path planning efficiency in an unstructured scene are solved, and the autonomous operation capacity and safety of the mechanical arm in the scenes of accompanying, intelligent sorting and the like are remarkably improved.

The invention provides a deep learning-based junction point relation network training and cell occlusion judgment method, which comprises the following steps of: performing instance segmentation on a cell microscopic image to obtain a cell instance mask, and extracting a cell boundary based on the cell instance mask; performing junction point detection on the cell boundary to obtain cell shielding junction points, and classifying the cell shielding junction points into cell T-shaped shielding points, cell Y-shaped convergent points or cell X-shaped cross points; cutting the local image patch by taking the cell shielding junction point as a center, and constructing a multi-channel sample containing the local image patch; adopting a unified relationship label to label the multichannel sample, wherein the unified relationship label comprises a shielding relationship, a same-layer relationship and an unjudgeable relationship; training a deep learning network by using the labeled multi-channel sample, wherein the deep learning network outputs probability distribution of a junction point relationship category; and inputting a multi-channel sample of a to-be-detected cell microscopic image into the trained deep learning network, and outputting a junction point relationship category and a confidence coefficient thereof.

The invention relates to the technical field of belt conveyor auscultation, and discloses a belt conveyor auscultation abnormity management method based on comparative learning and time sequence modeling, comprising the following steps: step 1, acquiring operation sound and equipment parameters, extracting frame features and acquiring a voiceprint baseline; 2, the carrier roller synchronization frequency is calculated, and period and sideband fingerprint prototype vectors are constructed; step 3, inputting the frame feature and the fingerprint prototype vector into a contrast learning network to obtain a contrast embedded vector and a time sequence feature vector; 4, the time sequence modeling network outputs three types of probabilities and classification original output values; step 5, calculating an abnormal score, generating an abnormal event and determining a grade; step 6, calculating a priority score and distributing a work order; and 7, backfilling the disposal information, calculating the total cost and storing the total cost. According to the invention, accurate detection, hierarchical management and operation and maintenance closed-loop processing of abnormal operation of the belt conveyor are realized.

The invention belongs to the field of automatic driving algorithms, and particularly relates to a vehicle track prediction method based on implicit map feature expression. Comprising the following steps: step 1, a track-map semantic relation alignment preprocessing mechanism based on a map topological structure; 2, a negative sample generation method based on map clustering preprocessing; 3, constructing a track-map semantic feature comparison learning network; and 4, establishing a universal fusion mechanism for the semantic features of the implicit map. Compared with an existing trajectory prediction method depending on a high-precision map, the method has the advantages that high-precision trajectory prediction can be realized without high-cost map information, and the robustness and generalization of the model in a map missing or distorted scene are improved. Meanwhile, the method can be seamlessly combined with various map-free prediction models, and has universality and expansibility.

The invention relates to the field of tunnel engineering, in particular to a tunnel surrounding rock stability intelligent evaluation system based on multi-source image fusion, which comprises an image acquisition module, an image processing module, an image recognition module, a time sequence analysis module, a risk monitoring module and a surrounding rock evaluation module, the image processing module adopts a pyramid feature matching method and a feature similarity algorithm to realize registration and multi-scale superposition fusion of different image source features; the image recognition module recognizes surrounding rock abnormal features by using a deep learning network model; the time sequence analysis module maps the multi-time-sequence surrounding rock state characteristics to a high-dimensional manifold based on a Riemannian manifold model, and calculates a deformation rate and an acceleration through a geodesic theory and curvature analysis; the risk monitoring module calculates a real-time monitoring index and sends out an early warning signal; the surrounding rock evaluation module evaluates the surrounding rock grade through the surrounding rock stability index and analyzes the support safety risk.

The invention discloses a coal cutter roller cutting interference state identification method based on an infrared image, and the method employs an infrared camera to shoot the surrounding environment, effectively prevents the recognition precision from being limited by complex illumination environments, such as underground low illumination, high dust, and strong water mist, and guarantees the stability and definition of a target contour. Then analyzing and processing the infrared image through a deep learning network, determining four corner points of the hydraulic support face guard in the area from the infrared image by adopting a Shii-Tomasi corner point detection algorithm, further accurately calculating the spatial attitude of the hydraulic support face guard, and constructing a three-dimensional trajectory model of the coal mining machine roller; the Kalman filtering is adopted to obtain the predicted poses of the two, and then the cutting interference state is judged through early warning grading, so that the response speed and the safety guarantee capability of the fully-mechanized coal mining system are remarkably improved, the dynamic prediction and real-time response of the cutting interference state are realized, and the efficiency and the precision of cutting interference state recognition are ensured.

The invention relates to an intelligent customer service digital human-oriented robust multi-mode emotion understanding method, and belongs to the field of artificial intelligence and human-computer interaction. The method is implemented by an intelligent customer service system, and comprises the following steps: S1, acquiring user data in real time; s2, extracting modal features by using a deep learning network; s3, for the deficiency of modal features, adopting a noise condition fractional network based on a diffusion model for recovery; s4, using the reinforcement learning network to optimize the fused weight corresponding to the modal features; s5, combining the weight of the fusion, and achieving the fusion of the modal features through an attention mechanism; and S6, taking the fused modal features as input, establishing an emotion recognition model by using a deep learning network, and completing an emotion recognition task. The method can effectively solve the problem of data missing in real environment interaction of the intelligent customer service digital person, has high emotion recognition accuracy, provides stable and high-quality service for the user, and has good robustness.

The invention relates to the technical field of power equipment fault diagnosis, and discloses a partial discharge type identification method based on deep learning, and the method comprises the steps: enabling a discharge physical equation to be embedded into a neural network structure through a physical constraint element learning network, and building a basic model with physical consistency; correcting physical model parameters by using a Bayesian fusion algorithm, and fusing priori knowledge of simulation data and observation information of an actual measurement sample; enabling the generative adversarial network to generate an enhanced training sample meeting the law of conservation of physics through the correction parameters and the conditions; a physical information guide network model is adopted to rapidly adapt to sample feature information and finely adjust model parameters, and rapid and accurate identification of discharge types under extreme working conditions is achieved. According to the method, the technical problem that the identification model is difficult to establish when the samples are extremely scarce under the extreme working condition is solved, and the technical effect of establishing the high-accuracy identification model within the hour-level time is achieved.

The invention relates to the technical field of infrared temperature measurement of power equipment, and provides an infrared temperature measurement compensation method combining physical transmission and depth residual error, which comprises the following steps: acquiring an infrared image and physical parameters related to radiation transmission; establishing a differentiable forward physical model, and calculating radiation brightness; performing first-order physical inversion according to the differentiable forward physical model to obtain a baseline temperature of the target area; constructing a deep residual learning network, taking the infrared image, the baseline temperature and the physical parameters as input, outputting a pixel-level temperature residual estimation value, and correcting the baseline temperature to obtain a corrected temperature; closed-loop joint optimization is constructed and executed, and the closed-loop joint optimization is based on the consistency constraint of the simulated radiance and the actually measured radiance, and parameters of the target distance and the deep residual learning network are synchronously updated; and based on the corrected temperature after closed-loop joint optimization, calculating the temperature difference between the hot spot area and the reference area, carrying out distance normalization processing on the temperature difference, and outputting a standardized temperature difference result.