300results about How to "Guaranteed robustness" patented technology

The invention discloses a power efficiency optimization method based on intelligent load management. The method comprises the following steps: S1, generating an initial load characteristic matrix; s2, decomposing the initial load characteristic matrix into a short-term behavior characteristic matrix and a long-term trend characteristic matrix through pseudo-symmetry decoupling; s3, multi-level load demand prediction is carried out through a fusion structure of the multi-scale convolution and the recurrent neural network; s4, coding a cognitive gene sequence, triggering gene mutation and recombination operation when significant change occurs, and generating a load distribution strategy; s5, a load distribution strategy is transmitted into a causal reasoning and self-supervision feedback control module, and if the deviation between feedback data and a prediction result exceeds a threshold value, a self-learning mechanism is triggered; s6, dynamically distributing the load priority, and generating a decision path of load management; and S7, performing adaptive optimization through a multi-layer feature feedback closed-loop optimization system. The intelligent load management method is adopted, and the power efficiency is optimized through multi-layer feature decoupling and closed loop.

The invention discloses a method and system for improving satellite remote sensing inversion precision by fusing hyperspectral proximity, and belongs to the technical field of water environment and water ecology monitoring. Acquiring a water body hyperspectral near-sensing reflectivity spectral data set, a satellite remote sensing surface reflectance data set and a water environment actual measurement data set, and performing preprocessing; a synchronous data set containing # imgabs0 # and water quality and a synchronous data set containing # imgabs1 # and water quality are constructed; based on the synchronous data of the # imgabs2 # and the water quality, synchronous data of the # imgabs3 # and the water quality are generated, and synchronous data of the # imgabs4 # and the # imgabs5 # are obtained; the # imgabs6 # is normalized to obtain # imgabs7 #, and the quantitative conversion relation between the # imgabs7 # and the # imgabs8 # is constructed band by band to eliminate errors; generating a fused satellite-ground data set, and constructing a training water quality parameter inversion model; and updating corresponding methods and data sets in real time by repeating specified steps, and optimizing the model. According to the method, the inversion result is displayed and downloaded according to the query instruction, the satellite spectrum precision and the satellite-ground synchronous data set scale are improved, and the method is used for high-precision water environment monitoring.

The embodiment of the invention discloses a bridge crack width monitoring method. The method comprises the following steps: acquiring a continuous video image of a bridge surface; generating a pixel-level segmentation mask image of the crack by using the target detection model; when the target detection model is pre-trained, optical flow features between adjacent frames in the first data set are extracted through an optical flow calculation algorithm, and deep fusion is performed on the optical flow features and the image features; contour pixel points of the crack are extracted based on the crack segmentation mask pattern, and the width change information and the width change rate of the crack are obtained; according to the width change information and the width change rate, performing trend prediction on the crack width change by using a time sequence prediction model to obtain a prediction result of the crack width change; in the training process of the time sequence prediction model, a basic model and a complexity capture module are designed, and the weight of a main model is optimized by using a weighted fusion strategy. According to the invention, efficient and accurate crack width monitoring is realized through crack identification and prediction.

The invention relates to an underwater image enhancement method and device based on generative adversarial network multi-scale fusion, and the method comprises the steps: obtaining an original underwater image, and carrying out the preprocessing of the original underwater image; inputting a preprocessing result into an underwater image enhancement model based on generative adversarial network multi-scale fusion to obtain an enhanced underwater image; the underwater image enhancement model comprises a generator and a discriminator, and the generator is obtained by replacing the last downsampled convolution module in an encoder of the U-Net network with a multi-scale feature fusion module; and the multi-scale feature fusion module is used for extracting multi-scale fusion features by adopting multi-scale convolution, residual connection, dense connection and a global attention mechanism. The method can adapt to different underwater scenes and enhance the image quality, the color cast and brightness recovery of the underwater image are improved, the robustness of the image is kept, and color distortion and noise artifacts are avoided.

The invention provides a robot-assisted wound suturing method and system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the robot-assisted wound suturing method comprises the steps: obtaining multi-modal perception data of a wound area, constructing multi-modal three-dimensional representation of a wound, and determining a plurality of suturing points based on the multi-modal three-dimensional representation of the wound; for each suture point, calculating a candidate suture point, and for the adjacent suture points, generating a needle tip track of a needle tip in the wound, and sampling on the needle tip track to obtain a suture scheme adapted to the wound form and the tissue attribute; based on the obtained stitching scheme, an optimal control sequence of the stitching robot is obtained, meanwhile, a robust compensation control law is constructed through coordinate transformation and a feedback control law, and the optimal control sequence and the robust compensation control law are integrated into a stitching robot real-time control system; the output of a real-time control system of the suturing robot is converted into an actuator control instruction, and the suturing robot is controlled to assist in wound suturing.

The invention discloses a tunnel fire thermodynamic response reasoning method based on MGraphFomer. The method comprises the following steps: constructing a tunnel structure fire thermodynamic response database; according to tunnel classification design and structural damage judgment principles, structural thermal response grade judgment standards of different types of tunnels are constructed in a self-adaptive manner by means of a branch convolutional neural network, and structural thermal response grades of different types of tunnels are combined according to the structural thermal response grade judgment standards. Obtaining a structural unit mechanical response coarseness judgment model; based on a heterogeneous graph processing method of double encoders and a structural unit mechanical response coarse-grained judgment model, a structural unit multi-scale response fine-grained model is constructed, and a tunnel fire structure thermal response digital-intelligent prediction model is obtained; and the tunnel fire structure thermal response is predicted by using the tunnel fire structure thermal response digital intelligent prediction model, and reasoning of the tunnel fire structure thermal response is realized. The thermal response prediction method is high in accuracy and higher in generalization ability.

The invention provides a multi-modal large model content security assessment method and device, and the method comprises the steps: inputting a multi-modal harmful data set into a to-be-assessed multi-modal large model, and obtaining an output response of the multi-modal large model; initializing a plurality of large-scale language models, and endowing the plurality of large-scale language models with different role features to form a heterogeneity evaluation system; performing independent evaluation on the multi-modal large model in the aspect of content security by utilizing a plurality of large language models; a plurality of large language models are used for jointly reviewing independent evaluation results, and consistent evaluation suggestions are achieved through cooperative discussion among the models; and obtaining a final evaluation result by using a hard voting mechanism and a soft voting mechanism. According to the method, a plurality of large language models with different role features are introduced for cooperative assessment, and multi-view and multi-level security risk assessment is realized. And through a mode of combining independent evaluation and collaborative review, diversified view angles are effectively integrated, and the reliability and consistency of evaluation results are remarkably improved.

The invention relates to a fault prediction method for a mining high-voltage power distribution device. Temperature, current, voltage and gas leakage sensors are installed in the mining high-voltage power distribution device; integrating the multi-dimensional data by using a multi-sensor data fusion algorithm to form a real-time operation portrait of the power distribution device; extracting time sequence characteristics of temperature and voltage key variables by combining an operation rule of the high-voltage power distribution device; capturing abnormal fluctuation of the key characteristic parameters by using a sequence analysis model, and identifying a specific fault type in the high-voltage environment; constructing a feature mode database based on a time sequence, and matching new data with a historical fault mode; applying a feature selection algorithm to extract representative fault prediction features from the sensor data; establishing a fault prediction risk level model according to the hazard degrees of different fault modes; based on the fault risk level predicted in real time, an optimized self-adaptive maintenance scheme is generated, and the maintenance period and strategy are dynamically adjusted in combination with the actual operation state of the equipment.

The invention relates to the technical field of power grid side large-scale wind and light and energy storage configuration, in particular to a wind and light storage capacity distribution robust optimization configuration method, device and equipment and a medium. Establishing a power grid side wind and light energy storage optimization configuration model considering inertia constraint; uncertain factors in the optimal configuration model are considered, and a probability distribution fuzzy set based on a mixed norm is constructed; establishing a two-stage distribution robust optimization configuration model based on data driving; converting the two-stage distribution robust optimization configuration model into a mixed integer linear programming model by using a linearization technology; and solving the mixed integer linear programming model by using a column and number generation algorithm to obtain an optimal configuration and operation scheme of each device. According to the method, the economical efficiency and load fluctuation can be balanced, and the problems of frequency instability and the like of a large-scale wind and light energy grid-connected power grid are avoided by considering inertia constraint.

The invention provides a substation safety management and control method and system based on a three-dimensional reconstruction model, and relates to the technical field of safety management and control, and the method comprises the steps: obtaining real-time point cloud data inputted by a laser point cloud sensor and a real-time image collected by a camera; comparing the real-time point cloud data with a background point cloud model, and extracting a changed dynamic region by using a background difference algorithm; the dynamic reconstruction module is used for identifying a first position of a first dynamic target in the dynamic region based on the real-time point cloud data; calculating a second position of a second dynamic target contained in the real-time image; if the first position is successfully matched with the second position and the sparse structure of the real-time point cloud data is successfully matched with dense pixels of the real-time image, determining that the first dynamic target and the second dynamic target are both dynamic flight targets; and determining the position of the dynamic flight target according to the first position and the second position, and marking in the three-dimensional model. The safety management and control precision of the transformer substation can be improved.

The invention discloses a high-frequency vehicle track map matching method, which comprises the following steps of: (1) dividing a track T to be matched into a plurality of sub-tracks based on a reference line distance, and extracting key track points in the sub-tracks; (2) extracting a road where the key track point is located and a road type of the road, filtering the extracted road based on the road type, and storing filtered road data by adopting an R-Tree structure; and (3) matching an optimal matching path of the track T to be matched from the road data stored in the R-Tree structure through a hidden Markov probability matching model. From the perspective of reducing the scale of original data, trajectory and road network data are simplified. Then, space and direction analysis is carried out on the simplified trajectory, a micro-region splicing strategy is adopted to accelerate shortest path search, and a path deviation function is designed to improve the topology of a matching result; in addition, the extraction process of track motion key points is refined by fusing more road semantic information, so that a better matching result is obtained.

The invention provides an abnormal street lamp detection system and method based on a YOLOv8 model, the system comprises an image acquisition module, a target detection module and an abnormal street lamp identification module, the image acquisition module is used for acquiring a monitoring image on a road in real time, the target detection module adopts the YOLOv8 model to carry out target detection on image data, and the abnormal street lamp identification module is used for identifying the abnormal street lamp. Identifying a street lamp area in the image, positioning a street lamp in the image and marking the position of the street lamp; and the abnormal street lamp identification module analyzes the detected street lamp state by using image features, performs multiple times of abnormal detection on multiple street lamp images collected in the same time period by adopting a multiple-time identification strategy, and judges a final result through a voting mechanism, thereby identifying the abnormal street lamp. According to the invention, automatic detection and state analysis are carried out on the street lamps by using the YOLOv8 model, so that real-time monitoring and identification of abnormal street lamps are realized, and the intelligent management level of an urban road illumination system is improved.

The invention discloses a tunnel structure response prediction and protection optimization method based on a mechanical response model, and the method comprises the steps: carrying out the training and optimization of an explosion structure response neural network through an explosion structure mechanical response prediction database, and obtaining a tunnel explosion structure mechanical response model; based on a tunnel explosion structure mechanical response model, tunnel structure responses under different tunnel anti-explosion geometric protection means are obtained, and a geometric protection damage data set is constructed by calculating tunnel physical damage and functional damage; training and optimizing the geometric protection damage neural network by using the geometric protection damage data set to obtain a geometric protection damage model; and based on the geometric protection damage model, geometric protection means optimization is realized according to a damage minimization principle. The method is high in efficiency, strong in generalization ability and high in precision.

The invention relates to a satellite navigation and positioning method and device, and belongs to the technical field of satellite navigation and positioning. Resolving the GNSS observation value based on a single-difference GNSS sliding window optimization model to obtain a floating point solution corresponding to a positioning result; the sparsity of a matrix in a graph optimization problem is ensured based on a single-difference GNSS sliding window optimization model; further, gross error detection is carried out on the GNSS observation value based on sliding window data, a test result is obtained, and the robustness of the algorithm in a complex environment is ensured; when the test result is that the test is passed, determining a single-difference ambiguity variable according to the floating point solution, and performing inter-satellite single-difference on the single-difference ambiguity variable to obtain floating point double-difference ambiguity; the floating point double-difference ambiguity is fixed based on a partial ambiguity fixing method, the satellite navigation positioning process of a GNSS fixed hold solution (Hold solution) under a graph optimization model is realized, and the precision and reliability of satellite navigation positioning are improved.

The invention relates to the technical field of scale prediction, in particular to an electric vehicle scale prediction method based on an AI large model and multi-source information fusion, and the method comprises the following steps: collecting related data affecting the scale of an electric vehicle from a plurality of heterogeneous data sources; fusing the features from different data sources by using a multi-modal fusion technology; based on the fused features, constructing an AI large model with an uncertainty quantification capability; and carrying out hierarchical prediction based on the AI large model, and respectively predicting the scales of the electric vehicle in different areas, different vehicle types and different charging infrastructure scenes. According to the method, uncertainty measurement of the prediction result is provided by analyzing the model parameters and the output probability distribution, so that the prediction result is more credible and explainable, comprehensive evaluation under different scenes is provided, and powerful support is provided for decisions in the aspects of regional planning, vehicle model production, charging infrastructure construction and the like.

The invention discloses a defect detection system, method and device based on artificial intelligence image processing, and particularly relates to the field of image defect detection.A data acquisition module is used for obtaining image data from industrial equipment, a sensor and a camera and conducting preliminary formatting processing on the image data. The model convergence time is shortened, the utilization efficiency of training resources is improved, the generalization ability of the model is enhanced, and the performance of the model in processing a rare scene is improved; according to the method, high accuracy can be kept while computing resource consumption can be remarkably reduced, it is ensured that the defect detection system can also operate efficiently in resource-limited equipment, different hardware conditions are adapted, the application scene requirement of the system is expanded, the operation problem caused by environmental conflicts is reduced, the system safety is enhanced, and the system performance is improved. And the influence on system operation caused by external malicious software attack or resource leakage is avoided.

A straight line fitting method for image detection in a complex interference environment comprises the following steps: firstly, carrying out interference point deletion on a feature point set of an environment image, then carrying out multiple times of random sampling, forming a total point set by using associated points through distance operation, obtaining a plurality of straight line models through initial fitting, and carrying out linear fitting on the linear models; and then calculating an interior point proportion in the total point set according to the linear model, finally selecting the total point set with the maximum number of interior points, and continuously reducing a residual threshold in multiple iterative calculations to improve an interior point proportion threshold by dynamically adjusting a feedback mechanism of the residual threshold when the interior points are calculated. In the initial stage of iteration, more points can enter the model to ensure the diversity and robustness of the model, and the threshold value is gradually tightened along with the convergence of the model, so that the interior points conforming to the model are screened out more accurately. And finally, performing secondary sampling on the inner points and performing straight line fitting on an inner point set subjected to secondary sampling again to eliminate potential error inner points and further improve the fitting precision.

The invention provides a satellite-borne atomic clock remote discontinuous frequency control method. The method comprises the following steps: acquiring a clock difference between a controlled satellite-borne atomic clock and a space reference clock; performing state estimation on the state quantity of the clock correction by taking the clock correction as input through a Kalman filtering algorithm; according to the clock difference state quantity output by the Kalman filtering algorithm, calculating an optimal frequency control value through an LQR control algorithm; and inputting the optimal frequency control value into the controlled satellite-borne atomic clock to realize frequency control. In this way, the problems that in the spaceborne atomic clock frequency control process, measurement noise is large, and control reference is discontinuous can be solved.

The invention relates to the technical field of motor control and parameter identification, and discloses a permanent magnet synchronous motor parameter identification method based on a self-adaptive forgetting factor, which comprises the following steps of: acquiring an input signal and an output signal of a permanent magnet synchronous motor in real time, and constructing a regression vector by using the acquired input signal and output signal; calculating the information entropy of the current moment according to the constructed regression vector; performing normalization processing on the information entropy, and calculating a forgetting factor at the current moment according to the normalized information entropy; and substituting the forgetting factor into a parameter updating formula of a recursive least square method for parameter identification. According to the method disclosed by the invention, the adaptive capacity to dynamic change of motor parameters can be improved, the interference of noise on identification precision is effectively suppressed, the credibility of an identification result is improved, and the robustness and long-term stability of a control system are ensured.

The invention discloses a generative model redrawing attack-oriented image robust watermark traceability method and system and a computer readable storage medium, and belongs to the field of digital information security and artificial intelligence content governance. The method comprises the following steps: in a watermark embedding stage, performing error correction coding and digital signature processing on traceability information containing identity information and a timestamp to generate a load to be embedded, and embedding the load to be embedded and a synchronization template for geometric synchronization into a host image in a function separation manner by using a deep neural network; in the training stage, combined optimization is carried out on the watermark embedding and extracting process by introducing generative model redrawing and microsimulation of image distortion attack, so that the robustness of the watermark under a complex attack condition is improved; in the extraction and verification stage, under the condition that an original image is not needed, geometric synchronous correction and blind extraction of a watermark load are carried out on an image to be analyzed, and authenticity confirmation of traceability information is completed through error correction decoding and digital signature verification. By adopting the technical scheme of the invention, the problems of insufficient traceability information robustness, difficult source confirmation and incomplete evidence chain in a generative model redrawing scene in the prior art are solved, and verifiable traceability and credible evidence generation of the image source information are realized.

The invention discloses a video recognition method based on dynamic space-time full-scale feature fusion, which comprises the following steps: a video data set used for training a deep neural network model is acquired, the video data set comprises a plurality of video sequences, and each video sequence is composed of a plurality of time frames; training a pre-defined video recognition deep neural network model by using the video data set, wherein the video recognition deep neural network model comprises a parallel space-time adapter and a self-prompt generator; and applying the trained video recognition deep neural network model to a video recognition task, inputting a to-be-recognized video sequence, and outputting a classification prediction result of the video. The method has the advantages that through cooperative work of the parallel space-time adapter and the self-prompt generator, efficient modeling of dynamic changes in the video is achieved, the accuracy and robustness of video recognition are remarkably improved, and the method is suitable for various video recognition tasks such as supervised learning, few-sample learning and zero-sample recognition.

The invention provides a batch lamp fault detection method and device based on unmanned aerial vehicle inspection and a medium, and relates to the technical field of computers. The batch lamp fault detection method comprises the following steps: acquiring a to-be-detected lamp effect image at a geographic coordinate of an inspection point through an unmanned aerial vehicle, and matching a corresponding reference lamp effect image group according to the geographic coordinate of the inspection point, and performing feature fusion on each reference lamp effect image including various lamp effects in the reference lamp effect image group and the lamp effect image to be detected to obtain a target lamp effect fusion image, inputting the target lamp effect fusion image into a pre-trained fault area position detection model to obtain an initial lamp fault area, screening a target lamp fault area from the initial lamp fault area, and obtaining a target lamp fault detection result. And determining the position of the fault lamp through the target lamp fault area. According to the invention, the comprehensiveness of batch lamp detection can be improved, the fault detection time of batch lamps can be shortened, the fault lamp detection efficiency can be improved, and the accuracy of the detection result can be improved.

The invention provides a driving track determination method and device, a terminal and a storage medium, and belongs to the technical field of vehicles. The method comprises the following steps: acquiring state data of a target vehicle; creating an initial behavior decision tree based on the state data; based on at least one of a first heuristic rule and a second heuristic rule, pruning the initial behavior decision tree to obtain a target behavior decision tree, the first heuristic rule being used for representing that the number of behavior changes of the target vehicle in the future time period does not exceed a preset number of times, and the second heuristic rule being used for representing that the number of behavior changes of the target vehicle in the future time period does not exceed a preset number of times; the second heuristic rule is used for representing that the target vehicle meets a lane changing condition; and for each behavior sequence in the target behavior decision tree, based on a plurality of behaviors in the behavior sequence and the state data, fitting the moment in the future time period and the vehicle position to obtain a driving track corresponding to the behavior sequence. According to the technical scheme, the driving track of the target vehicle can be quickly determined.

The invention relates to a large language model processing method and device, equipment, a storage medium and a program product, and relates to the technical field of natural language processing. According to the method and the device, the adjusted large language model can be used for expanding the risk data set, and the original generation capability of the large language model can be prevented from being degraded. The method comprises the following steps: identifying security activation parameters in a large language model for a sample by using data; inputting the multiple groups of risk data into the large language model and recording a current gradient slice corresponding to the security activation parameter after the large language model is processed; according to the similarity between the current gradient slice and the abnormal gradient slice, identifying valid data in the multiple groups of risk data to obtain a data training set; performing low-rank fine tuning on the large language model according to the data training set to obtain fine tuning parameters; and updating the original weight of the security activation parameter based on the fine tuning parameter and the security activation parameter, and adjusting the model parameter according to the updated weight, the adjusted large language model being used for expanding the risk data set.

The invention provides an NoC low-delay data transmission method based on a dynamic routing algorithm, relates to the technical field of data transmission, and aims to solve the technical problems that an existing NoC routing algorithm cannot adapt to a link dynamic load, is lack of congestion trend prejudgment, is slow in path search convergence and is insufficient in QoS differential scheduling. The method comprises the following steps: constructing a dynamic sensing module containing a double-branch LSTM time sequence prediction model, collecting states such as link bandwidth and queue length in real time, and predicting a congestion trend in 50-100ms in the future; constructing a multi-objective evaluation function taking delay as a core, and dynamically adjusting the weight according to the QoS level; solving an optimal path by adopting an improved ant colony algorithm which introduces a congestion penalty and pruning strategy; during transmission, dynamic reselection is triggered through a pre-calculation + fast matching mechanism; and iteratively optimizing the prediction model based on incremental learning. According to the method, congestion is avoided in advance, the path search efficiency and the transmission reliability are improved, the delay is remarkably reduced compared with a traditional algorithm, and the method is adaptive to delay sensitive scenes such as a high-performance processor and an AI chip.

The invention discloses a rotating target detection method and system based on phase shift coding and directional feature enhancement, and belongs to the technical field of image recognition. The method comprises the steps of obtaining an image data set and performing enhancement processing; inputting into an ARC network to extract features of a target in any direction; performing adaptive feature optimization on the features through an EMA module, encoding optimized feature direction information by using ARF, generating direction sensitive features for bounding box regression, and extracting directional features for classification through a rotation invariant pooling layer; linearly mapping the angle of the rotating frame into a phase of a corresponding frequency by using a phase shift encoder, predicting a phase shift cosine value, and decoding to obtain a final phase; and finally, removing overlapped detection frames by applying a non-maximum suppression method, and reserving an optimal detection result. Wherein a Pareto optimization theory is used for adaptively weighting a loss item, model parameters are updated through a multi-gradient descent algorithm, and the obtained final model can improve the accuracy and efficiency of detection of a target containing a multi-directional small object.

The invention discloses a photovoltaic power station minute-level power prediction method and system fused with a satellite cloud picture, and belongs to the technical field of data processing and prediction, and the method comprises the steps: obtaining a continuous satellite cloud picture sequence of an area where a target photovoltaic power station is located; obtaining cloud cluster individuals and time-space trajectory data, and extracting time-varying multi-dimensional vital sign parameters of the cloud cluster individuals; inputting the multi-dimensional vital sign parameters into a cloud cluster state evolution prediction model to obtain a cloud cluster state evolution trajectory; obtaining a three-dimensional digital model of the target photovoltaic power station, and generating a super-local irradiance distribution diagram sequence of the target photovoltaic power station based on the state evolution trajectory; and power station electrical characteristic parameters of the target photovoltaic power station are acquired, and a power prediction sequence of the target photovoltaic power station is calculated according to the super-local irradiance distribution diagram sequence. By adopting a means of tracking and predicting multi-dimensional vital sign evolution of cloud cluster individuals and combining a power station three-dimensional model to perform light and shadow simulation calculation, the accuracy of power prediction can be improved.

The invention discloses a depression detection method based on image-text social media analysis, relates to the technical field of automatic detection of depression, and solves the technical problem in the prior art that whether a user is a depression patient or not is detected by analyzing character and picture data published by the user on a social media platform. Specifically, the method comprises four steps of data capture, post screening, image-text representation and fusion classification. Generally speaking, the method comprises the following steps: firstly, capturing a posted document or posted post of a user in a certain time period from a social media platform, including characters and picture data in the posted document; then, performing depression related symptom prediction on the posted document, and screening out the most related K posts; then, respectively carrying out text and picture representation on the screened posts; and finally, fusing the characterization vectors of the two modes for depression detection.

The invention discloses a tunnel fire temperature field and function damage effect forward and reverse modeling digital intelligent learning method, and the method comprises the steps: carrying out the training and optimization of a temperature field neural network through a temperature field data set which comprises a training data set and a verification data set, obtaining a temperature field digital intelligent learning model, and achieving the digital intelligent learning of the tunnel fire temperature field; and constructing a tunnel damage data set according to a learning result of digital intelligent learning of the tunnel fire hazard temperature field, obtaining a forward modeling data set and an inversion data set based on the tunnel damage data set, respectively substituting the forward modeling data set and the inversion data set into the forward modeling neural network and the inversion neural network, carrying out data learning and iteration, and constructing a tunnel function damage forward modeling model and a fire source information inversion model. The method is high in prediction efficiency, strong in generalization ability and high in precision.