53results about How to "Efficient reconstruction" patented technology

The invention discloses a point cloud data processing method and device, a storage medium and electronic equipment. The method comprises the following steps: acquiring initial point cloud data to be sent; dividing the initial point cloud data into a group of point cloud blocks; performing graphical representation on the basis of the point cloud data in each point cloud block in the group of point cloud blocks to generate a group of graph structure data, one graph structure data in the group of graph structure data corresponding to one point cloud block in the group of point cloud blocks, and the graph structure data being used for representing features of the corresponding point cloud block; and encoding the group of graph structure data to obtain a group of data packets, and sending the group of data packets to a target terminal so as to perform decoding reconstruction at the target terminal to obtain target point cloud data. The technical problems that the processing efficiency of the point cloud data is low and the effect cannot be quickly achieved in the related technology are solved.

The invention belongs to the field of computer graphics, and discloses a fast parallel multi-scale smooth symbol distance curved surface reconstruction method and system, and the method comprises the steps: carrying out the local fitting of point cloud data through employing a moving least square method, and calculating the Gaussian curvature of a point cloud; constructing a curvature adaptive octree based on Gaussian curvature, wherein the octree is used for dynamically adjusting local resolution and capturing geometric details to improve curved surface reconstruction precision; global fitting is carried out through an implicit function, and a zero contour surface is obtained through an octree contour surface extraction algorithm; and multiple threads are introduced in the process of iteratively solving the linear system to carry out parallel computing. The adaptive octree based on Gaussian curvature can effectively capture complex geometric details, and the algorithm realizes high-precision curved surface reconstruction under different scales. In addition, the local and overall calculation efficiency is remarkably improved through parallel calculation, and the high efficiency of the algorithm in large-scale data processing is ensured. The method provided by the invention provides an efficient and accurate solution for complex curved surface reconstruction.

The invention discloses a confrontation and defense method based on multi-view image super-resolution, and the method comprises the steps: obtaining an attacked image, and enabling the attacked image and a clean image to form a clean image and attacked image pair; inputting the attacked image and the auxiliary view angle image into a network for coding to obtain implicit representation; obtaining a homography matrix from an auxiliary view angle image to an attacked image, and performing spatial transformation on the auxiliary view angle image on a feature level to realize spatial alignment; implicit representation is fused in a feature space through a decoding function, a clean pixel value is obtained by polling each pixel position, and defensive output after the network purifies the attacked image is obtained. The method does not depend on the limitation of a single view angle image, reduces the dependence on the consistency of observation angles, effectively achieves the purification of adversarial disturbance, improves the visual quality of the image, enhances the defense capability of the network for potential adversarial attacks, and can be flexibly applied to an image processing robustness task in a complex real environment.

The invention discloses a drug target combinability prediction method based on variational coding, and belongs to the field of bioinformatics. According to the method, efficient extraction of drug and target features, construction of potential feature space of the drug and target features and prediction of the interaction relationship of the drug and target features are realized through a dual-stage coupled end-to-end neural network. The model is mainly composed of two stages, in the first stage, potential representations of drug and target features are obtained through variational compression and feature extraction, and in the second stage, a drug target action matrix is reconstructed through an up-down sampling comprehensive deep neural network. Through the architecture, the chemical characteristics of the drug and the biological characteristics of the target can be effectively integrated, and the accuracy of interactive relationship prediction is improved. The invention provides a novel and efficient solution for deep analysis of biological big data at the same time.

The invention discloses an indoor visual positioning method and positioning system based on an occlusion clearing algorithm, and belongs to the technical field of occlusion image processing, image restoration and indoor positioning. The method is characterized by comprising the following steps: step 1, querying an image in a database, extracting feature information of the image in the database, and storing the feature information; step 2, shielding object detection is carried out, and shielding objects in all query images are removed; 3, performing image restoration by adopting the image restoration model, and restoring a query image; and step 4, indoor positioning is carried out. According to the indoor vision positioning method and positioning system based on the occlusion clearing algorithm, target detection, image restoration and feature matching technologies are combined, and the positioning accuracy and robustness are improved. According to the method and the device, the blocking objects in all the query images are detected and removed, so that the pixels of the blocking objects do not interfere with subsequent processing, and meanwhile, a clear target area is provided for an image recovery process.

The invention provides a magnetic resonance image reconstruction method and device, an imaging system and a storage medium, and the method comprises the steps: inputting to-be-reconstructed undersampling K space data into an image reconstruction model, and obtaining a reconstructed image; and the image reconstruction model is trained by taking model training loss value minimization as a target: performing inverse Fourier transform and down-sampling on the full-sampling K-space data to obtain a full-sampling image and under-sampling K-space data, and performing randomization, interpolation and inverse Fourier transform on the under-sampling K-space data in sequence to obtain a frequency domain training image; and fusing an undersampling training image obtained by inverse Fourier transform of the undersampling K space data and an undersampling image obtained by sequentially carrying out multiple times of second downsampling and interpolation on the undersampling training image to obtain an image domain training image, and calculating a loss value of the training image obtained by fusing the frequency domain and the image domain and the full-sampling image. According to the invention, through adaptive filling and reconstruction of the image domain and the frequency domain, efficient, rapid and high-quality magnetic resonance image recovery is realized.

The invention discloses a turbine blade surface field rapid prediction method based on a multi-channel self-attention network, and relates to the technical field of data prediction. The technical problems that in the prior art, an existing turbine blade surface field prediction technology is long in consumed time, high in calculation cost and insufficient in precision, and obvious temperature and pressure distribution nonuniformity exists are solved. According to the technical scheme, the turbine blade surface field rapid prediction model building method based on the multi-channel self-attention network comprises the steps that turbine blade surface field data are collected; collecting a preset multi-channel self-attention network architecture; performing pretreatment; performing feature extraction and modeling on the turbine blade surface field according to the processed multi-channel self-attention network architecture; and according to the feature extraction and modeling results, establishing a'mapping relation from boundary conditions to a temperature field and a pressure field on the surface of the turbine blade '. The method is suitable for predicting the temperature field and the pressure field on the surface of the turbine blade.

The present invention discloses a method and related device for topology identification of an active distribution network based on multi-task graph learning, belonging to the technical field of power system monitoring, and the method comprises using a plurality of autoencoders to respectively complete the missing of different measurement features in the distribution network; representing the topology of the distribution network in the form of a graph, training the topology identifier using graph data, performing message aggregation from the target node to the source node according to the multi-head attention coefficient, and obtaining the feature matrix of the whole graph edge; inputting the feature of each line into a fully connected neural network layer, and mapping the feature into the judgment value of the state of each line, judging the disconnection state of the line, and realizing topology identification; using multi-task learning to jointly learn and optimize the missing of the complement measurement feature and the topology identification process, determining the target model, inputting the real-time measurement data into the target model, and outputting the topology identification result. The present invention can reduce the dimension, reconstruct the incomplete data, and accurately identify the topology structure of the distribution network under the condition of limited measurement data.

The invention discloses a decentralized application-oriented smart contract program generation method, device and equipment and a storage medium, and belongs to the field of block chain smart contracts, and the method comprises the steps of persisting an executable code and state data of a fragmented smart contract in block data; in response to a calling instruction for the fragmented smart contract, extracting the executable code and the state data persisted in the block data to execute the fragmented smart contract, and updating the state data persisted in the block data according to execution of the fragmented smart contract; and in response to a reconstruction instruction for the fragment smart contract in execution, extracting persistent executable codes and state data from the block data, and reconstructing the fragment smart contract in execution through the extracted executable codes and state data. The problems of complex cross-fragment calling, low execution speed and difficult fault recovery of the smart contract are solved, and the performance and stability of the block chain system in decentralized application processing are improved.

The invention discloses a neural radiation field incremental optimal view selection method based on hybrid uncertainty estimation. The method comprises the following steps: inputting an RGB image set; calculating a 3D position and an observation direction; a step of initializing NeRF training; based on the 5D coordinate (x, d) of the residual image set, threshold sampling is carried out through a modified NeRF network, the color and opacity of the RGB image are calculated, the color and opacity are integrated into Beta distribution, and rendering uncertainty is calculated; inputting a 3D coordinate x of a residual image set, judging whether a track is a plane or a non-plane through a classifier, and estimating position information by using a Voronoi graph algorithm to obtain position uncertainty; an optimal view selection step: performing normalization and summation on the rendering uncertainty and the position uncertainty, and calculating the hybrid uncertainty of each RGB image; and selecting the image with the highest hybrid uncertainty, adding the image into a training set, and repeating the process until a specific reconstruction effect or a preset image number limit is achieved.

The invention provides an ice and snow albedo data reconstruction method based on deep learning fusion multi-source data, and relates to the technical field of remote sensing, and the method comprises the steps: obtaining albedo data and geographical auxiliary data; processing the albedo data by using a non-local space-time filtering method and a Whittaker filtering method to obtain space-time auxiliary data; an R-TNet deep learning neural network model is constructed; training an R-TNet deep learning neural network model through albedo data, geographical assistance data and space-time assistance data; obtaining albedo data to be recovered; and inputting the geographic auxiliary data and the space-time auxiliary data corresponding to the missing pixels of the albedo data to be recovered into the trained R-TNet deep learning neural network model, and recovering the albedo data to be recovered of the missing pixels. According to the technical scheme of the invention, the neural network is combined to deeply mine the internal nonlinear relationship of the variables, and the time-space continuous and accurate albedo data is effectively reconstructed.

The invention discloses an index sequence generation method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring first gene expression data and at least one codebook entry, encoding the first gene expression data, extracting local information in the first gene expression data, and mapping the local information into a potential space to obtain potential representation of the local information; performing quantization processing on the potential representation based on the at least one codebook entry to generate a first codebook index matrix and at least one quantization vector; decoding and reconstructing the first codebook index matrix, outputting second gene expression data, performing adversarial training on a decoder and at least one quantization vector, and outputting optimization parameters; and performing autoregression modeling processing on at least one new codebook entry to obtain a second codebook index matrix, and processing the second codebook index matrix by using the optimized decoder to obtain a data sample. According to the method, the calculation burden of high-dimensional data modeling is reduced, and the data capturing capability of the model is improved.

The invention relates to an image reconstruction model construction method and device, an image generation method and device and computer equipment. The method comprises the following steps: acquiring two training X-ray images, a noise-added CT image and a to-be-trained image reconstruction model; taking the current noise adding step identifier as guide data, taking the two training X-ray images and the noise adding CT image as condition data, and inputting the condition data into the to-be-trained image reconstruction model to obtain model prediction noise; calculating a model noise loss value corresponding to the difference between the representation model prediction noise and the image real noise of the noise-added CT image; the model noise loss value comprises skeleton label information for the skeleton region; and training the to-be-trained image reconstruction model according to the model noise loss value to obtain a trained image reconstruction model. By adopting the method, the image of the target object can be more effectively reconstructed, so that the three-dimensional skeleton space information of the target object can be obtained on the basis of low-radiation damage, and the definition and medical diagnosis value of the reconstructed image are further remarkably improved.

The present application relates to the technical field of computer graphics, computer vision and three-dimensional reconstruction, and particularly relates to an immersive scene reconstruction rendering method and system based on adaptive 3D Gaussian sputtering, which acquires scene initial point cloud data and camera parameters, constructs a 3D Gaussian point cloud model containing multiple attributes and generates sky point cloud; based on the mechanism of pulse neurons, the opacity of Gaussian points is adaptively controlled, and according to the gradient information and geometric attributes of Gaussian points, splitting, copying or pruning operations are performed to optimize the distribution of point cloud; through a deep learning network, the camera pose is corrected, large-scale scenes are divided into blocks for parallel training, and a bilateral grid post-processing is combined with multiple loss functions to optimize rendering, and the sky point cloud is subjected to exclusive constraint, so that the technical defects of the prior art are solved, and the immersive scene application requirements in the fields of VR / AR, digital twin and the like are met.

The present invention proposes a NeRF three-dimensional reconstruction method and system based on 3D point cloud. The system integrates 3D point cloud data and image information collected by laser radar and high-definition camera, and inputs them into the improved NeRF model for training to build a high-precision three-dimensional model of the scene. The improved NeRF model integrates explicit and implicit scene representations, uses hash grid coding to independently model volume density and color attributes, and uses a multi-layer perceptron (MLP) to refine the color attributes, thereby improving the model construction efficiency and rendering quality. During the automatic parking operation stage, the vehicle control system performs accurate path planning based on the 3D model output by the improved NeRF model to ensure the safety and accuracy of automatic parking. The implementation of the intelligent parking lot assisted automatic parking system optimizes the automatic parking process and improves operational efficiency and user experience.

The invention discloses a radar signal PRI sequence reconstruction method based on trend filtering, and the method comprises the steps: firstly constructing a PRI sequence filtering model, proposing a maximum posterior probability estimation problem, then updating model parameters through employing an E-M algorithm, solving the maximum posterior probability estimation problem, obtaining PRI sequence filtering results, fusing the obtained PRI sequence filtering results, and carrying out the reconstruction of the PRI sequence. And finally, performing interpolation correction on the PRI sequence, and outputting a final reconstruction result of the PRI sequence. The PRI sequence probability model provided by the method has better capability of expressing radar signal PRI features, can effectively reconstruct radar signal PRI sequence patterns containing abnormal values, remarkably improves PRI sequence reconstruction performance, has important significance in the field of radar signal receiving and analysis, and has good application prospects. The method is suitable for PRI parameter reconstruction of a non-ideal detection signal sequence under false pulse and leakage pulse conditions.

The application discloses a kind of based on structure-detail separation bidirectional recurrent neural network super-resolution method, belong to video super-resolution technical field.The application mainly includes the following steps:1, the spatial feature, structure feature and detail feature of each frame picture of low-resolution video are extracted;2, the forward and backward optical flow of each frame picture is calculated;3, the spatial feature, structure / detail feature and forward optical flow of each frame picture are input into network, and the forward structure / detail feature of current frame is obtained;4, the spatial feature, structure / detail feature, backward optical flow and forward structure / detail feature of each frame picture are input into network, and the backward structure / detail feature of current frame is obtained;5, the spatial feature, structure / detail feature and forward and backward structure / detail feature of each frame picture are calculated to obtain reconstruction structure / detail feature, and high-resolution result is obtained by calculating the result with spatial feature.The video super-resolution method based on the application can obtain high-quality output with more details.

A training method for an MRI reconstruction network based on speedup ratio segmentation, firstly, by deeply segmenting the target speedup ratio, a series of MRI data with different speedup ratios are constructed; secondly, based on a sequential training method, from low speedup ratio to high speedup ratio, the mapping relationship between each undersampled MRI image and the fully sampled MRI image is sequentially learned, and only a small number of iterative trainings are performed for each speedup ratio; finally, multiple iterative trainings are performed for the target speedup ratio. The method provided by the present invention can effectively utilize small-scale training samples to reconstruct high-multiple undersampled MRI images, and has strong practicality.

The invention relates to a fault recovery method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: if a virtual network element in a network meets a fault condition, determining an actual restart number of a target time point; and initiating restart of the virtual network element corresponding to the actual restart number of the target time point at the target time point, and determining the restart success rate of the virtual network element at the target time point. And based on the restart success rate of the target time point and a preset number adjustment strategy, determining an actual restart number of a next target time point of the target time point, and re-executing restart of the virtual network element corresponding to the restart number of the target time point at the target time point. And determining a restart success rate of the virtual network element at the target time point until the virtual network element is determined not to meet the fault condition, so that a restart storm of the virtual network element after the local side equipment in the network has a fault can be avoided, and rapid recovery of the fault is realized.

The invention relates to the technical field of artificial intelligence, comparative learning and semantic communication, and provides a semantic communication method and system based on improved comparative learning, and a terminal. The method is completed based on a trained auto-encoder model, and comprises the following steps of: configuring a plurality of texts with different lengths, and sequentially inputting the plurality of texts with different lengths into a semantic communication encoder to obtain a semantic encoding sequence; the semantic coding sequence is transmitted into a fuzzy semantic coding sequence through a channel, and the fuzzy semantic coding sequence is input into a semantic communication decoder for semantic decoding; constructing a contrast learning positive sample pair and a contrast learning negative sample pair; based on the positive sample pair and the negative sample pair, an improved contrast learning loss function is adopted to train an auto-encoder model, and when a loss value does not decrease in continuous preset times of training, the training is completed; and deploying the trained auto-encoder model to a channel, and performing semantic communication. According to the method, alignment and uniformity of comparative learning can be ensured, and semantic communication has better anti-noise robustness.

The invention discloses a two-stage progressive rain removal method and system, and aims to solve the problem that the vehicle detection precision of an intelligent connected vehicle is reduced in a rainy day scene. According to the method, a special vehicle detection data set MCGVD is constructed, the MCGVD is divided into a training set and a test set, and a two-stage progressive rain removal network TSPNet including an improved U-Net feature extraction network and a feature fusion mechanism is used for training. Inputting the test set into the trained TSPNet, performing channel dimension raising through 1 * 1 convolution, extracting shallow rain stripe information, extracting deep rain stripe information through an improved encoder-decoder architecture, and finally recovering to obtain a vehicle image, thereby realizing rainy day vehicle detection. According to the system, through the lightweight feature extraction and weighting module and the high-efficiency feature transmission and fusion module, the rain removal effect and the image quality are improved, the performance of the vehicle detection system under complex weather conditions is remarkably improved, and the system has important practical application value.

The application provides a frequency variational autoencoder-based image super-resolution reconstruction method, device and equipment, and relates to the technical field of image processing. The method comprises the following steps: performing frequency domain decomposition on an input low-resolution image through a low-frequency extractor of a frequency variational autoencoder, extracting low-frequency features, and performing up-sampling interpolation on the low-resolution image to obtain a low-frequency image consistent with a target high-resolution size; sampling latent variables from a relevant distribution of low frequencies and high frequencies learned by a relevant encoder of the frequency variational autoencoder, wherein the relevant distribution represents the statistical correlation between the low-frequency information of the low-resolution image and the high-frequency information of a high-resolution image; inputting the latent variables and the low-frequency features into a high-frequency decoder of the frequency variational autoencoder for decoding to generate a high-frequency detail image; and superimposing and fusing the high-frequency detail image and the low-frequency image to obtain a final high-resolution reconstructed image.

The invention relates to the field of radiation field detection, and discloses a radiation field three-dimensional reconstruction method based on a GA-BP neural network. The radiation field three-dimensional reconstruction method based on the GA-BP neural network comprises the following steps: S1, performing a simulation experiment on a radioactive source based on Geant4 simulation software to obtain three-dimensional radiation field simulation data; meanwhile, carrying out a real experiment to obtain actually measured data; s2, constructing a BP neural network model; optimizing the BP neural network model based on a GA algorithm to obtain an initial GA-BP model; s3, training samples in a database are selected to train the initial GA-BP model, and an optimized GA-BP model is obtained; and selecting other untrained data of the database, and inputting the data into the optimized GA-BP model to complete reconstruction of the three-dimensional radiation field. By the adoption of the radiation field three-dimensional reconstruction method based on the GA-BP neural network, the reconstruction accuracy and speed can be remarkably improved, and powerful technical support is provided for accurate mastering of nuclear facility space radioactivity distribution.

This invention belongs to the field of blade trajectory monitoring technology, specifically involving a method for dynamic monitoring of the three-dimensional trajectory of wind turbine blades by fusing imagery and laser point cloud data. The steps include: fixing a camera and scanner with a rigid connection device for use as a dynamic monitoring system for the three-dimensional trajectory of the wind turbine blade; calibrating the camera; calibrating the relative position and attitude between the camera and scanner; acquiring image data and laser point cloud data during normal operation of the wind turbine; performing motion compensation correction on the laser point cloud data; calculating the three-dimensional coordinates corresponding to the monitoring point; obtaining a time-series image-side coordinate sequence of the monitoring point; and calculating the three-dimensional coordinate sequence corresponding to the time-series image-side coordinate sequence of the monitoring point, i.e., the three-dimensional motion trajectory of the monitoring point on the wind turbine blade during operation. This invention uses imagery and laser point cloud technology to monitor the three-dimensional trajectory of wind turbine blades, requiring no auxiliary markers, eliminating complex calibration, offering flexible deployment and high accuracy, and efficiently supporting blade safety assessment and maintenance decisions.

The application discloses a five-dimensional seismic data interpolation method and system based on coordinates, constructs a point-by-point seismic data interpolation model based on a NeRF theory, modifies an MLP framework by combining an additional convolution layer on the basis of the point-by-point interpolation model, constructs a profile-by-profile seismic data interpolation model, introduces kernel norm regularization to construct an objective function, reads original five-dimensional seismic data D obs and inputs the five-dimensional seismic data with missing traces into the network to realize reconstruction of the five-dimensional seismic data. The application eliminates the need for additional labeled data, utilizes unique characteristics of seismic data, enables a convolution network decoder to output data profile by profile, can significantly improve data processing efficiency by 40 times, and increases kernel norm regularization in the objective function, so that the noise resistance and robustness of the model can be effectively improved.

The invention discloses a point cloud geometric compression method based on deep learning, and the method comprises the following steps: (1) segmenting an initial point cloud according to the shape, so as to ensure that each block corresponds to different parts of an object; (2) applying a k-Means clustering algorithm to the segmented point cloud data to further refine the blocks so as to enable the points between the blocks to be relatively balanced; (3) inputting the subdivided point cloud data into a Point Net + + network for feature extraction so as to capture local geometric features; (4) constructing a graph structure in the processed point cloud data to enhance feature interaction between blocks; (5) fusing the extracted point cloud features by using a graph attention mechanism, and focusing on parts with similar shapes in adjacent blocks; and (6) adding point information into the features, and carrying out arithmetic coding and transmission on the features. And (7) finally, gradually reconstructing each block according to the features, and interpolating the point cloud blocks according to the point information to obtain a final reconstructed point cloud. Through the method, the geometric compression and reconstruction performance of the point cloud is remarkably improved.

The invention discloses a video super-resolution method and system based on strip-shaped and global time-space attenuation attention, and belongs to the technical field of video super-resolution. The method comprises the following steps of: 1, extracting spatial features of an original video frame image after enhancement and performing up-sampling; 2, calculating spatial strip attention and time strip attention on the upsampling spatial features; 3, processing the spatial strip-shaped attention and the time strip-shaped attention by using depth separable convolution and a residual structure to obtain final attention outputs, and performing fusion processing on the two attention outputs to obtain fusion features; and 4, calculating the global space-time attenuation attention of the spatial feature map, and fusing the spatial features, the fusion features and the global space-time attenuation attention to obtain a final high-resolution reconstruction result. Based on the video super-resolution method provided by the invention, the PSNR / SSIM can be improved on a common data set, and sharp detail information can be recovered.

The present invention belongs to the field of laser radar technology, and specifically discloses a three-dimensional imaging method of adaptive scanning. The present invention adopts an adaptive scanning method to solve the problem of long laser radar acquisition time under large background conditions, and adopts an innovative regularized denoising model to implement three-dimensional reconstruction, thereby obtaining a better depth map. The main steps include: Step 1, using an adjustable spot size laser radar to realize pre-scanning and spot prior; Step 2, generating a spatial blur kernel and a scanning area matrix; Step 3, determining a specific scanning area for single-photon data acquisition according to the scanning matrix; Step 4, constructing a regularized denoising optimization function model; Step 5, subjecting the experimental data to regularized denoising to realize three-dimensional reconstruction and restore the depth map.

The invention discloses an incremental grid reconstruction method based on adaptive parameters, and the method comprises the steps: firstly obtaining point cloud data, carrying out the linear point cloud segmentation of an obtained point cloud based on a point index, dividing the point cloud into m subsets, and enabling the adjacent subsets to have an overlapping region; then density estimation is carried out on the segmented point cloud subsets, and downsampling and normal estimation are carried out; and finally, constructing a triangular mesh for each point cloud subset by adopting an incremental method. According to the method, through point cloud segmentation and overlapping region introduction, the problem of computing resource limitation is solved, and the consistency and continuity of a processing result are ensured through the design of the overlapping region; the dynamic adjustment of down-sampling and normal estimation is realized through adaptive parameters obtained by density estimation, so that the retention of important geometric features is ensured, and the overall processing efficiency is improved; dynamic vertex combination and patch reconstruction are realized by defining a vertex mapping function and a distance threshold.

The invention discloses an adipic acid reaction chamber crystallization anomaly monitoring method based on attention fusion, which comprises the steps of S1, constructing an image deblurring model based on space frequency attention fusion, S2, acquiring an adipic acid reaction chamber image, and segmenting the adipic acid reaction chamber image into N non-overlapped image blocks, s3, extracting semantic features implied in each image block according to the image deblurring model, and reconstructing a deblurred image; s4, performing deblurring processing on the deblurred image Idebur to obtain a clear image; s5, clear image features and target attention features are extracted through a multi-head self-attention module, and representative adipic acid reaction chamber abnormal image features are obtained based on heterogeneous feature attention fusion; and S6, building an adipic acid reaction chamber abnormal target detection model through bounding box loss, classification loss and confidence coefficient loss functions, and carrying out abnormal real-time monitoring. According to the invention, monitoring and early warning can be automatically carried out, and the safety of the production process is effectively improved.