530 results about "Point set" patented technology

The invention discloses a risk assessment and early warning method and system applied to water conservancy project construction, and belongs to the technical field of water conservancy project construction safety. Multi-source data of a construction area is collected, and a multi-dimensional construction information flow set divided according to time, procedures and work points is constructed; constructing a space-time risk causal map based on the construction units, and expressing time sequence association and risk propagation paths among the construction units; mapping the information flow to a graph structure, training a multi-task graph neural network model, and obtaining a risk score and a future risk evolution trend of each construction unit; performing clustering analysis on the risk state, identifying a high-risk work point set, and identifying a risk linkage chain based on a causal map; when the risk score exceeds a dynamic threshold value or a closed propagation structure exists, a grading early warning signal is triggered; according to the method, dynamic identification, trend prediction and intelligent early warning of complex construction risks are realized, and the intelligent level of construction safety management is improved.

A three-dimensional point cloud segmentation method and apparatus based on a locally weighted curvature and a two-point method, which method and apparatus relate to the field of point cloud processing. The segmentation method comprises: calculating a normal vector of each sample point in a three-dimensional point cloud image (S101); calculating a locally weighted curvature of each sample point on the basis of the normal vector, and comparing the locally weighted curvature with a preset curvature threshold value, so as to obtain a comparison result (S102); when the comparison result indicates that the locally weighted curvature of the sample point is greater than or equal to the preset curvature threshold value, marking the sample point as a boundary feature point, so as to obtain a boundary feature point set (S103); fitting a point cloud planar model on the basis of the boundary feature point set, and performing iterative calculation on the point cloud planar model, so as to obtain a target inner point set (S104); and segmenting the three-dimensional point cloud image on the basis of the boundary feature point set and the target inner point set (S105).

The invention relates to the technical field of industrial internet, in particular to a mine network operation and maintenance management system and method for coal mine development, and the system comprises a path passage evaluation module, an equipment state recognition module, a scheduling priority module, an operation path conflict evaluation module and a scheduling task issuing module. According to the method, a weighted time consumption model is constructed through task positions and path section passing difficulty parameters to realize path time consumption dynamic estimation and optimal passing path combination, and equipment state identification adopts operation parameter interval assignment and weighted summarization to realize multi-dimensional risk judgment. In task sorting, time consumption and risk levels are normalized to establish a combination interval to dynamically mark the emergency degree, the scheduling sequence is adjusted in combination with task point distribution, a path point set intersection is extracted in path conflict processing, the proportion of overlapped sections is calculated, and task positions are adjusted according to a threshold value to guarantee scheduling continuity; and task issuing uniformly generates a scheduling instruction set containing a path equipment state and a task type, so that the scheduling precision and the execution consistency are improved.

The invention discloses a PINN-based high-precision hydrodynamic numerical simulation method and system, and the method comprises the steps: firstly building a computational domain, setting reasonable geometric parameters and boundary conditions, and constructing a dimensionless Navier-Stokes control equation set; then designing a deep neural network architecture with space-time coordinate input and flow field variable output, and adopting a loss function combining physical constraint and data driving; the core innovation lies in providing a timing sequence sensing RAR-D adaptive sampling strategy, dividing a time domain into a plurality of time frames, performing residual error evaluation in each frame, constructing a probability density function related to residual errors, and balancing priority sampling and overall coverage of a high residual error region; adam and L-BFGS optimizers are adopted to carry out network training, the weight of a loss function is dynamically adjusted, and a sampling point set is periodically updated; and finally, the solution precision is verified through multi-dimensional flow field visualization analysis. Therefore, the prediction precision of the complex flow field is effectively improved, and the calculation efficiency is remarkably improved.

The invention provides a power cable comprehensive online monitoring system, and relates to the technical field of data processing, and the system comprises the steps: extracting a partial discharge distribution point set in a multi-dimensional feature fusion data set, and calculating the weight factor of each distribution point; based on the weight factors, a weighted point set concave packet generation algorithm is adopted to construct a partial discharge source probability distribution boundary, and a boundary feature point set is obtained; inputting the boundary feature point set and the multi-dimensional feature fusion data set into a convolution-long and short-term memory hybrid neural network to generate a diagnosis result; and a third-level early warning signal is generated based on a diagnosis result, early warning information is pushed to a remote operation and maintenance terminal and a mobile APP through an MQTT protocol, a digital work order service is automatically triggered to generate a maintenance task, and positioning verification is executed. According to the invention, the accuracy of monitoring and fault diagnosis is improved.

The invention provides a grouting simulation method and system based on time sequence interaction and self-adaptive grid reconstruction, and relates to the technical field of grouting engineering.The grouting simulation method comprises the steps that monitoring data of the grouting construction process are obtained, the monitoring data are input into a grouting diffusion prediction model, a grouting pressure prediction value is output, and a series of prediction responses are conducted based on the grouting pressure prediction value. Obtaining a grouting pressure predicted value, a boundary point set of the grouting diffusion area, the area and the volume of the grouting diffusion area and a diffusion response result of the pressure distribution diagram; constructing a geometric model of the grouting area based on the diffusion response result, and performing grid unit division and grouting diffusion area mapping on the geometric model; and performing local grid encryption and reconstruction on the grid units according to a density control rule to obtain a structured grid data model, adaptively updating the structured grid data model by using real-time monitoring data, and performing real-time visual grouting optimization control. According to the invention, the high-precision prediction response of grouting is improved.

The invention relates to the technical field of three-dimensional model reconstruction, in particular to an existing building three-dimensional model reconstruction system based on point cloud data, and the system comprises a point cloud data processing module which carries out the octree recursive space subdivision of input point cloud data, builds a hierarchical index structure, and generates a multi-view fusion point set. In the method, octree recursive space subdivision is performed on the input point cloud data to establish a hierarchical index structure, so that the key area of the building surface can be positioned more accurately in subsequent data processing, data redundancy is reduced, and the precision and efficiency of three-dimensional model construction are improved; then voxel depth field data is obtained by calculating a truncation symbol distance function value of voxels, the description accuracy of a building surface structure is improved, and an initial surface grid is formed through contour surface extraction; meanwhile, plane slicing is conducted on the multi-view fusion point set in the preset building main axis direction, and contour details of all building structure sections are captured.

The invention discloses a communication equipment fault intelligent diagnosis method and device, equipment and a medium. The method comprises the steps of collecting equipment operation logs in real time, and extracting an abnormal point set; clustering the abnormal points, and matching a hardware fault, a software error and a network congestion type in combination with a preset mode library; dimensionality reduction is carried out on temperature, voltage and current sensor data based on principal component analysis, and a time-synchronized fusion feature vector is constructed; analyzing features and abnormal modes, and outputting potential fault points; associating the classification model to generate a fault priority by extracting the dynamic characteristics of the network flow; and a dependency network is constructed in combination with Bayesian reasoning, so that accurate fault positioning is realized. The system supports incremental learning, updates an abnormal mode library and automatically expands a knowledge base. Through multi-dimensional data fusion and spatio-temporal feature joint modeling, the problems that a traditional method depends on manual rules and is high in false alarm rate are solved, and the method is suitable for complex scenes such as a 5G base station and a data center.

The invention discloses a BIM-based construction engineering visual management method and system, and relates to construction engineering management, and the method comprises the steps: collecting a data set containing geometric information, attribute information and associated information; dividing the data set into a plurality of hierarchical subsets according to the hierarchical relationship of the building engineering BIM model; an improved DBSCAN clustering algorithm is adopted to perform clustering analysis on each hierarchy subset; classifying and archiving the data set according to the data cluster, and generating a semantic association data collaboration array A according to a classifying and archiving result; carrying out dimension reduction processing on the semantically associated data collaboration array A; constructing a spatial index of the dimension reduction data array by adopting a density-based spatial index method; and dividing the building engineering BIM model into a plurality of partitions by using a distributed computing framework, and performing distributed matching on the partitions and the spatial index to obtain a data matching point set. For the problem that the BIM-based building engineering high-dimensional data clustering precision is low in the prior art, the analysis precision is improved.

The invention provides a water-based adhesive coating control method and system based on artificial intelligence optimization, and the method comprises the steps: obtaining a historical data set composed of process parameters of a coating process and coating quality parameters, carrying out the clustering, and obtaining a global induction point set according to a clustering result; determining a stage induction point subset according to the target parameter value of the current coating stage and the boundary of the target parameter, and updating the sparse Gaussian process regression model by using the global induction point set and the stage induction point subset; calculating a quality fluctuation index, determining the length of a prediction time domain based on the quality fluctuation index, constructing an optimization problem in the prediction time domain by adopting a multi-level opportunity constraint mode, and when the deviation value between the actual value of any key process parameter and the prediction trajectory based on the sparse Gaussian process regression model exceeds a deviation threshold value, determining that the prediction trajectory does not exceed the deviation threshold value. And solving the optimization problem to obtain an optimal control action sequence in the prediction time domain, and determining a final control action from the optimal control action sequence and sending the final control action to an execution mechanism.

The invention discloses a special-shaped shutter fragment splicing drawing method based on space coordinates, and relates to the field of intelligent construction.According to the method, a laser radar scanning device is used for obtaining a window face initial point cloud, and three-dimensional point set representation under a local space coordinate system is established; then executing a multi-level boundary point extraction process including boundary edge identification, mutation point detection and structure segmentation, and constructing a candidate boundary point set Bpre; on the basis, a five-level space structure partitioning strategy is applied, partitioning point distribution rules are set for regions with different geometric characteristics, and finally a window boundary point set B is formed; the point set has high adaptability in the aspects of structural continuity, spatial distribution density, sampling precision and boundary integrity, and the geometric reduction degree of subsequent splicing modeling is remarkably improved; according to the scheme, the problem that a traditional method is difficult to process local complex curvature and structure jump of a special-shaped window is solved, and continuous mapping of structure boundary information to spliced pattern construction is achieved.

The invention discloses a welding seam self-adaption B spline curve fitting method for a 3D dense out-of-order point set, and aims at solving the problem of complex curved surface welding seam track generation in robot welding and breaking through the limitation of the prior art through a three-stage self-adaption fitting mechanism. The method comprises the following steps: firstly, self-adaptive screening and projection sorting are carried out by utilizing a straight line fitting error, so that the disordering problem of discrete points of a quadratic curved surface intersection line is effectively solved, and the strong dependence of a traditional method on a CAD model is overcome; secondly, in combination with a region growing method guided by an end tangent vector, robust conversion from an out-of-order point set to an ordered control point is realized through dynamic neighborhood search and projection value sorting, and path continuity of an open / closed curve is guaranteed; and finally, mean sampling quadratic fitting is adopted, control point distribution is optimized while noise is restrained, and a smooth B-spline track meeting the movement requirement of the welding gun is generated. According to the method, direct fitting of the 3D dense out-of-order point set under the model-free condition is achieved for the first time, and the generation efficiency and adaptability of the irregular welding seam track are remarkably improved.

The invention discloses a dragging positioning spraying track planning method for a local area of the surface of special equipment and a storage medium, and belongs to the technical field of industrial robot automation. The method comprises the steps of obtaining point set data of a to-be-sprayed area; calculating a convex hull based on the point set, and solving a minimum enclosing rectangle by using a rotary caliper method to determine an accurate rectangle spraying boundary; constructing a rectangular amplitude limiting sedimentary model through spraying testing and parameter identification; generating a grid track with a buffer area based on the model and the boundary, and optimizing the spraying speed by taking the minimization time as a target; and for the curved surface, performing B-spline curved surface reconstruction and extracting a normal vector to control the attitude of the spray gun. According to the method, the complex multi-parameter trajectory optimization problem is simplified into single-speed optimization through the rectangular amplitude limiting deposition model, the problems that in local spraying, trajectory planning is complex, edge control is not precise, a coating is not uniform and the like are effectively solved, and the spraying efficiency and quality are remarkably improved.

The invention relates to the field of assembly automation and flexible manufacturing, and discloses an assembly method for a prefabricated firewall of a transformer substation, and the method comprises the following steps: setting a plurality of space positioning reference points at the edge of a wallboard; acquiring three-dimensional coordinate data to form a first point set and a second point set; calculating an initial rigid body transformation parameter between the wallboards according to the first point set and the second point set; establishing an energy functional for describing a total error; carrying out minimization solution on the energy functional by adopting a variational method to obtain an optimal disturbance vector field; according to the optimal disturbance vector field, a plurality of fine adjustment actuators of the wallboard are controlled, posture fine adjustment is conducted, and physical assembly connection of the wallboard and the target wallboard is completed. According to the invention, accurate attitude estimation and error correction in the wallboard assembly process are realized; the robustness and geometric fitting capability of the assembly process are improved; the assembly precision and efficiency are remarkably improved, and the problem that traditional redundancy control distribution is unstable is solved.

The invention discloses a dual-path three-dimensional point cloud compression reconstruction method fusing multi-scale geometric features. The method comprises the following steps: extracting representative skeleton anchor points from original point clouds by adopting farthest point sampling; predicting an octree node occupancy state by using an anchor encoder, and performing entropy coding to generate a skeleton bit stream; a skeleton condition feature encoder is introduced, probability distribution of local features is predicted by using skeleton context priori, cross-path context association is constructed, and a feature bit stream is generated; in the decoding stage, the sparse skeleton point set is restored from the skeleton bit stream, local features are recovered from the feature bit stream in combination with the skeleton structure context, fine-grained reconstruction is carried out on the point cloud through a geometric fusion point cloud up-sampling module, and finally the high-fidelity point cloud is obtained. According to the invention, the point cloud compression effect and reconstruction quality can be improved. The method can be widely applied to the point cloud field.

The invention belongs to the technical field of point cloud data processing, and particularly relates to a robust point cloud registration method and device based on common-view query embedding. The method comprises the following steps: using KPConv-FPN as a feature encoder to carry out multi-scale sampling on an input point cloud pair and extracting point-by-point features; inputting the thickest sampling point and the features of the thickest sampling point into an inter-block common-view feature interaction module, and carrying out common-view feature interaction and common-viewpoint segmentation; inputting the enhanced features and common viewpoints into a block matching module with filtering so as to extract a corresponding relation under a coarse scale; on the basis of a rough matching result, inputting the upper sampling points and the features thereof into a point matching module to generate a dense corresponding point set; calculating a group of transformation candidates by using the corresponding point sets; and finally, registering by using a local-to-global registration scheme. According to the method, the problems of high training cost, extra calculation overhead and low overall efficiency of point cloud registration caused by a traditional two-stage design are solved.

The invention provides a multi-objective optimization method fusing a classifier and an active learning strategy, and belongs to the technical field of artificial intelligence and optimization design, and the method specifically comprises the steps: obtaining a sample data set containing a plurality of target performance function values; performing non-dominated sorting on the sample data set to extract a Pareto frontier point set, and extracting geometric features from the point set as a geometric feature vector sequence; using the geometric features and the corresponding convex marks to train a classifier model of a Transform architecture, wherein the classifier model supports a multi-head attention mechanism and a position coding structure; judging the convex confidence coefficient of the current Pareto front by using a classifier model, and obtaining a candidate scheme; and inputting the candidate scheme into a double-attention mechanism prediction model based on context reasoning, and predicting the multi-target performance of the candidate scheme. According to the method, the experiment frequency can be reduced, the target performance can be accelerated to be achieved, and the intelligence and efficiency of the multi-target optimization process are improved.

The invention discloses a tool path and blank cutting volume calculation method based on point set representation, and the method converts a blank and a tool path into point sets, calculates the cutting volume through the intersection of the point sets, and simplifies the traditional geometric calculation. Meanwhile, the density of the point set can be flexibly adjusted, the calculation efficiency and precision of tool path optimization are improved, and dependence on an STL file and a third-party plug-in is avoided. The method is suitable for large-scale tool path optimization tasks, and the flexibility and controllability of the tool path optimization process can be improved.

The invention discloses a data processing method and device, computer equipment and a cluster, and relates to the field of computers. A plurality of sliding windows can obtain a plurality of tangency points by processing the to-be-stored data at a time, and the calculation time consumed by the calculation device for obtaining the tangency points is shortened. If the number of the sliding windows is too large, the multiple sliding windows slide in parallel to search for the tangency points of the to-be-stored data, the storage space of the cache may be insufficient, and the cache occupancy rate is high, or the multiple sliding windows slide in parallel to determine multiple tangency points, and the calculation time consumption of the tangency points may be long; if yes, the computing device adjusts the degree of parallelism according to at least one of the computing consumed time of the first tangent point set and the cache occupancy rate so as to adjust the number of sliding windows, so that the computing consumed time of tangent points is reduced, cache occupancy is reduced, reasonable tangent points of the to-be-stored data are obtained as soon as possible, and the length of data blocks selected by the multiple sliding windows is adjusted; the accuracy of the computing device for acquiring the tangent point is improved, the acquisition rate of data repeated with the stored data in the to-be-stored data is improved, and the data deduplication rate is improved.

A robot interference checking motion planning technique using point sets. The technique uses CAD models of robot arms and obstacles and converts the CAD models to 3D point sets. The 3D point set coordinates are updated at each time step based on robot and obstacle motion. The 3D points are then converted to 3D grid space indices indicating space occupied by any point on any part. The 3D grid space indices are converted to 1D indices and the 1D indices are stored as sets per object and per time step. Interference checking is performed by computing an intersection of the 1D index sets for a given time step. Swept volumes are created by computing a union of the 1D index sets across multiple time steps. The 1D indices are converted back to 3D coordinates to define the 3D shapes of the swept volumes and the 3D locations of any interferences.

The invention provides a motion video key clip extraction method based on behavior analysis. The method comprises the following steps: firstly, performing decoding and frame standardization on a match video stream, detecting and positioning a coach target, and obtaining and expanding a bounding box; extracting a candidate skeleton key point set in the extended region, selecting the skeleton with the most key points as the skeleton of the coach, and performing cross-frame tracking; analyzing the skeleton in real time, triggering hand fine analysis and extracting hand key points when a scoring trend appears, comprehensively judging based on geometry, kinematics and time sequence rules, and generating a trigger signal when a specific scoring gesture is detected; and the background thread extracts fragments before and after the triggering moment from the video stream as key fragments, and stores the key fragments in association with metadata such as event types, coach identities and competition states.

The invention discloses a wind power equipment fault monitoring method and system based on operation data, and the method comprises the steps: firstly obtaining historical operation data containing a fault type, employing a preset relevance algorithm to quantify the relevance between each operation parameter and a fault, mapping the relevance to a two-dimensional plane, constructing a structured point set, and carrying out the calculation of the structured point set; and the visibility and the processability of the multi-dimensional data are enhanced. In the key influence factor screening process, the operation parameter with the maximum correlation degree is selected as a reference point, the method has definite physical fault characterization significance, can focus on core parameters highly related to a target fault mechanism, and improves the accuracy of a candidate factor set from the source, so that the key influence factors are further screened. And the exclusive fault monitoring sub-model is constructed, so that high pertinence of model input can be ensured, the recognition accuracy and classification definition of various faults are greatly improved, and the fault detection accuracy of the wind power equipment is improved.

The invention relates to the technical field of data storage, and particularly provides a point cloud digital twinning-oriented efficient data storage and retrieval method, which comprises the following steps of: converting an original disordered point set into a dynamic voxel unit with probability density by analyzing a spatial distribution entropy value of a point cloud; generating a hierarchical voxel network with entropy weight marks; based on a hierarchical voxel network, extracting a voxel state transition rule across a time sequence, and encoding dynamic change features into a directed weighted graph; map nodes of the directed weighted map represent voxel stability, and a topological map with a space-time coupling characteristic is formed; training an implicit neural network to generate a continuous semantic vector field by using the space-time coupling characteristic of the topological graph; during retrieval, a target feature region is directly positioned through vector field gradient descent and then is reversely mapped to an original voxel level, and discrete retrieval is converted into vector navigation in a continuous space. According to the method, the contradiction between storage compression and retrieval efficiency is unified.

The invention provides a task allocation method and device for calculation tasks. The method comprises the steps of obtaining task attributes and resource state information of to-be-allocated calculation tasks; calculating an initial position of each initial individual in the initial population based on a good point set generation function, task attributes and resource state information; based on a multi-objective optimization algorithm and an optimization objective, searching to obtain an optimal individual corresponding to the initial population; and based on the optimal individual, obtaining an optimal allocation scheme of the to-be-allocated calculation task. According to the method provided by the invention, an initial position of each initial individual in an initial population is calculated through a good point set generation function, application task attributes and resource state information; based on a multi-objective optimization algorithm and an optimization objective, searching to obtain an optimal flame corresponding to the initial population; based on the optimal flame, the optimal distribution scheme is obtained, the practicability and diversity of the distribution scheme are improved, and then the task performance, reliability and stability of the cloud computing platform are improved.

The invention belongs to the technical field of automatic driving, and relates to a Transform attention mechanism-fused point cloud target detection method and system, and the method comprises the steps: 1, point cloud preprocessing: mapping an irregular sparse point set into a cylindrical voxel tensor of a fixed structure; 2, PillarVFE feature coding is carried out, and point-channel joint feature optimization is realized while the lightness of the network is maintained; 3, generating a pseudo image by the BEV backbone network, and mapping the pilar features to a two-dimensional aerial view plane after feature coding enhancement to form a dense BEV feature map; 4, outputting by a detection head, completing multi-scale feature extraction and context sampling fusion, and finally predicting a three-dimensional bounding box by the detection head; according to the method, the three-dimensional detection performance is remarkably improved in sparse, small-target and long-distance scenes while high reasoning efficiency is kept, and the method has good engineering practicability and research and popularization value.

The invention discloses a video key frame extraction method fused with self-supervised deep learning, and the method comprises the following steps: carrying out the standardized sampling of an input video according to a fixed interval, extracting a SuperPoint local key point and a Video MAE global semantic feature, and generating a dense descriptor and a semantic vector; constructing multi-dimensional change indexes such as local matching degree and global similarity; constructing a soft distribution matrix and a matching point set based on the fusion features; dynamically judging the key frame through a self-adaptive multi-threshold rule; and outputting the key frame set. The method fuses local and global spatio-temporal information, has robust feature extraction and key frame discrimination capabilities under a weak supervision condition, and can effectively improve the efficiency and precision of video compression, abstract and event detection.

The invention discloses a human body three-dimensional modeling method and device, electronic equipment and a storage medium, and relates to the technical field of three-dimensional modeling, and the method comprises the steps: generating an initial sampling point set according to SMPL template vertexes; establishing a dynamic mapping relation between the target surface point set and the initial sampling point set to obtain a mapping matching point set; an error compensation field is constructed to adjust the mapping matching point set, and then a geometric distribution feature map of the human body is obtained; and splicing the geometric distribution characteristic graph and the SMPL vertex graph to obtain a three-dimensional model of the human body. According to the method, the geometric distribution feature map and the SMPL vertex map are spliced, and then the three-dimensional model of the human body is generated according to the spliced feature map obtained by splicing, so that decoupling of the geometric distribution feature map and the SMPL vertex map is realized, the calculation efficiency is improved, and the application cost is reduced.

The invention discloses a radar interference resource allocation method based on a grey wolf algorithm. The method mainly solves the problems that in the prior art, interference resource allocation efficiency is low, combination explosion occurs, and engineering implementation is not facilitated. The scheme comprises the following steps: 1) constructing an interference task scene, and constraining a solution space through a coding strategy optimization method; 2) constructing an interference resource allocation objective function, and initializing an allocation scheme according to a good point set theory; 3) on the basis of the grey wolf algorithm, a nonlinear convergence factor, a dynamic weight and a reverse learning method are introduced to complete updating of a grey wolf position, namely updating of a distribution scheme is completed; and 4) repeating the iterative updating process in the step 3) until the algorithm converges, and taking the wolf pack position corresponding to the currently obtained interference benefit as a final interference resource allocation result. The method can effectively improve the optimization speed in the resource allocation process, avoids falling into a local optimal solution, and can be used for improving the resource allocation task efficiency in a radar cooperative interference scene.

The invention discloses a mechanical arm tail end posture constraint path optimization method based on constraint manifold. A discrete constraint manifold point set which is high in posture precision, robust in singular point, uniform in space distribution and constrained in kinematics is generated based on a Monte Carlo method, a weighted damping pseudo-inverse method, an attenuation type learning rate gradient descent iteration method and voxel density optimization; based on measurement of a laser tracker, the relation between static environment information and a target point relative to a mechanical arm base coordinate system is constructed; a coarse path is obtained through a sampling method, and a feasible path on a constraint manifold meeting attitude constraint and having no collision is generated based on a search strategy of Euler distance difference + maximum joint angle variation + collision detection + index reconstruction criterion; and in combination with a multi-target weighted CHOMP algorithm, the feasible path is optimized into a collision-free and smooth path which meets the requirements of attitude constraint and kinematics constraint. According to the method, the problems of insufficient consideration of attitude constraint, low efficiency, local convergence and the like of traditional path planning are solved.

The invention relates to the technical field of machine learning training fault tolerance, and discloses a LoRA training fault tolerance method based on state awareness. The method comprises the steps that training indexes are collected through a sensor network, a time sequence knowledge base is constructed, and sampling frequency is adjusted in a self-adaptive mode; historical data is utilized to generate a state change trend baseline, and abnormity is preliminarily identified through deviation degree comparison. And for the abnormal behavior, performing verification and risk assessment by fusing the gradient frequency domain characteristics and the loss curve form of the abnormal behavior. Encoding the anomalies into a high-dimensional point set, analyzing and identifying a continuous homologous structure of the high-dimensional point set by adopting topological data, and judging a systematic gradient anomaly mode according to topological characteristics; and deducing a compensation coefficient according to geometric attributes of the topological characteristics, calibrating a risk estimation value, and dynamically reconstructing a monitoring strategy. According to the method, systematic training faults can be accurately identified from a data structure level, adaptive intelligent fault-tolerant control is realized, and the reliability and efficiency of a training process are improved.