9687results about How to "Improve generalization ability" patented technology

The invention discloses a decision-making method and model for off-line reinforcement learning and continuous on-line fine tuning, and the method comprises the steps: obtaining original text-image pair data and historical track data, and obtaining an enhanced state sequence through feature matching verification and standardization processing; performing sequence prediction by using a multi-level attention mechanism, and obtaining pre-training model parameters through knowledge distillation and task adaptability enhancement; on the basis of stability evaluation and plasticity control, online knowledge migration is carried out on the model, and optimized model parameters are obtained by combining experience playback optimization; and finally, performing multi-dimensional performance evaluation and resource optimization. According to the method, through multi-modal feature fusion, dynamic knowledge migration and an adaptive optimization mechanism, the problems of insufficient feature representation, disastrous forgetting, low optimization efficiency and the like are effectively solved, and the learning efficiency and generalization ability of the model are improved.

The invention provides an intelligent task alarm rule self-learning method and system based on support priority, and relates to the technical field of intellectualization, and the method comprises the steps: constructing a bidirectional feature extraction network to collect task operation data, and combining historical alarm data to generate a task exception association model; and carrying out migration mapping on the model and a business scene knowledge base by utilizing a migration learning algorithm, generating a task priority quantitative index, and constructing a self-adaptive priority scoring system. And training an alarm rule generator based on the model and a scoring system, generating a multi-level alarm strategy tree, and realizing differential alarm. And the incremental learning engine monitors a task state in real time, dynamically adjusts an alarm response strategy and feeds back data to optimize a rule. A federal learning mechanism is adopted to cooperatively train a model, and the generalization ability and adaptability of alarm rules are improved, so that more accurate and efficient task alarm is realized, the false alarm rate is reduced, and the operation and maintenance efficiency is improved.

The invention provides an intranet service quality optimization method and system based on deep reinforcement learning, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting multi-dimensional performance index data, carrying out the preprocessing, constructing an analytic hierarchy model, outputting the performance score of each layer, integrating the performance score of each layer, and outputting an intranet service quality score. Constructing a simulation environment, adjusting hyper-parameters of the adaptive deep reinforcement learning model, continuously learning in an actual intranet environment, determining multi-dimensional rewards, performing multi-target optimization, and generating a multi-target optimization result; and setting a hierarchical performance index threshold, if a hierarchical optimization strategy is triggered, adjusting a network topology structure, adjusting resource allocation and service configuration, adjusting weights of different indexes in the service quality evaluation model, updating a reward function, solving an optimal service decision, generating an optimal service decision sequence, and executing the optimal service decision sequence.

The invention relates to a business credit evaluation and supervision method based on a multi-modal coevolution algorithm, and the method comprises the steps: collecting and cleaning multi-modal data, guaranteeing the data quality, uniformly storing the multi-modal data as a structural database, extracting features through cross-modal comparative learning, and enabling the features to be embedded into a dynamic multi-modal knowledge graph, a modal sharing network is used to extract and fuse modal sharing and specific features, consistency of modal feature distribution is optimized through adversarial learning, a global co-evolution credit scoring model is initialized, a dynamic multi-objective algorithm is used to optimize accuracy, robustness and interpretability of the model, and a credit scoring result is obtained. A causal reasoning model is introduced to analyze driving factors of enterprise credit changes, a causal relationship is established, a closed-loop feedback system is designed, a credit evaluation model is updated, adaptability is enhanced, an interpretable credit evaluation report is generated based on a generative adversarial network, and transparency and credibility are enhanced. According to the method, the accuracy and adaptability of credit assessment are improved, and the robustness and interpretability of an assessment result are ensured.

The invention relates to the technical field of electric power short-term load prediction, in particular to an industrial park short-term load prediction method and system based on a Diffusion-Transform-AR model. Acquiring non-stationary historical load data and meteorological data of the industrial park, and performing missing value processing on the data; performing correlation analysis on the data by using the maximum information coefficient, selecting meteorological data with relatively high correlation, and reprocessing time data; performing time sequence decomposition on the load data by using a seasonal trend decomposition method to obtain trend term, seasonal term and residual term features; constructing a diffusion Transform prediction model, and taking the residual term, the meteorological data and the time data as input prediction residual; constructing an autoregression prediction model and taking the trend term as an input prediction trend; according to the method, the advantages of various models are combined, time sequence features are deeply mined, the prediction precision and the model generalization ability are effectively improved, and the method is suitable for complex industrial park load prediction scenes.

The invention provides an intelligent networked automobile dynamics simulation method and system based on a physical engine, and relates to the technical field of intelligent networked automobiles, and the method comprises the steps: obtaining vehicle parameters and road condition information of an intelligent networked automobile, optimizing a multi-agent cooperative control model through a self-adaptive optimization algorithm based on the vehicle parameters and the road condition information, and carrying out the optimization of the multi-agent cooperative control model. Constructing a digital twinborn model, and mapping the optimized multi-agent cooperative control model to an actual intelligent networked automobile control system; the method comprises the following steps: establishing a distributed edge computing network of an intelligent networked automobile group, constructing a decentralized knowledge base by utilizing a block chain technology, forming a decision-making experience data set, designing a multi-level attention mechanism to extract target information based on the decision-making experience data set and multi-modal data of a current traffic scene, and generating a final vehicle control instruction. And transmitting the final vehicle control instruction to a vehicle execution system to realize dynamic adaptive group cooperative control of the intelligent networked vehicle.

The invention discloses a power load prediction method based on a two-channel cross attention network, and belongs to the technical field of time sequence prediction in deep learning. The method comprises the following steps of obtaining a power load data set, performing preprocessing on the power load data set, decomposing power load data into a plurality of local components by applying VMD, iteratively optimizing VMD parameters through an APSO algorithm, and screening a subsequence with the lowest complexity to be combined with original features to serve as model input; and then constructing a dual-channel cross attention network model, processing time sequence data by using a parallel mechanism of a BiLSTM channel and a Transform channel, enhancing the feature expression ability through an improved cross attention layer, and performing power load prediction by using an MLP fused KAN network. According to the method, the time sequence prediction of the power load is carried out by constructing the two-channel cross attention network, the prediction precision and robustness are high, and the accuracy of load time sequence prediction is remarkably improved.

The invention provides a crop yield prediction method and system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining continuous historical meteorological data, soil data, crop planting data and remote sensing image data of a designated region for many years, and carrying out the alignment according to a time sequence, and constructing a multi-dimensional spatio-temporal data cube; a fusion prediction model is trained, the fusion prediction model is used for extracting growth feature vectors in the crop growth and development process and carrying out yield prediction, and model parameters of the fusion prediction model are dynamically adjusted through an adaptive learning rate optimization algorithm; the method comprises the following steps: acquiring real-time meteorological data, soil data and remote sensing image data of a specified area, inputting the data into a pre-constructed fusion prediction model, introducing an agricultural knowledge graph into the fusion prediction model to restrain and correct a prediction result of the fusion prediction model, and reasoning to obtain crop growth and estimated yield of each grid of the specified area. And performing spatial interpolation on the estimated yield of each grid to generate a yield distribution diagram of the specified area.

The invention relates to the technical field of data processing, and discloses a coal and gas outburst dynamic monitoring and intelligent early warning method and device. The method comprises the following steps: collecting and preprocessing multi-source sensor data arranged in a target area to obtain a multi-dimensional data set; performing multivariable time sequence analysis on the multi-dimensional data set to obtain coal seam gas outburst risk prediction data; based on coal seam gas outburst risk prediction data, performing dynamic risk assessment and spatial interpolation processing on the multi-dimensional data set to obtain real-time risk distribution data of the target area; performing multi-dimensional threshold analysis on the real-time risk distribution data to obtain preliminary early warning data; performing multi-modal information fusion on the preliminary early warning data to obtain target early warning data; and performing intelligent decision tree analysis on the target early warning data to obtain a target early warning scheme. According to the invention, the accuracy and reliability of dynamic monitoring and intelligent early warning of coal and gas outburst are improved.

The invention provides a three-dimensional point cloud data segmentation method and system based on multi-scale point features, and relates to the technical field of point cloud processing, and the method comprises the steps: obtaining original point cloud data, constructing a three-dimensional Hash index table, carrying out the self-adaptive hierarchical division through a local maximum value of a density distribution curve, and obtaining a local maximum value of a density distribution curve; noise point elimination is carried out based on a region density count value and a geometric distribution coefficient, an enhanced feature vector containing a main feature direction, a main curvature value and the like is constructed for non-eliminated points, a local geometric descriptor is constructed by adopting four spherical search regions with increasing radiuses, a three-branch space attention network is input to extract spatial features, and the spatial features are extracted; and fused feature data are obtained through a multi-level geometric feature extraction network and a bidirectional feature propagation network. And obtaining a final segmentation result through segmentation energy function optimization and a region merging strategy.

The invention relates to the technical field of intelligent traffic prediction, in particular to a time-space correlation traffic flow prediction method based on deep learning, which comprises the following steps: S1, collecting traffic flow data and related external environment data in real time, and establishing an initial traffic network topology structure; s2, calculating the dynamic weight of each node; s3, dynamically adjusting the weight of each edge in the traffic network according to the node weight change in the S2 so as to update the topological structure of the traffic network; s4, performing feature extraction on the spatial dependency relationship among the traffic nodes, and extracting time evolution features of the traffic flow data at the same time; s5, constructing an adaptive space-time fusion model; s6, traffic flow prediction is carried out, and adaptive space-time fusion model parameters are optimized; according to the method, through dynamic fusion of the spatial-temporal characteristics and adaptive adjustment of the model parameters, the precision and real-time performance of traffic flow prediction are effectively improved, and the adaptability and generalization ability of the model in a complex dynamic traffic scene are enhanced.

The invention relates to the technical field of multi-axis linkage numerical control machining, and discloses a multi-axis linkage numerical control machining state monitoring method and system.The method comprises the steps that vibration, acoustic emission and cutting force signals in the multi-axis linkage numerical control machining process are collected and preprocessed, and first multi-source sensor data are obtained; performing feature extraction and compressed sensing to obtain a low-dimensional feature subset; performing fusion judgment to obtain an abnormal state detection result; training a deep learning model comprising a backbone network and a plurality of branch networks to obtain a state evaluation model; and inputting second multi-source sensor data collected in real time into the state evaluation model for state evaluation to obtain a processing state classification result and a parameter estimation value. According to the method, cooperative state evaluation and prediction simulation of multiple machine tools are realized.

The invention provides a method and system for realizing electronic channel risk control processing based on a code insertion technology, and relates to the technical field of electronic channel risk control, and the method comprises the steps: encrypting user information and channel identification information through a dual Hash algorithm to generate a dynamic session identification code, and forming a target coding sequence through employing a self-adaptive segmentation algorithm and a business feature flag bit. And constructing a dynamic risk control rule model by using a graph neural network, extracting user multi-dimensional features, carrying out weighted fusion to generate a user risk score, comparing the user risk score with an adaptive risk threshold, analyzing abnormal behaviors based on a probability graph model, and generating an intelligent risk early warning report. And calling a reinforcement learning risk control strategy engine to determine an optimal disposal strategy, and issuing a multi-factor interaction verification instruction. And after verification, selectively activating a blocking mechanism according to a verification result, synchronizing interaction data and a result to a distributed risk control data center, and continuously optimizing the risk control model based on federal learning.

The invention provides a prediction model construction method and system for personalized equipment operation and maintenance, and relates to the technical field of data processing, and the method comprises the steps: collecting equipment operation data and environment data, extracting the change characteristics of the equipment operation state, carrying out the correlation analysis of the environment data based on historical equipment fault records, and obtaining a prediction model. Performing normalization processing, performing time sequence analysis, generating a joint feature matrix, segmenting the joint feature matrix into time slices of a fixed window according to a time sequence, classifying the time slices into a normal state and an abnormal state, performing feature contribution degree analysis, generating a feature weight matrix, and performing feature extraction; re-combining the equipment features in the time slice feature data set with the environment factors according to feature weights, generating fusion feature data layer by layer, performing supervised learning training on the fusion feature data, constructing an equipment prediction model, and outputting the equipment prediction model as a prediction model of personalized equipment operation and maintenance; according to the invention, the operation and maintenance effect of personalized equipment is improved.

The invention relates to the technical field of image processing, and discloses an industrial article appearance detection method and system based on machine vision. The method comprises the steps that a workpiece image is collected, and a standardized image is obtained through Gaussian filtering, histogram equalization and Otsu algorithm preprocessing; feature vectors are extracted through edge detection, Hough transform and a gray-level co-occurrence matrix; training a defect identification model by adopting a CNN of an attention mechanism; inputting an image for real-time detection; fusing the detection data and the environmental parameters by using a D-S theory; and establishing a database based on a distributed architecture to generate a quality report. According to the method, the detection accuracy is improved through multi-source information fusion, and efficient management and analysis of mass detection data are realized by adopting a distributed architecture.

The invention belongs to the technical field of image classification neural networks, and discloses a lightweight image classification neural network architecture system based on knowledge distillation. The invention provides a lightweight image classification neural network architecture system based on knowledge distillation. The lightweight image classification neural network architecture system comprises a data set preprocessing module, a teacher model preprocessing module, a subnet model construction module, a personalized training module, a real loss calculation module, a distillation loss calculation module, a total loss calculation module and a student model construction module. The invention provides a lightweight neural network architecture system based on knowledge distillation, and designs a brand new training mode by using the superiority of knowledge distillation for teacher model weight extraction: the knowledge distillation of teacher and student models is combined with a deep compression method; the system for constructing the lightweight network can be used for training various complex image classification neural networks so as to improve the network performance.

The invention relates to the technical field of data processing, and discloses a power distribution method based on AI artificial intelligence and an intelligent power distribution cabinet. The method comprises the following steps: carrying out real-time data acquisition on key nodes of the power distribution system through an intelligent sensor network to obtain original operation data of the power distribution system, and carrying out cleaning and standardization processing on the original operation data to obtain a standardized data set; performing analysis modeling on the standardized data set through multi-level hybrid model training to obtain an AI analysis model of the power distribution system; processing and analyzing the power distribution data collected in real time to obtain a load prediction result and a dynamic balance strategy; performing abnormal mode identification and multi-path reasoning on the system operation parameters to obtain equipment health state evaluation and maintenance suggestions; and based on the equipment health state assessment and maintenance suggestions, performing multi-objective optimization calculation on renewable energy source supply and power consumption demands to obtain an intelligent energy scheduling scheme. According to the invention, the efficiency and accuracy of the power distribution method based on AI artificial intelligence are improved.

The invention relates to the technical field of environmental sound event detection, in particular to an environmental sound event detection method based on multi-modal data fusion, which utilizes a microphone array, a camera and a sensor to synchronously acquire audio, video and environmental data to form a multi-modal data source. And the collected multi-modal data is preprocessed, so that the data quality and the accuracy of subsequent analysis are improved. And key features are extracted from the processed data, and the features are fused to form multi-modal features. And a deep learning model is constructed, and multi-modal features are used for training, so that a model capable of identifying an environmental sound event can be obtained. And performing optimization such as pruning, quantification and knowledge distillation on the trained model, and deploying the model on edge equipment to realize real-time preliminary feature extraction and event detection. The problems that an environment sound event detection method in the prior art is poor in robustness, low in real-time performance and limited in generalization ability are solved.

The invention provides a building construction monitoring method, which comprises the steps of obtaining multi-view video image data of a construction site, and performing fusion processing on the multi-view video image data by adopting an image splicing algorithm to obtain a panoramic three-dimensional scene model of the construction site; acquiring an image frame in the panoramic three-dimensional scene model, and analyzing the image frame by adopting a pre-trained work type identification model to obtain position information of operating personnel and equipment of each work type; acquiring position information of the operating personnel of each type of work, and analyzing postures of the operating personnel by adopting a pre-trained human body posture estimation model to obtain action state information of the operating personnel; and acquiring the position information and the action state information of the operating personnel of each type of work, and judging the specific operation state of the operating personnel of each type of work in combination with a pre-established work type operation knowledge base.

The invention relates to a temperature control method, device and equipment of semiconductor equipment and a storage medium. The method comprises the following steps: performing finite element mesh division and modal coordinate transformation on a plurality of temperature control areas of semiconductor equipment to obtain a temperature control area dynamic model; collecting measurement data of a temperature sensor and performing time-frequency domain transformation to generate a temperature prediction data sequence; gaussian mixture modeling is carried out to obtain prior distribution, and temperature parameter probability distribution is obtained; performing sub-domain division on the temperature control range to obtain a temperature inter-domain feature mapping matrix; and a temperature interval thermal coupling coefficient matrix is calculated based on the temperature control area dynamic model, a temperature control objective function is constructed in combination with the temperature prediction data sequence and the temperature inter-domain feature mapping matrix, and a temperature control quantity sequence is solved by adopting a model prediction control algorithm. According to the invention, the accuracy and dynamic response speed of temperature control are improved, and the stability of temperature control and the energy utilization efficiency are ensured.

The invention discloses a multi-working-condition-oriented equipment fault diagnosis method based on unbalanced learning, and the method comprises the following steps: collecting the operation data of equipment under different working conditions through a plurality of sensors, carrying out the preprocessing, feature extraction and feature selection of the collected original operation data, and obtaining a feature set; processing the feature set by adopting an unbalanced learning technology and a cost-sensitive learning algorithm to obtain a balanced data set; constructing a multi-working-condition fault diagnosis model based on learning, and optimizing model parameters by adopting a method of combining cross validation and grid search according to a balanced data set in a model training process; deploying a multi-working-condition fault diagnosis model obtained through training, carrying out the online diagnosis of equipment faults, and continuously updating and optimizing the model through a continuous learning mechanism. According to the method, the unbalanced learning technology is applied to multi-working-condition fault diagnosis, the accuracy and robustness of diagnosis can be remarkably improved, and real-time monitoring and predictive maintenance of the equipment health state are achieved.

The invention relates to the technical field of image classification, and discloses a breast image classification evaluation method and system based on deep learning, and the method comprises the steps: carrying out the preprocessing of sample breast image data, and obtaining comprehensive feature data; performing feature enhancement based on the comprehensive feature data to obtain enhanced feature data; performing multi-modal bidirectional feature fusion on the enhanced feature data to obtain multi-modal fusion feature data; based on the multi-modal fusion feature data, constructing and training a semi-supervised local feature reservation dense connection context-aware graph convolutional network to obtain a graph volume integral class model; according to the multi-modal fusion feature data, constructing and training a multi-branch residual attention reconstruction network to obtain a multi-scale feature fusion model; through the image volume integral class model and the multi-scale feature fusion model, classification evaluation and prediction result integration are carried out on the real-time mammary gland image data to obtain a target mammary gland image classification result, and the method improves the accuracy and reliability of mammary gland image classification.

The invention provides an unmanned aerial vehicle accessory intelligent identification and access method and system based on a sensor, and relates to the technical field of unmanned aerial vehicle accessory management, and the method comprises the steps: collecting the three-dimensional contour information of unmanned aerial vehicle accessories, generating standard point cloud data, carrying out the feature extraction, and carrying out the feature alignment and interaction, generating a high-dimensional fusion feature matrix of the target unmanned aerial vehicle accessories; based on the feature pyramid network, constructing multi-scale feature representation, generating hierarchical feature descriptors, performing similarity calculation, generating a similarity score matrix, performing optimal matching, establishing an adaptive verification rule base and performing accessory verification; calculating an optimal action sequence of an access path based on model information, mapping the optimal action sequence into a motion track parameter of a mechanical arm, collecting multi-point stress data, predicting the optimal clamping force, establishing an accessory service life prediction model, performing prediction in combination with historical data, and generating an accessory state evaluation report and a maintenance suggestion.

The invention belongs to but not limited to the technical field of emotion recognition, and discloses a multi-modal classroom emotion recognition method and system based on modal adaptive learning, and the method comprises the steps: building a multi-modal data collection system, collecting the multi-modal data, such as texts, audios and videos, in a classroom teaching process in real time, and carrying out the real-time collection of the multi-modal data; and preprocessing and feature extraction are carried out on data of different modes. For emotion expression feature differences of different classroom scenes, a multi-modal data fusion method based on adaptive modal scoring (AMS) is provided, and a time sequence modeling and emotion classification model is constructed. The multi-modal fusion and time sequence modeling mechanism simulates the multi-dimensional perception of the teacher on the emotional states of the students in the class, and enhances the interpretability of the emotion recognition system. According to the system, weighted fusion is carried out on data features from texts, audios and videos through fusion of multi-modal data, and complete emotion feature representation is formed.

The invention relates to the technical field of sensor detection, and discloses a detection method and device based on a multi-mode gas sensor, and the method comprises the following steps: carrying out blending processing on a metal oxide semiconductor and a conductive material to obtain a bimodal nano composite material, and creating a multi-mode gas detection module, the preprocessed multi-mode signal data is acquired and obtained; acquiring and analyzing data of the temperature sensor and the humidity sensor to obtain environmental factor compensation parameters; performing adaptive calibration processing on the preprocessed multi-modal signal data according to the environmental factor compensation parameters to obtain calibrated multi-modal signal data; constructing a multi-input neural network model based on the calibrated multi-modal signal data, and training to obtain a gas recognition model; multi-mode signals collected and calibrated in real time are analyzed and judged based on the gas recognition model, gas type and concentration information is obtained, and high-precision and high-reliability gas detection is achieved.

The invention provides a structured data modeling analysis method based on Transform, and relates to the technical field of data analysis, and the method comprises the steps: carrying out the piecewise linear transformation and adaptive quantum coding of numeric features and category features through a double-flow feature encoder, and carrying out the fusion through a gating mechanism to obtain fusion features; inputting the enhanced training sample into a feature extraction network comprising a hierarchical attention module and a self-calibration module, and extracting global feature representation through a feature affinity graph and a causal attention mechanism; and constructing an information bottleneck feature filter to select a key feature subset, performing prediction and reconstruction by adopting a double-branch network, and analyzing the feature contribution degree by using a structured gradient method. According to the method, heterogeneous features can be effectively processed, the model performance is improved, and meanwhile an interpretable feature importance evaluation mechanism is provided.

The invention belongs to the technical field of learner cognitive state interpretation, and particularly discloses a brain-language-vision large model-based cognitive state interpretable method, which comprises the following steps of performing multi-modal data collection for the cognitive state of a learner; performing learner cognitive state characterization based on a multi-scale map convolutional network; brain-vision-language mixed multi-modal time sequence feature fusion based on hierarchical mutual information maximization is carried out; embedding knowledge graph nodes for enhancing the complex task understanding ability of the multi-modal large model; training a brain-vision-language multi-modal large model guided by knowledge; recognizing a learner cognitive state based on a multi-modal large model; the cognitive state of the learner can be explained and analyzed under the support of multi-mode owned reasoning. According to the cognitive state interpretable method based on the brain-language-vision large model, the input multi-modal fusion features are converted into high-quality prompts so as to guide the large language model to extract implicit knowledge, and important support is provided for reasoning generation.

The invention discloses a method and a device for constructing a sample efficient world model based on Transform. The method comprises the following steps: separating historical interaction data to obtain historical action sequence data and historical original state data; converting the original state data into multi-scale feature data and carrying out denoising enhancement processing to generate enhanced state data; performing adaptive quantization coding on the enhanced state data to obtain discrete latent variable data; causal structure data are constructed based on historical data, attention weights are designed, and causal enhancement representation data are generated; calculating uncertainty data based on the time sequence association data and generating multi-step prediction sequence data; decomposing the long-term prediction target to obtain layered value data and calculating strategy gradient data; and screening the new interaction data and updating model parameters. Through the methods of multi-scale feature extraction, causal structure learning, hierarchical value evaluation and the like, the prediction accuracy and the sample utilization efficiency of the world model are improved.

The invention provides a multi-modal large model construction and operation method and system for reservoir dam safety. The method comprises the following steps: constructing a training set, wherein the training set comprises a single sample; multi-modal data as model input: sensor data arranged on the dam, historical hidden danger records, engineering technology documents formed in various stages of planning, design, construction and operation of the dam, information resources related to dam safety, geometric and material information of the BIM model, timestamps and spatial position information; the probability value of occurrence of the hidden danger event serving as the training label is obtained; the training set is adopted to train a multi-mode pre-training model based on a Transform framework; performing real-time prediction and early warning on potential safety hazards of the dam; and calculating the actual influence degree, the comprehensive risk score, the dynamically set early warning threshold and the emergency measure suggestion of the hidden danger event according to the probability value of whether the hidden danger event occurs or not output by the model. The scientificity and effectiveness of dam safety monitoring are comprehensively improved.

The invention discloses an unmanned aerial vehicle cooperative air combat decision-making method based on GRU-MAPPO deep reinforcement learning, and relates to an air combat game maneuvering strategy. The method comprises the following steps: starting from a complex air combat environment, analyzing a motion model, a maneuvering action, an air combat situation and a reward function of an unmanned aerial vehicle, and establishing a multi-unmanned aerial vehicle air combat environment model which comprises an unmanned aerial vehicle particle motion model, an unmanned aerial vehicle maneuvering action model, an air combat situation evaluation model and an unmanned aerial vehicle reward function model; for a completely cooperative multi-agent environment, a partially observable Markov decision chain framework is established, and a working mode of centralized training and distributed execution is adopted; for unmanned aerial vehicle collaborative decision making, a near-end strategy optimization PPO algorithm is expanded to the multi-agent field, and a GRU is used to replace a full-connection neural network in an original algorithm, so that a GRU-MAPPO algorithm suitable for multi-agent deep reinforcement learning is obtained; a simulation experiment shows that the method can improve the collaborative decision-making capability and the air combat winning rate of the unmanned aerial vehicle.