35 results about "Gradient network" patented technology

The invention relates to the technical field of process optimization control, and discloses a silk-covered enameled winding process optimization control method and system based on intelligent analysis. The method comprises the steps that multi-area production parameters are collected in real time and preprocessed; constructing parameter and product quality mapping through principal component dimensionality reduction, grey correlation and support vector regression; training an Actor-Critic structure by using a depth deterministic strategy gradient network to obtain an adjustment strategy; a non-dominated sorting genetic algorithm is combined with analytic hierarchy process to carry out multi-objective optimization, and a control scheme for balancing mass, efficiency and energy consumption is realized. According to the application, on the basis of considering the complex coupling relationship among the process parameters, multi-objective dynamic balance optimization of product quality, production efficiency and energy consumption is realized, the process parameters can be adaptively adjusted according to the production state and the task demand, and the stability and the optimization degree of the silk-covered enameled winding production process are improved.

The invention discloses a multi-target bridge unmanned aerial vehicle inspection path optimization method and system, and relates to the technical field of crossing of bridge operation and maintenance and unmanned aerial vehicle cluster control. The method comprises the following steps: constructing a bridge three-dimensional point cloud model, extracting key inspection points, and establishing a multi-objective optimization function including path length, time consumption, coverage rate, risk coefficient, cluster safety and formation retention; a multi-target reward mechanism is fused through a double-delay depth deterministic strategy gradient network, and a deep learning model is constructed and used for generating an unmanned aerial vehicle cluster obstacle avoidance strategy; training a deep learning model; and based on real-time data, performing multi-target bridge unmanned aerial vehicle routing inspection path optimization by using the trained deep learning model, and adjusting a weight coefficient of a multi-target optimization function through an attention network dynamic weight distribution technology in the optimization process. According to the invention, multi-objective optimization of routing inspection safety, efficiency, quality and cluster cooperation is realized, and the method is suitable for automatic routing inspection of various complex steel structure bridges.

The invention discloses a fault identification method for generating vibration data based on frequency domain transformation under a finite sample. The method comprises the following steps: collecting multiple types of fault vibration signals and extracting frequency domain spectrum characteristics; a forward diffusion stochastic differential equation based on mirror image Brownian motion is constructed, and original spectral distribution is converted into Gaussian prior distribution through progressive noise injection; training a gradient network by adopting a fractional matching criterion, and realizing spectrum feature reconstruction by combining a reverse stochastic differential equation and a Langevin dynamic sampling strategy; and the generated sample and original data are spliced and then input into a spectrum feature lexical element Transform model, local resonance frequency band feature extraction is enhanced through a multi-head attention mechanism, and finally fault identification is realized through coding, decoding and linear classification. According to the method, the problem of low fault identification precision of the rotating machinery in a data scarce scene is effectively solved, and a high-reliability data enhancement and feature decoupling scheme is provided for an industrial equipment intelligent diagnosis system.

The invention relates to the technical field of fire safety monitoring and early warning, in particular to a fire risk dynamic assessment and early warning method based on deep reinforcement learning, comprising the following steps: acquiring multi-source sensing data of a target area; a dynamically updated fire risk assessment state space is constructed based on multi-source sensing data, and a real-time state in the state space is processed and a risk assessment decision is generated through a deep reinforcement learning agent optimized based on an accumulated award function weighted by an accuracy award, a false alarm punishment and a missing alarm punishment; the deep reinforcement learning agent adopts a deep deterministic strategy gradient network structure; mapping the continuous output of the risk assessment decision into a discrete dynamic fire risk level; and triggering an early warning signal matched with the dynamic fire risk level according to the dynamic fire risk level. Through multi-source data standardization, dynamic state space, reinforcement learning optimization and grading early warning binding, accurate dynamic evaluation, reliable early warning and efficient emergency response are realized.

The invention provides an AGV scheduling method and related equipment integrating path planning and traffic control, the method comprises the steps of obtaining task information, vehicle dynamic factors and topological map data of an AGV vehicle, the vehicle dynamic factors comprising a cargo state, a storage location state, spinning prohibition and point location prohibition; based on the task information, the vehicle dynamic factors and the topological map data, a depth deterministic strategy gradient network is adopted to optimize the weight, and an improved A star algorithm formed by the sum of basic cost and real-time congestion cost is adopted to plan the driving path of the AGV; continuously monitoring the running state of the AGV cluster in the process that the AGV runs based on the planned path; and conflict prediction is carried out based on the operation state of the AGV cluster, and traffic control is carried out on the AGV cluster by using a preset conflict prevention mechanism and a heuristic deadlock removing strategy. According to the method, vehicle dynamic factors are considered, an A star algorithm is optimized, traffic control is configured, and the AGV scheduling efficiency is improved.

The invention relates to the technical field of intelligent fuel supply, in particular to an intelligent optimization method and system for fuel supply of a thermal power plant based on multi-source information fusion. The method comprises the following steps: collecting multi-source heterogeneous data and carrying out space-time alignment and normalization processing; constructing a coal as fired calorific value soft measurement model by adopting a physical constraint corrected LSTM network; generating a transportation delay predicted value and probability distribution thereof through an XGBoost regression model; calculating a comprehensive risk score in combination with a Bayesian risk network; solving a Pareto optimal solution set under the constraint of an inventory safety threshold by adopting a depth deterministic strategy gradient network, and generating a fuel scheduling instruction; and updating the as-fired coal calorific value soft measurement model on line. Through combination of the physical constraint corrected LSTM network and multi-source time sequence data fusion and regularization training, the fire coal calorific value prediction precision is remarkably improved, and the problems that a traditional method depends on off-line assay, response is lagged, and the real-time dynamic state of the combustion process is difficult to reflect are solved.

The invention relates to the technical field of intelligent control, and discloses a special line painting control method and system based on an intelligent algorithm. The method comprises the following steps: collecting painting line sensor data to construct a process digital model; training the depth deterministic strategy gradient network based on the model, and generating a multi-parameter cooperative control instruction; executing the control instruction and generating a feedforward signal through multi-stage comparison of quality data; the feedforward signal and the control instruction are fused, and a real-time control strategy is formed through network correction. According to the method, the problem of digital modeling based on multi-dimensional parameter mapping in special line painting control is solved, and the problem of insufficient precision of painting process multi-parameter cooperative control is overcome. Through an intelligent algorithm of a depth deterministic strategy gradient network and a feedforward control signal fusion mechanism, the problems of lag detection of quality abnormity and insufficient dynamic response in the painting process are effectively solved.

The invention discloses a network data security and privacy protection method and device, and the method comprises the steps: obtaining network behavior data features in real time, calculating a behavior entropy value, judging whether to trigger privacy protection or not based on the behavior entropy value, and continuing a next step if the privacy protection is triggered; dynamic privacy budgets of different features are generated based on the behavior entropy in combination with an LSTM model; generating a feature sensitivity grade through a dynamic privacy budget; performing mask processing on the gradient network behaviors of the corresponding features based on feature sensitivity classification; the main technical scheme and effects are as follows: 1, through a dynamic privacy quantification engine, a network behavior entropy value is calculated in real time, and a risk is predicted by using an LSTM model, so that a dynamically changing privacy budget is generated; in this way, the privacy protection intensity can be adaptively adjusted according to the real-time threat situation, and the rigid mode of fixed budget in the existing differential privacy technology is thoroughly broken through.

The application relates to the technical field of intelligent control, and discloses a special line painting control method and system of an intelligent algorithm. The method comprises the following steps: collecting painting line sensor data to construct a process digital model; training a deep deterministic policy gradient network based on the model to generate a multi-parameter collaborative control instruction; executing the control instruction and generating a feedforward signal through multi-level comparison of quality data; fusing the feedforward signal and the control instruction, and forming a real-time control strategy through network correction. The application solves the problem of digital modeling based on multi-dimensional parameter mapping in special line painting control, and overcomes the problem of insufficient precision of multi-parameter collaborative control of the painting process. Through the intelligent algorithm of the deep deterministic policy gradient network and the feedforward control signal fusion mechanism, the problems of lagging detection and insufficient dynamic response of quality abnormalities in the painting process are effectively solved.

The invention belongs to the technical field of intelligent automobile lighting, and particularly relates to a dynamic glare suppression method and system based on multi-automobile cooperative perception, a vehicle and a computer readable storage medium. According to the method, a glare sensitivity distribution diagram which is sent by an opposite vehicle and based on the visual angle of the vehicle is received through V2X vehicle wireless communication, vehicle-vehicle information sharing is achieved, and the sensing limitation of a single vehicle is broken through; constructing a human eye glare sensitivity model based on the CIE standard, converting the physical brightness into subjective glare perception to generate a smooth weight map, and performing curve and ramp dynamic geometric compensation on the map to eliminate perspective distortion; and finally, generating a pixel-level dimming instruction by using the multi-agent depth deterministic strategy gradient network in combination with the lighting requirement of the vehicle, and driving the pixelated headlight to execute accurate brightness adjustment. According to the method, the problems of multi-vehicle meeting game conflicts, rough control strategies and failure of complex road conditions are effectively solved, and collaborative and accurate glare suppression and remarkable improvement of subjective comfort are achieved.

The application discloses a chiller group control optimization method based on layered reinforcement learning, and aims at the problems of poor dynamic adaptability of traditional rule control, high difficulty of single reinforcement learning in processing mixed action space and existence of equipment operation risk, wherein the control task is divided into high-layer discrete decision and low-layer continuous control; the high layer adopts a deep Q network to output the number of chiller units in operation and a water outlet temperature setting value, and the safety timer is used to limit the high-layer action update frequency; the low layer adopts a deep deterministic policy gradient network to output a chiller load distribution ratio and the operation frequency of each variable frequency equipment, and the safety mapping is used to guarantee that the equipment constraint is in line with the regulations; the high and low layers are cooperatively optimized through window cumulative rewards, and the multi-objective reward function is combined to balance the energy saving, stability and safety; and the application can effectively reduce the energy consumption of the chiller room, reduce the number of chiller unit start and stop, and improve the cooling control stability and system energy efficiency.

The invention provides a kiln state optimization method and system based on deep learning, and relates to the technical field of state optimization, and the method comprises the steps: collecting the operation parameters of a kiln in real time through a sensor, and the operation parameters comprise temperature, pressure, gas concentration and energy consumption data; a deep reinforcement learning model is constructed, the deep reinforcement learning model comprises a state encoder and a strategy gradient network, the deep reinforcement learning model is trained based on the historical operation parameters and the historical adjustment amount, and the state encoder adopts a convolutional neural network and a long-short-term memory network to perform feature extraction on the historical operation parameters; the strategy gradient network is used for learning the mapping relation between the extracted features and the historical adjustment amount; and inputting real-time data into the trained deep reinforcement learning model to obtain a real-time adjustment amount, and adjusting kiln operation parameters according to the real-time adjustment amount. According to the method, the reliability of the kiln state optimization result based on deep learning can be improved.

The invention discloses a measurement data dynamic correction method based on deep reinforcement learning, and the method comprises the following steps: carrying out the self-adaptive correction of an error caused by environment change and noise in measurement data through combining a depth deterministic strategy gradient network and a Bayesian linear regression model. According to the method, a state vector is constructed through acquisition time, equipment state and environmental parameter information, an initial correction action is generated by using a depth deterministic policy gradient network, and confidence interval parameters are output through Bayesian linear regression to evaluate correction uncertainty. And performing weighted adjustment on the correction action according to the confidence coefficient and then correcting the measurement data, and generating a joint reward signal based on the error change and the confidence coefficient for updating a strategy network and a Bayesian model so as to realize stable optimization and strategy output in a dynamic environment. The invention aims to solve the problem of measurement errors caused by environment change and noise interference.

The invention discloses a multispectral ambient light adaptive compensation method and system based on reinforcement learning, and belongs to the technical field of image processing and intelligent control, and the method comprises the following steps: a multispectral environment perception step: collecting a visible light image and an infrared image through an RGB-IR four-channel optical perception system, carrying out multispectral feature fusion and estimating environment illumination intensity; a reinforcement learning decision-making step: outputting parameter adjustment actions of light source intensity and exposure time based on the depth deterministic strategy gradient network; light compensation execution: adjusting light source and camera parameters according to the action vector; and an image quality evaluation step: calculating the gradient intensity, gradient uniformity and noise standard deviation of the structured light stripes as reward signals and feeding back the reward signals to the decision network, thereby forming an intelligent closed loop of perception-decision-execution-evaluation, and effectively solving the problem of unstable image acquisition quality caused by illumination fluctuation and metal reflection in a complex industrial environment.

The invention discloses a camouflage detection method, system and device based on an improved depth gradient network, and a medium, and belongs to the technical field of picture detection, and the method comprises the steps: obtaining a camouflage target image; processing the camouflage target image by using the improved context encoder to obtain an initial feature; processing the initial features according to an integrated attention mechanism to obtain enhanced features; performing context exploration on the foreground and the background of the enhanced feature prediction to obtain refined features; and outputting a camouflage target detection result according to the refining features, and verifying the authenticity of the detection result through indexes. According to the invention, the problem that the accuracy of the current camouflage target detection algorithm and model cannot reach the expectation when some camouflage targets in actual situations and real life are detected in the prior art is solved.

The application relates to a multi-AUV bearing-only perception cooperative hunting control method based on reinforcement learning, which comprises the following steps: multiple autonomous underwater vehicles interact in a hunting environment; each autonomous underwater vehicle acquires its own state and passive sonar bearing-only observation of a target at the current time, receives delayed communication information from a neighbor autonomous underwater vehicle, constructs observation and state input for strategy learning, and stores samples obtained through interaction into an experience replay pool; under a centralized training and decentralized execution (CTDE) framework, a multi-agent deep deterministic policy gradient network sharing parameters is constructed; sequence samples are sampled from the experience replay pool, and trained network parameters are obtained through joint optimization of each network; the trained network parameters are loaded, each autonomous underwater vehicle generates input for a strategy network based on its own state, passive sonar bearing-only observation and delayed communication information, and realizes multi-AUV cooperative hunting control based on output actions.

The invention discloses a power transmission line icing intelligent identification method and system based on multi-source data, and relates to the technical field of icing intelligent identification, and the method comprises the steps: estimating the rough sag of a lead, carrying out the differential interference processing of an SAR complex image, calculating a winding gradient, constructing a target function, and obtaining an optimal integer flow field; and calculating the bus density of the wire after icing, generating the icing weight on the wire, and converting the icing weight into the equivalent icing thickness. According to the method, the quantitative inversion capability of the continuous distribution icing thickness along the power transmission line is improved through the combination of the low-orbit satellite SAR differential interference deformation inversion result and the fiber grating strain and span data, and the wrapping gradient network flow optimization unwrapping method is combined with the expected vertical deformation field constructed based on the rough sag and atmospheric phase correction. And the robustness of interferometric phase unwrapping and the inversion precision of the deformation quantity are improved.

The invention provides a commercial vehicle double electric drive axle torque distribution model generation method and device, and belongs to the technical field of motor control, and the method comprises the steps: obtaining multi-dimensional working condition variable data at the current moment in the vehicle operation process; inputting the multi-dimensional working condition variable data into an Actor network of a depth deterministic strategy gradient network model to obtain a double-bridge torque distribution proportion; based on the torque distributed to the double electric drive axles of the vehicle, multi-dimensional working condition variable data and instant rewards at the next moment are obtained; constructing a training sample based on the multi-dimensional working condition variable data at the current moment and the next moment, the double-bridge torque distribution proportion and the instant reward; and training the depth deterministic strategy gradient network model based on the training sample to obtain a double-electric-drive-axle torque distribution model. According to the method, the technical problem that an existing double-electric-drive-axle torque distribution strategy is low in adaptive capacity and generalization capacity of a new scene can be solved.

This invention discloses a hybrid mathematical-analog unmanned swarm brain-like collaborative navigation method, comprising: Step 1: establishing a motion model and a perception model for the unmanned intelligent agent swarm; Step 2: constructing a navigation algorithm based on a mathematical model using a dynamic window algorithm; Step 3: constructing a dual-delay deep deterministic gradient network with temporal correlation by combining a long short-term memory network; Step 4: obtaining the coordinates of the nearest obstacles for each intelligent agent in the unmanned swarm and performing density clustering on the coordinate set; Step 5: calculating the obstacle density ρ based on a local information map. obs Combined with intelligent agent S i Step 6: Calculate the dynamic fusion weights based on the closest distance to the obstacle. i This invention employs a hybrid fusion of mathematical model-based methods and deep reinforcement learning to improve decision-making speed. It dynamically fuses the two navigation algorithms based on environmental complexity, further enhancing the performance of deep reinforcement learning-based navigation algorithms.

The invention relates to the technical field of energy transfer data analysis, and discloses a gate-water flow coupled vibration space-time energy transfer visual analysis system and method, and the system comprises an energy fingerprint extraction system, an energy potential gradient network construction system and a dynamic energy potential topology mapping system. The method comprises the steps that a space-time continuous energy density distribution field is generated through a space-frequency field reconstruction technology, non-uniform sampling data is converted into a scalar field with energy fingerprint characteristics, energy fingerprint codes are aggregated into energy infinitesimal elements with directivity, and a space-time energy infinitesimal element array is formed; generating an energy potential gradient network based on the dynamic coupling relationship, and identifying a main energy path and a branch energy path to form a hierarchical energy potential gradient network structure; and mapping the energy potential gradient network to an affine space-time coordinate system, and generating a dynamic energy potential topological curved surface according to path energy flux intensity and directivity. According to the invention, multi-parameter synchronous observation in a complex energy transfer process is realized.

The application discloses a method for generating nonlinear collaborative financial product alpha factors by using reinforcement learning, relates to the fields of financial quantification and reinforcement learning, and comprises the following steps: step 1, outputting a token of BEG, inputting a current environment state as a feature into a policy gradient network, and outputting a next token; step 2, updating the current environment state, inputting the current environment state into the policy gradient network again, and outputting a next token; step 3, repeating step 2 until a complete alpha factor formula is generated; step 4, placing the complete alpha factor formula into an alpha factor pool, assigning a random weight, inputting all factors in the alpha factor pool into a nonlinear neural network for regression training, and determining the weight of each alpha factor; step 5, composing a total alpha factor with a collaborative effect, and taking a mutual information coefficient as a reward function; step 6, updating the policy gradient network according to the reward function by using a Monte Carlo method; and repeating steps 1 to 5 until an effective prediction result is obtained.

This invention discloses a power-on / off timing control system for a motor test power supply based on data acquisition. The system includes a real-time data acquisition module, a timing logic closed-loop control module, a power drive execution module, a fault protection module, and a host computer interaction module. The real-time data acquisition module acquires electrical parameters and status data; the timing logic closed-loop control module abstracts the multi-channel power supply as graph nodes and the electrical coupling relationships between channels as edges, dynamically calibrating the power-on / off timing nodes of the multi-channel power supply and generating timing control commands by using a graph neural network and a deep deterministic policy gradient network to minimize the electrical impact penalty function; the power drive execution module manages the power supply on / off according to the commands; the fault protection module performs feedforward protection; and the host computer interaction module configures rules. This system transforms static open-loop timing into dynamic closed-loop timing, enabling timing nodes to match the actual electrical transition process and eliminating transient electrical impacts generated during power-on / off switching.

The invention discloses an intelligent drilling machine autonomous decision-making method based on reinforcement learning, and the principle of the method comprises the steps: firstly, precisely capturing the distance information of an obstacle from original radar data through a clustering algorithm, and effectively recognizing potential obstacles in the surrounding environment through the process; then, the collected sensing data is processed, more effective and representative feature input is extracted, and the accuracy and efficiency of subsequent decision model training are improved; on the basis of the feature input, a reward function related to the obstacle and the target point is designed, and the intelligent drilling machine can be guided to quickly and safely reach the preset target point in a complex and changeable environment; and finally, a double-delay depth deterministic strategy gradient network model is designed, and by adopting the advanced reinforcement learning algorithm and combining with manually designed feature input, it is ensured that drilling machines of different sizes show good generalization ability, so that the adaptability and practicability of the whole system are improved.

The application discloses a bionic gradient network composite structure resisting impact damage of external objects and a design method thereof, and belongs to the field of composite materials. In order to solve the problems of low composite degree, single function and lack of impact resistance design of part bearing structure of a carrier vehicle of an existing impact resistance structure, a new design scheme is provided. The bionic gradient network composite structure realizes gradient distribution of impact resistance performance through compounding of a skeleton structure and a filler phase, can quickly absorb impact energy of external objects when meeting the impact, and simultaneously realizes lightweight design. The design method comprises reading impact hot zone information from user input data, generating the bionic gradient network composite structure, and optimizing through weight analysis and service life analysis. The application can be widely applied to the fields of aerospace, vehicles and ships, and improves the impact resistance and service life reliability of a protective structure.

The invention discloses an evidence perception document level relation extraction method based on dynamic semantics and reinforcement learning, belongs to the technical field of natural language processing and artificial intelligence, and aims to solve the key technical problems of insufficient task interaction, poor static threshold adaptability, limited long document modeling capability, representation bottleneck and the like in traditional document level relation extraction. According to the method, a dynamic semantic enhancement module DynaSpan and a reinforcement learning strategy are fused, and collaborative optimization of relation classification and evidence prediction is realized by dynamically adjusting an evidence judgment threshold value; the DynaSpan module adopts a learnable parameter matrix to finely model entity interaction, and captures a complex semantic relationship; and the strategy gradient network dynamically optimizes an evidence retrieval threshold according to real-time feedback, so that the accuracy and flexibility of evidence screening are improved. According to the method, the performance and efficiency of relation extraction are remarkably enhanced, the method is particularly suitable for long text complex relation extraction scenes, an efficient and accurate solution is provided for the field of natural language processing, and the method has wide application prospects.

The application discloses a damping force closed-loop control method, device, medium and product, relates to the technical field of suspension control, and the method comprises the steps of constructing a data-driven model; obtaining a simulation estimated damping force based on characteristic data by using the data-driven model; obtaining a simulation expected damping force based on the characteristic data and a simulation expected force by using boundary constraints; obtaining a simulation control current based on the simulation expected damping force, the simulation estimated damping force and the characteristic data by using a deep deterministic policy gradient network; determining whether the characteristic data fed back by a simulation semi-active suspension system meets a set requirement, and then obtaining a force tracking controller; integrating the data-driven model, the boundary constraints and the force tracking controller into a force tracking module; obtaining a control current based on original characteristic data and an expected force by using the force tracking module; and the control current is used for controlling the semi-active suspension system. The application can improve the force tracking precision of the semi-active suspension system.

The application relates to the technical field of intelligent driving and discloses an automatic driving vehicle control method and system based on intelligent sensing of a driving environment. The method comprises the following steps: acquiring an RGB image sequence and a corresponding depth image sequence in front of a vehicle at multiple continuous moments; performing multi-modal feature extraction on the RGB image sequence and the depth image sequence; generating a bird's-eye view semantic feature; combining the spatio-temporal feature map and the bird's-eye view semantic feature map through a multi-modal multi-view decision state generation module, and performing time series modeling through a gated recurrent unit to output a latent feature vector; and inputting the latent feature vector as a current environment state into a deep deterministic policy gradient network to output a control instruction. The application can not only accurately identify and classify various obstacles on a road, but also can analyze and predict the motion trajectory of the obstacles in real time to support decision making and safety control of the system in a complex traffic scene.

The invention relates to a reinforcement learning-based multi-AUV pure orientation perception cooperative hunting control method, and the method comprises the steps: enabling a plurality of AUVs to interact in a hunting environment, obtaining the states of the AUVs at the current time, and carrying out the passive sonar pure orientation observation of a target, receiving delay communication information from a neighbor autonomous underwater vehicle, constructing observation and state input for strategy learning, and storing a sample obtained by interaction to an experience playback pool; constructing a parameter-shared multi-agent depth deterministic policy gradient network under a centralized training distributed execution CTDE framework; sampling sequence samples through an experience playback pool, and jointly optimizing each network to obtain trained network parameters; network parameters obtained through training are loaded, each autonomous underwater vehicle generates input used for a strategy network based on the state of the autonomous underwater vehicle, passive sonar pure orientation observation and delay communication information, and multi-AUV cooperative hunting control is achieved based on the output action.

This invention discloses a multi-UAV formation and obstacle avoidance control method based on safety reinforcement learning in low-altitude environments. This method designs a safety filter based on a control obstacle function to filter potentially unsafe or unreasonable actions of the UAVs during flight. Simultaneously, it combines parameter sharing technology and an observation embedding layer network to construct a max-pooling multi-agent deep deterministic policy gradient network to train the nominal policy for formation control. During the online execution phase, the max-pooling multi-agent deep deterministic policy gradient network extracts key features from the UAV's local observation information and uses the feature vector as network input, outputting a nominal action. The safety filter, based on this nominal action and the locally observed obstacle information, generates control commands for the UAVs that comply with safety constraints through a quadratic programming solution. This invention effectively improves the autonomous formation and obstacle avoidance capabilities of multi-UAV systems in unknown environments, enabling autonomous cooperative formation and obstacle avoidance flight of multi-UAV systems.

The present application relates to the technical field of unmanned aerial vehicle control, and particularly relates to a dynamic hovering anti-disturbance control method for unmanned aerial vehicles based on a neural network, comprising: constructing a time series convolution network as a disturbance observer to estimate total disturbance according to unmanned aerial vehicle state quantities and wind speed measurement values within a historical time window; constructing a deep deterministic policy gradient network as a controller to generate motor speed instructions according to state errors and total disturbance; jointly training the two networks through domain randomization in a simulation environment; deploying the trained networks to a flight controller, estimating total disturbance by the time series convolution network and generating speed instructions by the deep deterministic policy gradient network to drive the motor in real-time flight; and setting a safety constraint layer to perform amplitude limiting on the speed instructions or switch to a backup controller when state parameters exceed a threshold. The present application realizes high-precision hovering control under strong wind disturbance and guarantees safety and real-time performance.