746results about How to "Lighten the computational burden" patented technology

The invention discloses a method for realizing high-efficiency low-consumption micro-aerobic hydrolytic acidification of petrochemical wastewater by regulating and controlling the aeration rate in combination with a space-time diagram neural network and reinforced learning, and relates to a method for realizing high-efficiency low-consumption micro-aerobic hydrolytic acidification of petrochemical wastewater by regulating and controlling the aeration rate. The method comprises the following steps of: 1, collecting and cleaning data; 2, constructing a graph data set; 3, training the ST-GNN model; 4, constructing a deep Q reinforcement learning model; 5, the trained ST-GNN model is embedded into an automatic control system of a sewage treatment plant, and the model outputs the optimal aeration rate of the current time step and predicts the micro-aerobic hydrolytic acidification effect through interaction of the current working condition and historical data; and adjusting operation parameters of aeration equipment according to the output of the model, and carrying out hydrolytic acidification treatment. The method is high in regulation and control efficiency and low in operation error, and can be applied to the field of digital and intelligent management of sewage treatment plants.

The invention belongs to the technical field of power grid regulation and control, and particularly relates to a park-level adjustable load-oriented resource regulation and control method and system, and the method comprises the steps: formulating an initial energy scheduling scheme according to the predicted energy demand in combination with the output characteristics of distributed energy in a park and the load condition of a power grid; and training a load regulation and control strategy model by adopting a reinforcement learning algorithm based on the initial energy scheduling scheme, and generating a personalized load regulation and control instruction for each energy consumption unit. According to the scheme, the real-time energy consumption data of the plurality of energy consumption units in the park, the output characteristics of the distributed energy and the power grid load condition can be comprehensively considered, a scientific and reasonable energy scheduling and load regulation and control strategy is formulated, and efficient utilization and green low-carbon development of park energy are realized. Meanwhile, the method also needs to have the characteristics of data sharing, accurate prediction, intelligent scheduling, rapid response and the like so as to adapt to complex and changeable energy requirements and environmental protection requirements of the park.

The invention discloses an industrial equipment health monitoring and fault early warning system based on edge computing, belongs to the technical field of edge computing, and aims to solve the technical problem of how to efficiently realize health monitoring and fault early warning of equipment in real time and identify different types of faults. Each edge computing node is used for taking local data as input, predicting a fault category and a probability value corresponding to the fault category through a local model, outputting the fault category and the probability value as a prediction result, encrypting the local data and the prediction result through the cloud edge collaboration module and then sending the local data and the prediction result to the central server layer, the fusion model comprises three fault diagnosis models, namely a first fault model constructed based on LSTM, a second fault analysis model constructed based on a rule expert model and a third fault analysis model constructed based on a physical model; and the central server layer is used for fusing prediction results based on a self-adaptive weight mechanism.

The invention discloses an intelligent security and protection monitoring method and system based on image recognition. The method comprises the following steps: S1, acquiring multi-modal data; s2, performing preprocessing to generate complete target information; s3, performing real-time detection on the generated complete target information, and performing primary processing on a conventional scene; s4, based on an improved visual SLAM technology, obtaining three-dimensional information of multiple cameras; s5, when the edge device detects a potential abnormal condition, performing abnormal behavior analysis on the data by using a Swin Transform model, and optimizing a detection result in combination with the three-dimensional information of the multiple cameras; s6, performing adaptive learning on historical data by using a depth deterministic strategy gradient algorithm, and dynamically adjusting a threshold value of abnormal behavior detection; and S7, after an abnormal condition is detected, processing data through differential privacy protection and an end-to-end encryption technology, automatically controlling security and protection equipment, and giving a real-time alarm. According to the invention, image recognition and deep learning technologies are fully utilized, and accurate real-time monitoring in a complex scene is realized.

The invention relates to the field of electric meter box monitoring, and discloses an anti-electricity-stealing method and system for an intelligent electric meter box, which is used for solving the problem of false alarm caused by harmonic interference when anti-electricity-stealing early warning is carried out, and comprises the following steps: collecting time domain signals of current and voltage in the intelligent electric meter box in a historical detection time period; converting the time domain signal into a frequency domain signal, extracting harmonic information of each harmonic, and establishing a normal harmonic feature model to obtain a target harmonic range; the method comprises the following steps: acquiring time domain signals of current and voltage in an intelligent electric meter box in real time, extracting harmonic information, performing filtering processing according to a target harmonic range, performing first abnormality judgment by using an abnormality detection algorithm, performing multi-parameter joint evaluation if abnormality is judged for the first time, and performing second abnormality judgment according to a multi-parameter joint evaluation result. And electricity larceny early warning is carried out according to the second abnormal judgment result, so that false alarm and missing alarm are effectively reduced, and the adaptability of the system to a complex electricity larceny mode is remarkably enhanced.

The invention discloses a high-resolution remote sensing image change detection method based on multi-scale convolution decoding, electronic equipment and a storage medium, and belongs to the technical field of remote sensing image change detection. In order to meet the high-resolution remote sensing image change detection requirement, the method comprises the following steps: constructing a remote sensing image data set; a first stage and a second stage are formed based on a CNN module, a third stage and a fourth stage are formed based on a MambaVision module and a Transform module, and four dual-time-phase multi-scale feature maps are respectively obtained through processing of the four stages; constructing a time feature interaction module TFIM; a multi-scale convolution decoder EMCAD is constructed; and constructing a loss function, carrying out training, testing and evaluation by using the constructed remote sensing image data set, and applying the obtained trained optimal model to high-resolution remote sensing change detection. According to the invention, the detection capability of tiny changes in a high-resolution image is improved.

The invention provides an Internet of Things data processing method and system based on a block chain, and relates to the technical field of block chains, and the method comprises the steps that an Internet of Things device carries out the digital signature of an authentication request through employing a pre-stored private key, and generates a signed authentication request; the smart contract in the block chain network receives the signed authentication request, and the smart contract generates an authentication token including an authentication result, a validity period and an authorization level based on the historical authentication record, the credit score, the current timestamp and the challenge value; the Internet of Things equipment receives the encrypted authentication token and the intelligent contract address; the Internet of Things equipment configures own access authority and function limitation according to the authentication result, the validity period and the authorization level; and the Internet of Things equipment takes the intelligent contract address as an identity credential within the period of validity, and is used for subsequent secure communication and data exchange with other Internet of Things equipment or application programs.

The invention discloses a transformer substation inspection robot positioning and mapping method based on point-line feature fusion. According to the method, a binocular camera is used for collecting point-line features in an image frame, and high-quality long-line features are quickly obtained based on an adaptive criterion optimized EDLines algorithm; the spatial depth of point-line features is obtained through triangularization of a binocular camera, and redundant line features are filtered based on a length response threshold strategy of the spatial depth of a transformer substation. Based on a point-line pole feature pole plane constraint method, IMU information and visual information are combined, gyroscope zero offset is independently estimated and optimized, IMU pre-integration information is updated, and a more accurate initial posture of the inspection robot is obtained. A point-line feature dictionary based on a transformer substation scene and a large-scale equipment target detection model in a transformer substation are made, closed-loop detection is performed through transformer substation equipment road signs and point-line features, and rapid relocation correction is performed on sliding windows of a current key frame and a previous fixed key frame number. And finally, the pose of the substation inspection robot and a substation scene map are optimized through global BA, and the problem of large-scale scene error drift accumulation is solved.

The invention provides an Internet of Things gateway control method fusing local communication and AI computing power, and relates to the technical field of data processing, the method comprises the following steps: receiving an original data packet sent by at least one terminal device; protocol analysis, noise filtering and format conversion are carried out on the received data, an AI analysis model corresponding to the equipment type is loaded, standardized data are input into the AI analysis model, an analysis result is sent to target terminal equipment, and the analysis result and a communication result are stored in a gateway cache module; according to a preset time interval, the analysis result is transmitted to a remote cloud server, and synchronization state information and synchronization error log information are generated; according to the synchronization error log information, synchronization error analysis data is generated, network synchronization optimization and data anomaly positioning are executed, and a synchronization optimization result is generated; according to the data type and the error type, abnormal data are processed, and synchronous updating is executed; according to the invention, the data processing efficiency and accuracy are improved.

The invention discloses a self-adaptive event triggering safety control method of a multi-agent system under spoofing attack, and relates to event triggering control of the multi-agent system. A communication topological graph, an adjacent matrix and a Laplacian matrix of the system are defined based on a graph theory, it is assumed that communication topology between multi-agent systems is randomly switched, and the switching process follows a continuous time Markov process; assumptions on the communication topological graphs are relaxed, it is required that the combined graph of all possible communication topological graphs is undirected connected, and it is not required that each switching sub-graph is connected; modeling a spoofing attack, setting an attack signal upper bound, and defining a system security control objective function under the spoofing attack; an event-triggered distributed adaptive security control strategy is designed, the state of a neighbor agent is estimated by using an observer technology, it is verified that the control strategy can ensure the system security under the Markov random switching communication topology, and the effectiveness of the method is verified by using a Lyapunov stability theory, a stochastic analysis and linear matrix inequality method and simulation.

The invention discloses an unmanned aerial vehicle cluster intelligent formation method, device and equipment based on distributed reinforcement learning, and the method comprises the steps: constructing a reinforcement learning environment of an unmanned aerial vehicle cluster, which comprises a state space, an action space and a reward function, according to a flight task scene of an unmanned aerial vehicle, the flight mission scene comprises execution of air refueling and long-distance formation transportation; building a training framework of a distributed reinforcement learning algorithm based on the reinforcement learning environment, and performing network construction based on the training framework to obtain a neural network including an action decision network and a formation target evaluation network; performing distributed training on the neural network by adjusting hyper-parameters of the neural network and utilizing large-scale distributed data to obtain a formation model; and loading the formation model to an actual unmanned aerial vehicle cluster system, and adjusting a formation strategy according to a real-time flight environment and task requirements to carry out a formation flight task. The cooperative flight capability and the self-adaptability of the unmanned aerial vehicle cluster can be improved.

The invention discloses a water quality multi-parameter detection method based on an ultraviolet-visible absorption spectrum technology, which comprises the following steps: preparation before detection, water sample pretreatment, spectral feature extraction, measurement of the absorbance spectrum of a water sample by using an ultraviolet-visible spectrograph to obtain original spectral data; the method comprises the following steps: using a spectral feature extraction algorithm to obtain an average change rate S, adjusting the weight of spectral data, establishing a partial least square model, using a weighted spectral data set as input, combining the model and using a comprehensive prediction algorithm to obtain a comprehensive prediction value and a result report of water quality parameters. The method has the advantages that the influence of redundant features is reduced, the target water quality parameters are detected, the detection precision is improved, the calculation burden is reduced, the detection efficiency is improved, the comprehensive prediction value can utilize the strong terms of each model, the problems of over-fitting and under-fitting possibly occurring in a single model are reduced, and the prediction accuracy of the whole model is improved.

The invention discloses a power grid communication equipment fault identification method and system based on multi-source heterogeneous data characterization learning. The method comprises the following steps: firstly, collecting various complex multi-source heterogeneous data containing fault power grid communication equipment and preprocessing the complex multi-source heterogeneous data; extracting deep semantic features by using a pre-trained feature extraction network and obtaining a feature vector sequence; capturing time features of the feature vector sequence through a bidirectional long-short-term memory network, and obtaining complete sequence feature representation through a full connection layer; performing low-rank multi-modal fusion on the complete sequence feature representation in combination with cross-modal Transform to obtain three types of output splicing construction joint representation, and obtaining a fusion feature by using a self-attention mechanism; and finally, constructing a fault identification model by using the fusion features. Through innovative data preprocessing and data fusion technologies and in combination with an advanced deep learning model, fault-related characteristics can be automatically learned from multi-source heterogeneous data, and intelligent identification of the fault state of the power grid communication equipment is realized.

The invention discloses a fixed-wing unmanned aerial vehicle predetermined time cooperative path following formation control method, and relates to an unmanned aerial vehicle technology. A multi-unmanned aerial vehicle cooperative path following control strategy based on preset time stability, neural network adaptive control and preset performance control is provided; according to the method, the multiple fixed-wing unmanned aerial vehicles can realize preset time convergence of path following errors and preset time preset performance convergence of control signal tracking errors under the conditions of external disturbance and actuator faults, so that a high-precision and quick-response cooperative path following formation control task is completed; and meanwhile, the control precision and robustness are improved.

The invention provides a data access control method and device based on a block chain and improved CP-ABE, and belongs to the technical field of block chains. The method comprises the following steps: an authority generates key parameters including a system public key and a master key, and uploads the master key to a block chain; the user sends an attribute set and an identity label of the user to an authority, the identity label of the user is hidden, an attribute private key, a search private key, a conversion key and a recovery key are generated for the user, the conversion key is sent to the cloud server, and the attribute private key, the search private key and the recovery key are sent to the user, and the user comprises a data owner and a data visitor; the data owner carries out offline encryption, online encryption and index encryption on the file in sequence, and index information is linked; the data accessor performs data access; and the data accessor decrypts the data. According to the method, the problems in data storage and access control are effectively solved, the integrity of user data is protected, and the calculation burden of a data access user is reduced.

The invention discloses a method and device for training a medical large language model, and the method comprises the steps: obtaining a training medical data set, the training medical data set comprises a plurality of training prompt words, and each training prompt word comprises a training medical text; extracting medical text features of each training medical text; performing fine tuning on the reference large language model by using a low-rank adaptive method LoRA, determining a rank for a low-rank matrix according to the medical text features of each training medical text so as to obtain a first low-rank matrix and a second low-rank matrix, and fusing the first low-rank matrix and the second low-rank matrix with the reference large language model to obtain a fine tuning large language model; and training the fine-tuning large language model by using the training prompt word to obtain a trained medical large language model which is used for classifying the medical text. According to the method, the calculation burden during fine tuning of the large language model is reduced, and the generalization ability and adaptability of the large language model are improved.

The invention discloses a segmentation federal learning model training method based on a heterogeneous system. The system comprises a plurality of clients, a central server and an edge server. The global model segments the client local model and the server local model based on each client device condition and deploys to corresponding clients and central servers, respectively. And in each round of training, the client performs parallel training, forwards propagates local data, outputs crushed data, transmits the crushed data to the server local model of the central server, continuously forwards propagates to obtain a loss value, and backwards propagates a gradient to update parameters of the server local model and the client local model. And each client transmits the updated model to the edge server, and the edge server and the central server complete parameter exchange of a public layer and then perform model gathering to generate a global client model and a global server model. The global client model distributes client local models to the clients according to each client break point. The method can adapt to equipment conditions of heterogeneous equipment, and the model training efficiency is improved.

The invention provides a high-resolution positioning method for noise source imaging. The high-resolution positioning method comprises the following steps: designing a sensor array layout and selecting sensors; a multi-channel acquisition synchronous acquisition sensor and records signals to obtain acquisition data; performing signal processing on the acquired data by using adaptive filtering, machine learning calibration, time-frequency analysis and intelligent beam forming technologies to obtain standard data; based on the standard data, performing high-resolution direction estimation by using an optimized multi-signal classification algorithm, and removing a sidelobe effect by using an optimized CLEAN algorithm to complete processing of the standard data; and based on the processed standard data, reconstructing a spatial distribution image of the noise source by using an inversion technology, and dynamically adjusting the resolution of the reconstructed spatial distribution image by using an adaptive algorithm to realize high-resolution positioning of noise source imaging. According to the invention, high-resolution positioning can be realized, the reliability and precision of noise detection are improved, and noise control and management are more effective.

The invention belongs to the technical field of multi-robot path planning, and particularly relates to a consistent path planning method for multi-robot collaborative operation. Firstly, an environment map is converted into a grid map, and a starting point and a terminal point of each robot are set; then, calculating a priority score of each robot, layering all the robots according to the priority scores to obtain inter-layer priorities, and determining intra-layer priorities of the robots according to the priority scores; finally, according to the starting point and the ending point of the robot, a plurality of optional paths are generated for the robot through the Yen's algorithm, the motion paths of the robot are distributed according to the priority in the layer, and primary distribution of the motion paths is achieved; and according to the inter-layer priority, re-allocating the primarily allocated motion paths of the robots, and finely adjusting the primarily allocated motion paths to ensure that the motion paths of all the robots do not conflict. According to the method, a strategy of combining hierarchical path allocation, intra-layer priority ranking and inter-layer conflict detection is adopted, path conflicts can be effectively reduced, meanwhile, a new path generated through re-planning is continuation of a previous path or a subset of the previous path, continuity and consistency of path planning are guaranteed, and path changes are reduced.

The invention provides a traffic flow prediction method based on a space-time sequence, and relates to the technical field of traffic prediction, and the method comprises the steps: obtaining open-source network traffic flow data, and carrying out the preprocessing of the data, and obtaining a traffic flow sequence; inputting a traffic flow sequence into a data embedding layer to obtain high-dimensional representation of input features, performing periodic embedding and adaptive space-time embedding, and connecting the three kinds of embedding to obtain hidden space-time representation; inputting the hidden space-time representation into a space-time encoder for modeling of space dependence and time dependence to obtain a hidden state; and inputting the hidden state into an output layer, and outputting to obtain a traffic flow prediction result. According to the method, the deep features of the traffic flow data can be extracted more effectively, the complex spatial-temporal change rule of the traffic flow is captured, and the accuracy of urban traffic flow prediction is further improved.

The invention provides a stabilization control method and system for a fuel filling ship at a zero navigational speed. The method provided by the invention comprises the following steps: acquiring rolling data and wind wave data of a refueling ship under a target sea condition in a fuel refueling process in real time; the rolling data and the wind wave data are processed based on a radial basis function neural network, and the precession angular velocity of the stabilization gyro device under the target sea condition is obtained through calculation; the stabilization gyro device is controlled to rotate at the precession angular speed, the stabilization gyro device generates a gyroscopic effect and generates rolling moment, the rolling moment is converted into stabilizing moment for resisting rolling through precession motion, and the rolling angle of the filling ship is reduced; and the stabilizing moment is opposite to the vibration direction of the rolling. According to the method and system, rolling control of the filling ship at the low speed or the zero navigational speed is achieved, the stabilization torque can be adjusted in real time according to sea wave changes, and the stability and safety of the filling ship in the fuel filling operation process are improved.

The invention belongs to the field of power electronic power conversion, and provides a coupling type three-level inverter direct current side capacitance current ripple suppression method and system, and the method comprises the steps: determining a coupling type three-level inverter space vector diagram region division method according to a virtual voltage vector duty ratio distribution factor value range; according to the amplitude, the phase angle and the three-phase output current of the reference voltage vector, the sector and the area where the reference voltage vector is located are judged; according to the sector and the area where the reference voltage vector is located, selecting a virtual voltage vector and two basic voltage vectors which are closest to each other, and respectively calculating duty ratios of the virtual voltage vector and the two basic voltage vectors; designing a controller based on a dead-beat control method according to a sector and an area where a reference voltage vector is located, a direct-current side capacitor voltage deviation given value and a direct-current side capacitor voltage deviation actual value, and updating a small vector duty ratio distribution factor to realize independent control of capacitor voltage; and designing a switch sequence based on the updated small vector duty ratio and the sector and the area where the reference voltage vector is located.

The invention belongs to the related technical field of multi-agent control, and particularly discloses a predetermined performance consistency tracking control method and system for a nonlinear multi-agent system, and the method comprises the steps: constructing a Lyapunov function based on the consistency tracking error of the multi-agent system, and obtaining a control law; consistency tracking control of the multi-agent system is realized according to the control law; wherein the system state quantity in the control law is estimated through a distributed state observer; the distributed state observer is designed through the following method: according to a universal approximation principle, a nonlinear function in a multi-agent system dynamic model is approximated based on a fuzzy logic system, and then the distributed state observer is obtained by combining the multi-agent system dynamic model. According to the method, the precision and robustness of the consistency tracking process of the multi-agent system can be remarkably improved.

The invention relates to a clustering-based industry subdivision electricity sale quantity prediction method and device and a storage medium, which are applied to the technical field of electricity market transactions, and comprise the steps of obtaining representation vectors in a historical preset period of time of an electricity user, so that the electricity consumption behavior of the user can be represented more accurately, the accuracy of a prediction model is improved, and the user experience is improved. Meanwhile, vectorization representation is carried out on the attributes of the users, and the attributes are spliced with power consumption behavior representation vectors, so that more comprehensive user features can be provided through the integration, and the generalization ability of the model is enhanced; according to the method, high-dimensional original time sequence data is converted into low-dimensional feature vectors, so that the calculation burden of a clustering algorithm is reduced, the clustering efficiency is also improved, users are divided into different class clusters, modeling prediction is carried out for the electrical load of each group of users, clearer model interpretation can be provided, and the user experience is improved. And the electricity utilization characteristics of different user groups can be understood conveniently.

The invention provides a method for training an end-to-end automatic driving strategy, and the method comprises the following steps: S1, data collection and preprocessing, S2, defining a loss function and an optimizer, and constructing a model, S3, data division, S4, model training, S5, evaluation on a verification set, S6, model adjustment and optimization, and S7, test set evaluation. S8, weight pruning and evaluation: calculating and evaluating the importance of each weight by adopting a gradient norm calculation mode, evaluating the model after weight pruning by using the test set in the step S7 according to an importance evaluation result, starting from a relatively small proportion, gradually increasing and removing unimportant weights, and if the performance of the model does not reach the standard at the moment, evaluating the model after weight pruning by using the test set in the step S7. According to the method, through weight pruning, the applicability of a model system is wider, and the training efficiency of the system is higher.

The invention provides a bidirectional DC / DC converter parameter identification method based on physical information machine learning, and relates to the technical field of parameter identification. The method comprises the steps of firstly completing establishment of a system dynamic model of the bidirectional DC / DC converter, secondly utilizing an implicit Runge-Kutta method to construct a relationship between a data intermediate state and an initial state as well as a relationship between the data intermediate state and an end state, applying the relationship to the system dynamic model, then inputting historical training data into a deep neural network to obtain the intermediate state of the data, and finally establishing a system dynamic model of the bidirectional DC / DC converter. And substituting the physical model as an input to realize seamless coupling of the physical model into training of the deep neural network. And then the trained PINN is used to realize online parameter identification. The hybrid design can seamlessly combine the data of the BDDC physical model, and effectively modulate the training stage, so that the training stage is forced to be executed to the underlying physical principle and constraint of the converter, and the required training data is greatly reduced.

The invention relates to a multi-agent training method based on a heterogeneous dynamic graph attention mechanism, and the method comprises the steps: initializing agents, building a scene and a heterogeneous relation graph, building a deep reinforcement learning model based on the heterogeneous dynamic graph attention mechanism, carrying out the training, generating the aggregation features of the agents, and carrying out the multi-agent collaborative decision making through the trained model. And multi-agent control is realized. According to the method, a graph attention mechanism is used for aggregating the state feature vectors of the intelligent agents, the state feature vectors keep low dimensionality and contain related information of other intelligent agents, and the calculation burden in the subsequent strategy optimization process is reduced; a graph attention mechanism is subdivided into node-level attention and semantic-level attention, rich semantic relations between heterogeneous agents are captured, and more efficient collaboration or competition between the agents is achieved; a dynamic attention mechanism is introduced, the model is allowed to dynamically adjust the weight of neighbor nodes according to the characteristics of query nodes, the complex relation between agents can be better expressed, and the expression ability of the model is enhanced.

The invention discloses an unmanned aerial vehicle auxiliary operation state adjustment method and system based on a Markov chain, and belongs to the technical field of information processing, and the method comprises the steps: S1, constructing an initial experience tuple of the Markov chain based on historical trimming data of an unmanned aerial vehicle; training a reinforcement learning model through the initial experience tuple to determine an initial strategy set; s2, constructing an uncertainty model of a cable attribute and an environment space; determining a space-time adjustment factor according to the uncertainty model; s3, updating the initial state space and the initial action space according to the space-time adjustment factor to obtain a state updating space and an action updating space; s4, reconstructing the initial experience tuple according to the state update space and the action update space to obtain a strategy adjustment path, and obtaining a target strategy set according to the strategy adjustment path; the unmanned aerial vehicle responds to the target strategy set to execute a state adjustment strategy; the unmanned aerial vehicle can make an accurate and rapid response to a complex working environment, and the working efficiency of the unmanned aerial vehicle in the complex environment is remarkably improved.

The invention discloses a port multi-agent task allocation method and system based on reinforcement learning, and belongs to the field of logistics management, and the method comprises the steps: constructing a task set for various operation tasks in a port multi-task environment; aiming at different types of port intelligent equipment, constructing a dynamic model of the heterogeneous multi-agent; constructing a global environment model based on the task set and the dynamical model; constructing a collaborative multi-agent task allocation mechanism, and performing task allocation and scheduling by using the task allocation mechanism; and constructing a collaborative reinforcement learning model based on the task set, the dynamic model, the global environment model and the task allocation mechanism, and training a task execution strategy of the intelligent agent by using the collaborative reinforcement learning model. Through the hierarchical design of the high-layer strategy network and the low-layer strategy network, a complex task can be decomposed into a plurality of sub-tasks, management and optimization are facilitated, a multi-task environment can be effectively processed, and an intelligent agent can flexibly cope with different task sets.

The invention provides a mouth shape generation method and device for voice driving, and relates to the technical field of natural language processing. The method comprises the following steps: based on an ffmpeg tool, carrying out image processing according to original video data to obtain spliced frame image data and facial feature points; according to the original audio data, performing feature extraction through a deep audio feature extractor to obtain audio features; performing feature fusion through an audio-video sequence feature fusion device to obtain fusion features; performing video generation through a lip action generator according to the facial feature points and the fusion features to obtain synthesized video data; optimizing a lip action generator according to the original video data and the synthesized video data; and performing video generation based on the depth audio feature extractor, the audio-video sequence feature fusion device and the optimized lip action generator to obtain target synthesized video data. The invention relates to a mouth shape generation method aiming at voice driving, which is high in video resolution and sufficient in facial texture detail reservation.