92 results about "Empirical data" patented technology

According to the BIM-based construction process monitoring and early warning system provided by the invention, the data acquisition dimension and precision are remarkably improved through multi-source sensing fusion of millimeter-wave radar, multispectral imaging and voiceprint recognition; feature vector voxel units carrying material characteristics and process constraints are adopted, so that risk early warning has space-time relevance and process interpretability; through a composite risk calculation model containing environmental interference correction and time-varying gradient, the problem that a traditional threshold value method is poor in adaptability to complex working conditions is solved; the holographic early warning mechanism realizes upgrading from sound-light alarm to touch sense-three-dimensional projection cooperative interaction; the double-circulation self-optimization system not only guarantees real-time control, but also realizes block chain evidence storage of empirical data, and finally forms a construction monitoring closed-loop system with space-time perception, intelligent decision, accurate early warning and sustainable evolution capabilities.

The invention relates to the technical field of parameter optimization, in particular to an automatic kernel network parameter optimization method, which comprises the steps of constructing an enhanced deep Q network model, and integrating the enhanced deep Q network model with a priority playback buffer area, a meta learning module, a Bayesian optimizer and a neural architecture search module; using performance index data to train an enhanced deep Q network model, the training process including using a priority playback buffer to store and sample empirical data, using a meta-learning module to perform task adaptation, and monitoring training indexes of multiple dimensions to evaluate the convergence state of the model; selecting a kernel parameter adjustment action according to the current state through the trained enhanced deep Q network model; executing the selected kernel parameter adjustment action, and evaluating a parameter adjustment effect based on the multi-target reward function; and updating the enhanced deep Q network model according to an evaluation result, wherein the priority playback buffer area and the Bayesian optimizer are utilized in the updating process.

The embodiment of the invention provides an unmanned ship cooperative hunting control method, and belongs to the technical field of control, and the method specifically comprises the steps: distributing a main hunting target for each chasing unmanned ship; for each chasing unmanned ship, an original observation vector containing a variable-length interaction sequence is formed according to the distributed main hunting target of the chasing unmanned ship; a variable-length interaction sequence in the original observation vector is converted into a standardized state vector of a fixed dimension through a two-stage attention mechanism; inputting the standardized state vector into a strategy network of the unmanned ship, and outputting a continuous action control instruction; and calculating an instant reward based on empirical data obtained after executing an action control instruction, and performing iterative updating on a strategy neural network and a corresponding value neural network by adopting a centralized training and decentralized execution mode based on a multi-agent depth deterministic strategy gradient framework to obtain a control model to generate a cooperative control scheme. Through the scheme disclosed by the invention, the control precision and adaptability are improved.

The invention discloses a government affair material auditing method and system based on historical case analysis, and relates to the technical field of intelligent government affair platforms, and the method comprises the steps: obtaining archived historical case data and a to-be-audited target government affair material in a government affair auditing system, carrying out the structural modeling of the historical case data, and obtaining a to-be-audited target government affair material; forming a case database containing content features, audit conclusions and business categories, and realizing computable representation of empirical data; based on the multi-dimensional feature extraction of the target material, combining the reliability weight of the business category obtained by clustering, the similarity weight between the target material and the historical case and the importance weight of each auditing dimension; a multi-layer weighted voting mechanism is adopted, a material comprehensive auditing score is generated through fusion calculation integrating three types of weights, and a machine auditing conclusion is automatically output according to a score threshold value. Therefore, the conversion from artificial experience to data-driven decision is realized, and the accuracy, consistency and intelligent level of government affair material auditing are effectively improved.

The invention discloses a spacecraft Hall propulsion device test working medium flow stability control system, which comprises a sensing module for detecting working condition state parameters of working medium supply equipment, and an adaptation module for exploring current working condition control parameters by combining detection data of the sensing module and similar working condition control parameter empirical data fusion, the adjusting module is used for receiving the control parameters and dynamically adjusting the adjusting weight of the adjusting device according to different phase change periods, an experience library unit and a self-adaption unit are arranged in the adaption module, the experience library unit provides experience data of similar working conditions and control parameters, and the self-adaption unit provides self-adaption data of the control parameters. And the autonomous adaptation unit fuses and explores control parameters by constructing a neural network adaptation model, so that working condition automatic identification and control parameter autonomous adaptation under the environment that the working medium type or the working medium state rapidly changes are realized; the problems of low control precision and slow response speed caused by manual presetting of control parameters of different working media in a traditional control technology are solved, and stable control of the working medium flow is realized.

The invention discloses an intelligent load distribution system and method for a centrifugal compressor unit. The system comprises a data sensing module, an intelligent decision-making module, an instruction execution module and a model updating module. The method comprises the steps that operation parameters of multiple parallel centrifugal compressors are synchronously collected in a high-precision mode through an industrial real-time network supporting hardware timestamps, and time-aligned system state observation data are obtained; constructing a state vector based on the data and a load prediction result, inputting a deep reinforcement learning agent fused with compressor physical characteristic constraints, and generating a load distribution instruction and a unit start-stop suggestion; after safety boundary verification is conducted on the original action, the guide vane opening degree is adjusted through a feedforward speed planning and fuzzy self-adaption PID composite servo control algorithm, and starting and stopping operation is executed; meanwhile, empirical data are collected online, and model parameters are periodically and finely adjusted. According to the invention, efficient, safe, accurate and self-evolution compressor group control is realized.

The invention discloses a railway roadbed settlement prediction method and system, and belongs to the technical field of railway engineering. According to the method, basic data such as a roadbed structure, filler parameters, geological conditions and historical monitoring data are collected, a differential prediction model is matched according to roadbed strength grades, risk grading and monitoring resource dynamic allocation are carried out after parameter optimization and multi-index precision evaluation, and iterative optimization is realized based on an empirical database. The system comprises a data acquisition module, a model matching module, a settlement simulation module, a precision evaluation module, a post-processing module and a data storage module, and all the modules cooperate to achieve full-process automation. According to the method and the system, the problem of insufficient adaptability of a traditional single model is solved through accurate model adaptation, the prediction efficiency and the data reliability are improved by means of automatic processing, the safety and the cost are balanced through risk hierarchical management and control, and the adaptability of the system to complex working conditions is continuously enhanced through an iteration mechanism.

The invention discloses an alternating-current and direct-current hybrid power transmission network rack optimization evaluation method and system, which are used for solving the technical problems that a traditional alternating-current and direct-current hybrid power transmission network rack evaluation method depends on static parameters or empirical data, cannot dynamically respond to cross-regional load fluctuation and the like, is difficult to cope with risks, and causes low quality of a rack scheme. The method comprises the following steps: generating comprehensive evaluation indexes of a plurality of grid schemes according to acquired operation data of an AC / DC hybrid power transmission network by adopting power system simulation software; constructing a fuzzy evaluation matrix of each grid scheme according to the comprehensive evaluation index of each grid scheme by adopting an improved descending semi-trapezoidal membership function and a Gaussian membership function; calculating a target weight corresponding to the comprehensive evaluation index of each grid scheme; and according to the target weight corresponding to the comprehensive evaluation index of each grid scheme and the fuzzy evaluation matrix, calculating the comprehensive score of each grid scheme, and selecting the grid scheme with the maximum comprehensive score as the target grid scheme of the AC / DC hybrid power transmission network.

The invention discloses an intelligent agent optimization method and system based on task similarity clustering, and belongs to the technical field of intelligent power grids, and the method comprises the steps: constructing a simulation model of a power system according to a topological structure of the power system and power equipment data; constructing a Markov decision process model of the power system in combination with the simulation model and a power task scheduling target of an intelligent agent; based on the Markov decision process model, collecting historical operation data of the simulation model as an empirical data set, pre-training an intelligent agent through the empirical data set, outputting similarity among power tasks and grouping the power tasks to obtain a plurality of similar task groups; and performing grouping optimization on the pre-trained agents through the similar task groups to obtain a multi-task execution model. Therefore, by implementing the method and the device, the problem of poor model applicability caused by the fact that the constructed intelligent agent can only execute the electric power task in the specific electric power environment in the prior art can be solved.

The invention relates to a target-driven digital model orthogonal analysis and dual-adaptation fusion method and system, and the method comprises the steps: obtaining an overall demand, and combining the overall demand with empirical data and a knowledge graph to generate a model demand; obtaining a model composition file and a model description file which is determined based on a model demand and is used for dividing model categories, generating a meta-model description file according to the model composition file and the model description file, and generating a meta-model based on the meta-model description file; analyzing input and output data logic of the meta-model, scheduling parameters according to meta-model interface specifications, and redefining an interface by adopting a nonlinear iteration and aggregation optimization method; correlation adaptation and proxy packaging are carried out on a result of transmission logic numerical value fitting, and numerical value verification is completed in combination with a precision constraint condition; and adapting the incidence relation of the meta-agent model, generating a business-level numeralization agent model, scheduling parameters to perform precision verification on the business-level numeralization agent model, and generating a unified and operable system model.

The invention provides a resource adjustment method, a resource adjustment device and computing equipment. The resource comprises a reasoning instance group and a training instance group, the reasoning instance group is used for processing the first input data through the AI model to obtain response data, the training instance group is used for updating model parameters of the AI model according to empirical data, and the empirical data is obtained according to the response data and stored in the buffer. The method comprises the steps that first information and second information of a buffer are obtained, the first information is used for representing the load of a reasoning instance group, and the second information is used for representing the load of a training instance group; and adjusting configuration information of a target instance group according to the first information and the second information, wherein the target instance group comprises a reasoning instance group and / or a training instance group. According to the scheme, the load between the instance groups is sensed through the buffer, and the configuration information of the instance groups can be adjusted in time under the condition that the load between the instance groups is unbalanced, so that the load between the instance groups is balanced.

The invention discloses a method, a device and a system for planning an SEEG electrode implantation path. The method comprises the following steps: processing medical image data of the brain of an epileptic to obtain a target brain anatomical image; matching and searching a target experience module in an experience database according to the clinical description data of the epileptic; the experience database is established according to prior experience, a plurality of experience modules are arranged in the experience database, and each experience module comprises a plurality of combinations of a first target range and a second target range of a fixed electrode; and determining a target path of each electrode based on the target brain anatomy image and the target experience module. According to the method, the experience database containing a plurality of experience modules is established, the target experience module is screened in combination with the clinical description data of the patient, the target path of each electrode is further screened in combination with the target brain anatomy image of the patient, the proper path of the SEEG electrode is accurately and efficiently planned for the individual patient, and the accuracy of the SEEG electrode is improved. The method provided by the invention can reduce the learning and use threshold of doctors, and has high popularization value.

The invention provides a turning method of an intelligent car, the car and a control system. When a front-end drive enters a curve, navigation information is collected in real time, a front-drive turning instruction is generated, and meanwhile the front-drive turning instruction and the total mileage of front wheels serve as empirical data to be stored in a learning instruction queue. When the rear-end driver enters the same curve, the rear-end driver does not depend on data which is collected by a camera of the rear-end driver and possibly lags behind any more, empirical data corresponding to the front-end driver is taken out from the learning instruction queue to be reproduced, and collaborative smooth turning of the front-end driver and the rear-end driver is achieved. Through a path memory and reproduction mechanism, a front driving system and a rear driving system which are spatially separated are skillfully synchronized on a time axis, and simultaneous processing is changed into successive following processing, so that the inherent turning delay problem of a dual-camera system is solved.

The invention discloses a multi-agent learning method, device and equipment based on a dynamic weighted average field. The method comprises the following steps: selecting actions according to dynamic weighted average actions corresponding to agents; obtaining current position information of the intelligent agent, calculating an aggregation factor, and dynamically updating the weight of the intelligent agent by adopting an improved PSO algorithm in combination with the aggregation factor; updating a dynamic weighted average action according to the weight and the action, executing a joint action to obtain a next state, updating a reward according to the action and the state, and updating an individual weight optimal value and a group weight optimal value according to the weight and the reward; empirical data including the current state, the joint action, the reward, the next state and the dynamic weighted average action are sampled from the empirical playback buffer area, the Q function of the intelligent agent is updated, and the target network of the intelligent agent is updated according to the updated Q function and the specified learning rate. According to the invention, while multiple agents are guided to tend to group performance optimization, it is also ensured that the individual agents have sufficient exploration ability.

A system and method for determining a balance point of a building that has undergone or is about to undergo modifications (such as shell improvements) are provided. A balance point of the building before the modifications can be determined using empirical data. Total thermal conductivity of the building before and after the modifications is determined and compared. Indoor temperature of the building is obtained. The balance point temperature after the modifications can be determined using a result of the comparison, the temperature inside the building, and the pre-modification balance point temperature. Knowing post-modification balance point temperature allows power grid operators to take into account fuel consumption by that building when planning for power production and distribution. Knowing the post-improvement balance point temperature also provides owners of the building information on which they can base the decision whether to implement the improvements.

Control unit for estimating the traction of a vehicle, comprising the following: a monitoring module that monitors wheel torque values; a determination module for determining maximum torque values ​​by obtaining and storing empirical data while driving on a surface, wherein the maximum torque values ​​are the wheel torque values ​​immediately before wheel slippage; a prediction module that predicts a traction threshold based on the maximum torque values, where the traction threshold is based on a stochastic model of the maximum torque values; and a control module that outputs a signal indicating the traction threshold.

The invention provides an intelligent resource allocation method and system for a railway cloud platform, and the method comprises the steps: carrying out the preprocessing of resource index data, and obtaining a resource feature sequence; establishing a three-layer LSTM prediction model, defining a current moment resource feature sequence according to the resource feature sequence, and performing forward calculation in combination with the three-layer LSTM prediction model to obtain a future moment resource feature sequence; defining a historical moment resource feature sequence according to the resource feature sequence, and establishing a comprehensive system state vector based on the historical moment resource feature sequence, the current moment resource feature sequence and the future moment resource feature sequence; respectively establishing a double-DQN decision model and an empirical data tetrad, and performing optimization training on the double-DQN decision model through the empirical data tetrad; performing forward calculation based on the double-DQN decision model, and outputting an optimal resource allocation decision; and performing decision verification on the optimal resource allocation decision. The overall resource utilization rate of the railway cloud platform can be improved.

The present application relates to the field of transportation information engineering, and more particularly to a traffic hub crowd counting method and system based on multi-scale cross-region graph convolution driving, wherein the method comprises: constructing a crowd counting sequence; constructing three time series density map identification evaluation standards of spatial consistency, time sequence continuity and scale adaptability; proposing a multi-scale cross-region graph convolution network model to train and test the traffic hub crowd counting; and performing instance verification of the traffic hub crowd counting. The proposed model is applied to a series of empirical data sets for training and testing, which proves that it can achieve better demand prediction performance than the baseline model.

This invention belongs to the field of robot localization and relates to an adaptive hierarchical robot localization method and apparatus for degraded scenarios. The method includes: real-time monitoring of the data quality acquired by heterogeneous sensors, mapping the states of the heterogeneous sensors to a unified quantization space, and outputting unique degradation level identifiers D1 to D4; for the D1 environment, uncertainty quantification is performed on feature matching, and the most contributing feature subset is selected; for the D2 environment, the degraded subspace is located through singular value decomposition; for the D3 environment, the expected error of the heterogeneous sensor engine is estimated through a performance prediction network, and continuous confidence weighting is used instead of hard switching; for the D4 environment, a cross-platform generalized pure inertial odometry is provided through a three-level architecture of pre-training-fine-tuning-online adaptation; and historical events are stored as empirical data. This achieves refined differentiation and directional processing of degraded scenarios and constructs an experience-driven closed-loop optimization mechanism.

The invention is suitable for the technical field of testing, and provides an experience generation method and a test script verification method and device.The method comprises the steps that target experience features and logic attributes associated with the target experience features are determined based on selection operation in an experience input interface; generating empirical data based on the target empirical features and the logic attributes associated with the target empirical features, the empirical data being used for verifying a test script; wherein the experience input interface provides a first type of control and a second type of control, the first type of control provides a function of inputting a target experience feature, and the second type of control provides a function of selecting a logic attribute associated with the target experience feature from a plurality of candidate logic attributes; the logic attribute is used for representing a logic relation of the related target empirical features. Through the method, the accuracy of the generated empirical data is improved.

The invention relates to a rapid emergency decision-making method and system for an electric power sudden natural disaster, and belongs to the technical field of electric power disaster emergency decision-making. The method comprises the following steps: deducing disaster intensity by superposing early warning information and a real-time state of a power grid, obtaining a predicted fault probability set, obtaining empirical data by matching historical cases, and generating a defensive plan; feature extraction is carried out on multi-source heterogeneous monitoring data collected in real time, a diagnosed fault equipment set is obtained based on evidence fusion, and then the real-time situation of a power grid is constructed in combination with a CIM model; and performing plan primary selection on the defensive plan and the real-time situation of the power grid through multi-dimensional quantitative comparison, performing processing decision based on a matching result, and outputting an emergency disposal scheme after simulation verification. Based on digital twinning and multi-source data fusion, equipment-level disaster damage prediction, minute-level situation awareness and second-level scheme generation are realized, and the power grid disaster prevention and rapid recovery capability is continuously improved through adaptive optimization.

This invention discloses a method for link adaptation in retransmission systems and mobile environments. The method includes: constructing and initializing a reinforcement learning neural network; acquiring information, including current user feedback on channel conditions and HARQ process information; monitoring whether the channel environment has switched; constructing an observation state vector based on the information; inputting the observation state vector into the reinforcement learning neural network and outputting the MCS corresponding to the current HARQ process; transmitting data packets containing the MCS to the user and receiving user feedback; determining whether the current HARQ process has ended based on the feedback information; if so, calculating the retransmission throughput reward of the current HARQ process, aligning and integrating the state, action, and reward, storing it in an experience pool, and then proceeding to subsequent steps; training the reinforcement learning neural network based on gradient descent using several empirical data points. This invention solves the problem that traditional algorithms cannot adapt to actual 5G networks and complex, ever-changing environments.

The invention relates to recommending matches using machine learning. Systems and methods for recommending matches between persons are provided. Data processing is performed using an artificial intelligence technology. A supervised machine learning engine is trained from empirical data about existing relationships that have been evaluated about the quality of the relationship. When input data of attributes of two candidates is provided, a quality of the candidate relationship is calculated as an output of a supervised machine learning engine. A likelihood of a successful relationship between the two candidates is predicted by comparing a quality of the calculated candidate relationship with a threshold. Predictions in the learning task may be made through a neural network. The user is notified of candidate matches that may become successful relationships.

The invention discloses a communication module instruction fuzzy test method and system. The method comprises the steps of analyzing a historical communication module instruction, determining an instruction parameter type, and determining test record data based on historical fuzzy test data. And determining variation experience data of the instruction parameter type according to the test record data. And on the basis of the variation experience data, constructing a variator scheduling model corresponding to each instruction parameter type. And based on the previously constructed variator scheduling model, predicting a target variator of each instruction parameter type in the to-be-tested instruction, and executing a fuzzy test through the target variator. According to the technical scheme provided by the invention, the hit rate and convergence rate of fuzzy testing are remarkably improved by reusing historical experience, intelligent scheduling of variation strategies is realized, the testing resources are enabled to act on variation combinations which are easier to trigger abnormal states in a centralized manner, the exploration capability of deep states and potential problems of the communication module is enhanced, and the development efficiency of the communication module is improved. And the test efficiency and precision of the fuzzy test process are obviously improved.

A method for maximizing the selectivity (S) of an epoxidation catalyst in an ethylene oxide reactor system, the method comprising: receiving measured reactor selectivity (S) from an ethylene oxide production system. meas ), measured reactor temperature (T) meas The ethylene oxide production system is configured to convert a feed gas containing ethylene and oxygen into ethylene oxide in the presence of the epoxidation catalyst and a chloride-containing catalyst moderator in the ethylene oxide reactor system, along with one or more operating parameters. The epoxidation catalyst contains silver and a promoting amount of rhenium (Re), and the reactor selectivity (S) is measured. meas The measured reactor temperature (T) meas The method also includes real-time and historical operating data points generated over time by the ethylene oxide production system, and the one or more operating parameters include data points generated over time by the ethylene oxide production system. The method further includes using a processor to perform the following steps: (a) for each time point, using a model to calculate the optimal moderating agent level (M) for the epoxidation catalyst. opt The model estimation selectivity (Sest) and model estimation temperature (Test) are considered. The model estimation selectivity (Sest) is... est The model-estimated temperature (Test) is determined based on at least one of one or more operating parameters at the time point, the at least one operating parameter excluding the moderating agent level, and the model is based at least in part on historical empirical data associated with the epoxidation catalyst, the ethylene oxide production system, or both. The method also includes using the processor to perform the following steps: (b) for each time point, determining the measured reactor selectivity (S). meas ) and the selectivity estimated by the model (S) est The difference (ΔS) between the measured reactor temperature (T) and the measured reactor temperature (T) meas (c) The difference (ΔT) between the temperature (Test) estimated by the model and the temperature (Test); (d) Fitting the curve to the Δselectivity (ΔS) data points as a function of the corresponding ΔT data points to obtain the fitted curve; (e) Based on the fitted curve and the real-time value of ΔS (ΔS real‑time The real-time values ​​of ΔT and ΔT real‑time Determine the real-time relative effective moderating agent level (RCl). eff real‑time (e) based on this real-time RCL eff The method also includes using a processor to perform the following step: (f) displaying the executable suggestion on a display.

The present application relates to the field of transportation environment engineering and carbon emission assessment, and particularly relates to a method for constructing a highland gradient zone vehicle CO2 emission factor calculation model, equipment and medium. Through the method, a route-level CO2 emission factor distribution map is obtained, and different quantitative calculation models are formed for different altitude intervals. Specific and quantitative design basis is provided for low-carbon highway design. Through empirical data, it is first revealed that the CO2 emission rate of the highland gradient zone has a significant altitude segmentation effect, and a method of constructing a calculation model by zones is innovatively proposed. The method overcomes the defects of poor adaptability and inaccurate prediction of traditional single models in complex highland environments, and by establishing differentiated quantitative models (such as a low-altitude quadratic function and a high-altitude S-type function) for different altitude intervals, the final zoned model provides a core theoretical tool and decision basis for accurate accounting of highland highway carbon emissions, optimization of low-carbon routes and green transportation construction.

The invention discloses a space-based observation configuration optimization method and system based on GDOP optimization. The method comprises the steps that an observation satellite obtains observation information of a target; screening a satellite set capable of participating in observation; enumerating all main and auxiliary satellite combinations meeting task requirements; introducing reinforcement learning, and screening out candidate combinations with high potential; calculating the GDOP value of each candidate combination in parallel in real time; judging and switching the optimal combination; and recording all decision data, efficiency indexes and energy consumption states of the current combination as empirical data. The system comprises a target sensing module, a satellite visibility judgment module, a dynamic resource allocation evaluation module, a main and auxiliary satellite combination enumeration module, a reinforcement learning screening module, a GDOP real-time evaluation module, a configuration switching decision module and a data storage module. According to the invention, accurate real-time positioning can be realized under the condition of dynamic change of the target. The method can be widely applied to the fields of satellite navigation and target observation.

The invention provides an intersection overflow autonomous management control method based on a large model, and the method comprises the steps: building a basic database, a traffic management knowledge base and a police management and control strategy base, determining the mutual association relation, and estimating whether the next phase is likely to overflow or not according to the overflow conditions of three historical signal periods of signal control data. If an overflow event does not occur in three historical signal periods, it is considered that overflow does not occur in the next phase, otherwise, overflow occurs in the next phase, and whether the phase release lane leads to an overflow exit lane is judged according to the allowing lane and the overflow exit lane ID, if yes, the decision moment is determined, and if not, the decision moment is determined; and constructing prompt words for the next phase based on the database, recording an intersection state, training an overflow management and control large model, and realizing intersection overflow autonomous management and control by using the overflow management and control large model. According to the method, the overflow control large model is trained in combination with empirical data such as traffic control experience, traffic control rules and control strategies, and the response speed for the intersection overflow problem is increased.

A quantum dot auto-annotator system includes a processor and a non-transitory computer-readable medium. Stored on the medium is a data structure for a binarized threshold map representing charge transitions in a multi-dimensional parameter space of a quantum dot device. A model-building module contains logic for generating a plurality of polygonal models from the binarized threshold map, where each polygonal model corresponds to a polytopal domain. The medium further includes a statistical inferencing module with logic for clustering the polygonal models into one or more orientation-based domains based on geometric orientations of the plurality of polygonal models. A global state determination module then executes logic for assigning a probabilistic state vector to pixel locations within the orientation-based domains to generate an annotated charge stability diagram.

The invention discloses a traffic signal control method based on regional hierarchical multi-agent reinforcement learning, and belongs to the technical field of intelligent traffic systems. According to the method, a target traffic network is divided into a plurality of areas, each area is configured with a management layer agent, each intersection in the area is configured with a control layer agent, and a double-layer layered architecture is constructed; the management layer intelligent agent periodically generates a regional control target based on regional joint state observation, and the target remains unchanged in a plurality of bottom layer control periods; the control layer intelligent agent collects local state observation in each control period, receives the regional control target, dynamically corrects the target based on a local congestion state, and outputs a signal lamp phase control action; the control layer and the management layer respectively calculate level rewards, empirical data are stored in an empirical playback buffer area, and an offline strategy training mode is combined with a double-Q network, Huber loss, delay update and an action entropy regularization mechanism to collaboratively optimize two layers of strategy functions.