A Terminal Interoperation and Hazard Response System Based on Urban Agglomeration Smart Sensing

Abstract

The present invention designs a terminal interoperability and hazard response system based on urban group intelligence perception, which includes an efficient interconnection module for heterogeneous terminals, a data aggregation and collaborative processing prototype system, and a hazard situation response system. Among them, the efficient interconnection module for heterogeneous terminals is a multi-mode heterogeneous terminal efficient interconnection mechanism with low coupling between hardware control and data interaction. The data aggregation and collaborative processing prototype system includes two independent control systems for the central node and the edge node. At the same time, a multi-agent collaborative algorithm is designed based on reinforcement learning to coordinate and control the heterogeneous terminals in the whole system to work collaboratively. To solve the problem of difficult interaction of large-scale smart city terminals at the present stage, a distributed terminal interoperability mechanism is proposed to ensure the efficient and reliable transmission and interconnection of data between heterogeneous terminals, and at the same time improve the crisis response ability of the urban system.

Description

Technical Field

[0001] The present invention relates to the field of data collaboration and interoperability technology, and in particular to a terminal interoperability and hazard response system based on urban group intelligence perception. Background Art

[0002] In the scenario of precise management of smart cities, terminals are key devices that connect the sensor network layer and the transmission network layer in the Internet of Things, and are important tools for collecting, transmitting and processing data.

[0003] At present, most of the terminal devices in the market are single-function terminals, such as image transmission service terminals in automobile monitoring, power monitoring terminals, and logistics RFID terminals, etc. The types and functions of current IoT terminal devices are significantly differentiated, the data types and communication protocols used by different terminal devices are difficult to communicate with each other, and different scenarios have different requirements for terminal devices.

[0004] To achieve accurate and efficient smart city management, a large number of different types of terminals need to be efficiently interconnected. However, a survey of mobile sensing terminal intercommunication platforms shows that due to the fragmentation of the software and hardware of the above-mentioned terminal devices, the existing mainstream platforms often only support a specific set of mobile sensing terminals. In the smart city precision management scenario, there is still a lack of multi-terminal interconnection mechanisms that can effectively be compatible with heterogeneous single-function terminals and general intelligent terminals. The fragmentation of terminal devices has seriously hindered the advancement of smart cities and the development of the Internet of Things. In addition, the fragmentation of terminal devices has also affected joint control and collective decision-making. Therefore, it is imperative to design a universal and efficient interconnection mechanism and hazard response system that is compatible with multi-mode heterogeneous terminal devices. Summary of the invention

[0005] Technical problem: Based on the above problems and requirements, the present invention designs a multi-mode heterogeneous terminal efficient interconnection and hazard response system with low coupling between hardware control and data interaction. First, by analyzing the functions and characteristics of common terminals on the market, the present invention depicts an abstract model of edge terminals, designs an abstract hardware layer to separate from upper-layer control and data interaction, extracts terminal hardware features and abstracts them into virtual intelligent agents with the capabilities of movement, perception, computing, and communication, and compatible with and extends different terminal hardware devices through different instantiations of the virtual intelligent agents. Specifically, taking advantage of the strong versatility and high interface compatibility of general intelligent terminals, the present invention binds single-function perception terminals without wireless communication capabilities to general intelligent terminals, endows them with basic communication capabilities, and abstracts them into a virtual intelligent agent to achieve control and interoperability of the terminals. In addition, in view of the characteristics of data interaction of intelligent terminals, the present invention designs a unified data interaction mechanism with good compatibility and scalability. At the same time, aiming at the serious problem of heterogeneous communication protocols in the current Internet of Things, the present invention designs a general data interaction mechanism based on the TCP / IP protocol, making it independent of specific underlying communication protocols, with high flexibility, easy to maintain and implement. This data interaction mechanism adopts a hybrid architecture. In the presence of a control center, it allows terminal devices to interoperate with other terminal devices, which can reduce the communication pressure and computing pressure on the center. In terms of design, the present invention also introduces a priority mechanism, enabling terminal devices to independently control and process control requests and data requests from other terminals. For hazard response, the present invention defines the problem as a multi-agent collaborative game problem, introduces a deep reinforcement learning method for solution, and optimizes the algorithm for the suitable scenario.

[0006] Technical solution: The present invention designs a terminal interoperability and hazard response system based on urban group intelligence perception, which is characterized by including an efficient interconnection module for heterogeneous terminals, a data aggregation and collaborative processing prototype system, and a hazard situation response system.

[0007] By analyzing the functions and characteristics of common terminals on the market, the present invention depicts an abstract model of edge terminals, designs an abstract hardware layer to separate from upper-layer control and data interaction, extracts terminal hardware features and abstracts them into virtual intelligent agents with the capabilities of movement, perception, computing, and communication, and compatible with and extends different terminal hardware devices through different instantiations of the virtual intelligent agents.

[0008] In view of the characteristics of data interaction of intelligent terminals, the present invention designs a unified data interaction mechanism with good compatibility and scalability.

[0009] Aiming at the serious problem of heterogeneous communication protocols in the current Internet of Things, the present invention designs a general data interaction mechanism based on the TCP / IP protocol, making it independent of specific underlying communication protocols, with high flexibility, easy to maintain and implement

[0010] This data interaction mechanism adopts a hybrid architecture. In the presence of a control center, it allows terminal devices to interoperate with other terminal devices, which can reduce the communication pressure and computing pressure on the center.

[0011] A priority mechanism is introduced, enabling terminal devices to autonomously control and process control requests and data requests from other terminals.

[0012] In view of the differences among different terminal devices, a hierarchical architecture based on agents and an information interaction specification across platforms and languages are designed. By extracting the common features and common patterns of different types and functions of terminals, terminal devices are abstracted into virtual agents with the capabilities of movement, perception, and computing. Different instantiations of the virtual agents are used to be compatible with the control methods of different terminal hardware devices, achieving low coupling between the control of terminal hardware devices and the system function design.

[0013] A hybrid architecture of "centralized management - distributed interoperability" and modular design are adopted to interconnect the central node and edge nodes. The central node consists of a terminal proxy module, an instruction processing module, a task model management module, an interconnection communication module, a mobile collaborative perception module, and a permission management module.

[0014] The main function of the terminal proxy module of the central node is to provide an interface for terminal control for the central node. In order to be compatible with the underlying control methods of different hardware mobile sensing terminals, according to the unified coding protocol for status control of the terminal control module, the terminal proxy module encodes the terminal control information into a JSON packet and sends it to the corresponding terminal through the interconnection communication module to achieve remote control and status monitoring of the terminal.

[0015] The main function of the instruction processing module of the central node is to parse and execute the instructions submitted by the interconnection communication module and the task model management module to achieve control of the central node and terminal nodes. The specific workflow of this module is mainly divided into two parts: instruction fetching and storage and instruction execution.

[0016] When the module fetches instructions, the instruction processing module continuously reads the control messages in the instruction queue and submits them to the execution module for parsing. When the module stores instructions, it first stores the instructions in the instruction queue, where the instruction queue is a first-in-first-out queue with priority control. The priority of the instruction can be determined when storing the instruction, or specified in the instruction word. By default, the storage module presets the priority according to the instruction type.

[0017] When executing instructions, the module decodes the message to obtain the control instruction according to the communication protocol defined by the interconnection communication module. And according to the different instructions, it calls the agents of different modules to execute the instructions to control and respond to terminal devices.

[0018] The task model management module of the central node allows users to customize tasks and models and provides an API interface for deploying these tasks to each terminal device or the central node. Specifically, the task model management module is divided into two parts: model distribution and model control agent.

[0019] In the model distribution part, the task model management module allows the central node to distribute user-customized tasks and models to the target edge nodes to achieve intelligent control of the edge nodes. The model distribution process is implemented by the task model management module submitting corresponding instructions to the interconnection communication module. The model distribution and deployment process is disassembled into multiple file transfer instructions and model execution instructions, which are sent to the corresponding nodes through the interconnection communication module to control the model distribution process.

[0020] In the model control agent part, the task model management module is responsible for monitoring and managing the models deployed on the edge nodes and providing API services for upper-layer applications. The API services encapsulate the instruction details of the corresponding functions (but hide some instructions from developers to avoid incorrect calls). It is mainly implemented by calling the terminal agent module to submit corresponding instructions to the interconnection control center. Developers can send partial control instructions to the target nodes by calling the interconnection communication module to meet the implementation needs of collaborative tasks. The main function of the interconnection communication module of the central node is to enable communication between the central node and multiple terminal devices. It is mainly based on the TCP protocol, listens for the information sent by the terminal devices, and classifies the information into two categories for interaction and processing.

[0021] The interconnection communication module defines the encoding method for communication between the central node and the terminals and designs a communication protocol for the communication between the central node and the terminals. Each data packet in the protocol consists of a header and a main body. The header is assigned a field to define six different packet types, and the main body of the packet is responsible for transmitting the main information in the message, such as specific control commands, status information, priority, etc. For the convenience of function and device expansion, the packet types and contents in the protocol can be customized and added by developers. After receiving a message from the terminal, the central node will simply classify the message according to the protocol, submit the status message to the terminal agent module to update the status information of the terminal device, and if it is a control message, it will be pushed into the instruction queue of the instruction processing module.

[0022] To achieve efficient data interaction, the data sending and receiving process is controlled by the instruction processing module but is separated from the status control communication process. The data exchange is established through a four-way handshake. The sender first sends the connection information for data transmission. After receiving it, the receiver attempts to establish a connection and send an ACK message. After receiving the ACK message, the sender sends a data INFO message, including the data size, MD5 checksum, etc. After the receiver is ready, it sends an ACK message again to start the data sending and receiving process. The end of the sending and receiving process automatically ends according to the size of the received data, and the validity of the data sending and receiving process is ensured by calculating the MD5 checksum.

[0023] The mobile collaborative sensing module of the central node provides collaborative sensing services through collaborative algorithms.

[0024] The permission management module of the central node provides permission management services and security authentication services for edge nodes.

[0025] Permission management service. Before an edge node attempts to control the terminal device of another node, it needs to first obtain the corresponding permission. The specific process is as follows: The edge node submits a request instruction for obtaining the permission of the target node to the center. After receiving the request, the permission management module determines whether it is legal. After passing the verification, the center sends an instruction to the target node, allowing the target node to accept the control instruction from the requesting node, realizing secure interoperability between edge nodes.

[0026] Security authentication service. The central node maintains a list of legitimate edge nodes. When a node accesses, the permission management module determines whether it is legal. When a new node is added, it needs to be written into the list of legitimate nodes.

[0027] The implementation of the edge node is composed of four parts: a terminal control proxy module, an instruction processing module, a model control proxy module, and an interconnection communication module. Among them, the instruction processing module and the interconnection communication module have the same design structure as the central node.

[0028] The main function of the terminal control module of the edge node is to provide an interface for controlling hardware devices. After researching the functions and characteristics of major mobile sensing terminal devices, the terminal control module abstracts the mobile sensing terminal devices into proxy agents with mobility, sensing, and computing capabilities.

[0029] The model control proxy module of the edge node provides an API interface for controlling terminal devices for locally user-defined scripts. The main services provided by the model control proxy module are model control services and communication interaction services.

[0030] The model control service allows the instruction processing module to call and run locally user-defined scripts, or receive and run tasks and models deployed remotely by the center. At the same time, it provides a list of currently running models, allowing the instruction processing module to control and close a specific model or script.

[0031] The communication interaction service allows user scripts to communicate and interact with the local communication interconnection module. It is based on the local SOCKET connection and provides communication interaction support for the terminal control service. The terminal control service provides part of the interface API for controlling terminal devices to user scripts. When customizing scripts, users need to instantiate the proxy module of the terminal control service and access or control the terminal hardware devices through the proxy to perform tasks such as data collection. For the convenience of controlling and managing terminal devices, when implementing the terminal control service, protocol packets are sent from the local to the terminal communication interconnection module, and the communication interconnection module and instruction execution module of the terminal implement the control of the terminal devices by user scripts.

[0032] The application service module is based on the basic function module, is user-oriented, and provides API interfaces for upper-layer applications. The application service module is mainly divided into three parts: model distribution and deployment, multi-source data aggregation, and terminal collaborative control. In addition, service expansion can also be performed based on the basic function module according to application needs.

[0033] Model distribution and deployment. Model distribution and deployment is a further encapsulation of the task model management module in the basic function module, providing advanced API interfaces for model deployment to users. The implementation of the model distribution and deployment service is divided into several steps. First, the central node distributes the defined tasks and the required data processing models to each terminal device for deployment through the central task model management module. The terminal devices run the received scripts and perform tasks such as perception collection and data processing.

[0034] The data aggregation service is a further encapsulation of the data transmission of each sensing terminal device. It provides advanced API interfaces for calling data from multiple sensors, allowing user programs to initiate data call requests to multiple sensing terminals simultaneously, and aggregating multi-source data through the communication interconnection module and model distribution module in the basic function module to support the execution of tasks by upper-layer applications.

[0035] The terminal collaborative control module is a further encapsulation of the basic control functions of multiple terminal devices. It provides advanced API interfaces for multi-terminal collaborative control to user applications. This function is implemented based on the terminal control module and terminal proxy module in the basic function module, allowing user programs to call multiple terminals to cooperate in executing a specific task simultaneously.

[0036] In the face of abnormal situations, the sensors are abstracted as independent agents, and the entire hazard response problem is transformed into a multi-agent collaborative game problem.

[0037] A unique multi-agent reinforcement learning algorithm is designed to solve this collaborative game problem. Each terminal agent has its own policy network, while the Q-function network is shared by all agents.

[0038] Each agent has its own observation o i and action a i , where the observation has different manifestations according to the nature of the terminal agent. Concatenate the observation o i and action a i of each agent, then pass the concatenated result through a one-layer MLP respectively, and calculate the state e i = g i (o i , a i ).

[0039] Each agent can query the observations o = (o1,... o n ) and actions a = (a1,... a n ) of other agents, and incorporate this information into its own state and action estimates, from which the state set e = (e1,... e n ) of the agent cluster can be obtained.

[0040] For the state-action value function of agent i as shown in Equation (1), where f i is encoded using two layers of fully connected layers, and the intermediate layer variables of the fully connected layers are calculated through a multi-relational head as shown in Equation (2).

[0041] Among them, v j is the value vector of e j , which is obtained by linear transformation using the shared matrix W v . h is a non-linear activation function (Leaky ReLU), and r j l represents the l+1-ary relational weight.

[0042]

[0043]

[0044] The advantage function of multi-agent is based on a baseline design. By comparing the value function of a specified action with the value function of the average action, it is judged whether the increase in the return obtained by the current action is attributed to the actions of other agents.

[0045] Considering the different terminal functions, our baseline does not need to assume that each agent has the same action space, and does not require a global reward.

[0046] Beneficial effects:

[0047] (1) The design abstracts the implementation of the hardware layer to separate it from upper-layer control and data interaction, extracts the characteristics of terminal hardware, and abstracts them into virtual intelligent agents with the capabilities of movement, perception, computing, and communication. Different terminal hardware devices can be compatible and extended through different instantiations of the virtual intelligent agents. (2) In addition, the present invention designs a unified data interaction mechanism with good compatibility and scalability for the characteristics of data interaction of intelligent terminals. (3) At the same time, aiming at the serious problem of heterogeneous communication protocols in the current Internet of Things, the present invention designs a general data interaction mechanism based on the TCP / IP protocol, making it independent of specific underlying communication protocols, with high flexibility, easy to maintain and implement. This data interaction mechanism adopts a hybrid architecture. In the presence of a control center, it allows terminal devices to interoperate with other terminal devices, which can reduce the communication pressure and computing pressure on the center. (4) A multi-terminal intelligent agent collaborative control algorithm is designed, which can respond more systematically and quickly to abnormal situations and make optimal decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is the architecture diagram of the entire system.

[0049] Figure 2 is an example diagram of various terminals.

[0050] Figure 3 is a display diagram of a robot for guiding the flow of people in the subway.

[0051] Figure 4 is the Q-function calculation model of the designed multi-intelligent agents. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. The following drawings and embodiments are used to illustrate the present invention, but not to limit the scope of the present invention.

[0053] As Figure 1 shown, the present invention provides a terminal interoperability and hazard response system based on urban group intelligent perception, which is characterized by including an efficient interconnection and intercommunication module for heterogeneous terminals and a data aggregation and collaborative processing prototype system.

[0054] By analyzing the functions and characteristics of common terminals on the market, characterizing the abstract model of edge terminals, designing the abstract hardware layer to separate the implementation from upper-layer control and data interaction, extracting the characteristics of terminal hardware, and abstracting them into virtual intelligent agents with the capabilities of movement, perception, computing, and communication. Different terminal hardware devices can be compatible and extended through different instantiations of the virtual intelligent agents.

[0055] As Figure 1As shown in the figure, the data aggregation and collaborative processing prototype system provided by the present invention is characterized by adopting a hybrid architecture of "centralized management - distributed interoperability" and modular design, and interconnecting the central node and the edge nodes.

[0056] As Figure 1 shown in the figure, the central node consists of a terminal proxy module, an instruction processing module, a task model management module, an interconnection communication module, a mobile collaborative sensing module, and a permission management module.

[0057] As Figure 1 shown in the figure, the main function of the terminal proxy module of the central node is to provide an interface for terminal control for the central node. In order to be compatible with the underlying control methods of different hardware mobile sensing terminals, according to the unified coding protocol for the status control of the terminal control module, the terminal proxy module encodes the terminal control information into a JSON packet and sends it to the corresponding terminal through the interconnection communication module to achieve remote control and status monitoring of the terminal.

[0058] As Figure 1 shown in the figure, the main function of the instruction processing module of the central node is to parse and execute the instructions submitted by the interconnection communication module and the task model management module to achieve control of the central node and the terminal nodes.

[0059] As Figure 1 shown in the figure, the task model management module of the central node allows users to customize tasks and models and provides API interfaces for deploying these tasks to each terminal device or the central node.

[0060] As Figure 1 shown in the figure, the main function of the interconnection communication module of the central node is to achieve communication between the center and multiple terminal devices. It is mainly based on the TCP protocol, listens to the information sent by the terminal devices, and classifies the information into two categories for interaction and processing.

[0061] As Figure 1 shown in the figure, the main function of the mobile collaborative sensing module of the central node is to sense and aggregate more data under the constraint of limited energy consumption. It is mainly based on collaborative algorithms and provides API interfaces for automatically collaborating and executing the sensing tasks customized by users.

[0062] As Figure 1 shown in the figure, the permission management module of the central node provides permission management services and security authentication services for the edge nodes.

[0063] As Figure 1 shown in the figure, the implementation of the edge node is divided into four parts: a terminal control proxy module, an instruction processing module, a model control proxy module, and an interconnection communication module. Among them, the instruction processing module and the interconnection communication module have the same design structure as the central node.

[0064] As Figure 1As shown in the figure, the main function of the terminal control module of the edge node is to provide an interface for hardware device control. After investigating the functions and characteristics of the main mobile sensing terminal devices, the terminal control module abstracts the mobile sensing terminal devices into proxy intelligent agents with the capabilities of movement, perception, and calculation.

[0065] As Figure 1 shown in the figure, the model control agent module of the edge node provides an API interface for controlling terminal devices for user-defined local scripts.

[0066] As Figure 1 shown in the figure, the application service module is based on the basic function module, user-oriented, and provides an API interface for upper-layer applications. The application service module is mainly divided into three parts: model distribution and deployment, multi-source data aggregation, and terminal collaborative control. In addition, service expansion can also be carried out based on the basic function module according to application needs.

[0067] In addition, the efficient interconnection and interoperability module for heterogeneous terminals in the present invention designs a unified data interaction mechanism with good compatibility and scalability according to the characteristics of data interaction of intelligent terminals.

[0068] Aiming at the serious problem of heterogeneous communication protocols in the current Internet of Things, the present invention designs a general data interaction mechanism based on the TCP / IP protocol, making it independent of the specific underlying communication protocol, with high flexibility, easy to maintain and implement.

[0069] This data interaction mechanism adopts a hybrid architecture. In the presence of a control center, it allows terminal devices to interoperate with other terminal devices, which can reduce the communication pressure and computing pressure of the center.

[0070] A priority mechanism is introduced, enabling terminal devices to independently control and process control requests and data requests from other terminals.

[0071] Aiming at the differences between different terminal devices, a hierarchical architecture based on intelligent agents and an information interaction specification across platforms and languages are designed. By extracting the common features and common patterns of different types and functions of terminals, the terminal devices are abstracted into virtual intelligent agents with the capabilities of movement, perception, and calculation. By different instantiations of the virtual intelligent agents, the control methods of different terminal hardware devices are compatible, realizing low coupling between the control of terminal hardware devices and the design of system functions.

[0072] As Figure 4 shown in the figure, the present invention designs a unique multi-agent reinforcement learning algorithm to solve this collaborative game problem. Each terminal intelligent agent has its own policy network, while the Q function network is shared by all intelligent agents.

[0073] As Figure 4 shown in the figure, each intelligent agent has its own observation oi and action a i , where the observations are manifested differently according to the nature of the terminal agents. Concatenate the observations o i and action a i of each agent, then pass the concatenated result through a one-layer MLP respectively, and calculate the state e i = g i (o i , a i ).

[0074] As Figure 4 shown, each agent can query the observations o=(o1,…o n ) and actions a=(a1,…a n ) of other agents, and incorporate this information into its own state and action estimates, thereby obtaining the state set e=(e1,…e n ) of the agent cluster.

[0075] As Figure 4 shown, for the state-action value function of agent i as shown in formula (3), where f i is encoded using two layers of fully connected layers, and the intermediate layer variables of the fully connected layers are calculated through the multi-relational head as shown in formula (4).

[0076] Among them, v j is the value vector of e j , and is obtained by linear transformation using the shared matrix W v . h is a non-linear activation function (Leaky ReLU), and r j l represents the l+1-ary relation weight.

[0077]

[0078]

[0079] Embodiment

[0080] The task of guiding the flow of people in the subway station has high requirements for the convergence of multi-source data and the cooperation of multiple terminals, and is a typical application of the data convergence and cooperative processing prototype system. With the development of rail transit, the proportion of the subway's traffic volume in the total traffic volume is increasing day by day. To ensure the convenience and safety of subway passengers' travel and improve the transportation efficiency of the subway, the prediction and guidance of the passenger flow in the subway are becoming more and more important. However, the environment in the subway station is complex, the passenger flow distribution is uneven, and the change speed is relatively fast. Single-source data and a single terminal are not sufficient to complete this complex task of guiding the passenger flow.

[0081] As Figure 3 shown, the robot will provide a display interface for passengers, and recommend elevator exits with less passenger flow on the robot display interface to avoid congestion on the platform and in the carriages.

[0082] As Figure 3 shown, the heat map of the passenger flow in the current concourse and on the platform will also be displayed on the robot display interface, enabling passengers to quickly understand the congestion situation in the station and making them as receptive as possible to the system-recommended routes.

[0083] The execution of the passenger flow guidance task in the subway station is achieved in three stages, namely model distribution and deployment, data aggregation and calculation, and mobile terminal collaborative control.

[0084] The model distribution and deployment service distributes and deploys the defined passenger flow guidance tasks and data processing models to each terminal device through the central node, and the terminal device runs the received script to perform tasks such as data collection.

[0085] The multi-source data aggregation service aggregates the real-time data collected by multiple sensing terminals, such as video surveillance data, temperature and humidity sensor data, CO2 sensor data, access card data at the turnstile for entering and leaving the station, and access card data at adjacent stations, to the node responsible for the calculation task.

[0086] The calculation node integrates and calculates the data obtained from each sensing terminal, combines it with the predicted passenger OD information, fuses them, predicts the heat map of the passenger flow in the station and potential hazards, and recommends indoor routes based on the prediction results. Then the results are submitted to the mobile terminal and control node responsible for the passenger flow guidance task.

[0087] After obtaining the data from the calculation node, the control node requests control authority from the central node through the terminal collaborative control service according to the deployed scheduling model, and then controls the target nodes, such as the mobile robot terminal and the display screen in the platform. The mobile terminal moves to the designated location and guides the crowd to wait in front of the carriage with fewer waiting people through voice or images, and the display screen shows the personnel density distribution in the carriage in real time, and they cooperate to execute the passenger flow guidance task.

Claims

1. A terminal interoperability and hazard response system based on urban agglomeration intelligent perception, characterized in that It includes an efficient interconnection and interoperability module for heterogeneous terminals, a data aggregation and collaborative processing prototype system, and a hazard situation response system. The efficient interconnection and interoperability module for heterogeneous terminals analyzes the functions and characteristics of common terminals on the market, depicts an abstract model of edge terminals, designs an abstract hardware layer to separate upper-layer control and data interaction, extracts terminal hardware features and abstracts them into virtual intelligent agents with mobility, perception, computing, and communication capabilities, and compatible and extends different terminal hardware devices through different instantiations of the virtual intelligent agents. The data aggregation and collaborative processing prototype system designs a hierarchical architecture based on intelligent agents and an information interaction specification across platforms and languages for the differences between different terminal devices. By extracting the common features and common patterns of different types of terminals with different functions, the terminal devices are abstracted into virtual intelligent agents with mobility, perception, and computing capabilities, and the control methods of different terminal hardware devices are made compatible through different instantiations of the virtual intelligent agents, realizing low coupling between the control of terminal hardware devices and the system function design. The hazard situation response system can quickly respond to abnormal situations and guide intelligent agents to perform corresponding obstacle avoidance and evacuation work by utilizing the fast data collaboration and terminal control of multiple terminals in urban space. The efficient interconnection and interoperability module for heterogeneous terminals designs a unified data interaction mechanism with good compatibility and scalability for the characteristics of intelligent terminal data interaction. Aiming at the serious problem of heterogeneous communication protocols in the current Internet of Things, a general data interaction mechanism is designed based on the TCP / IP protocol, making it independent of the specific underlying communication protocol, with high flexibility, easy to maintain and implement. The data interaction mechanism adopts a hybrid architecture. In the presence of a control center, it allows terminal devices to interoperate with other terminal devices, which can reduce the communication pressure and computing pressure on the center. The efficient interconnection and interoperability module for heterogeneous terminals introduces a priority mechanism, enabling terminal devices to independently control and process control requests and data requests from other terminals. The data aggregation and collaborative processing prototype system adopts a hybrid architecture of "centralized management - distributed interoperability" and modular design to interconnect central nodes and edge nodes. The central node consists of a terminal agent module, an instruction processing module, a task model management module, an interconnection communication module, a mobile collaborative sensing module, and a permission management module; the terminal agent module of the central node provides an interface for terminal control of the central node. To be compatible with the underlying control methods of mobile sensing terminals with different hardware, according to the state control of the terminal control module's unified coding protocol, the terminal agent module encodes the terminal control information into a JSON packet and sends it to the corresponding terminal through the interconnection communication module to achieve remote control and status monitoring of the terminal; the instruction processing module of the central node parses and executes the instructions submitted by the interconnection communication module and the task model management module to achieve control of the central node and terminal nodes; the task model management module of the central node allows users to customize tasks and models and provides an API interface for deploying these tasks to each terminal device or the central node; the interconnection communication module of the central node realizes communication between the center and multiple terminal devices. It is mainly based on the TCP protocol, listens to the information sent by the terminal devices, and classifies the information into two categories for interaction and processing; the mobile collaborative sensing module of the central node senses and aggregates more data under the constraint of limited energy consumption. Based on the collaborative algorithm, it provides an API interface for automatically collaborative execution of the sensing tasks customized by users; the permission management module of the central node provides permission management services and security authentication services for edge nodes; The edge node consists of four parts: a terminal control agent module, an instruction processing module, a model control agent module, and an interconnection communication module. Among them, the instruction processing module and the interconnection communication module have the same design structure as the central node; the main function of the terminal control module of the edge node is to provide an interface for hardware device control. After investigating the functions and characteristics of the main mobile sensing terminal devices, the terminal control module abstracts the mobile sensing terminal devices into agent intelligent bodies with mobile, sensing, and computing capabilities; the model control agent module of the edge node provides an API interface for controlling terminal devices for locally customized scripts by users; The application service module is based on the basic function module and is oriented to users, providing an API interface for upper-layer applications. The application service module is mainly divided into three parts: model distribution and deployment, multi-source data aggregation, and terminal collaborative control; in addition, service expansion can also be carried out based on the basic function module according to application needs; When facing abnormal situations, the hazard response system abstracts sensors as independent intelligent bodies and transforms the entire hazard response problem into a multi-agent collaborative game problem; and designs a unique multi-agent reinforcement learning algorithm to solve this collaborative game problem; For the hazard response system, each terminal agent has its own policy network, while the Q-function network is shared by all agents; each agent can query the observations and actions of other agents and incorporate this information into its own state and action estimates; each agent has its own observation O i and action a i , where the observations are manifested differently according to the nature of the terminal agent, and the state e of each agent is calculated i ; for the state-action value function of agent i, two layers of fully connected layers are used for encoding, where the intermediate layer variable x i l is calculated through the multi-relational head; The advantage function of multi-agents is designed based on a benchmark. By comparing the value function of a specified action and the value function of the average action, it is judged whether the increase in the reward obtained by the current action is attributed to the actions of other agents; considering the different terminal functions, the benchmark does not need to assume that each agent has the same action space and does not require a global reward.

Images