Marine remote driving light field perception edge service model optimization method and system

Abstract

The invention relates to a ship remote driving light field perception edge service model optimization method and system, and the method comprises the steps: collecting the light field data of a ship remote driving environment at a ship end, and designing an observation end target viewpoint rendering scheme; constructing a framework of an edge service model in the shore-based driving control center, and describing an interaction mode of an observation terminal and an edge service node in the framework; respectively establishing a service time delay model and a tariff and overhead model for local execution and edge execution, and constructing a target optimization function and a plurality of constraint conditions; aiming at a plurality of observation terminals accessed by the same edge service node, a computing power and bandwidth resource allocation problem is converted into a convex optimization problem, a resource allocation closed-form solution is solved by using a KKT condition, and resource layer optimization is realized; the optimal task unloading strategy under the current time slot is generated based on Actor-Critic, and decision-making layer optimization is realized, so that the light field data transmission efficiency and real-time performance are improved, the service time delay and tariff overhead of an observation end are balanced, and the safety and stability of ship remote driving are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent ship navigation technology, and in particular to an edge service model optimization method and system for light field perception of remote ship driving. Background Art

[0002] Stereoscopic light field ship environment perception services in the field of remote ship navigation are typically characterized by low latency and high bandwidth. Furthermore, rendering the viewpoint image at the crew observation end of the control center requires significant computing power. If traditional centralized communication service models are maintained, the control center will be unable to perceive the ship's navigation environment in a timely manner due to core network congestion and other factors. The upgrade of sensing equipment with higher computing power also incurs high costs, which will hinder the further development of intelligent navigation services. With the continuous development of mobile communication technology, the emergence and widespread use of fifth-generation mobile communication technology (5G) will lay a solid foundation for the development of emerging intelligent industries. 5G communication technology has significantly improved performance compared to traditional 3G / 4G communication networks. Its typical application standards include higher data speeds, lower communication latency, a larger number of device connections, and more flexible service deployment models. These advantages enable the network to better meet the growing demands of communication and emerging applications, and greatly promote the development of intelligent technology across various industries. At the same time, with the increasing scale of access devices and application complexity, the network will generate enormous data transmission and computing processing demands. While traditional centralized cloud computing can address these challenges to a certain extent, it still has many shortcomings in terms of real-time performance and deployment flexibility. This has led to the emergence of new distributed edge computing architectures. The core concept of the edge computing service architecture is to deploy server nodes at the network edge, providing storage and computing resources. Terminal devices directly connect to the edge service nodes via a wireless access network, efficiently offloading local computing tasks and receiving the resulting data from the edge service nodes. By bringing computing and storage resources closer to the terminal, the edge computing service architecture reduces the transmission distance of application data and improves computing efficiency, making it ideal for emerging intelligent applications with high real-time requirements and large data volumes. By deploying edge service nodes at the shore-based control center, the task of rendering the navigation environment light field viewpoint image from the observation terminal device can be offloaded to the edge service node, which then receives the viewpoint data via the wireless access network link. This effectively reduces the core network load, better utilizes the access bandwidth advantages of the 5G network, and improves the efficiency of the light field perception service for remote ship operation. The workflow of the ship remote driving system based on the edge computing service architecture is as follows: after collecting environmental data, the ship-side light field camera transmits low-density viewpoints to the shore via the ship-shore network. By executing the viewpoint interpolation algorithm locally or at the edge service node, multi-directional perspectives are rendered for shore-based control personnel to observe the environment.

[0003] Light field acquisition often includes both spatial and angular information about a scene, and possesses higher data dimensions than traditional video, requiring more resources for storage and transmission. If the light field acquisition results from the ship are transmitted directly to a shore-based control center for observation, the high-dimensional viewpoint data of the light field requires significant network resources for transmission. Furthermore, as control center personnel continue to demand greater freedom of viewing angles and orientations, the manufacturing process and cost of the acquisition equipment also become important considerations. Currently, network infrastructure in most inland waterways is relatively weak, resulting in slow data transmission speeds. Furthermore, inland waterways are susceptible to natural factors such as weather and currents, making network connections susceptible to interference and causing unstable equipment operation. Timely network maintenance in remote inland waterways remains an issue, and prolonged network instability can hinder timely remote driving environment perception and pose significant safety risks. This makes efficient light field data transmission solutions crucial.

[0004] Existing research on light field data transmission is mostly limited to the field of viewpoint sparse reconstruction. The core idea is to encode only part of the viewpoint code stream at the acquisition end for transmission, decode the sparse viewpoints at the receiving end, and restore the complete viewpoint array. However, this method focuses on improving the compression performance ratio of the viewpoint array content, while ignoring service quality factors such as network fluctuations and viewpoint decoding delay in long-term environments. Since the ship-shore network infrastructure in most inland areas is relatively weak, coverage is limited, and it is easily affected by natural environmental factors, and real-time observation and perception of the navigation environment is also crucial to the safety of remote ship driving, traditional light field data transmission methods that only consider compression performance cannot take into account real-time issues. On the other hand, the existing light field data transmission architecture does not consider the computing power support issue at the receiving end, which still has a certain gap in practical application scenarios. There is an urgent need to study efficient light field data transmission methods for light field perception scenarios in remote ship driving environments. Summary of the Invention

[0005] The present invention addresses the problems of existing light field data transmission technology in remote ship driving scenarios, such as large network resource usage, low transmission efficiency, poor real-time performance, and failure to consider the computing power support of the receiving end. The present invention provides an edge service model optimization method for light field perception of remote ship driving. Through four-layer collaborative optimization of data layer compression transmission, architecture layer dynamic scheduling, resource layer optimization allocation, and decision layer closed-loop control, an overall closed-loop optimization of the edge service model is formed. This method can effectively reduce the core network load in a network-constrained ship-shore environment, improve the efficiency and real-time performance of light field data transmission, and balance the service delay and tariff overhead of the observation end, significantly improving the safety and stability of remote ship driving. The present invention also relates to an edge service model optimization system for light field perception of remote ship driving.

[0006] The technical solutions of the present invention are as follows:

[0007] A method for optimizing an edge service model for light field perception of remote ship driving, characterized by comprising the following steps:

[0008] S1. Ship-side light field data processing step: light field data of the ship's remote driving environment is collected on the ship, and a target viewpoint rendering scheme is designed for the observation end based on the correlation between light field viewpoints. The viewpoint rendering scheme includes rendering parameter information, single / multiple interpolation synthesis rules, and generation of a low-density light field viewpoint code stream through quality-graded encoding. The low-density light field viewpoint code stream includes the rendering parameter information, and the low-density light field viewpoint code stream is then transmitted to the shore-based driving control center via the ship-to-shore network;

[0009] Steps for building the edge service model architecture: Deploy edge service nodes at the shore-based control center, construct a light field perception business system consisting of N observation terminals and M edge service nodes as the edge service model architecture, and describe the interaction between the observation terminals and edge service nodes in the architecture. This includes: the architecture supports observation terminals in flexibly selecting edge service nodes for access, uses vectors to represent the access status and task offloading decisions of observation terminals at edge service nodes in time slots, and shares computing power and bandwidth resources among observation terminals connected to the same edge service node.

[0010] S3. Computational task quantification and decision-level model construction steps: At the shore-based control center, virtual viewpoint data is synthesized using a convolutional neural network interpolation algorithm based on the low-density viewpoint code stream transmitted in step S1 and the rendering parameter information it contains. Simultaneously, based on the architecture established in step S2 and the synthesized virtual viewpoint data, service latency models and tariff cost models are established for local and edge execution, respectively. With minimizing long-term latency and tariff cost as the long-term optimization goal, a target optimization function and several constraints are constructed, including edge server computing resource constraints, edge server bandwidth constraints, task offloading constraints, and long-term latency constraints. The Lyapunov stability framework is used to transform the long-term latency constraints into cumulative management of virtual latency queues, and the long-term latency optimization problem is transformed into a single-slot real-time optimization problem.

[0011] S4. Resource Allocation Solution Step: Based on the architecture established in step S2 and the service latency model and cost model established in step S3, the computing power and bandwidth resource allocation problem is transformed into a convex optimization problem for multiple observation terminals connected to the same edge service node. The KKT condition is used to solve the closed-form solution for resource allocation, achieving resource-level optimization. This closed-form solution serves as the resource constraint for subsequent deep reinforcement learning.

[0012] S5. Task offloading decision optimization step: Based on the Actor-Critic deep reinforcement learning framework, the real-time network environment parameters of the current time slot, the service delay model and tariff expenditure model established in step S3, and the closed-form solution of resource allocation obtained in step S4 are input to generate the optimal task offloading strategy for the current time slot. By executing the optimal task offloading strategy, the virtual delay queue is updated to meet the long-term delay constraint, and the decision-level optimization is achieved, thereby forming a closed-loop optimization of the edge service model.

[0013] Preferably, in step S1, the rendering parameter information in the rendering scheme includes the correlation between light field viewpoints and the interpolation step size of adjacent viewpoints in the horizontal / vertical direction; the single / multiple interpolation synthesis rule defines the synthesis of the observation end target viewpoint that is not directly encoded and transmitted by the viewpoint interpolation technology, including the operation and order of single interpolation and multiple interpolation; the quality hierarchical coding includes:

[0014] Evaluate the importance of viewpoints in light field data and prioritize viewpoints based on driving scenario requirements;

[0015] A layered coding strategy is adopted to retain high-frequency information for high-priority viewpoints and compress low-priority viewpoints to generate viewpoint streams with different quality levels.

[0016] Preferably, in step S2, an N×M-dimensional binary vector is used to represent the access status of each observation terminal to the edge service node in time slot t, where N represents the number of observation terminals and M represents the number of edge service nodes; when an element in the vector is 1, it indicates that the corresponding observation terminal chooses to access the corresponding edge service node, and when it is 0, it indicates that it has not accessed; an N-dimensional binary vector is used to represent the task offloading decision of each observation terminal, where when an element in the vector is 1, it indicates that the corresponding observation terminal offloads its computing task to the edge server for execution, and when it is 0, it indicates that it is executed locally.

[0017] Preferably, in step S3, synthesizing virtual viewpoint data by using a convolutional neural network interpolation algorithm includes:

[0018] Traverse the viewpoint data decoding status of all I×J observation end positions;

[0019] If the viewpoint at the current position is in the state to be decoded, search for the adjacent viewpoints in the horizontal / vertical direction;

[0020] If there is a viewpoint at an adjacent position that has been decoded, the adjacent viewpoint is used as the input of the convolutional neural network to interpolate the current target position viewpoint;

[0021] If there is no decoded viewpoint at the adjacent position, continue traversing the next viewpoint position;

[0022] Repeat the above steps until all viewpoints at all positions are decoded, and record the number of viewpoint interpolation times required for decoding each viewpoint.

[0023] Preferably, in step S3, the service delay model and the tariff cost model respectively include the computing delay during local execution and edge execution, the data transmission delay, the computing power leasing cost, and the data download traffic fee.

[0024] Preferably, in step S4, using the KKT condition to solve the closed-form solution of resource allocation includes:

[0025] Construct a Lagrangian function and introduce computing power resource constraints and bandwidth resource constraints;

[0026] The closed-form solution for resource allocation is obtained by taking partial derivatives of the computing power allocation variable and the bandwidth allocation variable and setting them to zero.

[0027] Preferably, in step S5, based on the Actor-Critic deep reinforcement learning framework, its Actor module extracts system state features and generates task offloading decisions through a deep convolutional neural network DCNN, and its Critic module evaluates the pros and cons of the decision and optimizes the resource allocation strategy based on the reward function; in each time slot, according to the current network state and resource allocation situation, the optimal task offloading strategy for the current time slot is generated; the virtual delay queue is updated by executing the optimal task offloading strategy to meet the long-term delay constraint; the experience samples are stored in the experience pool, and the DCNN network parameters are updated regularly to adapt to changes in the network environment.

[0028] An edge service model optimization system for light field perception of remote ship driving is characterized by comprising a ship-side light field data processing module, an edge service model architecture building module, a computing task quantification and decision layer model building module, a resource allocation solution module, and a task offloading decision optimization module, which are connected in sequence.

[0029] The ship-side light field data processing module collects light field data of the ship's remote driving environment on the ship side, designs an observation-side target viewpoint rendering scheme based on the correlation between light field viewpoints, and generates a low-density light field viewpoint code stream through quality-graded coding. The low-density light field viewpoint code stream includes the rendering parameter information, and then transmits the low-density light field viewpoint code stream to the shore-based driving control center via the ship-to-shore network.

[0030] The edge service model architecture building module: deploys edge service nodes in the shore-based control center, builds a light field perception business system consisting of N observation terminals and M edge service nodes as the architecture of the edge service model, and describes the interaction between the observation terminals and edge service nodes in the architecture, including: the architecture supports the observation terminal to flexibly select the edge service node to access, uses vector sum to represent the access status and task offloading decision of the edge service node of the observation terminal in the time slot, and shares its computing power and bandwidth resources for the observation terminals accessing the same edge service node;

[0031] The computing task quantification and decision-making layer model construction module: in the shore-based control center, based on the low-density viewpoint code stream transmitted by the ship-side light field data processing module and the rendering parameter information contained therein, virtual viewpoint data is synthesized through a convolutional neural network interpolation algorithm; at the same time, based on the architecture constructed by the edge service model architecture construction module and the synthesized virtual viewpoint data, a service delay model and a tariff cost model for local execution and edge execution are established respectively, and a target optimization function and several constraints are constructed with minimizing long-term delay and tariff cost as the long-term optimization goal, including edge server computing resource constraints, edge server bandwidth constraints, task offloading constraints and long-term delay constraints. The Lyapunov stability framework is used to convert the long-term delay constraints into cumulative management of virtual delay queues, and the long-term delay optimization problem is converted into a real-time optimization problem for a single time slot;

[0032] The resource allocation solution module: Based on the architecture built by the edge service model architecture building module and the service delay model and tariff expenditure model established by the computing task quantification and decision-making layer model building module, for multiple observation terminals connected to the same edge service node, the computing power and bandwidth resource allocation problem is converted into a convex optimization problem, and the KKT condition is used to solve the closed-form solution of resource allocation to achieve resource layer optimization. The closed-form solution of resource allocation serves as the resource constraint condition for subsequent deep reinforcement learning;

[0033] The task offloading decision optimization module: based on the Actor-Critic deep reinforcement learning framework, inputs the real-time network environment parameters of the current time slot, the service delay model and tariff expenditure model established by the computing task quantification and decision-making layer model construction module, and the closed-form solution of resource allocation obtained by the resource allocation solution module, generates the optimal task offloading strategy for the current time slot, updates the virtual delay queue by executing the optimal task offloading strategy to meet the long-term delay constraint, realizes decision-making layer optimization, and thus forms a closed-loop optimization of the edge service model.

[0034] Preferably, in the edge service model architecture building module, an N×M-dimensional binary vector is used to represent the access status of each observation terminal to the edge service node in time slot t, where N represents the number of observation terminals and M represents the number of edge service nodes; when the element in the vector is 1, it indicates that the corresponding observation terminal chooses to access the corresponding edge service node, and when it is 0, it indicates that it has not accessed; an N-dimensional binary vector is used to represent the task offloading decision of each observation terminal, where when the element in the vector is 1, it indicates that the corresponding observation terminal offloads its computing task to the edge server for execution, and when it is 0, it indicates that it is executed locally.

[0035] Preferably, in the task offloading decision optimization module, based on the Actor-Critic deep reinforcement learning framework, its Actor module extracts system state features and generates task offloading decisions through a deep convolutional neural network DCNN, and its Critic module evaluates the pros and cons of the decision and optimizes the resource allocation strategy based on the reward function; in each time slot, based on the current network state and resource allocation situation, the optimal task offloading strategy for the current time slot is generated; the virtual delay queue is updated by executing the optimal task offloading strategy to meet the long-term delay constraint; the experience samples are stored in the experience pool, and the DCNN network parameters are updated regularly to adapt to changes in the network environment.

[0036] The technical effects of the present invention are as follows:

[0037] The present invention relates to an edge service model optimization method for light field perception of remote ship driving. First, the ship-side light field data processing step collects light field data of the ship remote driving environment at the ship side, and designs an observation end target viewpoint rendering scheme based on the correlation between light field viewpoints. The rendering scheme includes determining rendering parameter information such as the interpolation step size of adjacent viewpoints in the horizontal and vertical directions, defining single / multiple interpolation synthesis rules, and generating a low-density light field viewpoint code stream through quality graded coding. The low-density light field viewpoint code stream includes the above-mentioned rendering parameter information such as the interpolation step size, and then transmits it to the shore-based driving control center via the ship-shore network. Through the single / multiple interpolation rules, the fineness of the synthesized viewpoint is flexibly controlled (such as single interpolation ensures real-time performance, and multiple interpolation improves clarity), and the quality is improved. The combination of quality grading strategy and rendering parameters meets the safety and economy requirements of remote driving and realizes dynamic QoS guarantee. Low-density light field viewpoint code stream is generated through quality grading coding to achieve data compression and efficient transmission, effectively reducing the amount of redundant data, reducing the bandwidth requirement of ship-to-shore network transmission, and improving data transmission efficiency. The design based on the correlation between light field viewpoints makes the generated viewpoint code stream more suitable for transmission in ship-to-shore networks with limited bandwidth, enhancing the adaptability of the network environment. The generated low-density viewpoint code stream provides structured input for the shore-based virtual viewpoint interpolation, ensuring that the shore-based control center can perform further processing and rendering. Through compression and optimized coding, the data transmission delay is shortened, providing a real-time basis for remote driving environment perception. The steps for building the edge service model architecture are to deploy edge service nodes in the shore-based control center, build a light field perception business system consisting of N observation terminals and M edge service nodes as the architecture of the edge service model, and describe the interaction mode between the observation terminals and edge service nodes in the architecture, supporting the observation terminals to flexibly select edge service nodes for access, improving the flexibility and adaptability of the system, and being able to dynamically adjust the access strategy according to different network status and task requirements, adapt to network fluctuations and load changes, avoid node overload or resource idleness, and observation terminals connected to the same edge service node share its computing power and bandwidth resources to achieve resource sharing and collaboration, improve resource utilization efficiency, reduce resource waste, reduce service costs, and enhance the overall performance of the system. The access status of the observation terminal and the task offloading decision are represented by vector sum, and the interaction rules between the observation terminal and the edge service node are clearly defined, providing clear guidance for subsequent task processing and resource allocation. In addition, the model architecture is scalable, allowing the number of edge service nodes (M) and the scale of observation terminals (N) to be dynamically expanded to adapt to remote driving scenarios of ships of different sizes.The computing task quantification and decision-making layer model construction steps synthesize virtual viewpoint data through the convolutional neural network interpolation algorithm, which can generate high-quality virtual viewpoints, provide richer visual information for the shore-based driving control center, and enhance the visual experience of remote driving. A service delay model and a tariff expenditure model are established, and a target optimization function is constructed with the goal of minimizing long-term delay and tariff expenditure. Accurate delay and tariff modeling are performed, and the task processing cost is quantified, providing an optimization basis for resource allocation and task offloading. Various constraints are considered, including edge server computing resource constraints, bandwidth constraints, task offloading constraints, and long-term delay constraints, ensuring the feasibility and practicality of the optimization process. The Lyapunov stability framework is used to convert long-term delay constraints into cumulative management of virtual delay queues, and the long-term delay optimization problem is converted into a single-time slot real-time optimization problem, avoiding the computational burden of complex global planning, ensuring the system's rapid response in a dynamic network environment, and improving the real-time and adaptability of the optimization process. The resource allocation solution step transforms the computing power and bandwidth resource allocation problem into a convex optimization problem, and uses the KKT condition to solve the closed-form solution of resource allocation, ensuring the mathematical optimality of resource allocation, realizing the optimization of the resource layer, and improving the efficiency and fairness of resource allocation. The low computational complexity of convex optimization meets real-time requirements and avoids the delay problem of traditional heuristic algorithms. The obtained closed-form solution of resource allocation serves as the resource constraint condition for subsequent deep reinforcement learning, provides important input for task offloading decision optimization, and enhances the overall optimization capability of the system. The task offloading decision optimization step is based on the Actor-Critic deep reinforcement learning framework, combined with the real-time network status (bandwidth, computing power, queue length) and the closed-form solution of resource allocation, to generate the optimal task offloading strategy under the current time slot, balance the delay and tariff overhead, and can dynamically adjust the task offloading decision according to the real-time network environment parameters and resource allocation. It improves the adaptability and performance of the system, updates the virtual delay queue by executing the optimal task offloading strategy, that is, forms a closed-loop feedback control, and feeds the actual delay after the strategy execution back to the next time slot optimization process through the virtual delay queue update mechanism, ensuring the stable satisfaction of the long-term delay constraint, ensuring the long-term stability of the system, realizing decision-making layer optimization, forming a closed-loop optimization of the edge service model, and enhancing the efficiency and reliability of the ship remote driving light field perception service. The present invention realizes low-latency, low-cost, and high-stability operation of the ship remote driving light field perception service through the four-layer collaborative optimization of data layer compression transmission, architecture layer dynamic scheduling, resource layer optimization allocation, and decision layer closed-loop control, providing key technical support for the large-scale application of intelligent shipping.

[0038] The present invention is a light field data transmission perception service for the field of remote control of ships. A light field virtual viewpoint rendering scheme based on viewpoint interpolation synthesis is designed, which can support a free-viewing viewing experience through interpolation synthesis when only sparse light field viewpoints are received. An edge computing service architecture is introduced to assist the observation terminal in virtual viewpoint rendering. When the ship only transmits sparse light field viewpoints to the shore, the observation terminal can offload the target viewpoint rendering task to the edge node or execute it locally to alleviate the huge computing load of the terminal and provide good support for the real-time light field free-viewing perception of the ship-side environment. A light field viewpoint rendering task offloading and resource allocation strategy based on deep reinforcement learning and convex optimization is designed. The task offloading scheme can be flexibly selected for each observation terminal device according to the load balancing of the edge service node, so as to maximize the utilization efficiency of computing power and bandwidth resources, thereby balancing the delay and tariff overhead of the observation end light field perception service in the long-term remote control scenario of ships, and improving the overall service efficiency of the system.

[0039] In the ship-side light field data processing step, the present invention can further optimize data transmission efficiency. Through viewpoint importance evaluation and hierarchical coding strategies, it prioritizes retaining high-frequency information of high-priority viewpoints that are critical to the driving scene (such as channel obstacles and close-range ships), reduces the amount of transmitted data for low-priority viewpoints (such as distant views or static backgrounds), reduces the bandwidth occupancy of the ship-shore network, and can ensure perception accuracy. High-quality coding of core viewpoints ensures that the shore-based control center obtains key environmental information, avoiding remote driving decision-making errors due to excessive compression.

[0040] In the step of building the edge service model architecture, an N×M-dimensional binary vector can be used to represent the edge service node access status of each observation terminal at time slot t. When the element in the vector is 1, it indicates that the corresponding observation terminal has selected the corresponding edge service node, and when it is 0, it indicates that it has not connected. This enables efficient resource management, accurately describes the access relationship between the observation terminal and the edge node, simplifies resource allocation logic, and reduces scheduling complexity. Using an N-dimensional binary vector to represent the task offloading decision of each observation terminal, where the element in the vector is 1, the corresponding observation terminal offloads its computing task to the edge server for execution, and when it is 0, it executes it locally, can clarify the decision boundary, force tasks to only choose local or edge execution, avoid policy ambiguity, and improve system determinism.

[0041] During the computational task quantification and decision-level model construction steps, interpolation calculations are performed only on necessary locations through traversal and adjacent viewpoint search mechanisms, avoiding redundant calculations for decoding the full viewpoint array and improving viewpoint synthesis efficiency. Convolutional neural network (CNN) interpolation algorithms (such as U-Net) are used to synthesize virtual viewpoints, preserving high-frequency details (such as textures and edges) in light field data, enhancing rendering quality, and supporting high-fidelity rendering from free viewpoints. Separate modeling of local execution latency (computational latency) and edge execution latency (computational + transmission latency) accurately reflects the real-time differences between different strategies and precisely quantifies costs. Through computing power rental costs and data traffic fees, commercial operating costs are embedded in the optimization objectives to ensure the economic feasibility of the strategies.

[0042] In the resource allocation solution step, the closed-form solution of resource allocation can be solved through Lagrangian function and KKT conditions to ensure the global optimality of computing power and bandwidth allocation under constraints; the low computational complexity (polynomial time) of convex optimization solution meets the real-time resource scheduling requirements of the edge service model.

[0043] In the task offloading decision optimization step, the Actor-Critic deep reinforcement learning framework includes an Actor module and a Critic module. The Actor module extracts system state features (network bandwidth, computing power, queue length) through a deep convolutional neural network (DCNN) to generate a task offloading strategy adapted to the real-time environment. The Critic module evaluates the strategy based on a reward function, combining experience replay and parameter updates to gradually improve the generalization and robustness of the strategy. The virtual delay queue update mechanism feeds back the strategy execution results to the next time slot optimization to ensure the stable satisfaction of long-term delay constraints.

[0044] The present invention also relates to an edge service model optimization system for light field perception of remote ship driving. The system corresponds to the above-mentioned edge service model optimization method for light field perception of remote ship driving, and can be understood as a system for realizing the above-mentioned edge service model optimization method for light field perception of remote ship driving. The system is provided with a ship-side light field data processing module, an edge service model architecture building module, a computing task quantification and decision layer model building module, a resource allocation solution module and a task offloading decision optimization module. Each module is connected in sequence and works together with each other. Through hierarchical optimization (data compression → architecture building → resource allocation → dynamic decision), the core network load and computing delay are reduced layer by layer to ensure the real-time perception of the remote driving environment; through the efficient processing and transmission of ship-side light field data, the flexible architecture building of the edge service model, the optimization of resource allocation and the dynamic optimization of task offloading decision, the overall remote ship driving is improved. The efficiency of the driving light field perception system is improved, reducing data transmission delays and resource waste; the design and optimization of each module enables the system to better adapt to different network environments and task requirements, improves the system's flexibility and adaptability, and enhances its stability and reliability in complex environments; through high-quality synthesis of virtual viewpoints and optimized task offloading decisions, it provides the shore-based control center with richer visual information and a smoother visual experience, improving the overall user experience of remote ship driving; the coordinated optimization of resource allocation and task offloading strategies minimizes computing power leasing and data traffic costs, reduces the system's operating costs, and improves the system's economy and practicality; through the deep reinforcement learning (DRL) framework and the management of virtual delay queues, the closed-loop optimization of the edge service model is achieved, ensuring the long-term stability and performance optimization of the system, and improving the overall efficiency of the ship's remote driving light field perception service. The dynamic access and resource sharing mechanism of the edge computing architecture, combined with the online learning capabilities of DRL, can effectively address challenges such as fluctuations in ship-shore network bandwidth and unbalanced node loads. The modular design supports flexible expansion of edge nodes and observation terminals, making it suitable for remote driving scenarios of inland or ocean-going vessels of different sizes. It is also generalizable, and the technical framework can be migrated to other intelligent navigation applications that require low latency and high computing power, such as unmanned ship cluster collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of the edge service model optimization method for light field perception of remote ship driving of the present invention.

[0046] Figure 2 This is a flow chart of the edge service model optimization method for light field perception of remote ship driving of the present invention.

[0047] Figure 3 Flowchart for light field viewpoint rendering in the computational task quantization and decision layer model building steps.

[0048] Figure 4Flowchart for building the decision layer model in the computational task quantification and decision layer model building steps.

[0049] Figure 5 Flowchart for resource layer optimization in the resource allocation solution step.

[0050] Figure 6 Flowchart of decision layer optimization in the decision optimization step for task offloading. DETAILED DESCRIPTION

[0051] The present invention will be described below with reference to the accompanying drawings.

[0052] To meet the needs of shore-based control center personnel for real-time, free-viewpoint observation of the ship's operating environment, this paper proposes an edge service model optimization method for light field perception for remote ship operation based on an edge computing architecture, based on the correlation between light field viewpoints. This method ensures the stability of remote ship light field environment perception services in long-term, time-varying network environments. Remote ship operation: The ship is unmanned, and the remote pilot controls the vessel remotely via the network from the shore-based control center. Ship environment perception: The remote pilot needs to be able to observe the actual image of the ship's movement through the ship's camera equipment to provide accurate control operations. In a scenario where ship-to-shore network environments have imperfect communication infrastructure and require efficient transmission of light field perception data for remote ship operation, the light field video data collected onboard is encoded using quality-graded encoding to generate different-density viewpoint streams. The low-density viewpoint stream is first transmitted via the core network to the shore-based control center's receiving end. The rendering task of the observer's target viewpoint is then scheduled and executed locally or offloaded to an edge service node. After the offloaded viewpoint rendering is completed, it is transmitted back to the observation equipment via the edge access network to meet the observer's needs for remote observation of the ship's environment. The present invention is aimed at the remote driving scenario of ships in the field of intelligent navigation. It designs an edge service model optimization method based on deep reinforcement learning (DRL) for light field data perception, and studies the viewpoint rendering task offloading and bandwidth / computing resource allocation issues under multiple edge nodes of multiple control centers, so as to balance the observer perception delay and service cost overhead as much as possible in a long-term time-varying network environment. Specifically, the primary requirement of the present invention is to support the real-time perception of the environment during the ship's navigation process by the shore-based control center in a network-constrained ship-shore environment. In this process, the requirements that need to be implemented by the design of the present invention are:

[0053] First, the observation terminal in the control center can obtain a free-viewing experience of the ship's driving environment. This requires designing a virtual viewpoint rendering solution based on the redundant characteristics of the collected light field data, which is not directly collected or encoded and transmitted.

[0054] Second, the target light field viewpoint content at the observation end can be decoded in real time to view the ship's driving environment. When the observation terminal equipment is not sufficient to support real-time perception rendering, it is required to design an edge service model and task offloading mode for ship remote driving environment perception based on the advantages of the new distributed edge computing architecture, which can rely on the relatively sufficient computing power in the edge server to improve the rendering efficiency of the observation viewpoint.

[0055] 3. Offloading viewpoint rendering tasks to edge service nodes can alleviate the huge computing power burden faced by observation terminals. However, it also requires paying the corresponding edge service computing power rental and data download traffic fees. This requires the design of a reasonable task computing scheduling model that can achieve a good balance between perceived latency and application fee overhead.

[0056] Fourth, the computing power and access bandwidth resources of edge servers are relatively abundant but always limited. This requires the formulation of a reasonable computing power / bandwidth resource allocation strategy. At the same time, combined with the task scheduling model mentioned in the third point, it is possible to ensure the stability of the ship-side light field service when the network environment fluctuates.

[0057] In an environment where the core network bandwidth, computing power and storage resources of the shore-based control center are limited, the present invention implements a scenario where the ship-side light field video is reliably delivered to the shore. The light field data is quality graded and encoded based on the difference in importance of each area in the picture to the remote control perception. After transmission through the core network, viewpoint interpolation decoding is performed on the remote driving end, thereby completing the remote perception of the surrounding environment during the ship's navigation. In order to reduce the data transmission pressure of the core network and the viewpoint interpolation calculation load of the control end, while taking into account the delay requirements of scene perception, the edge computing mode of the present invention studies the viewpoint decoding calculation offloading and resource allocation problems in the scenario of multiple edge service nodes and multiple shore-based control centers, so that it can be dynamically adjusted to adapt to the common network bandwidth fluctuations during the long-term continuous navigation of the ship, while balancing the service fee overhead and perception response delay of the shore-based control end to the greatest extent. Among them, viewpoint interpolation: in the field of three-dimensional scene acquisition such as light field, the information contained between viewpoints is highly similar. Based on this, a series of intermediate viewpoints can be synthesized through the content of adjacent viewpoints. Edge computing mode: By deploying servers with computing and storage capabilities at the edge of the network close to the terminal, relevant application data can be directly obtained from the edge server and computing-intensive tasks can be efficiently offloaded. This can effectively reduce the communication load on the core network link and provide services for a wider range and a larger number of terminal applications.

[0058] The edge service model optimization method of the light field perception of remote ship driving of the present invention is as follows: Figure 1As shown: First, a rendering scheme for the target viewpoint at the observation end is designed based on the correlation between viewpoints of the original light field data. Light field acquisition is usually stored in an array-type data structure, with uniform horizontal / vertical parallax between the perspectives in the horizontal and vertical directions. The intermediate viewpoints that are not directly stored can be obtained through motion interpolation of adjacent viewpoints. Each interpolation can synthesize the viewpoint content at the middle position. The interpolated viewpoint can also be used to synthesize viewpoints at other positions. Therefore, the target viewpoint that is not directly encoded and transmitted can be obtained through single / multiple viewpoint interpolation. After the viewpoint rendering scheme is determined, an edge computing communication service model (i.e., edge service model) with multiple service nodes and multiple observation terminals is constructed. By considering the service charges for perception delay, edge computing power leasing, and viewpoint data download traffic, the system's target optimization function is established to minimize delay and tariff overhead. The network conditions and observation end requirements in a long-term environment are constantly changing and difficult to predict. First, the long-term objective optimization function of the above-mentioned system is converted into an optimization problem for a single time slot, and then the computing task offloading strategy for each time slot is output based on the deep reinforcement learning model; wherein, the present invention uses the Actor-Critic model structure (Actor-Critic deep reinforcement learning framework) for single time slot optimization, and takes the computing tasks, computing power, network environment and accumulated state of the delay queue of the current time slot as the input of the model. The Actor module extracts the input vector state features and maps out a candidate set of task offloading decisions. The Critic module executes the optimal resource allocation strategy based on the task offloading decision, and outputs the optimal task offloading decision and the corresponding resource allocation strategy in the candidate set as the optimization result of the current time slot system; the optimization result is placed in the experience pool for subsequent network parameter training and updating, and the delay queue is updated based on this result as the network input of the next time slot for satisfying long-term delay constraints. Specifically, the flow chart is as follows Figure 2 As shown, the following steps are included:

[0059] S1. Ship-side light field data processing steps: light field data of the ship's remote driving environment is collected on the ship side, and an observation-side target viewpoint rendering scheme is designed based on the correlation between light field viewpoints. The viewpoint rendering scheme includes rendering parameter information, single / multiple interpolation synthesis rules, and generation of a low-density light field viewpoint code stream through quality graded encoding. The low-density light field viewpoint code stream contains rendering parameter information, and then the low-density light field viewpoint code stream is transmitted to the shore-based driving control center via the ship-shore network.

[0060] Furthermore, the rendering parameter information in the designed rendering scheme may include the correlation between light field viewpoints, the interpolation step size of adjacent viewpoints in the horizontal / vertical direction, etc. The single / multiple interpolation synthesis rules define the synthesis of observation end target viewpoints that are not directly encoded and transmitted through viewpoint interpolation technology, including the operations and order of single interpolation and multiple interpolation. The quality graded coding includes: evaluating the importance of viewpoints on light field data, and dividing the viewpoint priorities according to driving scene requirements; adopting a layered coding strategy to retain high-frequency information for high-priority viewpoints and compress low-priority viewpoints to generate viewpoint code streams of different quality levels. The low-density light field viewpoints that are directly encoded and transmitted have high correlation in the horizontal and vertical directions.

[0061] The present invention considers a light field array containing I*J viewpoints. i，j Represents the viewpoint content of the i-th row and j-th column of the light field array. Figure 3 As shown in the figure, the virtual viewpoint data is synthesized through the convolutional neural network interpolation algorithm, or the light field viewpoint rendering process, including:

[0062] Step S101: traverse the decoding status of viewpoint data at all I*J observation end positions;

[0063] Step S102: If the viewpoint at the current position is in a state to be decoded, go to step S103; if the viewpoint at this position has been decoded, go to step S101 and continue traversing the next viewpoint position;

[0064] Step S103: For the viewpoint at the position to be decoded, search for the viewpoints at the adjacent positions in the horizontal / vertical directions. i，j-y ,viewpoint i，j+y ) or (viewpoint i-x，j ,viewpoint i+x，j ), where y and x represent the number of viewpoints that are different from the viewpoint at the position to be decoded in the horizontal / vertical direction between the adjacent viewpoint;

[0065] S104: If there is a viewpoint at an adjacent position that has been decoded, go to step S105; if not, the precondition for decoding the current viewpoint is still not met, go to step S101, and continue traversing the next viewpoint position;

[0066] Step S105: Use the adjacent viewpoints as the input of the convolutional neural network to interpolate and obtain the current target position viewpoint.

[0067] For step S101, it is necessary to loop and execute multiple times until all viewpoints at all positions are decoded and Viewpoint i，j The number of viewpoint interpolation times required for upper viewpoint decoding is recorded as In,m After this step, the number of interpolation calculations required for rendering virtual viewpoints that are not directly transmitted can be quantitatively described under the viewpoint density of the current ship-side light field acquisition and encoding transmission.

[0068] S2. Steps for building the edge service model architecture: deploy edge service nodes at the shore-based control center, build a light field perception business system consisting of N observation terminals (referred to as terminals) and M edge service nodes (referred to as edge nodes) as the architecture of the edge service model, and describe the interaction between the observation terminals and edge service nodes in the architecture, including: dividing the continuous time range into discrete time slots, each terminal in each time slot will generate different light field viewpoint requests, the architecture supports the observation terminal to flexibly select the edge service node to access and offload the virtual viewpoint rendering calculation task, use vectors to represent the access status and task offloading decision of the edge service node of the observation terminal in the time slot, and share its computing power and downlink communication bandwidth resources with the observation terminals accessing the same edge service node. Specifically, it includes the following steps:

[0069] Step S201: Construct a light field perception service system including N observation terminals and M edge nodes as the architecture of the edge service model. The M edge computing service nodes are distributed near the terminal devices and have relatively rich computing resources, represented by Ms = {ms1, ms2, ..., ms M}, and its coverage includes N observation terminals that generate ship-side light field environment perception requests, expressed as Ud = {ud1, ud2, ..., ud N}.

[0070] Step S202: The light field data collected by the ship is encoded into a low-density light field viewpoint stream file and transmitted via the core network to the observation terminal and edge node. If the terminal's target viewpoint is not included in the stream file, interpolation calculation is required to synthesize the target viewpoint. Depending on the system status of the current time slot, local calculation or offload calculation can be selected.

[0071] Step S203: Each terminal can select any edge node to access, using the binary vector a n，t ={a n，1，t ,…,a n，m，t} indicates the terminal ud n In the edge service node access state at time slot t, if a n，m，t =1, it means that the n-th observation end chooses to access the m-th edge service node; using binary b n，t ∈{0, 1} to represent the task offloading decision of the n-th user after accessing the edge service node, where b n，t =1 means that the n-th observation end offloads its computing task to the edge server for execution, b n，t=0 means local execution; observation terminals connected to the same edge node share its computing power and downlink bandwidth resources.

[0072] S3. Computational task quantification and decision-making layer model construction steps: At the shore-based control center, based on the low-density viewpoint code stream transmitted in step S1 and the rendering parameter information it contains, virtual viewpoint data is synthesized through the convolutional neural network interpolation algorithm. The virtual viewpoint at the target position can be obtained through single / multiple adjacent viewpoint interpolation; at the same time, based on the architecture built in step S2 and the synthesized virtual viewpoint data, service delay models and tariff overhead models are established for local execution and edge execution respectively, and the target optimization function and several constraints are constructed with minimizing long-term delay and tariff overhead as the long-term optimization goal, including edge server computing resource constraints, edge server bandwidth constraints, task offloading constraints and long-term delay constraints. The Lyapunov stability framework is used to transform the long-term delay constraints into the cumulative management of the virtual delay queue, and the long-term delay optimization problem is transformed into a real-time optimization problem for a single time slot.

[0073] This step quantitatively describes the computing tasks generated by each terminal, and establishes service delay and tariff overhead representation models for computing tasks executed locally and on edge servers, respectively, to set the target optimization function and constraints. In practical applications, the network and task information of the system at future moments is usually unknown, and the Lyapunov stability framework is used to transform the long-term delay optimization problem into an optimization problem for each time slot.

[0074] A light field perception task generated by the observation terminal n at time slot t is used express, and They represent the number of interpolations required for viewpoint rendering when the n-th observation end selects the m-th edge service node for access and the computing task is executed locally or on the edge server. and It indicates the number of viewpoints that need to be transmitted in the network when the n-th observation terminal chooses to access the m-th edge service node and executes tasks locally or on the edge server. Figure 4 The process shown is:

[0075] Step S301: If the terminal ud n Access ms m When the computing task is executed locally, the CPU computing frequency of this terminal is expressed as (in cycles / s), the main frequency is fully occupied when executing computing tasks locally, the amount of calculation required for each interpolation algorithm execution is represented by c (in cycles), the amount of data corresponding to the transmission of a single viewpoint stream between the edge node and the observation end is represented by d (in Mbits), and the edge service node export bandwidth allocated to it is Tr n，m，t , the task processing delay when executed locally is:

[0076]

[0077] Local execution does not involve edge server computing power usage. The operator charges a fee of f (in $) per megabyte of data downloaded. The local execution fee only includes the data download traffic fee:

[0078]

[0079] Step S302: If the terminal ud n Access ms m The computing task is offloaded to the edge node, and the upper limit of the edge node computing resources and the maximum downlink bandwidth are expressed as C m，total and B m，total，t The computing frequency and downlink bandwidth allocated to computing tasks when they are executed at the edge are and Tr n，m，t , then the corresponding task processing delay is:

[0080]

[0081] The edge service provider charges a unit price of p (in $) for each computing cycle, and the edge execution includes computing power rental Fc n，t and data download There are two parts to the fee:

[0082]

[0083] Step S303: Based on the calculation model definition of step S301 and step S302, observe the terminal ud n Total service delay t n，t and data download traffic cost Ft n，t It can be expressed as:

[0084]

[0085] The optimization goal of the system can be to meet the long-term terminal perception service delay constraint γ n Under the premise of limited computing / bandwidth resources, the two tariff costs of computing power leasing and interpolation synthesis viewpoint data download traffic are minimized. The optimization objective function can be expressed as:

[0086]

[0087] Correspondingly, ω is used to measure the importance of the two costs of computing power leasing and data downloading on the performance of the total cost. The edge server computing resource and bandwidth constraints C1 and C2, task offloading constraint C3, and long-term latency constraint C4 are:

[0088]

[0089] Step S304: In the above step S303, the perceived service delay constraint relationship is characterized as a long-term optimization problem. However, in practice, future tasks and network information cannot be obtained in advance. The Lyapunov stability framework is applied to transform the long-term optimization problem into a single-time slot real-time optimization problem. The long-term delay constraint condition is transformed into the virtual delay queue accumulation at each moment, where Y n (t) represents the backlog level of the delay queue at time t, V is an adjustable constant, and Y n (t) and V jointly adjust the optimization ratio between delay constraint and service cost:

[0090]

[0091] St stands for subject to, meaning that something satisfies or is subject to the constraints of C1, C2, and C3. This indicates that the formula satisfies, or is subject to, C1, C2, and C3.

[0092] S4. Resource allocation solution step: Based on the architecture built in step S2 and the model established in step S3, for multiple observation terminals connected to the same edge service node, the computing power and bandwidth resource allocation problem is transformed into a convex optimization problem, and the KKT condition is used to solve the closed-form solution of resource allocation, thereby obtaining a resource allocation plan and realizing resource layer optimization. The closed-form solution of resource allocation is used as the resource constraint condition for subsequent deep reinforcement learning. Figure 5 The process shown is:

[0093] Step S401: The access node and task offloading decision set a in step S404 are t and b t Treated as a constant;

[0094] Step S402: Based on the previous step, the continuous resource allocation solution Tr is to be solved t and C t It can be regarded as a convex problem to be solved, and the function terms and constraints contained in the original optimization problem are not coupled with each other, so it can be further decomposed into two independent resource allocation sub-problems (satisfying C1 and C2 respectively):

[0095]

[0096] Step S403: Apply the KKT (Karush–Kuhn–Tucker) condition to obtain the closed-form optimal solution for computing / bandwidth resource allocation. and

[0097]

[0098] The closed-form solution to resource allocation using KKT conditions includes:

[0099] Construct a Lagrangian function and introduce computing power resource constraints and bandwidth resource constraints;

[0100] The closed-form solution for resource allocation is obtained by taking partial derivatives of the computing power allocation variable and the bandwidth allocation variable and setting them to zero.

[0101] S5. Task offloading decision optimization step: Based on the Actor-Critic deep reinforcement learning framework, the real-time network environment parameters of the current time slot, the service delay model and tariff expenditure model constructed in step S3, and the closed-form solution of resource allocation (or resource allocation plan) obtained in step S4 are input to generate the optimal task offloading strategy for the current time slot. By executing the optimal task offloading strategy, the virtual delay queue is updated to meet the long-term delay constraint, and the decision-making layer is optimized, thereby forming a closed-loop optimization of the edge service model.

[0102] Specifically, based on the Actor-Critic deep reinforcement learning framework, its Actor module extracts system state features and generates task offloading decisions through deep convolutional neural network DCNN, and its Critic module evaluates the pros and cons of the decision and optimizes the resource allocation strategy based on the reward function; in each time slot, according to the current network state and resource allocation, the optimal task offloading strategy for the current time slot is generated; by executing the optimal task offloading strategy, the virtual delay queue is updated to meet the long-term delay constraint; the experience samples are stored in the experience pool, and the DCNN network parameters are updated regularly to adapt to changes in the network environment. Figure 6 The process shown is:

[0103] Step S501: The light field perception service optimization in the field of ship remote control uses an actor-critic based reinforcement learning algorithm model. The RL state space representation at each time slot includes S t ={I L (t), I E (t), D L (t), D E (t), C L (t), C E (t), Tr(t), Y(t)}, the corresponding action space representation includes The RL network agent extracts the input state feature map and outputs the corresponding action. L (t) and I E (t) represents the number of viewpoint interpolations required when the observation end task is connected to different edge nodes locally or offloaded to the edge server to perform the computing task, D L (t) and D E (t) represents the number of viewpoints that need to be transmitted in the network when the observation end task is connected to different edge nodes locally or offloaded to the edge server to perform the computing task, C L (t) represents the computing frequency resource of the observation end, C E (t) represents the available computing resources of the edge node, Tr(t) represents the available downlink bandwidth resources of the edge node, and Y(t) represents the cumulative length of the virtual delay queue of all observation terminals. Combining the optimization objectives of the service delay model constructed in step S3 and the tariff cost model, the reward function is defined as in Indicates that the agent performs an action The objective function value when Obj rl Indicates the objective function value when computing tasks are not offloaded.

[0104] Step S502: The Actor module consists of two parts: the DCNN network and the action quantizer. The DCNN network first takes the system state S t As input, use the sigmoid activation function to output an N*2M relaxation action matrix In order to enhance the network's strategic exploration capability, Gaussian noise is first added to the original predicted action result to obtain the action matrix containing noise Then, we design an action quantizer to expand it into a set of K candidate strategy solutions. Specifically, we first select the action matrix The largest element in each row is set to 1, and the remaining elements in the row are set to 0, which is used to represent the edge service node access and task execution offloading scheme selection of the observation end; the remaining K-1 candidate strategies are obtained by randomly selecting an observation end, setting the second largest element value position in the corresponding matrix row to 1, and keeping the schemes corresponding to the other observation ends unchanged, thereby obtaining the extended candidate strategies.

[0105] Step S503: Based on step S4, the optimal solution for resource allocation under each candidate strategy for edge access and task offloading is calculated. The Critic module can then quantitatively evaluate the pros and cons of each candidate solution and select the strategy action corresponding to the reward function as the optimal strategy solution for the current time slot system. Output.

[0106] Step S504: The experience sample (s t , ) is stored in the experience pool.

[0107] Step S505: Whether the current time slot is a network parameter update round, if so, go to step S506; if not, go to step S507.

[0108] Step S506: Randomly sample a batch of experience samples as "input-output" labels to update the network parameters of DCNN

[0109] Step S507: Based on step S503, the service delay of each terminal in the current time slot is obtained, and the cumulative length of the virtual delay queue is updated.

[0110] The present invention also relates to an edge service model optimization system for light field perception in remote ship control. This system corresponds to the aforementioned edge service model optimization method for light field perception in remote ship control, and can be understood as a system that implements the aforementioned edge service model optimization method for light field perception in remote ship control. In a ship-to-shore network environment with incomplete communication infrastructure and limited coverage, the present invention can also be understood as designing a real-time light field perception system for remote ship control environments based on an edge computing service architecture. This system no longer focuses solely on improving the compression ratio of light field viewpoint array decoding, but rather focuses on optimizing factors that affect service quality, such as latency and tariff overhead, in environments with long-term continuous network fluctuations. First, based on the correlation between light field viewpoints and the low-density viewpoints collected and transmitted back by the ship, the present invention designs a rendering scheme that supports free-viewpoint perception at the observation end. This scheme leverages the stereoscopic perception advantages of immersive video for remote environment observation services. Second, the present invention introduces an edge computing network architecture into the light field perception service. Edge servers with sufficient computing and storage resources can perform computational tasks offloaded from the observation terminal, alleviating the terminal's heavy computational load. Furthermore, if observation terminal tasks are centralized and offloaded to a single edge service node, the overall system performance would be significantly reduced. This paper, taking into account latency and cost factors that affect service quality, designs a light field viewpoint rendering task offloading and resource allocation strategy based on deep reinforcement learning and convex optimization to maximize the utilization of computing power and bandwidth resources. Consequently, an edge service model optimization method for light field perception services for remote ship control was designed, which ensures the stability of real-time light field environment perception services at the observation end in environments with long-term network fluctuations.

[0111] The system is equipped with a ship-side light field data processing module, an edge service model architecture building module, a computing task quantification and decision-making layer model building module, a resource allocation solution module and a task offloading decision optimization module, among which,

[0112] The ship-side light field data processing module collects light field data of the ship's remote driving environment on the ship side, designs an observation-side target viewpoint rendering scheme based on the correlation between light field viewpoints, and generates a low-density light field viewpoint code stream through quality-graded coding. The low-density light field viewpoint code stream includes the rendering parameter information, and then transmits the low-density light field viewpoint code stream to the shore-based driving control center via the ship-to-shore network.

[0113] The edge service model architecture building module: deploys edge service nodes in the shore-based control center, builds a light field perception business system consisting of N observation terminals and M edge service nodes as the architecture of the edge service model, and describes the interaction between the observation terminals and edge service nodes in the architecture, including: the architecture supports the observation terminal to flexibly select the edge service node to access, uses vector sum to represent the access status and task offloading decision of the edge service node of the observation terminal in the time slot, and shares its computing power and bandwidth resources for the observation terminals accessing the same edge service node;

[0114] The computing task quantification and decision-making layer model construction module: in the shore-based control center, based on the low-density viewpoint code stream transmitted by the ship-side light field data processing module and the rendering parameter information contained therein, virtual viewpoint data is synthesized through a convolutional neural network interpolation algorithm; at the same time, based on the architecture constructed by the edge service model architecture construction module and the synthesized virtual viewpoint data, a service delay model and a tariff cost model for local execution and edge execution are established respectively, and a target optimization function and several constraints are constructed with minimizing long-term delay and tariff cost as the long-term optimization goal, including edge server computing resource constraints, edge server bandwidth constraints, task offloading constraints and long-term delay constraints. The Lyapunov stability framework is used to convert the long-term delay constraints into cumulative management of virtual delay queues, and the long-term delay optimization problem is converted into a real-time optimization problem for a single time slot;

[0115] The resource allocation solution module: Based on the architecture built by the edge service model architecture building module and the service delay model and tariff expenditure model established by the computing task quantification and decision-making layer model building module, for multiple observation terminals connected to the same edge service node, the computing power and bandwidth resource allocation problem is converted into a convex optimization problem, and the KKT condition is used to solve the closed-form solution of resource allocation to achieve resource layer optimization. The closed-form solution of resource allocation serves as the resource constraint condition for subsequent deep reinforcement learning;

[0116] The task offloading decision optimization module: based on the Actor-Critic deep reinforcement learning framework, inputs the real-time network environment parameters of the current time slot, the service delay model and tariff expenditure model constructed by the computing task quantification and decision layer model construction module, and the closed-form solution of resource allocation obtained by the resource allocation solution module, generates the optimal task offloading strategy for the current time slot, updates the virtual delay queue by executing the optimal task offloading strategy to meet the long-term delay constraint, realizes decision layer optimization, and thus forms a closed-loop optimization of the edge service model.

[0117] Furthermore, in the edge service model architecture building module, an N×M-dimensional binary vector is used to represent the access status of each observation terminal to the edge service node in time slot t, where N represents the number of observation terminals and M represents the number of edge service nodes; when the element in the vector is 1, it indicates that the corresponding observation terminal chooses to access the corresponding edge service node, and when it is 0, it indicates that it has not accessed; an N-dimensional binary vector is used to represent the task offloading decision of each observation terminal, where when the element in the vector is 1, it indicates that the corresponding observation terminal offloads its computing task to the edge server for execution, and when it is 0, it indicates that it is executed locally.

[0118] Furthermore, in the task offloading decision optimization module, based on the Actor-Critic deep reinforcement learning framework, its Actor module extracts system state features and generates task offloading decisions through a deep convolutional neural network (DCNN), and its Critic module evaluates the pros and cons of the decision and optimizes the resource allocation strategy based on a reward function; in each time slot, the optimal task offloading strategy for the current time slot is generated according to the current network state and resource allocation; the virtual delay queue is updated by executing the optimal task offloading strategy to meet long-term delay constraints; the experience samples are stored in an experience pool, and the DCNN network parameters are regularly updated to adapt to changes in the network environment.

[0119] The edge service model optimization method and system for light field perception of remote ship driving of the present invention form a closed-loop optimization of the edge service model through four-layer collaborative optimization of data layer compression transmission, architecture layer dynamic scheduling, resource layer optimization allocation, and decision layer closed-loop control. In a network-constrained ship-shore environment, it can reduce the core network load and computing delay layer by layer, improve the efficiency and real-time performance of light field data transmission, and balance the service delay and tariff overhead of the observation end, significantly improving the safety and stability of remote ship driving.

[0120] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be included in the scope of protection of the patent for the present invention.

Claims

1. An edge service model optimization method for light field perception of remote ship driving, characterized by: The steps include: S1. Ship-side light field data processing step: light field data of the ship's remote driving environment is collected on the ship, and a target viewpoint rendering scheme is designed for the observation end based on the correlation between light field viewpoints. The viewpoint rendering scheme includes rendering parameter information, single / multiple interpolation synthesis rules, and generation of a low-density light field viewpoint code stream through quality-graded encoding. The low-density light field viewpoint code stream includes the rendering parameter information, and the low-density light field viewpoint code stream is then transmitted to the shore-based driving control center via the ship-to-shore network; Steps for building the edge service model architecture: Deploy edge service nodes at the shore-based control center, construct a light field perception business system consisting of N observation terminals and M edge service nodes as the edge service model architecture, and describe the interaction between the observation terminals and edge service nodes in the architecture. This includes: the architecture supports observation terminals in flexibly selecting edge service nodes for access, uses vectors to represent the access status and task offloading decisions of observation terminals at edge service nodes in time slots, and shares computing power and bandwidth resources among observation terminals connected to the same edge service node. S3. Computational task quantification and decision-level model construction steps: At the shore-based control center, virtual viewpoint data is synthesized using a convolutional neural network interpolation algorithm based on the low-density viewpoint code stream transmitted in step S1 and the rendering parameter information it contains. Simultaneously, based on the architecture established in step S2 and the synthesized virtual viewpoint data, service latency models and tariff cost models are established for local and edge execution, respectively. With minimizing long-term latency and tariff cost as the long-term optimization goal, a target optimization function and several constraints are constructed, including edge server computing resource constraints, edge server bandwidth constraints, task offloading constraints, and long-term latency constraints. The Lyapunov stability framework is used to transform the long-term latency constraints into cumulative management of virtual latency queues, and the long-term latency optimization problem is transformed into a single-slot real-time optimization problem. S4. Resource Allocation Solution Step: Based on the architecture established in step S2 and the service latency model and cost model established in step S3, the computing power and bandwidth resource allocation problem is transformed into a convex optimization problem for multiple observation terminals connected to the same edge service node. The KKT condition is used to solve the closed-form solution for resource allocation, achieving resource-level optimization. This closed-form solution serves as the resource constraint for subsequent deep reinforcement learning. S5. Task offloading decision optimization step: Based on the Actor-Critic deep reinforcement learning framework, the real-time network environment parameters of the current time slot, the service delay model and tariff expenditure model established in step S3, and the closed-form solution of resource allocation obtained in step S4 are input to generate the optimal task offloading strategy for the current time slot. By executing the optimal task offloading strategy, the virtual delay queue is updated to meet the long-term delay constraint, and the decision-level optimization is achieved, thereby forming a closed-loop optimization of the edge service model.

2. The edge service model optimization method for light field perception of remote ship driving according to claim 1 is characterized in that: In step S1, the rendering parameter information in the rendering scheme includes the correlation between light field viewpoints and the interpolation step size of adjacent viewpoints in the horizontal / vertical direction; the single / multiple interpolation synthesis rules define the synthesis of the observation end target viewpoints that are not directly encoded and transmitted through the viewpoint interpolation technology, including the operations and order of single interpolation and multiple interpolation; the quality hierarchical coding includes: Evaluate the importance of viewpoints in light field data and prioritize viewpoints based on driving scenario requirements; A layered coding strategy is adopted to retain high-frequency information for high-priority viewpoints and compress low-priority viewpoints to generate viewpoint streams with different quality levels.

3. The edge service model optimization method for light field perception of remote ship driving according to claim 1 is characterized in that: In step S2, an N×M-dimensional binary vector is used to represent the access status of each observation terminal to the edge service node in time slot t, where N represents the number of observation terminals and M represents the number of edge service nodes; when an element in the vector is 1, it indicates that the corresponding observation terminal chooses to access the corresponding edge service node, and when it is 0, it indicates that it has not accessed; an N-dimensional binary vector is used to represent the task offloading decision of each observation terminal, where when an element in the vector is 1, it indicates that the corresponding observation terminal offloads its computing task to the edge server for execution, and when it is 0, it indicates that it is executed locally.

4. The edge service model optimization method for light field perception of remote ship driving according to any one of claims 1 to 3 is characterized in that: In step S3, synthesizing virtual viewpoint data by using a convolutional neural network interpolation algorithm includes: Traverse the viewpoint data decoding status of all I×J observation end positions; If the viewpoint at the current position is in the state to be decoded, search for the adjacent viewpoints in the horizontal / vertical direction; If there is a viewpoint at an adjacent position that has been decoded, the adjacent viewpoint is used as the input of the convolutional neural network to interpolate the current target position viewpoint; If there is no decoded viewpoint at the adjacent position, continue traversing the next viewpoint position; Repeat the above steps until all viewpoints at all positions are decoded, and record the number of viewpoint interpolation times required for decoding each viewpoint.

5. The edge service model optimization method for light field perception of remote ship driving according to any one of claims 1 to 3, characterized in that: In step S3, the service delay model and the tariff cost model respectively include the computing delay during local execution and edge execution, the data transmission delay, the computing power leasing cost, and the data download traffic fee.

6. The edge service model optimization method for light field perception of remote ship driving according to any one of claims 1 to 3, characterized in that: In step S4, using the KKT condition to solve the closed-form solution for resource allocation includes: Construct a Lagrangian function and introduce computing power resource constraints and bandwidth resource constraints; The closed-form solution for resource allocation is obtained by taking partial derivatives of the computing power allocation variable and the bandwidth allocation variable and setting them to zero.

7. The edge service model optimization method for light field perception of remote ship driving according to any one of claims 1 to 3, characterized in that: In step S5, based on the Actor-Critic deep reinforcement learning framework, its Actor module extracts system state features through a deep convolutional neural network (DCNN) and generates task offloading decisions. Its Critic module evaluates the quality of the decisions and optimizes the resource allocation strategy based on a reward function. In each time slot, the optimal task offloading strategy for the current time slot is generated based on the current network state and resource allocation. The virtual delay queue is updated by executing the optimal task offloading strategy to meet the long-term delay constraint. The experience samples are stored in the experience pool, and the DCNN network parameters are updated regularly to adapt to changes in the network environment.

8. An edge service model optimization system for light field perception of remote ship driving, characterized by: It includes the ship-side light field data processing module, the edge service model architecture building module, the computing task quantification and decision-making layer model building module, the resource allocation solution module and the task offloading decision optimization module, which are connected in sequence. The ship-side light field data processing module collects light field data of the ship's remote driving environment on the ship side, designs an observation-side target viewpoint rendering scheme based on the correlation between light field viewpoints, and generates a low-density light field viewpoint code stream through quality-graded coding. The low-density light field viewpoint code stream includes the rendering parameter information, and then transmits the low-density light field viewpoint code stream to the shore-based driving control center via the ship-to-shore network. The edge service model architecture building module: deploys edge service nodes in the shore-based control center, builds a light field perception business system consisting of N observation terminals and M edge service nodes as the architecture of the edge service model, and describes the interaction between the observation terminals and edge service nodes in the architecture, including: the architecture supports the observation terminal to flexibly select the edge service node to access, uses vector sum to represent the access status and task offloading decision of the edge service node of the observation terminal in the time slot, and shares its computing power and bandwidth resources for the observation terminals accessing the same edge service node; The computing task quantification and decision-making layer model construction module: in the shore-based control center, based on the low-density viewpoint code stream transmitted by the ship-side light field data processing module and the rendering parameter information contained therein, virtual viewpoint data is synthesized through a convolutional neural network interpolation algorithm; at the same time, based on the architecture constructed by the edge service model architecture construction module and the synthesized virtual viewpoint data, a service delay model and a tariff cost model for local execution and edge execution are established respectively, and a target optimization function and several constraints are constructed with minimizing long-term delay and tariff cost as the long-term optimization goal, including edge server computing resource constraints, edge server bandwidth constraints, task offloading constraints and long-term delay constraints. The Lyapunov stability framework is used to convert the long-term delay constraints into cumulative management of virtual delay queues, and the long-term delay optimization problem is converted into a real-time optimization problem for a single time slot; The resource allocation solution module: Based on the architecture built by the edge service model architecture building module and the service delay model and tariff expenditure model established by the computing task quantification and decision-making layer model building module, for multiple observation terminals connected to the same edge service node, the computing power and bandwidth resource allocation problem is converted into a convex optimization problem, and the KKT condition is used to solve the closed-form solution of resource allocation to achieve resource layer optimization. The closed-form solution of resource allocation serves as the resource constraint condition for subsequent deep reinforcement learning; The task offloading decision optimization module: based on the Actor-Critic deep reinforcement learning framework, inputs the real-time network environment parameters of the current time slot, the service delay model and tariff expenditure model established by the computing task quantification and decision-making layer model construction module, and the closed-form solution of resource allocation obtained by the resource allocation solution module, generates the optimal task offloading strategy for the current time slot, updates the virtual delay queue by executing the optimal task offloading strategy to meet the long-term delay constraint, realizes decision-making layer optimization, and thus forms a closed-loop optimization of the edge service model.

9. The edge service model optimization system for light field perception of remote ship driving according to claim 8 is characterized in that: In the edge service model architecture building module, an N×M-dimensional binary vector is used to represent the access status of each observation terminal to the edge service node in time slot t, where N represents the number of observation terminals and M represents the number of edge service nodes; when an element in the vector is 1, it indicates that the corresponding observation terminal chooses to access the corresponding edge service node, and when it is 0, it indicates that it has not accessed; an N-dimensional binary vector is used to represent the task offloading decision of each observation terminal, where when an element in the vector is 1, it indicates that the corresponding observation terminal offloads its computing task to the edge server for execution, and when it is 0, it indicates that it is executed locally.

10. The edge service model optimization system for light field perception of remote ship driving according to claim 8 or 9, characterized in that: In the task offloading decision optimization module, based on the Actor-Critic deep reinforcement learning framework, its Actor module extracts system state features and generates task offloading decisions through a deep convolutional neural network (DCNN). Its Critic module evaluates the quality of the decisions and optimizes resource allocation strategies based on a reward function. In each time slot, the optimal task offloading strategy for the current time slot is generated based on the current network state and resource allocation. The virtual delay queue is updated by executing the optimal task offloading strategy to meet long-term delay constraints. The experience samples are stored in the experience pool, and the DCNN network parameters are updated regularly to adapt to changes in the network environment.

Images