Model health degree driven sensing routing method and system

Abstract

The invention discloses a model health degree driven sensing routing method and system, and belongs to the technical field of Internet of Things communication and edge computing. And each edge node constructs a multi-hop wireless network through a soft access point interface and a station interface, and operates an active routing protocol to maintain a next-hop routing table. The middleware layer is provided with an operation effect monitoring module for collecting operation indexes such as end-to-end time delay, packet loss rate, node residual electric quantity, reasoning task queue length and the like of a task flow; and the model health degree evaluation module calculates a model health degree index according to the reasoning result, the error rate and the confidence coefficient distribution change in the preset time window. And the strategy synthesis node enables the application service quality descriptor, the operation index and the model health degree to form an environment state vector, generates strategy configuration and issues the strategy configuration to the edge node. According to the method, dynamic balance can be carried out among time delay, energy consumption and intelligent task accuracy, and the comprehensive performance and robustness of an edge Internet of Things system are improved.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) communication and edge computing technology, specifically to a model health-driven perceptual routing method and system. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and edge computing, a large number of sensor nodes, smart terminals, and edge computing nodes are being deployed in urban infrastructure, industrial sites, energy systems, and smart homes. These nodes exhibit significant heterogeneity in terms of hardware computing power, storage capacity, power supply methods, and network link conditions, and typical forms include microcontroller development boards, single-board computers, and small edge servers.

[0003] In such systems, the traditional approach is to upload the collected data to the cloud or a central server via a multi-hop wireless network for unified processing. This model has the following problems:

[0004] (1) Long-path transmission results in end-to-end latency that is difficult to meet the requirements of real-time control or interactive applications;

[0005] (2) A large amount of raw data is forwarded in a multi-hop network, which increases bandwidth usage and energy consumption;

[0006] (3) Data that is sensitive to privacy or has a high level of security may be leaked during transmission.

[0007] To alleviate the aforementioned problems, edge computing and "in-network computing" concepts have been proposed. These concepts reduce core network load and end-to-end latency by performing some or all computing tasks on edge nodes close to the data source. In multi-hop wireless IoT, self-organizing networks based on Wi-Fi or Wi-Fi-like technologies are typically used to connect edge nodes in a tree or mesh structure and run active routing protocols to maintain end-to-end paths.

[0008] In existing technologies, some systems use middleware set up above the network layer to enable applications to influence routing behavior to a certain extent, such as prioritizing nodes with stronger computing power, achieving data sharing within a local area, or controlling the selection of parent nodes for newly accessed nodes. These solutions typically achieve "application awareness" through predefined policies or by having application developers manually configure routing-related parameters.

[0009] However, with the introduction of intelligent applications such as distributed neural network inference and federated learning, edge IoT systems no longer solely focus on traditional network metrics like latency and packet loss rate when making routing decisions. They also need to comprehensively consider the health of the intelligent model itself, i.e., whether inference accuracy, confidence distribution, and concept drift are within acceptable ranges under current network conditions and node load. Existing technologies mainly suffer from the following shortcomings:

[0010] 1. Applications require manual strategy selection.

[0011] Existing frameworks often provide fixed strategy libraries and interfaces, requiring application developers to explicitly specify which strategy to use and the specific parameter configurations, making it difficult to automatically adjust based on the running results.

[0012] 2. Closed-loop optimization lacking model health considerations

[0013] Most routing optimization methods only target latency, packet loss rate, or energy consumption, without incorporating the health of the intelligent model as a feedback signal into the routing control process. This results in a significant decrease in the accuracy of model predictions even if network performance indicators appear good.

[0014] 3. Lack of online adaptive policy synthesis mechanism

[0015] In edge environments, link conditions, node loads, and data distribution change significantly over time. Existing solutions often rely on offline configuration or static weights, and cannot dynamically weigh and switch between various strategy combinations through online learning.

[0016] Therefore, a new technical solution is needed to introduce a model health-driven online policy synthesis mechanism on the basis of existing multi-hop edge networks and application-aware middleware. This mechanism enables the system to automatically select appropriate routing policy combinations based on network status and model health during operation, thereby achieving comprehensive optimization of latency, energy consumption, and the accuracy of intelligent tasks. Summary of the Invention

[0017] The main objective of this invention is to overcome the following problems existing in current edge IoT application sensing routing methods:

[0018] 1. Routing behavior and the health status of the intelligent model are independent of each other, and cannot form an effective closed loop;

[0019] 2. Routing policy configuration relies on human experience and lacks online adaptive capabilities to adapt to dynamic changes in various applications and environments;

[0020] 3. It cannot flexibly balance latency, energy consumption, reliability, and model accuracy according to application requirements.

[0021] To address this, the present invention proposes a model health-driven perceptual routing method and system. By introducing a performance monitoring module, a model health assessment module, and an online strategy synthesis module in the middleware layer, the method enables automatic combination and parameter adjustment of perceptual routing strategies for various applications, transforming the routing control process from "manual configuration" to "data-driven online learning".

[0022] To achieve the above objectives, the present invention adopts the following technical solution: a model health-driven perceptual routing method, applied to a multi-hop edge IoT system composed of multiple edge nodes, comprising the following steps:

[0023] Step S101: Construct a multi-hop edge network and routing foundation.

[0024] Each edge node is configured with a soft access point interface and a site interface. The soft access point interface is used to create a wireless LAN and accept access from other edge nodes, while the site interface is used to access the soft access point interface of the upper-level edge node. An active routing protocol is adopted, using the soft access point address of each edge node as the node identifier to maintain the routing table. The routing table records the destination node address, the next-hop node address, the number of hops, and the sequence number, thereby forming a multi-hop edge network that supports dynamic node access and fault recovery.

[0025] Step S102: Collect network and node statistics during runtime.

[0026] Set up a runtime performance monitoring module in the middleware layer of each edge node to periodically collect runtime metrics that include at least one of the following:

[0027] (1) End-to-end latency of each task flow;

[0028] (2) Packet loss rate and retransmission count of the link or path;

[0029] (3) Node remaining power, battery voltage, or average transmission power;

[0030] (4) Load metrics such as inference task queue length and CPU utilization;

[0031] The aforementioned runtime metrics are then reported to the preset policy synthesis node via a preset control channel.

[0032] Step S103: Calculate the model health index.

[0033] For task flows involving edge intelligent reasoning, the application layer or coordination node calculates a model health index to characterize the model state based on the reasoning results and information such as reference labels, error rate, and confidence distribution within a preset time window, and reports it to the policy synthesis node.

[0034] Step S104: Construct the environment state vector.

[0035] In the middleware layer, the policy synthesis node combines the application service quality descriptor (including latency level, energy consumption budget, reliability level, privacy level, computing preference, etc.), the runtime metrics, and the model health metrics of each task flow to form a corresponding environment state vector, which is used as input to the online policy synthesis module.

[0036] When the computation preference in the application service quality descriptor is edge distributed inference, the online policy synthesis module prioritizes policy combinations that enable publish and subscribe policies and enhance the role of topology control policies; when the computation preference is edge centralized inference, the online policy synthesis module prioritizes policy combinations that primarily use intermediate node injection policies.

[0037] Step S105: Establish a strategy combination set.

[0038] Multiple application-aware routing policies are predefined in the policy synthesis node, and different application-aware routing policies and their parameter configurations are combined into a discrete policy combination set. Each policy combination corresponds to a candidate routing behavior pattern.

[0039] Step S106: Policy synthesis based on online learning.

[0040] An online policy synthesis module is set up in the middleware layer of the policy synthesis node. The online policy synthesis module uses an online learning algorithm to establish a mapping relationship between the environment state vector and the policy combination set.

[0041] Step S107: Generate and distribute the policy configuration.

[0042] The online policy synthesis module, based on the expected reward estimates of each policy combination, periodically or when a significant change in the environmental state vector is detected, selects the policy combination with the highest expected reward estimate as the target policy combination for each task flow; generates the corresponding policy configuration according to the target policy combination, including the injection of intermediate node set, topic partitioning scheme and parent node priority list, and distributes the policy configuration to the relevant edge nodes.

[0043] Step S108: Execute model health-driven perceptual routing.

[0044] When sending or forwarding data packets, the policy execution agent of each edge node encapsulates, replicates, or imposes topological constraints on the data packets according to the currently effective policy configuration, and calls the active routing protocol to perform the actual forwarding. During operation, as runtime metrics and model health metrics are updated, the online policy synthesis module continuously adjusts the policy configuration to achieve adaptive optimization of routing behavior with model health as an important feedback signal.

[0045] Optionally, the online learning algorithm preferably employs a multi-armed gambling machine algorithm, treating each application-aware routing strategy combination as an arm, and maintaining the selection frequency and expected reward estimate of each arm for each task flow. The online strategy synthesis module selects a strategy combination for each task flow each time:

[0046] Based on the upper confidence bound strategy or the ε-greedy strategy, a balance is struck between exploring underexperienced strategy combinations and estimating higher strategy combinations by taking advantage of the current expected reward;

[0047] The statistics for the corresponding arm of the selected strategy combination are updated using the instant reward value to improve the expected reward estimate.

[0048] Optionally, establishing a mapping relationship between the environment state vector and the policy combination set specifically includes:

[0049] For each task flow, select one from the set of policy combinations as the current action given the environment state;

[0050] After executing the selected strategy combination, the instant reward value is calculated according to the preset reward function, which is a weighted combination of end-to-end latency, packet loss rate, node remaining power and model health index.

[0051] The statistics of the corresponding strategy combination are updated using the instant reward value, thereby adjusting the expected reward estimate and realizing online evaluation of the merits of different strategy combinations.

[0052] The weights of the reward function are configured based on the latency level, energy consumption budget, reliability level, and computational preference in the application service quality descriptor, so that the latency weight is increased for task flows with high real-time requirements, and the model health weight is increased for task flows with high accuracy requirements.

[0053] Optionally, the model health index is a normalized real number, preferably a value in the range of 0 to 1. A higher value indicates a healthier model. The calculation process is as follows:

[0054] The accuracy rate or equivalent accuracy index of the inference results within a preset time window is statistically analyzed.

[0055] Based on the deviation between the accuracy rate and the target accuracy rate, as well as the degree of change in the confidence distribution, the health score is attenuated.

[0056] The processed health value is cropped or mapped to fall within the range of 0 to 1.

[0057] Optionally, the model health index is calculated based on at least one or more of the following:

[0058] The degree of change in the model output confidence distribution within a preset time window;

[0059] The degree of deviation between the input feature distribution and the preset training data distribution;

[0060] Uncertainty estimation results based on self-supervised tasks or consistency constraints.

[0061] Optionally, the various application-aware routing strategies include:

[0062] The intermediate node injection strategy is used to encapsulate the original data packet into an encapsulation structure with the outer destination address as the intermediate edge node and the inner destination address as the final destination node, and to select a set of intermediate edge nodes that meet the preset metric conditions according to the strategy configuration.

[0063] The publish and subscribe strategy is used to implement multicast distribution of data streams based on topics, enabling multiple edge nodes subscribed to the same topic to receive intermediate results or control information simultaneously.

[0064] The topology control strategy is used to adjust the parent-child relationship according to the parent node priority list in the strategy configuration when a new edge node joins the network or a parent node is reselected, thereby optimizing the overall network topology.

[0065] Optionally, the policy synthesis node is the root node in a multi-hop wireless network, or an edge server connected to the root node. The runtime metrics and the model health metrics are aggregated to the policy synthesis node through the multi-hop wireless network. The edge server is used to store online learning algorithm parameters and long-term statistical information to support long-term performance analysis and adjustment of policy combinations.

[0066] Optionally, the strategy combinations in the strategy combination set include at least:

[0067] A strategy combination aimed at latency optimization, wherein the strategy combination enables the intermediate node injection strategy and prioritizes intermediate edge nodes with fewer hops and stronger computing power.

[0068] A strategy combination aimed at energy consumption optimization, wherein the strategy combination weakens the effect of the intermediate node injection strategy and preferentially avoids edge nodes with remaining power below a threshold in the topology control strategy;

[0069] A strategy combination aimed at optimizing model health, wherein the strategy combination utilizes the publish and subscribe strategy to synchronize intermediate results among multiple edge nodes, and selects paths or inference nodes with higher health based on the model health index.

[0070] The present invention also provides a model health-driven perceptual routing system for implementing the above method, the system comprising multiple edge nodes interconnected via a wireless network and at least one policy synthesis node.

[0071] The edge nodes include:

[0072] The wireless communication module is used to access the multi-hop wireless network through the soft access point interface and the site interface, and to interact with other edge nodes.

[0073] The routing module is used to maintain a next-hop routing table identified by the soft access point address according to the active routing protocol, and to forward data packets according to the next-hop routing table to realize basic forwarding of data packets;

[0074] The runtime performance monitoring module is used to collect runtime metrics of the edge node and its task flow through a sliding time window mechanism, including at least one of end-to-end latency, packet loss rate, remaining power, and inference task queue length, and report them to the policy synthesis node, thereby reducing the impact of short-term fluctuations on the decision results of online learning algorithms.

[0075] The policy execution agent module is used to receive policy configurations from the policy synthesis node, encapsulate, copy, or impose topology constraints on data packets to be sent or forwarded to the target node according to the policy configurations, and call the routing module to complete the actual forwarding.

[0076] The application interface module is used to receive the application service quality descriptor set by the application layer for the task flow, including requirements for latency, energy consumption, reliability and privacy, and send it to the policy synthesis node.

[0077] The strategy synthesis node includes:

[0078] The model health assessment module is used to calculate model health indicators to characterize the model status based on the inference results of edge intelligent inference tasks and optional reference labels, and to statistically analyze information such as prediction accuracy, error rate, and confidence distribution changes within a preset time window.

[0079] The online policy synthesis module receives runtime metrics and model health metrics reported by each edge node, combines them with the application service quality descriptor to form an environment state vector, and selects target policy combinations from a predefined policy combination set for each task flow based on an online learning algorithm, and generates corresponding policy configurations and distributes them to the relevant edge nodes.

[0080] Policy synthesis nodes can be deployed on the root node of a multi-hop edge network or on an edge server connected to the root node.

[0081] Optionally, the specific operations of the model health assessment module in calculating the model health index include:

[0082] In labeled scenarios, the accuracy index is calculated based on the degree of matching between the inference results and the real labels within a preset time window, and the model health index is determined accordingly.

[0083] In unlabeled scenarios, the model health index is determined based on at least one of the following: changes in the model output confidence distribution, the degree of drift in the input feature distribution, and model uncertainty estimation based on a self-supervised task.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] 1. Introduce model health as a key feedback signal.

[0086] This invention introduces the health status of intelligent models into the routing control process for the first time. By using model health indicators to reflect factors such as inference accuracy, confidence distribution, and stability, routing decisions no longer rely solely on traditional network indicators, making them more suitable for deploying edge intelligent services.

[0087] 2. Implement data-driven online strategy synthesis

[0088] This invention uses an online policy synthesis module to treat multiple application-aware routing policies as composable candidate actions. It utilizes an online learning algorithm to continuously evaluate the performance of each policy combination and dynamically select the appropriate one during operation, thereby reducing the complexity of manual parameter tuning and policy design.

[0089] 3. Supports flexible trade-offs between latency, energy consumption, and accuracy.

[0090] This invention implements a differentiated optimization strategy for different task flows by simultaneously considering latency, packet loss rate, energy consumption, and model health in the reward function and setting weights according to the application service quality descriptor. This can meet the needs of complex scenarios where real-time control tasks and high-precision inference tasks coexist.

[0091] 4. The algorithm is lightweight and suitable for resource-constrained edge nodes.

[0092] The online learning component preferably adopts the multi-armed gambling machine algorithm, which only requires maintaining a small number of statistics to complete strategy evaluation and selection. It has low computational complexity and is suitable for deployment on resource-constrained microcontrollers or single-board computers.

[0093] 5. Clear structure, easy to implement and expand in engineering.

[0094] This invention uses the middleware layer as its core, clearly separating the modules for monitoring operational performance, assessing model health, and synthesizing online policies. This allows for the reuse of existing multi-hop routing and application-aware infrastructure, while also facilitating the future expansion of new policy types and evaluation metrics. Attached Figure Description

[0095] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0096] Figure 2 This is a schematic diagram of the multi-hop edge network and routing infrastructure in this invention.

[0097] Figure 3This is a schematic diagram illustrating the data flow for monitoring operational effectiveness and assessing model health in an embodiment of the present invention.

[0098] Figure 4 This is a schematic diagram of the internal structure and input-output relationship of the online strategy synthesis module of the present invention.

[0099] Figure 5 This is a schematic diagram of the online policy update process driven by the health status of the model in this invention.

[0100] Figure 6 This is a schematic diagram illustrating the data flow and policy synthesis in a distributed edge inference scenario according to the present invention.

[0101] Figure 7 This is a schematic diagram of the injection routing in a centralized edge inference scenario according to the present invention. Detailed Implementation

[0102] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the present invention; various equivalent modifications or substitutions can be made by those skilled in the art without departing from the spirit and scope of the present invention.

[0103] Example 1: Overall Architecture and Network Infrastructure

[0104] See Figure 1 and Figure 2 This embodiment provides an overall architecture for a model health-driven perceptual routing system.

[0105] The system includes various types of edge nodes, such as:

[0106] Resource-constrained microcontroller nodes are used to perform simple preprocessing and a small number of inference operations;

[0107] Single-board computer or industrial PC node, used to perform complex inference tasks and coordination functions;

[0108] One or more edge servers are used to perform the functions of policy synthesis nodes.

[0109] Each edge node integrates a wireless communication module, supporting the simultaneous activation of both the soft access point interface and the site interface. The system constructs a multi-hop wireless network using the following method:

[0110] 1. Start the soft access point interface on the root node and configure a fixed IP address range and SSID;

[0111] 2. The lower-level node scans the SSID through the site interface, selects a suitable upper-level node to connect to, and obtains the site interface address through DHCP;

[0112] 3. Each node maintains a routing table using an active routing protocol. Routing table entries include: destination node access point address, next-hop node access point address, hop count, and sequence number.

[0113] 4. Nodes periodically broadcast routing update messages, with sequence numbers used to indicate the freshness of the routing information, thereby ensuring loop-free routing and fast convergence.

[0114] Based on this, each node can also maintain a simple node lifecycle state machine, including initialization, searching for parent nodes, joining the network, normal operation, waiting for recovery, and reconnection, to improve the network's fault tolerance to node failures and migrations.

[0115] This part belongs to the existing multi-hop edge IoT network infrastructure, and this invention does not impose too many limitations on it.

[0116] Example 2: Monitoring Operational Results and Assessing Model Health

[0117] See Figure 3 This embodiment focuses on illustrating the workflow of the operation effect monitoring module and the model health assessment module.

[0118] 1. Monitoring of operational performance

[0119] Deploy and run a performance monitoring module in the middleware layer of each edge node, and periodically perform the following operations:

[0120] Record the timestamps of data packets sent or received by this node, and calculate the end-to-end latency based on the task flow identifier;

[0121] Count the number of retransmissions or use the confirmation mechanism to estimate the packet loss rate;

[0122] Read the battery voltage or remaining charge to estimate the remaining available energy of the node;

[0123] Statistical analysis of inference task queue length, average processing time, and CPU utilization.

[0124] The above metrics are packaged into runtime metric messages at fixed time intervals (e.g., 1 to 10 seconds), which include task flow identifier, node identifier, collection time and values ​​of each metric, and are reported to the policy synthesis node through control messages.

[0125] 2. Model health assessment

[0126] For task flows involving edge intelligent inference, two model health assessment methods can be designed:

[0127] (1) Scenarios with labels

[0128] When the system can obtain the true labels of some samples, the coordinating node calculates the consistency between the prediction results and the true labels within the time window, such as precision and recall, and can combine the trend of indicator changes in a recent period to obtain a model health value of 0 to 1.

[0129] (2) Unlabeled scenarios

[0130] When the system cannot directly obtain the real label, one or a combination of the following methods can be used:

[0131] Based on the confidence distribution of the model output, detect whether there are abnormal shifts, such as confidence being concentrated in an excessively high range or significantly reduced.

[0132] By performing drift detection on the input feature distribution, it can be determined whether the current data deviates from the training data distribution;

[0133] The model uncertainty is estimated using a simplified self-supervised index.

[0134] The model health assessment module will output a model health metric at the end of each window period, along with a task flow identifier and timestamp, and report it to the policy synthesis node.

[0135] In the preferred implementation, health can be given in the following form:

[0136] If the accuracy rate within the window is higher than the target threshold, the health score is close to 1.

[0137] If the accuracy falls below the threshold or a significant distribution drift is detected, the health score is reduced.

[0138] To avoid short-term fluctuations, an exponential moving average can be used to smooth the health curve.

[0139] Example 3: Design of Online Strategy Synthesis Module

[0140] See Figure 4 and Figure 5 This embodiment illustrates the internal structure of the online strategy synthesis module and the online learning process.

[0141] 1. State representation

[0142] The online policy synthesis module represents the current state of each task flow as a vector, including:

[0143] Application service quality descriptors: such as latency level (high / medium / low), energy consumption priority, reliability level, privacy level, computing preference (edge ​​distributed / edge centralized, etc.).

[0144] Runtime network metrics: such as recent average latency, latency jitter, packet loss rate, remaining battery power, path length, etc.

[0145] Model health metrics: such as the current window health value and the difference between the current window health value and the target health value.

[0146] The above vectors can be standardized or discretized during implementation to adapt to lightweight online learning algorithms.

[0147] 2. Action Space (Set of Strategy Combinations)

[0148] The system pre-configures several combinations of routing policies, and each policy combination is defined with the following information:

[0149] Whether to enable the injection strategy and its parameters (candidate set of intermediate nodes, maximum number of injection attempts, etc.);

[0150] Whether to adopt a publish / subscribe strategy and the granularity of topic segmentation;

[0151] Whether to adjust the parent node selection preference in the topology control strategy (e.g., prioritize high computing power nodes or low energy consumption nodes).

[0152] For example, the following combinations of strategies can be defined:

[0153] Combination A: Latency priority, enable injection strategy, and prioritize intermediate nodes with the fewest hops from the target node and the strongest computing power;

[0154] Combination B: Prioritize energy consumption, weaken injection strategy, adopt relatively short paths and avoid low-power nodes;

[0155] Combination C: Prioritize health, enable publish and subscribe strategies, synchronize intermediate results to multiple candidate inference nodes, and select the path with higher health.

[0156] Combination D: When the model health is too low, temporarily enable the centralized inference mode, inject data into the root node or edge server for re-inference.

[0157] 3. Online learning algorithms

[0158] In a preferred implementation, the online strategy synthesis module employs a multi-armed gambling machine algorithm:

[0159] Each strategy combination is considered as an "arm";

[0160] Maintain statistics for each arm in each task flow, including the number of times it was selected and the average reward estimate;

[0161] At each decision point, a trade-off is made between exploration and exploitation based on the upper confidence bound or an ε-greedy strategy:

[0162] With a certain probability, the strategy combination that has the largest current average reward estimate and the highest uncertainty is selected for the experiment;

[0163] In other cases, choose the strategy combination that is known to perform well.

[0164] 4. Reward Function Design

[0165] After executing a strategy combination, the online strategy synthesis module calculates an immediate reward value based on the task flow's performance within that period. The reward function can comprehensively consider the following factors:

[0166] Latency: The latency can be normalized and then negative, or scored in segments according to a threshold.

[0167] Packet loss rate: Convert the packet loss rate into a penalty item;

[0168] Energy consumption: Convert average energy consumption or the proportion of low-energy nodes on the path into a penalty item;

[0169] Model health: The health value is used as a positive term in the reward.

[0170] By assigning weights to the above items, an overall reward value is formed, which is used to update the statistics of the corresponding strategy combination. The weights can be configured differently according to the service quality descriptors of different task flows. For example, a latency weight can be added for real-time control tasks, and a health weight can be added for precision-sensitive inference tasks.

[0171] 5. Strategy Update Process

[0172] like Figure 5 As shown, the policy update process includes:

[0173] (1) Receive the latest environment state vector;

[0174] (2) Select the current strategy combination based on the online learning strategy;

[0175] (3) Generate policy configuration and distribute it to edge nodes;

[0176] (4) Collect runtime metrics and model health within an evaluation cycle;

[0177] (5) Calculate the immediate reward based on the reward function and update the statistics;

[0178] (6) Determine whether the strategy combination needs to be adjusted. If so, proceed to the next round of selection.

[0179] Through the above cyclical process, the system can gradually tend to select the strategy combination that performs better in the current environment during long-term operation.

[0180] Example 4: Distributed Edge Inference Scenario

[0181] See Figure 6 This embodiment provides an application example of the present invention in a distributed edge inference scenario.

[0182] 1. Scenario Assumptions

[0183] The system consists of multiple edge nodes with certain computing power, which jointly undertake a multi-layer neural network inference task. Each node is assigned neurons in different layers or branches. The input data is generated by several acquisition nodes, and the intermediate results are forwarded between the edge nodes in multiple hops.

[0184] 2. Strategy Configuration

[0185] For this inference task, the application service quality descriptor can be set as follows:

[0186] The latency level is "medium";

[0187] The energy consumption budget is "moderate";

[0188] Health score is weighted higher than latency and energy consumption.

[0189] The computational preference is "edge distributed".

[0190] The online policy synthesis module explores different policy combinations in the initial stage, such as different topic partitioning schemes and different sets of injection intermediate nodes. After running for a period of time, the module will gradually select the combination that can balance health and latency under the current network conditions.

[0191] 3. Health-driven routing adjustments

[0192] When the model health assessment module detects a decline in health, such as due to excessive load on some nodes or link deterioration, the online policy synthesis module automatically increases the exploration of health-sensitive paths, prioritizing intermediate nodes with more stable paths and lower packet loss rates, and even migrating some computations to other nodes. In this way, even when the network environment fluctuates, the overall accuracy of the inference task can be maintained as much as possible.

[0193] Example 5: Centralized Edge Reasoning Scenario

[0194] See Figure 7 This embodiment illustrates how the present invention drives injection routing through model health in a centralized edge inference mode.

[0195] 1. Scenario Assumptions

[0196] The system includes a high-performance edge server responsible for executing a complex inference task; other edge nodes handle only data acquisition and basic preprocessing. The task flow has relatively relaxed latency requirements, but the goal is to maintain high inference accuracy within a limited energy budget.

[0197] 2. Strategy Configuration

[0198] The application service quality descriptor can be set as follows:

[0199] The latency level is "low priority";

[0200] The energy consumption budget is set to "strict";

[0201] The weighting of health score is moderate.

[0202] The computational preference is "marginalized".

[0203] The online policy synthesis module tends to select policy combinations that prioritize centralized inference, have short injection paths, and avoid low-power nodes. Specifically:

[0204] Each acquisition node sends data packets to a group of candidate relay nodes through an injection strategy;

[0205] The relay nodes merge or compress the data before forwarding it to the edge servers;

[0206] The set of candidate relay nodes and path selection are dynamically adjusted based on changes in health status.

[0207] 3. Balance between health and energy consumption

[0208] When a decrease in model health is detected and there is no significant negative correlation with energy consumption, the online policy synthesis module can appropriately increase repeated inference or redundant paths to improve prediction stability. Conversely, when energy consumption is close to the preset upper limit and the health is still at a high level, the system can reduce the number of injected intermediate nodes or reduce the data sampling frequency to save energy.

Claims

1. A model health-driven perceptual routing method, characterized in that, The method, applied to a multi-hop edge IoT system consisting of multiple edge nodes, includes: Step S101: Configure a soft access point interface and a site interface on each edge node. The soft access point interface is used to create a wireless local area network and accept access from other edge nodes. The site interface is used to access the soft access point interface of the upper-level edge node. An active routing protocol is adopted to establish a next-hop routing table with the soft access point address of each edge node as the node identifier, thereby forming a multi-hop wireless network. Step S102: Set up a runtime performance monitoring module in the middleware layer of each edge node. The runtime performance monitoring module is used to periodically collect one or more runtime indicators, including end-to-end latency of each task flow, packet loss rate of link or path, remaining power of node, battery voltage or average transmit power, inference task queue length, and CPU utilization, and report the runtime indicators to the preset strategy synthesis node. Step S103: For task flows involving edge intelligent reasoning, the application layer or coordination node calculates a model health index to characterize the model state based on at least one of the following information: the reasoning results within a preset time window and the changes in reference labels, error rate, and confidence distribution. The model health index is a normalized value within a limited range. Step S104: In the middleware layer of the policy synthesis node, the application service quality descriptor of each task flow, the runtime metrics, and the model health metrics are combined to form an environment state vector. The application service quality descriptor includes at least one or more of the following: latency level, energy consumption budget, reliability level, privacy level, and computing preference. Step S105: In the policy synthesis node, a variety of application-aware routing policies are predefined, and different application-aware routing policies and their parameter configurations are combined into a discrete policy combination set, with each policy combination corresponding to a candidate routing behavior mode. Step S106: Set up an online policy synthesis module in the middleware layer of the policy synthesis node. The online policy synthesis module adopts an online learning algorithm to establish a mapping relationship between the environment state vector and the policy combination set. ； Step S107: The online strategy synthesis module, based on the expected reward estimate of each strategy combination, periodically or when a significant change in the environment state vector is detected, selects the strategy combination with the highest expected reward estimate as the target strategy combination for each task flow, generates the corresponding strategy configuration, and distributes the strategy configuration to the relevant edge nodes. Step S108: When sending or forwarding data packets, the policy execution agent of each edge node encapsulates, replicates, or imposes topological constraints on the data packets according to the currently effective policy configuration, and calls the active routing protocol to perform actual forwarding, thereby realizing adaptive optimization of perceptual routing behavior with model health as an important feedback signal.

2. The model health-driven perceptual routing method according to claim 1, characterized in that, The online learning algorithm employs a multi-armed gambling machine algorithm, treating each application-aware routing strategy combination as an arm. It maintains the selection frequency and expected reward estimate for each arm for each task flow. The online policy synthesis module selects a strategy combination for each task flow each time: Based on the upper confidence bound strategy or the ε-greedy strategy, a balance is struck between exploring underexperienced strategy combinations and estimating higher strategy combinations by taking advantage of the current expected reward; The statistics for the corresponding arm of the selected strategy combination are updated using the instant reward value to improve the expected reward estimate.

3. The model health-driven perceptual routing method according to claim 1, characterized in that, The establishment of a mapping relationship between the environment state vector and the policy combination set specifically includes: For each task flow, select one from the set of policy combinations as the current action given the environment state; After executing the selected strategy combination, the instant reward value is calculated according to the preset reward function, which is a weighted combination of end-to-end latency, packet loss rate, node remaining power and model health index. The statistics of the corresponding strategy combination are updated using the instant reward value, thereby adjusting the expected reward estimate and realizing online evaluation of the merits of different strategy combinations. The weights of the reward function are configured based on the latency level, energy consumption budget, reliability level, and computational preference in the application service quality descriptor, so that the latency weight is increased for task flows with high real-time requirements, and the model health weight is increased for task flows with high accuracy requirements.

4. The model health-driven perceptual routing method according to claim 1, characterized in that, The model health index is a normalized value within the range of 0 to 1, specifically including: The accuracy rate or equivalent accuracy index of the inference results within a preset time window is statistically analyzed. Based on the deviation between the accuracy rate and the target accuracy rate, as well as the degree of change in the confidence distribution, the health score is attenuated. The processed health value is cropped or mapped to fall within the range of 0 to 1.

5. The model health-driven perceptual routing method according to claim 1, characterized in that, The various application-aware routing strategies include at least: The intermediate node injection strategy is used to encapsulate the original data packet into an encapsulation structure with the outer destination address as the intermediate edge node and the inner destination address as the final destination node, and to select a set of intermediate edge nodes that meet the preset metric conditions according to the strategy configuration. The publish and subscribe strategy is used to implement multicast distribution of data streams based on topics, enabling multiple edge nodes subscribed to the same topic to receive intermediate results or control information simultaneously. The topology control strategy is used to adjust the parent-child relationship according to the parent node priority list in the strategy configuration when a new edge node joins the network or a parent node is reselected, thereby optimizing the overall network topology.

6. The model health-driven perceptual routing method according to claim 1, characterized in that, The strategy combinations in the strategy combination set include at least: A strategy combination aimed at latency optimization, wherein the strategy combination enables the intermediate node injection strategy and prioritizes intermediate edge nodes with fewer hops and stronger computing power. A strategy combination aimed at energy consumption optimization, wherein the strategy combination weakens the effect of the intermediate node injection strategy and preferentially avoids edge nodes with remaining power below a threshold in the topology control strategy; A strategy combination aimed at optimizing model health, wherein the strategy combination utilizes the publish and subscribe strategy to synchronize intermediate results among multiple edge nodes, and selects paths or inference nodes with higher health based on the model health index.

7. A model health-driven perceptual routing system, characterized in that, The system includes multiple edge nodes interconnected via a wireless network and at least one policy synthesis node. The edge nodes include: The wireless communication module is used to access the multi-hop wireless network through the soft access point interface and the site interface, and to interact with other edge nodes. The routing module is used to maintain a next-hop routing table identified by the soft access point address according to the active routing protocol, and to forward data packets according to the next-hop routing table; The runtime performance monitoring module is used to collect runtime metrics of this edge node and its carried task flow, and report the runtime metrics to the strategy synthesis node; The policy execution agent module is used to receive policy configurations from the policy synthesis node, encapsulate, copy, or impose topological constraints on data packets to be sent or forwarded to the target node according to the policy configurations, and call the routing module to complete the actual forwarding. The application interface module is used to receive the application service quality descriptor set by the application layer for the task flow, and send the application service quality descriptor to the policy synthesis node; The strategy synthesis node includes: The model health assessment module is used to calculate model health indicators to characterize the model status based on the inference results of edge intelligent inference tasks, as well as optional reference labels and statistical changes in prediction accuracy, error rate, and confidence distribution within a preset time window. The online policy synthesis module is used to receive the runtime metrics and model health metrics reported by each edge node, construct an environment state vector together with the application service quality descriptor, select target policy combinations from a predefined policy combination set for each task flow based on an online learning algorithm, and generate corresponding policy configurations and distribute them to the relevant edge nodes.

8. The model health-driven perceptual routing system according to claim 7, characterized in that, The specific operations for calculating model health indicators by the model health assessment module include: In labeled scenarios, the accuracy index is calculated based on the degree of matching between the inference results and the real labels within a preset time window, and the model health index is determined accordingly. In unlabeled scenarios, the model health index is determined based on at least one of the following: changes in the model output confidence distribution, the degree of drift in the input feature distribution, and model uncertainty estimation based on a self-supervised task.

Images