Wireless sensor network resource scheduling method and system for structural health monitoring
By employing deep reinforcement learning and a multi-objective nonlinear coupled reward function, the problem of multi-objective conflict in wireless sensor network resource scheduling under dynamic environments is solved, achieving network lifetime extension and efficient resource utilization, and is suitable for structural health monitoring scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINGDAO UNIV OF TECH
- Filing Date
- 2026-03-06
- Publication Date
- 2026-06-09
AI Technical Summary
Existing wireless sensor network resource scheduling methods are ill-suited to dynamic environments and cannot effectively balance multiple conflicting optimization objectives, leading to shortened network lifespan and resource waste.
By employing deep reinforcement learning methods, combining an Actor-Critic dual-network architecture and a multi-objective nonlinear coupled reward function, invalid actions are eliminated through an action masking mechanism, and hyperparameters are optimized using the PPO algorithm to achieve adaptive and efficient resource scheduling strategies.
It extends the network lifecycle, prevents premature node death, improves system stability and resource utilization efficiency, and adapts to the long-term monitoring needs of different structural health monitoring scenarios.
Smart Images
Figure CN122179910A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of wireless communication and artificial intelligence, specifically to a method and system for scheduling wireless sensor network resources for structural health monitoring. Background Technology
[0002] Wireless sensor networks (WSNs), as a key infrastructure of the Internet of Things (IoT), are widely used in environmental monitoring, smart agriculture, disaster early warning, and industrial automation. In recent years, WSNs have become the sensory nervous system for monitoring the health of infrastructure structures. With the expansion of network scale and the diversification of application needs, the scheduling of network resources such as spectrum, time slots, and energy faces significant challenges. Existing resource scheduling methods have shortcomings. Traditional scheduling strategies based on convex optimization or heuristic algorithms typically assume a static or known network environment, making it difficult to adapt to dynamic environments with time-varying channel states and random task arrivals, and the computational complexity increases exponentially with the number of nodes. Existing methods often focus on a single metric (such as maximizing throughput or minimizing energy consumption), making it difficult to simultaneously consider multiple conflicting optimization objectives such as latency, energy consumption, and reliability. Although deep reinforcement learning has been introduced into resource scheduling, traditional linear weighted summation reward functions struggle to handle the nonlinear coupling relationships between multiple objectives. For example, when battery power is extremely low, energy saving should have absolute priority, but linear weighting often leads agents to ignore node death risks in order to obtain throughput rewards, resulting in a shortened network lifetime.
[0003] Therefore, there is an urgent need for a wireless sensor network resource scheduling scheme that can adapt to dynamic environments, effectively balance multi-dimensional performance indicators, resolve multi-objective conflicts, and has efficient training and deployment capabilities, in order to meet the long-term, stable, and accurate monitoring needs in structural health monitoring scenarios. Summary of the Invention
[0004] The purpose of this invention is to solve the problems mentioned in the background art by proposing a wireless sensor network resource scheduling method and system for structural health monitoring.
[0005] The objective of this invention can be achieved through the following technical solutions:
[0006] In a first aspect, the present invention provides a wireless sensor network resource scheduling method for structural health monitoring, comprising the following steps:
[0007] S1: Initialize the wireless sensor network environment and the network parameters of the deep reinforcement learning model;
[0008] S2: Collect the current network state vector;
[0009] S3: Input the state vector into the Actor neural network, and remove invalid actions through the action masking mechanism before the Actor neural network outputs the action probability distribution;
[0010] S4: Based on the valid action output in step S3, a scheduling instruction is sent to the selected sensor node. After receiving the scheduling instruction, the sensor node performs data transmission operation.
[0011] S5: The wireless sensor network environment feeds back a reward value to the agent based on the transmission result of step S4;
[0012] S6: Store the experience tuples generated by the interaction of steps S2 to S5 into the experience replay pool. When the number of samples accumulates to a certain amount, use the PPO algorithm to update the gradients of the Actor and Critic networks.
[0013] In a preferred embodiment of the present invention, in step S1, the agent in the deep reinforcement learning model is constructed using the PPO algorithm and adopts an Actor-Critic dual-network architecture. The Actor neural network is used to output specific scheduling actions, and the Critic network is used to output the value assessment of the current state, calculate the advantage function, and guide the update direction of the Actor neural network, thereby completing the basic architecture initialization of the agent.
[0014] In a preferred embodiment of the present invention, step S1 further includes: pre-searching the key hyperparameters of the deep reinforcement learning model to establish the optimal initial training configuration, the specific process of which is as follows:
[0015] Define the search space and determine the hyperparameters to be optimized and their range;
[0016] A Gaussian process is constructed as a surrogate model to fit the mapping relationship between hyperparameters and model performance;
[0017] The acquisition function is used to select the next most promising combination of hyperparameters in the search space;
[0018] Evaluate its convergence performance by running a small number of training epochs using the selected hyperparameters;
[0019] The evaluation results are fed back to the Gaussian process to update the posterior distribution. When the maximum number of iterations is reached, the globally optimal combination of hyperparameters is output.
[0020] In a preferred embodiment of the present invention, in step S2, the state vector includes channel state information, remaining energy level, data queue length, and information age; the channel state information is the link quality from each sensor node to the gateway node; the remaining energy level is the remaining battery percentage of each sensor node; the data queue length is the number of data packets to be sent in the buffer of each sensor node; and the information age is the time elapsed since the last successful data upload by each sensor node.
[0021] In a preferred embodiment of the present invention, the specific process of eliminating invalid actions through the action masking mechanism in step S3 is as follows:
[0022] Generate an action mask vector based on the current liveness status of each sensor node;
[0023] The action mask vector is input into the action mask operation module of the Actor neural network for operation logic processing. For invalid actions, the Logits value is added to negative infinity, and for valid actions, the Logits value remains unchanged.
[0024] The masked Logits values are then input into the Softmax activation layer. The probability distribution output by the output layer is distributed only within the effective action space, i.e., the corrected effective actions are output. At the same time, the action mask vector, along with the state, action, and reward data, is stored in the experience replay pool.
[0025] In a preferred embodiment of the present invention, in step S4, the scheduling instruction includes the transmit power of each effective node, channel / subcarrier allocation, and time slot scheduling scheme.
[0026] In a preferred embodiment of the present invention, in step S5, the reward value is calculated using a multi-objective nonlinear coupled reward function, wherein the reward function is:
[0027]
[0028] Where ω1, ω2, and ω3 are the energy elasticity coefficients of each target, and γ is the coupling strength coefficient. For data freshness-related rewards, For energy-related rewards, For fairness-related rewards, For collaborative reward items, It is a small constant for regularization.
[0029] In a preferred embodiment of the present invention, in step S6, the empirical tuple includes the state vector S. t Effective Action A t Reward value R t The next state St+1 and action mask vector M t ;
[0030] The Critic network is based on the state vector S t Output value assessment V(S) t ), combined with the current reward value R t The advantage function is calculated to guide the Actor neural network to adjust the action probability distribution. At the same time, the action masking mechanism is combined to ensure that the updated network outputs only valid actions.
[0031] Secondly, the present invention provides a wireless sensor network resource scheduling system for structural health monitoring, comprising:
[0032] Sensor node layer: consists of several wireless sensor nodes;
[0033] Gateway node layer: includes at least one gateway node, used to aggregate data and execute scheduling instructions. The gateway node has a built-in trained deep reinforcement learning model, used to collect status information uploaded by the sensor node layer in real time, execute resource scheduling decisions, upload monitoring data and status information to the cloud platform layer through 4G / 5G network, and issue scheduling instructions to the sensor nodes.
[0034] Cloud platform layer: It communicates with the gateway node layer via 4G / 5G network, and is used to receive monitoring data and status data uploaded by the gateway node, and to store, analyze, visualize and display the data and provide early warning of anomalies; it is also used to remotely configure scheduling system parameters and monitor the operating status of the entire network.
[0035] User monitoring layer: Based on the mobile app monitoring platform, it is used to realize audio and video transmission, real-time sensor data display and early warning push functions, and also for historical data query.
[0036] Compared with the prior art, the beneficial effects of the present invention are:
[0037] 1. This invention overcomes the limitation of traditional linear weighted rewards in handling conflicting objectives by proposing a multi-objective coupled reward mechanism. This mechanism can adaptively adjust the priority of each optimization objective according to the real-time network status, effectively avoiding premature node death caused by blindly pursuing a single indicator (such as throughput) while ensuring service quality, thus extending the network lifecycle and adapting to the long-term monitoring needs of buildings.
[0038] 2. This invention optimizes the Actor neural network structure through an action masking mechanism, forcibly shielding invalid nodes with depleted energy and poor channel performance, avoiding waste of scheduling resources and training oscillations, improving model convergence speed and system stability, reducing training costs and real-time scheduling latency, and adapting to dynamic and complex network environments.
[0039] 3. This invention combines a Bayesian hyperparameter optimization strategy to achieve automatic optimization of DRL algorithm hyperparameters, eliminating the need for tedious manual debugging, improving model generalization and initial performance lower limit, and providing support for deployment in different structural health monitoring scenarios.
[0040] 4. The scheduling strategy of this invention based on deep reinforcement learning does not rely on a precise mathematical model. It learns through trial and error with the environment, which can effectively cope with the uncertainty of wireless channels and the randomness of task arrival. It can maintain high-efficiency resource scheduling performance in different structural health monitoring scenarios such as bridges, high-rise buildings, and robot training grounds. Attached Figure Description
[0041] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.
[0042] Figure 1 This is a distribution diagram of the network architecture in an embodiment of the present invention;
[0043] Figure 2 This is a flowchart of the resource scheduling method in an embodiment of the present invention;
[0044] Figure 3 This is a schematic diagram illustrating the interaction principle between the intelligent agent and the wireless sensor network environment in an embodiment of the present invention;
[0045] Figure 4 This is a schematic diagram of the weight change curve of the multi-objective nonlinear coupled reward function in an embodiment of the present invention;
[0046] Figure 5 This is a schematic diagram of the Actor neural network structure of the action masking mechanism in an embodiment of the present invention;
[0047] Figure 6 This is a flowchart of the hyperparameter optimization process based on Bayesian optimization in this embodiment of the invention. Detailed Implementation
[0048] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] Example 1
[0050] Please see Figure 1 As shown, the wireless sensor network resource scheduling system for structural health monitoring, taking the building structural health monitoring scenario as an example, includes a sensor node layer, a gateway node layer, a cloud platform layer, and a user monitoring layer.
[0051] Sensor Node Layer: This layer consists of several low-power wireless sensor nodes, specifically multiple sensor nodes deployed at critical structural locations within a building. These nodes, such as those deployed in key areas of target structures like bridges, high-rise buildings, and robot training grounds, are used to collect monitoring data on structural vibration, strain, cracks, and fatigue damage. This data is then wirelessly transmitted to the gateway node layer. The wireless sensor nodes feature sleep-wake modes and adaptive data transmission capabilities, and provide real-time feedback on their remaining energy and channel status to the gateway node layer, providing a foundation for multi-node collaborative monitoring. It should be noted that critical areas include main beams, piers, support piles, and the building facade.
[0052] Gateway node layer: Includes at least one gateway node (single gateway node for single-area deployment, multi-gateway node collaborative architecture for cross-area and large-scale deployment), used to aggregate data and execute scheduling instructions. The gateway node has a built-in trained deep reinforcement learning model, used to collect status information uploaded by the sensor node layer in real time, and to execute resource scheduling decisions. It uploads data to the cloud platform layer through 4G / 5G network, and issues scheduling instructions to the sensor nodes. At the same time, it realizes node management within the region and cross-area collaborative scheduling (under the multi-gateway node architecture), ensuring high efficiency and load balancing of data transmission.
[0053] Cloud platform layer: Connects to the gateway node layer via 4G / 5G network, used to receive monitoring and status data uploaded by the gateway nodes, and perform data storage, analysis, visualization, and anomaly warning; it can remotely configure scheduling system parameters (such as energy safety threshold E). safe The system can detect the coupling strength coefficient γ and monitor the overall network operation status, facilitating the addition of new sensor nodes and providing data support for structural health assessment and operation and maintenance decisions.
[0054] User Monitoring Layer: Serving as the interaction terminal between the system and users, this layer provides a convenient monitoring and control entry point, based on a mobile app monitoring platform. The mobile app monitoring platform supports audio and video transmission, real-time sensor data display, and early warning push notifications. It also supports historical data queries, enabling users to intuitively manage and respond to system anomalies, providing interactive support for operational and maintenance decisions.
[0055] Example 2
[0056] A wireless sensor network resource scheduling method for structural health monitoring, with a main process of closed-loop iterative operation, combines... Figures 2 to 6 As shown, the specific steps include:
[0057] Step S1: System initialization and hyperparameter optimization.
[0058] Initialize the wireless sensor network environment and the network parameters of the deep reinforcement learning model. The agent in the deep reinforcement learning model is constructed using the PPO algorithm, and an Actor-Critic dual network architecture is adopted (e.g., Figure 3 As shown), the Actor neural network (policy network) is responsible for outputting specific scheduling actions, while the Critic network (value network) is responsible for outputting the value assessment of the current state, used to calculate the advantage function and guide the update direction of the Actor neural network, completing the basic architecture initialization of the agent. Specifically, a Bayesian optimization module is set up to pre-search the key hyperparameters of the deep reinforcement learning model to establish the optimal initial training configuration. Its specific optimization process is as follows: first, define the search space, determine the hyperparameters to be optimized and their ranges (such as PPO shearing rate). 1∈[0.1,0.3]); secondly, a Gaussian process is constructed as a surrogate model to fit the mapping relationship between hyperparameters and model performance; then, the acquisition function is used to select the next most promising hyperparameter combination in the search space; subsequently, a small number of training rounds are run using the selected hyperparameters to evaluate its convergence performance; finally, the evaluation results are fed back to the Gaussian process to update the posterior distribution, and when the maximum number of iterations is reached, the globally optimal hyperparameter combination θ is output. * Used for formal training, including learning rate, γ, The process is as follows: Figure 6 As shown. Simultaneously, the model of the wireless sensor network resource scheduling system was constructed, clarifying the network environment composed of sensor nodes, gateway nodes, and communication links, laying the foundation for subsequent interactions.
[0059] Step S2: State awareness.
[0060] At each time slot t, the agent located at the edge gateway interacts with the wireless sensor network environment for the first time, collecting the current network state vector S. t (like Figure 3 As shown, the wireless sensor network environment feeds back the state vector S to the agent. t (As one of the core interactive data flows). The state vector specifically includes:
[0061] 1. Channel State Information (CSI): The link quality from each sensor node to the gateway node.
[0062] 2. Remaining Energy Level E: The percentage of remaining battery power in each sensor node.
[0063] 3. Data queue length Q: The number of data packets to be sent in the buffer of each sensor node.
[0064] 4. Information Age (AOI): The time elapsed since the last successful data upload by each sensor node, used to measure the freshness of the data.
[0065] The aforementioned state vector will serve as the core input for the agent's action decision-making, while also providing data support for subsequent action masking and reward calculation.
[0066] Step S3: Mask-based action selection.
[0067] The agent will collect the state vector S in step S2. t The input is fed into the Actor neural network, and before the Actor neural network outputs the action probability distribution, it is processed by an action masking mechanism (such as...). Figure 5 As shown in the figure, this action masking mechanism is used to eliminate invalid actions, avoid resource waste and training oscillations. Its specific implementation process is as follows: First, a binary action mask vector M is generated based on the current survival status of each sensor node (whether the remaining energy is sufficient and whether the channel quality meets the transmission requirements). t Let M t =[m t，1 m t,2 , ..., m t,i , ..., m t,n ], where m t,i ∈{0,1}, where n is the total number of sensor nodes. If sensor node i is out of power or in a dormant state, then m t,i =0 (marked as an invalid node); otherwise m t,i =1 (marked as a valid node);
[0068] Then the action mask vector M t Input the action masking module of the Actor neural network, according to The operation logic is processed, where This represents the Logits value of the raw output of the Actor neural network before it has been masked. This represents the corrected Logits value obtained after masking operations, for invalid actions (m t,i =0), its Logits value is added to negative infinity (using an extremely small negative number in the computer implementation), for valid actions (m t,i =1), its Logits value remains unchanged; then the masked Logits value is input into the Softmax activation layer, since the value of invalid actions is negative infinity, its calculated probability value Strictly equal to 0; the probability distribution of the final output layer is π(a|S). t The output is distributed only within the effective action space, meaning the output is a modified effective action A. t The effective action A t This will serve as the core instruction for subsequent resource scheduling, along with the action mask vector M. tIt will be stored in the experience replay pool along with data such as status, actions, and rewards for model updates.
[0069] Step S4: Perform scheduling and environment interaction.
[0070] The gateway node outputs the valid action A according to step S3. t Send scheduling instructions (such as) to the selected sensor nodes Figure 3 As shown, the agent outputs a valid action A to the environment. t As one of the core interactive data streams, the scheduling instructions specifically include the transmit power of each effective node, channel / subcarrier allocation, and time slot scheduling scheme. After receiving the scheduling instructions, the sensor nodes execute data transmission operations. As data transmission proceeds, the wireless sensor network environment changes accordingly, specifically manifested as a decrease in the remaining energy of each sensor node, changes in data queue length, fluctuations in channel state, and updates to data information age, ultimately transitioning to the next time-series state S. t+1 This completes the closed-loop interaction between the agent and the network environment, providing environmental feedback for subsequent reward calculations and model updates.
[0071] Step S5: Multi-objective coupled reward calculation.
[0072] Based on the transmission results from step S4, the wireless sensor network environment feeds back the reward value R to the agent. t (like Figure 3 As shown, the wireless sensor network environment feeds back the reward value R to the agent. t As one of the core interactive data flows), this invention utilizes a multi-objective nonlinear coupling reward function (such as... Figure 4 (As shown) Calculate this reward value to evaluate the current effective action A. t The advantages and disadvantages of traditional linear weighted rewards are analyzed to address the problem that traditional linear weighted rewards cannot handle multi-objective conflicts. The specific design and calculation process is as follows:
[0073] The reward function is defined as:
[0074]
[0075] Where ω1, ω2, and ω3 are the energy elasticity coefficients of each target, and satisfy ω1 + ω2 + ω3 = 1. γ is the coupling strength coefficient. For data freshness-related rewards, For energy-related rewards, For fairness-related rewards, For collaborative reward items, This is a small regularization constant used to prevent the calculation from being meaningless when the reward term is zero.
[0076] The core of this reward function lies in the state-dependent coupling adjustment mechanism, where the energy elasticity coefficient varies with the node's remaining energy E. t The change exhibits Sigmoid-type nonlinear curve characteristics: when the node's remaining energy E t Below the energy safety threshold E safe (like Figure 4 When the curve is in the low-position region (as shown by the dashed line), ω k →0, k=1, 2, 3; At this point, the system enters energy-saving protection mode, the weight of performance rewards is extremely low, and it is mainly driven by energy consumption penalties. The agent is forced to adopt a conservative strategy to extend the node's lifespan; when the node energy is sufficient (i.e., E... t >E safe When the curve rises rapidly and approaches 1, the system enters a performance-priority mode, encouraging high throughput and low latency transmission, making full use of energy to achieve high performance, and realizing a dynamic balance between node energy saving and transmission performance.
[0077] Step S6: Model update and iteration.
[0078] The empirical tuples (S) generated interactively from steps S2 to S5 are used to... t A t R t S t+1 M t The samples are stored in the experience replay pool. When a certain number of samples have accumulated, the PPO algorithm is used to update the gradients of the Actor and Critic networks. The Critic network updates the gradients based on the state vector S. t Output value assessment V(S) t ), combined with the current reward value R t The dominance function is calculated to guide the Actor neural network in adjusting the action probability distribution. Simultaneously, the action masking mechanism is used to ensure that the updated network outputs only valid actions. (e.g.) Figure 2 As shown, the update condition is determined. If the number of samples is greater than the number of samples used in each model update (Batch size), then the model is updated; otherwise, state awareness is performed. The model update includes collecting sample data, calculating the GAE advantage parameters, updating the PPO Actor-Critic parameters, and determining whether the maximum number of rounds has been reached for convergence. If yes, the optimal policy model is output; otherwise, the loop starts. The interaction and update process of steps S2 to S6 is repeated in each time slot t) until the model converges, that is, the reward value tends to stabilize and the performance of the resource scheduling policy (throughput, energy consumption, latency) reaches the preset target. At this time, the training of the deep reinforcement learning model is completed and it can be used for subsequent online resource scheduling deployment.
[0079] In summary, by adopting the above scheme, the method described in this invention demonstrates significant advantages over traditional polling scheduling and standard DQN algorithms in typical application scenarios such as structural health monitoring. These advantages include extending network lifetime, improving the capture rate of sudden events, and balancing network load. The method is suitable for the long-term and stable monitoring needs of building structural health monitoring scenarios and can be widely applied to various structural health monitoring scenarios. It has high engineering application value and broad market prospects.
[0080] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A wireless sensor network resource scheduling method for structural health monitoring, characterized in that, Includes the following steps: S1: Initialize the wireless sensor network environment and the network parameters of the deep reinforcement learning model; S2: Collect the current network state vector; S3: Input the state vector into the Actor neural network, and remove invalid actions through the action masking mechanism before the Actor neural network outputs the action probability distribution; S4: Based on the valid action output in step S3, a scheduling instruction is sent to the selected sensor node. After receiving the scheduling instruction, the sensor node performs data transmission operation. S5: The wireless sensor network environment feeds back a reward value to the agent based on the transmission result of step S4; S6: Store the experience tuples generated by the interaction of steps S2 to S5 into the experience replay pool. When the number of samples accumulates to a certain amount, use the PPO algorithm to update the gradients of the Actor and Critic networks.
2. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, In step S1, the agent in the deep reinforcement learning model is constructed using the PPO algorithm and adopts an Actor-Critic dual-network architecture. The Actor neural network is used to output specific scheduling actions, and the Critic network is used to output the value evaluation of the current state, calculate the advantage function, and guide the update direction of the Actor neural network, thus completing the basic architecture initialization of the agent.
3. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, Step S1 further includes: pre-searching the key hyperparameters of the deep reinforcement learning model to establish the optimal initial training configuration, the specific process of which is as follows: Define the search space and determine the hyperparameters to be optimized and their range; A Gaussian process is constructed as a surrogate model to fit the mapping relationship between hyperparameters and model performance; The acquisition function is used to select the next most promising combination of hyperparameters in the search space; Evaluate its convergence performance by running a small number of training epochs using the selected hyperparameters; The evaluation results are fed back to the Gaussian process to update the posterior distribution. When the maximum number of iterations is reached, the globally optimal combination of hyperparameters is output.
4. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, In step S2, the state vector includes channel state information, remaining energy level, data queue length, and information age; the channel state information is the link quality from each sensor node to the gateway node; the remaining energy level is the remaining battery percentage of each sensor node; the data queue length is the number of data packets to be sent in the buffer of each sensor node; and the information age is the time elapsed since the last successful data upload by each sensor node.
5. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, In step S3, the specific process of eliminating invalid actions through the action masking mechanism is as follows: Generate an action mask vector based on the current liveness status of each sensor node; The action mask vector is input into the action mask operation module of the Actor neural network for operation logic processing. For invalid actions, the Logits value is added to negative infinity, and for valid actions, the Logits value remains unchanged. The masked Logits values are then input into the Softmax activation layer. The probability distribution output by the output layer is distributed only within the effective action space, i.e., the corrected effective actions are output. At the same time, the action mask vector, along with the state, action, and reward data, is stored in the experience replay pool.
6. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, In step S4, the scheduling instructions include the transmit power of each valid node, channel / subcarrier allocation, and time slot scheduling scheme.
7. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, In step S5, the reward value is calculated using a multi-objective nonlinear coupled reward function, wherein the reward function is: ， Where ω1, ω2, and ω3 are the energy elasticity coefficients of each target, and γ is the coupling strength coefficient. For data freshness-related rewards, For energy-related rewards, For fairness-related rewards, For collaborative reward items, It is a small constant for regularization.
8. The wireless sensor network resource scheduling method for structural health monitoring according to claim 1, characterized in that, In step S6, the empirical tuple includes the state vector S t Effective Action A t Reward value R t The next state S t+1 and action mask vector M t ; The Critic network is based on the state vector S t Output value assessment V(S) t ), combined with the current reward value R t The advantage function is calculated to guide the Actor neural network to adjust the action probability distribution. At the same time, the action masking mechanism is combined to ensure that the updated network outputs only valid actions.
9. A wireless sensor network resource scheduling system for structural health monitoring, characterized in that... The wireless sensor network resource scheduling method for structural health monitoring, based on any one of claims 1-8, includes: Sensor node layer: consists of several wireless sensor nodes; Gateway node layer: includes at least one gateway node, used to aggregate data and execute scheduling instructions. The gateway node has a built-in trained deep reinforcement learning model, used to collect status information uploaded by the sensor node layer in real time, execute resource scheduling decisions, upload monitoring data and status information to the cloud platform layer through 4G / 5G network, and issue scheduling instructions to the sensor nodes. Cloud platform layer: It communicates with the gateway node layer via 4G / 5G network, and is used to receive monitoring data and status data uploaded by the gateway node, and to store, analyze, visualize and display the data and provide early warning of anomalies; it is also used to remotely configure scheduling system parameters and monitor the operating status of the entire network. User monitoring layer: Based on the mobile app monitoring platform, it is used to realize audio and video transmission, real-time sensor data display and early warning push functions, and also for historical data query.