System and method for deployment and performance optimization of a multi-gateway lorawan network

Abstract

This invention discloses a deployment and performance optimization system and method for multi-gateway LoRaWAN networks, belonging to the field of industrial wireless network technology. The system includes a gateway deployment module, a throughput optimization module, and an end-to-end latency optimization module, which form a collaborative closed loop through state awareness, parameter transmission, and dynamic feedback. The gateway deployment module determines gateway locations based on the spatial distribution of terminal nodes and guides the deployment scheme towards fewer gateways, wider coverage, and lower overlap through a comprehensive evaluation index (GUE). The throughput optimization module uses the A2C algorithm, with throughput as the reward, to output the optimal spreading factor and transmission power configuration. The end-to-end latency optimization module integrates an attention mechanism with a deep Q network, calculating attention weights based on device priority, queue length, and waiting time, and introducing action selection correction to achieve differentiated bandwidth allocation. This invention improves gateway utilization efficiency and network throughput while ensuring coverage and effectively reduces end-to-end latency.

Description

Technical Field

[0001] This invention relates to the field of industrial wireless network and Internet of Things (IoT) communication technology, specifically to a LoRaWAN (Low Power Wide Area Network) network system with a multi-gateway architecture, and particularly to a comprehensive performance improvement scheme for gateway layout strategy, air interface resource scheduling and data transmission latency optimization for this system, which is suitable for IoT application scenarios with high-density terminal access, wide coverage and multiple service priorities. Background Technology

[0002] With the large-scale deployment of Internet of Things (IoT) technology and the deepening application in vertical industries, the access of massive numbers of terminal devices has placed higher demands on communication technologies that offer low power consumption, wide coverage, and high reliability. As one of the mainstream technologies for Low Power Wide Area Networks (LPWAN), LoRaWAN has been widely used in smart cities, industrial monitoring, and agricultural IoT due to its advantages such as wide coverage, strong penetration, flexible networking, and low terminal power consumption.

[0003] In practical engineering deployments, to meet the demands for large-scale continuous coverage and system capacity, multi-gateway networking architecture has become an inevitable choice for large-scale LoRaWAN network deployments. However, with the increase in the number of gateways and the expansion of network scale, existing technologies mainly face the following technical challenges in engineering practice and resource scheduling: Gateway deployment strategy: The contradiction between coverage overlap and cost control. Existing deployment methods typically prioritize "eliminating coverage blind spots," tending to densely deploy gateways within planned areas. This approach not only leads to high hardware purchase and maintenance costs, but more seriously, excessive signal overlap exacerbates co-channel interference and signal collisions, reducing spectrum utilization efficiency. Although some existing solutions introduce coverage quality assessment metrics, their objective functions are often relatively simple, making it difficult to establish an effective comprehensive quantitative trade-off model between coverage rate, deployment cost, and signal overlap. Furthermore, most existing algorithms rely on pre-defined propagation models or grid searches, resulting in high computational complexity and failing to fully utilize the actual spatial distribution characteristics of terminal nodes, making it difficult to achieve "on-demand deployment" of gateway resources.

[0004] Throughput Optimization: Rigid Parameter Configuration and Insufficient Dynamic Adaptability. At the LoRaWAN physical layer, the spreading factor (SF) and transmit power (TP) of the terminal nodes directly determine the data transmission rate, communication distance, and anti-interference capability. However, existing throughput optimization methods largely rely on static preset rules or empirical formulas based on link budget (such as simply selecting the spreading factor adaptively based on the signal-to-noise ratio). These methods lack real-time awareness and dynamic response capabilities to the global network state (such as real-time gateway load, channel occupancy, and concurrent conflicts among multiple gateways). When network topology, service load, or channel conditions change dynamically, rigid parameter configurations easily lead to throughput limitations, failing to achieve an optimal balance between network capacity and coverage stability.

[0005] End-to-end latency assurance: Lack of differentiated service capabilities and crude queue management. In typical IoT business scenarios, different terminal nodes often carry business data of varying importance (such as critical alarm signals and routine meter reading data). However, existing latency optimization strategies typically employ a "best-effort" fair scheduling mechanism, failing to adequately differentiate between device priorities and real-time queue backlogs. With fixed bandwidth resources, this undifferentiated scheduling approach easily leads to high-priority data being blocked by low-priority background traffic, resulting in uncontrollable end-to-end latency and difficulty in guaranteeing the real-time transmission of critical data. While existing heuristic queue management algorithms are simple and easy to implement, they lack the ability to utilize historical experience and make adaptive decisions in complex dynamic environments.

[0006] Therefore, existing multi-gateway LoRaWAN network optimization solutions are relatively fragmented in terms of gateway deployment, throughput improvement, and end-to-end latency guarantee, lacking a systematic and comprehensive optimization solution that can consider the above three dimensions and adapt to large-scale dynamic network environments. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems by providing a deployment and performance optimization system and method for multi-gateway LoRaWAN networks, optimizing at three levels: gateway deployment strategy, throughput improvement, and end-to-end latency guarantee. At the gateway deployment strategy level, to address the issues of severe overlap and low gateway utilization in existing deployment strategies, a node density-based gateway deployment strategy is proposed to reduce costs and improve gateway efficiency (GUE). At the throughput optimization level, to address the problem that existing optimization strategies struggle to dynamically adjust spreading factors and transmission power based on network conditions, resulting in limited throughput, an A2C algorithm will be used for optimization. At the end-to-end latency optimization level, because existing optimization strategies cannot differentiate scheduling based on the priority and queue status of different devices, end-to-end latency is high, and the transmission time of important data is difficult to guarantee. This application, within a deep reinforcement learning framework, dynamically allocates bandwidth based on the priority and queue status of different devices to reduce queue backlog, thereby reducing end-to-end latency and ensuring the transmission of important data.

[0008] The technical solution of the present invention is as follows: A deployment and performance optimization system for a multi-gateway LoRaWAN network, comprising: The gateway deployment module is used to determine the gateway deployment location based on the spatial distribution of terminal nodes and output gateway topology information; The throughput optimization module is used to construct a state space based on the gateway topology information and output the spreading factor and transmission power configuration of each terminal node through the A2C algorithm. The end-to-end latency optimization module determines the service rate benchmark based on the spreading factor and transmission power configuration of each terminal node, through the deep integration of attention mechanisms. The network allocates bandwidth and sends feedback signals to the throughput optimization module to adjust decision weights when there is sustained high latency.

[0009] The above settings enable "on-demand deployment" of gateways, prioritizing their placement in densely populated node areas to maximize coverage with limited gateway resources and reduce hardware deployment costs. They also allow the configuration of spreading factors and transmission power to dynamically change with network conditions, enhancing the network's adaptability to dynamic environments and enabling refined, real-time management of network resources. Furthermore, they provide differentiated quality of service (QoS) assurance, giving high-priority devices an advantage in resource competition; simultaneously, a feedback mechanism enables the system to self-evolve, forming a closed-loop optimization. This transforms a static, well-planned layout into dynamic, superior performance, truly achieving a closed-loop improvement in overall network performance.

[0010] Furthermore, in the end-to-end latency optimization module, the attention weight is determined by a combination of device priority score, queue length score, and waiting urgency score; node Comprehensive attention score Calculate using the following formula: , in , , These are the weighting coefficients. , , They are nodes Priority score, queue length score, and waiting urgency score; The attention scores are then normalized using softmax to obtain the normalized weights for each node: .

[0011] The above settings enable the gateway to prioritize the nodes that most need service when allocating bandwidth resources, achieving optimal allocation of limited resources. Information from the three dimensions is quantified into a unified attention score, providing a quantitative basis for differentiated scheduling. The attention score is then transformed into a probability distribution, amplifying the difference between high and low scores to make the attention mechanism more effective.

[0012] Furthermore, the depth In the network's action selection, normalized weights are introduced. Value correction: The corrected action score is calculated using the following formula: , in For state Next action of value, For action Assigned to node bandwidth, For action Total bandwidth For nodes Normalized attention weights This is the penalty coefficient.

[0013] By explicitly introducing attention weights into the action selection process of deep reinforcement learning, the final decision takes into account both Q-value prediction and attention constraints, achieving a balance between Q-value optimization and attention constraints.

[0014] Furthermore, the gateway deployment module, throughput optimization module, and end-to-end latency optimization module form a three-level collaborative closed loop through state awareness, parameter transmission, and dynamic feedback: The gateway topology information output by the gateway deployment module serves as the initial channel reference in the state space of the throughput optimization module. The spreading factor and transmission power configuration output by the throughput optimization module are used as input parameters for calculating the service rate benchmark of each node in the end-to-end delay optimization module. When the end-to-end latency optimization module detects persistently high latency, it sends a feedback signal to the throughput optimization module. The throughput optimization module adjusts the decision weights of the spreading factor and transmission power based on the feedback signal, and feeds back the node density information of the latency anomaly area to the gateway deployment module to update the priority of gateway candidate points.

[0015] Through the above settings, the underlying deployment results directly provide topological constraints for upper-layer optimization, and a good gateway layout creates the channel condition foundation for throughput improvement. A causal relationship is established between physical layer parameter configuration and MAC layer queue scheduling, realizing a transmission chain of "parameter configuration → service rate → queue backlog prediction." This enables the system to possess an organic unity of "deployment setting the foundation, throughput increasing the upper limit, latency ensuring experience, and feedback promoting iteration," forming a complete closed-loop control. Compared with existing technologies that handle coverage planning, rate adjustment, and queue scheduling separately, the collaborative framework of this invention enables the network to achieve rapid joint response through cross-layer information transmission when encountering sudden traffic surges, channel degradation, or topology changes, achieving an overall optimization effect of 1+1+1>3.

[0016] Furthermore, in the throughput optimization module, the state space is defined as a multi-dimensional combination of the current throughput, the load status of each gateway, the channel occupancy rate, and the average signal-to-noise ratio of the nodes. The action space is defined as the combination of the spreading factor and the transmission power; The reward function is determined by throughput.

[0017] The above settings enable the model to comprehensively reflect the real-time state of the network, providing complete perceptual information for intelligent decision-making. Joint optimization of SF and TP is achieved, avoiding the performance limitations of adjusting a single parameter. The model makes decisions with the goal of maximizing throughput, ensuring a clear objective.

[0018] Furthermore, the A2C algorithm adopts a centralized control architecture, in which the network server collects the status information of each terminal node, makes unified decisions, and issues the optimal spreading factor and transmission power configuration to each terminal node. The algorithm interacts with the environment, selects an action, updates the node configuration, calculates the reward, and selects the optimal parameter configuration based on the reward value.

[0019] This enables centralized intelligent management and control, avoiding resource conflicts and suboptimal solutions caused by distributed decision-making. It allows the model to dynamically explore and converge to the optimal solution in complex and ever-changing environments through end-to-end learning.

[0020] Furthermore, in the gateway deployment module, the location with the highest node density in each sub-region is selected as the gateway candidate point; All gateway candidate points are grouped into candidate combinations, all combinations are traversed, and combinations that can cover more than 90% of terminal nodes are selected. Calculate the comprehensive evaluation index for each selected combination, and select the gateway combination with the highest index value as the deployment result.

[0021] Gateways are prioritized for deployment in densely populated node areas to maximize coverage efficiency with limited gateway resources. Under the hard constraint of achieving a preset high coverage rate, candidate selection is prioritized based on coverage.

[0022] Furthermore, the comprehensive evaluation index is GUE, which is calculated according to the following formula: , in The total area covered by the gateway. For the number of gateways, For intersection, union, and comparison, This is the overlap penalty coefficient.

[0023] It comprehensively reflects the overall performance of the deployment plan. The GUE increases with the total coverage area and decreases with the increase in the number of gateways and the degree of overlap, effectively guiding the optimization direction.

[0024] Furthermore, the sub-regions are divided as follows: , in Given the side length of the region, divide the region into... Sub-regions; For each sub-region, candidate gateway points are selected. If the terminal node coverage rate in a certain sub-region is less than 90%, the location with the second highest node density in that sub-region is selected as a supplementary candidate point until the overall terminal node coverage rate reaches more than 90%.

[0025] The algorithm adaptively determines the number of partitions based on region size, enabling it to adapt to various deployment scales from park-level to city-level while maintaining high computational efficiency. Through an iterative replacement mechanism, the number of gateways is minimized while meeting overall coverage constraints, achieving a balance between coverage and cost.

[0026] This application also includes a method for deploying and optimizing the performance of a multi-gateway LoRaWAN network, used in a system for deploying and optimizing the performance of a multi-gateway LoRaWAN network, comprising the following steps: Determine the gateway deployment location based on the spatial distribution characteristics of terminal nodes, and output gateway topology information; The state space is constructed based on the gateway topology information, and the spreading factor and transmission power configuration of each terminal node are determined by the A2C algorithm. The service rate benchmark is determined based on the spreading factor and transmission power configuration of each terminal node. Bandwidth is allocated through a deep Q network with a fused attention mechanism, and the decision weights for adjusting the spreading factor and transmission power are fed back during periods of sustained high latency.

[0027] Compared with existing technologies, the advantages of this invention are: 1. At the gateway deployment level, an effective balance between coverage efficiency and deployment cost is achieved. This application departs from the traditional approach of relying on dense deployment to eliminate blind spots, proposing a gateway deployment strategy based on the spatial distribution characteristics of terminal nodes. By designing a comprehensive evaluation metric, GUE, total coverage area, number of gateways, and coverage overlap (IoU) are incorporated into the same quantitative framework. This effectively guides the deployment scheme towards "fewer gateways, wider coverage, and lower overlap" while meeting the hard constraint of a preset high coverage rate. Compared to existing graphics-based deployment methods, the deployment strategy of this application has a lower coverage crossover ratio and higher gateway utilization efficiency, and exhibits good robustness in different experimental scenarios, thereby significantly reducing hardware deployment costs and the risk of co-channel interference.

[0028] 2. At the throughput optimization level, this application enhances the network's adaptability and decision-making stability in dynamic environments. It overcomes the limitations of traditional methods that rely on fixed signal-to-noise ratio formulas or local information by employing an A2C algorithm to jointly optimize the spreading factor and transmission power. This scheme uses a centralized control architecture to perceive the global network status in real time (including gateway load, channel occupancy, and node signal-to-noise ratio), enabling parameter configuration to dynamically adjust with network conditions. Experimental results show that this optimization strategy can make the spreading factor distribution of each node more uniform, effectively reducing the probability of packet collisions. Furthermore, when the network scale is large, its throughput performance is significantly better than traditional schemes based on signal-to-noise ratio adaptation. Simultaneously, the algorithm training process converges stably with small fluctuations, demonstrating its high reliability and practical deployment feasibility in complex dynamic environments.

[0029] 3. Regarding end-to-end latency protection, this application enhances differentiated service quality and the accuracy of resource allocation. Specifically designed for IoT scenarios with multiple service priorities, this application innovatively combines attention mechanisms with depth... Network integration. By comprehensively calculating multi-dimensional attention scores based on device priority, queue length, and waiting urgency, and utilizing Softmax normalization to generate resource allocation weights, bandwidth allocation decisions can proactively favor high-priority, high-backlog, and urgent data. Based on this, deep... Through experience replay and a stable network architecture, the network continuously learns the optimal allocation strategy from historical interactions. This mechanism not only effectively controls the queue backlog length of each node and significantly improves bandwidth resource utilization, but also, especially in large-scale network deployments, ensures that the end-to-end transmission latency of high-priority devices remains at a low level, enhancing the overall network latency performance stability and thus expanding the application boundaries of LoRaWAN technology in mission-critical IoT.

[0030] 4. At the system architecture level, a closed-loop collaborative global optimization system has been formed. This application organically integrates gateway deployment strategies, throughput optimization, and end-to-end latency optimization into a multi-layered collaborative framework. A reasonable gateway layout at the bottom layer provides good channel conditions and topology foundation for throughput and latency optimization at the upper layer; the improvement of throughput at the middle layer effectively alleviates network congestion and reduces the load pressure on latency optimization; and the differentiated bandwidth scheduling at the upper layer ensures the real-time performance of critical data from the business perspective. The three modules are interconnected and progressively advance, together forming a complete performance optimization closed loop for multi-gateway LoRaWAN networks, from physical topology planning to dynamic resource scheduling. Attached Figure Description

[0031] Figure 1 This application presents a multi-gateway LoRaWAN system architecture.

[0032] Figure 2 This is a schematic diagram of the gateway deployment strategy in this application.

[0033] Figure 3 This is a scenario diagram of the gateway coverage node in this application.

[0034] Figure 4 This is a comparison chart of the IoU metric in this application.

[0035] Figure 5 This is a comparison chart of the GUE metrics in this application.

[0036] Figure 6 This is a flowchart of the throughput optimization algorithm in this application.

[0037] Figure 7 A comparison chart showing the selection of SF nodes for this application.

[0038] Figure 8 This is a comparison chart of throughput values ​​for this application.

[0039] Figure 9 This is a schematic diagram of the delay optimization strategy in this application.

[0040] Figure 10 This is a comparison chart of the average queue length in this application.

[0041] Figure 11This is a comparison chart of bandwidth resource utilization in this application.

[0042] Figure 12 This is a comparison chart of delays for high-priority devices in this application.

[0043] Figure 13 This is a comparison chart of end-to-end delays in this application. Detailed Implementation

[0044] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0045] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0046] Please see Figure 1-13 A deployment and performance optimization system for multi-gateway LoRaWAN networks, such as Figure 1 As shown, it includes: The gateway deployment module is used to optimize gateway deployment performance. Throughput optimization module, used to improve throughput; End-to-end latency optimization module, used to reduce end-to-end latency.

[0047] In the gateway deployment module, existing deployment methods often over-deploy gateways to ensure no blind spots, resulting in high hardware costs and severe signal overlap. Therefore, a solution that can balance network performance and gateway deployment costs is needed to adapt to the current environment.

[0048] This application proposes a comprehensive optimization model that can simultaneously quantify cost, coverage, and overlap. The goal of this model is not simply to reduce the number of gateways, but rather to find an optimal balance between cost and performance while meeting the pre-set high coverage constraint. To this end, a comprehensive evaluation index, GUE, is proposed, calculated as follows: , in, This represents the total coverage area, i.e., the total range covered by all gateways; Indicates the number of gateways; It is the crossover ratio, which is the ratio of the overlapping area to the total coverage area. The smaller the value, the less overlap there is and the higher the gateway utilization. Under the same configuration, the larger the coverage overlap area, the higher the probability of signal collision. This is used to adjust the overlap penalty level; in this embodiment, the value can range from 0.1 to 10, which is suitable for most common deployment scenarios. If it is necessary to reduce the number of gateways, then... Choose the smaller value to reduce deployment costs; if it is necessary to reduce overlap to meet high coverage requirements, then... Take the larger value.

[0049] In terms of form, GUE increases with the total coverage area and decreases with the increase in the number of gateways and the degree of overlap. This characteristic allows GUE to effectively guide the algorithm towards optimization towards "fewer gateways, wider coverage, and lower overlap." Compared with existing gateway deployment evaluation metrics, GUE can comprehensively reflect the overall performance of the deployment scheme.

[0050] Regarding the selection of technical routes for gateway deployment strategies, existing methods have the following limitations: the objective function is singular, making it difficult to simultaneously consider the three dimensions of coverage, cost, and interference; the computational complexity is high, making it difficult to adapt to large-scale deployment scenarios; and it relies on a pre-set search space or model, lacking utilization of the actual distribution characteristics of nodes.

[0051] Based on the above analysis, an ideal gateway deployment strategy should meet the following requirements: First, it should fully utilize the spatial distribution characteristics of terminal nodes, prioritizing gateways in densely populated areas to achieve maximum coverage efficiency with limited gateway resources; second, while meeting coverage constraints, it should minimize the number of gateways and reduce coverage overlap to improve the GUE (Global Effectiveness) index; finally, the algorithm should have high computational efficiency and be able to adapt to various deployment scales from campus to city level. To address these requirements, this application designs a method based on node spatial distribution characteristics to optimize the deployment strategy.

[0052] The strategy aims to cover over 90% of terminals while selecting the gateway combination with the highest GUE value. Initially, settings are configured based on different region sizes; for larger regions, consideration is given to... equal parts, The calculation is as follows: , in, This indicates the side length of the region, such as when the side length of the region is 5km. If we choose 1, we consider dividing the region into 4 equal parts. For each small region, we use the algorithm once to select a candidate gateway location to try to cover the terminal nodes in that small region so that the coverage rate reaches more than 90%. If the terminal node coverage rate of that small region does not reach 90%, we continue to select the gateway location with the second highest node density in the surrounding area, and so on. As long as the number of terminal nodes in the entire region that are covered accounts for more than 90% of the total number of terminal nodes, it is acceptable.

[0053] In the throughput optimization module, existing methods lack adaptability to dynamic changes in network conditions, have a single optimization objective, struggle to simultaneously consider multiple metrics such as throughput and fairness, rely on preset rules, and lack the ability to learn from historical experience. Therefore, a solution is needed that can dynamically adjust multiple key parameters in real time to optimize the throughput of multi-gateway LoRaWAN networks.

[0054] Reinforcement learning offers a new approach to solving these problems. It does not rely on precise environmental models or pre-defined rules; instead, it learns optimal strategies through continuous interaction between the model and the environment, learning from trial and error. In multi-gateway LoRaWAN networks, the model can perceive the network state in real time, select SF (Service Provider) and TP (Transmission Provider) to adjust actions, and receive throughput rewards from the environment. The model continuously optimizes its decisions based on these reward signals. This learning mechanism naturally adapts to the dynamic changes in network states and can gradually converge to the optimum in complex environments.

[0055] Among numerous reinforcement learning algorithms, A2C belongs to the policy gradient method, which can directly optimize the policy in a continuous or high-dimensional action space, avoiding the difficulty of solving the maximization operation required in value function methods. Secondly, A2C combines the advantages of both policy networks and value function networks, introducing an advantage function to effectively reduce the variance of the policy gradient, making the training process more stable. Therefore, this application chooses the A2C algorithm to optimize throughput.

[0056] By constructing an environment that includes nodes, gateways, spreading factors, and transmission power options, and employing a centralized control architecture, a network server collects state information, makes unified decisions, and distributes optimal parameter configurations to each node. In this process, the state space is defined as a multi-dimensional combination of current throughput, gateway load, channel occupancy, and average signal-to-noise ratio of nodes to reflect the real-time state of the network. The action space is defined as a combination of spreading factors and transmission power, and the reward function is determined by throughput.

[0057] The algorithm optimizes node configuration by continuously interacting with the environment. During training, it updates the node configuration and calculates rewards after selecting an action, selecting the optimal parameter configuration based on all reward values. Finally, the algorithm applies the trained model to the actual network to achieve optimal node parameter configuration, thereby improving the throughput of the entire multi-gateway LoRaWAN network.

[0058] In end-to-end latency optimization modules, existing methods lack awareness of the overall network state. Their optimization objectives are singular, making it difficult to balance latency optimization, priority assurance, and resource utilization efficiency. Relying on preset rules or local information, they cannot learn from historical experience. Therefore, designing a latency optimization mechanism that can distinguish device priorities and comprehensively consider multi-dimensional information is of significant practical importance for reducing latency in multi-gateway LoRaWAN networks.

[0059] Deep reinforcement learning offers a novel approach to solving these problems. Unlike existing methods, deep reinforcement learning models the bandwidth allocation problem as a Markov decision process, approximating it through neural networks. The value function can automatically extract effective features from high-dimensional state inputs. The model learns the optimal bandwidth allocation strategy through continuous interaction with the environment. The environment provides reward signals related to latency, and the model continuously optimizes its decisions based on these rewards. This approach can adapt to dynamic changes in network state and achieve a balance among multiple optimization objectives. Simultaneously, the introduction of an attention mechanism provides prior guidance for differentiated services. By calculating the priority score, queue length score, and waiting urgency score of each node, attention weights are generated to guide bandwidth allocation towards higher-priority devices. Based on the above analysis, this application adopts a fusion of attention mechanism and deep learning. Network solutions to optimize end-to-end latency.

[0060] The purpose of the attention mechanism is to enable the gateway to focus on the nodes that most need service when allocating bandwidth resources, thereby achieving optimal allocation of limited resources. For each terminal node, its attention weight is determined by a combination of three factors: device priority score, queue length score, and waiting urgency score.

[0061] The device priority score reflects the importance of the services carried by the node, the queue length score reflects the data backlog of the node, and the waiting urgency score reflects the waiting time of data packets in the queue.

[0062] Define nodes Comprehensive attention score As shown in the following calculation formula: , Among them, the weighting coefficient , , To balance the importance of the three factors, optimization was performed based on practical application needs and multiple experiments. To transform the attention score into a normalized weight that can be used for resource allocation, this application employs the softmax normalization method. This method maps any real-valued score vector to a probability distribution, while amplifying the difference between high and low scores, making the attention mechanism more significant. The softmax normalization calculation is shown below: , Based on the attention mechanism that calculates the weights of each node, this application employs depth... The network implements specific bandwidth allocation decisions.

[0063] depth The network architecture uses a three-layer fully connected neural network. The input layer has 8 neurons corresponding to an 8-dimensional state vector, and the hidden layer... The first layer contains 128 neurons using the ReLU activation function, the second hidden layer contains 64 neurons using the ReLU activation function, and the output layer has 200 neurons corresponding to the Q-values ​​of 200 candidate actions. The use of the ReLU activation function helps alleviate the vanishing gradient problem and accelerates network convergence.

[0064] To improve training stability, an experience replay technique is employed, with the experience replay pool capacity set to 10,000 records to ensure sufficient historical experience is stored for the model to learn from. A discount factor of 0.99 is also set to reflect a moderate emphasis on future rewards.

[0065] In the action selection phase, an α-greedy strategy is employed to balance exploration and utilization. The initial exploration rate is set to 1.0, gradually decreasing to 0.01 as training progresses. This allows the model to fully explore the action space in the early stages and focus on utilizing learned knowledge in the later stages. More importantly, this application introduces attention weights into the action selection process, ensuring that the final action selection is balanced. The modified action score calculation formula, based on value prediction and attention constraints, is shown below: , Among them, the penalty coefficient Controlling the intensity of attentional constraints, The larger the value, the more the final action tends to conform to the attention distribution. This application takes... Equal to 0.5, in A balance must be struck between value optimization and attention constraints.

[0066] The decision-making process can be summarized in the following steps: First, the model perceives the current network state, which is obtained through periodic collection by the gateway or reporting by terminal nodes. Second, the attention weights of each node are calculated based on the attention mechanism. Then, candidate bandwidth allocation schemes are selected from the predefined action space, and the current state is input into the depth... The network obtains each candidate action Value prediction. Then, attention weights are used to predict the value of each candidate action. The value is adjusted, and the action with the highest score after adjustment is selected as the action to be executed. If it is in the exploration phase, an action is randomly selected to collect more experience. Then, the selected action is mapped to a specific bandwidth allocation configuration, and each node is notified of its allocated bandwidth resources through downlink control commands. The nodes adjust their sending behavior according to the allocation results.

[0067] After configuring, observe the new network state and reward value, and store the experience of this interaction in the experience pool. Finally, periodically sample small batches of samples from the experience pool for training the depth. The network parameters are updated to gradually approach the optimal value. Value function.

[0068] Calculate the current during training Values ​​and Objectives The mean squared error of the values ​​is used as the loss function, and the network weights are updated through gradient backpropagation. After a certain number of steps, the parameters of the main network are synchronized to the target network. Through repeated iterations of the above process, the model continuously optimizes the bandwidth allocation strategy in continuous interaction with the environment, and finally converges to the optimal strategy that can effectively reduce end-to-end latency.

[0069] In the specific implementation process described above in this application, it should be particularly noted that the gateway deployment module, throughput optimization module, and end-to-end latency optimization module are not independent and isolated components, but rather form a tightly coupled collaborative optimization closed loop through a "state awareness-parameter transmission-dynamic feedback" mechanism. The following provides a detailed explanation of the three-layer collaborative linkage relationship in this application: (I) Physical layer constraints and support for throughput optimization of gateway topology The output of the gateway deployment module (i.e. the final spatial distribution of the gateway) is not only used for static coverage evaluation, but also serves as a priori environmental parameter for the throughput optimization module, deeply participating in the state space construction of the A2C algorithm.

[0070] Specifically, after determining the gateway location, the gateway deployment module calculates the signal attenuation factor from each terminal node to its home gateway and the interference topology matrix between multiple gateways. This matrix, along with the overlap coverage information reflected by the GUE metric, is encoded into the initial baseline values ​​of "channel occupancy" and "load status of each gateway" in the A2C algorithm's state space. This deep integration of physical topology and algorithm input enables the A2C model to inherently perceive the channel capacity limit and interference risk brought about by the deployment topology when making joint decisions on the spreading factor (SF) and transmit power (TP).

[0071] For example, in sub-regions with high gateway density but low overlap, the model is incentivized to use a lower spreading factor to pursue higher throughput; while in gateway coverage edge regions, the model automatically tends to prioritize link connectivity, thus achieving joint optimization of "topology-link" at the global level, rather than simply local link optimization.

[0072] (ii) Causal transmission of parameter configuration and queue service capabilities The terminal node SF / TP configuration determined by the throughput optimization module constitutes the dynamic service rate benchmark in the end-to-end latency optimization module. This synergistic effect is achieved through a causal chain between the physical layer and the MAC layer: The optimal SF / TP combination output by the A2C algorithm directly determines the theoretical data transmission capacity (i.e., over-the-air transmission rate) of each terminal node per unit time. This rate information is synchronized in real time to the end-to-end latency optimization module and acts as a dynamic correction factor in the calculation of the "queue length score" in the attention score. Specifically: Positive drive: When the A2C model selects a higher-order SF (low rate) for a node, the latency optimization module predicts that the risk of data backlog for that node will increase. The attention mechanism automatically increases the urgency weight of that node, guiding bandwidth allocation to tilt towards it, thereby preventing queue overflow caused by slow physical layer transmission.

[0073] Reverse mitigation: When the A2C model selects a lower-order SF (high rate) for a node, the delay optimization module correspondingly reduces the queue backlog score weight of that node, releasing valuable bandwidth resources to other nodes with limited physical channels.

[0074] This cross-layer information transmission transforms throughput optimization and latency optimization from a "serial sequential relationship" to a "parallel predictive relationship," significantly improving the foresight of resource scheduling.

[0075] (iii) Delayed state feedback driven dynamic reconfiguration closed loop This application does not involve one-way topology and parameter transfer; it also includes a reverse feedback adjustment mechanism from the delay optimization module to the underlying module, forming a complete closed-loop control.

[0076] During its interaction with the environment, Deep Q Network (DQN) not only performs bandwidth allocation actions but also continuously monitors the real-time changes in the actual clearing speed of each node's queue and the end-to-end latency. When the system detects that multiple nodes in a certain area are experiencing persistent high latency and the queue backlog cannot be alleviated (i.e., the attention mechanism has repeatedly pushed the weights of nodes in this area to a high level, but the latency still does not converge), the system triggers an abnormal state report.

[0077] Based on this feedback, the network server determines that the current network capacity configuration of the region can no longer meet the business needs, and then executes the following two-level dynamic reconfiguration logic: Parameter-level rollback: Instructs the throughput optimization module to appropriately reduce the throughput reward weight for nodes in this area during the current period, and shift to a more robust SF / TP combination to prioritize communication success rate rather than absolute rate, thereby assisting the delay module in reducing the additional latency caused by retransmission.

[0078] Strategy-level optimization: If the above operations still fail to improve the situation, the system will record the long-term service outage status of the area and save the statistical data to the network operation and maintenance database. When the next periodic gateway deployment strategy update cycle arrives, this data will be used as a weighted correction item for "historical node density" to guide the gateway deployment module to adjust the selection priority of gateway candidate points (i.e., add a latency penalty item to the node density score, tending to redundantly deploy microgateways in the area).

[0079] (iv) Overall technical significance of synergy Through the aforementioned three-layer collaborative mechanism, this application achieves an organic unity of "deployment as the foundation, throughput as the upper limit, latency as the guarantee of experience, and feedback as the driving force for iteration." Compared with the existing technologies that handle coverage planning, rate adjustment, and queue scheduling in a fragmented manner, the collaborative framework of this application enables the network to achieve rapid joint response through cross-layer information transmission when encountering sudden traffic surges, channel degradation, or topology changes. Moreover, the decision of each module is based on the effective prior information provided by other modules, thereby achieving an overall optimization effect of 1+1+1 greater than 3.

[0080] In this embodiment, gateway deployment strategy optimization serves as the underlying foundation, and its deployment results directly provide network topology constraints for subsequent throughput and latency optimization. A reasonable gateway layout can provide good channel conditions for terminal nodes, creating a foundation for throughput improvement, while reducing propagation delay and collision probability, thus laying the groundwork for latency assurance.

[0081] Specifically, the gateway deployment strategy module based on node density analyzes the distribution of terminal nodes to obtain gateway deployment results, and then deploys the gateway based on the obtained deployment results.

[0082] like Figure 2 The diagram illustrates a gateway deployment strategy. The algorithm steps are as follows: First, candidate gateway locations need to be selected. When the area is small, a single gateway can achieve full coverage; when the area is large, consider deploying the gateway to a larger area. Divide the region into four equal parts. When the region's side length is 10 km, divide the region into four equal parts. Select the location with the highest node density in each region as a candidate gateway point, ensuring that the node coverage of each region is above 90%. Otherwise, select the location with the second highest node density and place the gateway there, and so on. After selecting the candidate gateway locations, combine all the gateways into various possible combinations. Iterate through all combinations, and any combination that covers more than 90% of the terminal nodes can proceed to the next round. Next, calculate the GUE value for each gateway combination and select the gateway combination with the highest GUE value.

[0083] like Figure 3 The diagram shows a scenario where a gateway covers the nodes. With a side length of 10 km, a single gateway cannot cover most nodes. Therefore, the area is divided into four equal parts, and a new deployment strategy is implemented in each sub-region to achieve a coverage rate of over 90%. At this point, almost all nodes are covered by at least one gateway, demonstrating good coverage.

[0084] This application selects suitable gateway locations based on the node distribution within a square region with side lengths varying between 1 and 16 km. After ensuring terminal node coverage of over 90%, it selects the gateway combination with the highest GUE value to achieve better coverage. The application also compares its self-designed gateway deployment strategy with a graph-based gateway deployment method.

[0085] like Figure 4 and Figure 5 The graphs show a comparison of IoU and GUE, respectively. It's evident that the node density-based method has a smaller intersection-union ratio (IoU) in gateway coverage compared to the graph-based method, resulting in higher gateway utilization and reduced deployment costs. Furthermore, the former exhibits a better GUE metric, demonstrating superior coverage. Statistical analysis shows that the standard deviation across multiple experiments is within 2%, validating the algorithm's robustness. Further independent samples... The test results showed that the difference between the two strategies in terms of GUE was statistically significant. This indicates that the advantage of strategies based on node density is not caused by random fluctuations. Furthermore, to verify the effectiveness of the indicator... Sensitivity, compared GUE values ​​of 0.1 and 10 were used. The results show that when... When the GUE is 10, the two methods are compared. Both decrease when the weight is 0.1, but the performance difference between the two remains unchanged, indicating that the conclusion is robust under different weights. In addition, when the terminal nodes are randomly distributed, the final GUE and IoU are similar to those when the nodes are uniformly distributed.

[0086] In this embodiment, the throughput optimization effect is directly affected by the quality of gateway deployment. A good gateway layout can provide better channel conditions for terminal nodes, enabling the use of lower spreading factors and thus increasing the throughput ceiling. At the same time, the decision results of the throughput optimization strategy will also affect the network load distribution and congestion status, thereby affecting end-to-end latency.

[0087] Specifically, such as Figure 1 As shown, the SF and TP selector modules generate all possible combinations of SF and TP. After the A2C model interacts with the environment for training, it compares the throughput reward values ​​under different SF and TP combinations to determine the optimal SF and TP parameter combination. The selected parameter combination is then sent to the terminal node via the LinkADRReq command to maximize network throughput.

[0088] like Figure 6 The diagram illustrates the flow of the throughput optimization algorithm. This algorithm breaks away from the dependence on fixed models and local information in traditional methods, endowing the network with an autonomous evolutionary capability through an end-to-end learning approach. The model can dynamically explore in complex and ever-changing environments and converge to the optimum. This not only enables refined and real-time management of network resources, but more importantly, as the dynamic execution layer in the collaborative optimization framework, it can fully explore and utilize the performance potential provided by the optimization deployment scheme in the previous stage, ultimately transforming a static, well-designed layout into dynamic, superior performance, thereby truly achieving a closed-loop improvement in the overall network performance.

[0089] This application compares the throughput of four methods: the Aloha scheme, an adaptive algorithm for selecting the spreading factor based on the signal-to-noise ratio, the enumeration method, and A2C. All experiments were independently repeated 10 times, and the results were recorded and the standard deviation was considered to reflect the average performance and stability of the methods.

[0090] like Figure 7 The chart shows a comparison of node selection SF (Signal Provider) values. It can be seen that the Aloha access scheme does not select SF values ​​evenly, and some nodes are not covered, resulting in lower throughput. The bars on the right represent the results of parameter settings using the A2C algorithm. The SF distribution is more even rather than concentrated at a single SF value, effectively reducing the probability of collisions, and all nodes are covered, leading to higher throughput.

[0091] like Figure 8The diagram shows a throughput comparison of four methods. To verify the stability of the results and ensure they are not caused by random factors, the average of 10 experiments was used, with a standard deviation within 3%. While keeping other factors constant, the method of parameter selection was varied. The Aloha-based access scheme tends to select higher power and lower spreading factors, while the method of adaptively selecting the SF based on the signal-to-noise ratio only considers the spreading factor parameter. The A2C algorithm considers both the spreading factor and transmission power, aiming to maximize throughput. It achieves better results in a shorter time. As training progresses, the average reward steadily increases, converging after approximately 300 rounds with small fluctuations, indicating good training stability. Furthermore, an enumeration method was attempted in a less complex network environment to obtain the optimal solution.

[0092] As shown in the figure, except for the Aloha-based access scheme, the other three schemes all achieved good throughput values ​​when the area side length was 1–6 km. When the area side length was greater than 9 km, the A2C algorithm outperformed the method of adaptively selecting the spreading factor based on the signal-to-noise ratio and the traditional method. Further independent samples... The test results show that, under the condition that the region side length is greater than 9 km, the throughput difference between the A2C algorithm and the comparison method is statistically significant. This verifies that the performance advantage of the A2C algorithm in this scenario is not caused by random fluctuations.

[0093] In this embodiment, latency optimization serves as an upper-layer quality of service (QoS) assurance mechanism, and its effectiveness is influenced by gateway deployment and throughput optimization. A well-planned gateway layout can reduce propagation latency and collision probability, creating favorable conditions for latency optimization. Throughput optimization, by adjusting the spreading factor, reduces transmission latency, complementing latency optimization. Simultaneously, an attention mechanism is introduced into latency optimization.

[0094] The DQN model with attention mechanism dynamically learns bandwidth allocation strategies based on device priority and queue status. The model selects the optimal bandwidth allocation strategy based on the current network status, and the environment provides reward signals related to end-to-end latency. Through iterative training, the model minimizes transmission latency.

[0095] like Figure 9 The diagram illustrates the latency optimization strategy. First, by introducing an attention mechanism, differentiated service quality assurance is achieved, enabling high-priority devices to gain an advantage in resource competition and expanding the application boundaries of LoRaWAN technology in mission-critical domains. Second, deep reinforcement learning is applied to bandwidth allocation decisions, giving the network the ability to adapt to dynamic environments. This allows for continuous optimization of allocation strategies under complex and changing network conditions, overcoming the limitations of traditional heuristic algorithms that rely on fixed rules and struggle to handle sudden traffic surges.

[0096] This application comprehensively examines the performance of two algorithms under different network ranges by focusing on four core indicators: average queue length, resource utilization, high-priority device latency, and average end-to-end latency. Furthermore, it reveals the optimization mechanism of the fusion attention mechanism algorithm through the analysis of the intrinsic relationship between these indicators.

[0097] like Figure 10 The chart shows a comparison of average queue lengths. The algorithm incorporating the attention mechanism significantly outperforms the baseline algorithm across all network ranges, and its advantage increases with the network size. The percentage reduction in queue length gradually increases from 7.7% at 1km to 40.6% at 16km, with a maximum reduction of 40.8%. This indicates that depth... With its powerful function approximation ability, the network can extract effective features from a high-dimensional continuous state space. At the same time, the attention mechanism gives the algorithm priority awareness, enabling it to accurately identify key nodes and allocate resources accordingly, thereby achieving effective control over queue backlog.

[0098] like Figure 11 The diagram shows a comparison of bandwidth resource utilization. The algorithm incorporating the attention mechanism achieves higher utilization than the baseline algorithm across all network scales, with the absolute improvement increasing from 4 percentage points at 1km to 17 percentage points at 16km. The baseline algorithm's utilization rate decreases by 1.75 percentage points per kilometer, while the algorithm in this application only decreases by 0.94 percentage points per kilometer, demonstrating its stronger robustness in large-scale networks. This improved resource utilization directly translates to increased effective throughput. Efficiently utilizing bandwidth resources means that more data can be transmitted in the same amount of time, reducing queue backlog.

[0099] like Figure 12 The image shows a latency comparison for high-priority devices. The high-priority latency of the algorithm incorporating the attention mechanism is consistently more than 60% lower than that of the baseline algorithm, with a maximum difference of 236ms. This capability enables multi-gateway LoRaWAN networks to not only meet the needs of large-scale data acquisition but also support real-time communication for critical tasks, expanding the application boundaries of low-power wide-area network technology.

[0100] like Figure 13 The diagram shows an end-to-end latency comparison. The deep reinforcement learning algorithm incorporating the attention mechanism exhibits significant latency optimization, with the average latency increasing from 105ms at 1km to 250ms at 16km, a much lower absolute value than the baseline algorithm's 390ms. Beyond 10km, the latency growth curve flattens out considerably, increasing slowly from 211ms to 250ms, indicating that the algorithm has learned effective strategies for adapting to larger networks. Furthermore, the latency variance of this algorithm is significantly lower than that of the baseline algorithm, demonstrating stronger stability in its latency performance.

[0101] Compared with existing technologies, the three optimization schemes proposed in this application provide new ideas and methods for performance optimization of multi-gateway LoRaWAN networks from three aspects: gateway deployment, throughput improvement, and end-to-end latency guarantee.

[0102] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A deployment and performance optimization system for a multi-gateway LoRaWAN network, characterized in that, include: The gateway deployment module is used to determine the gateway deployment location based on the spatial distribution of terminal nodes and output gateway topology information; The throughput optimization module is used to construct a state space based on the gateway topology information and output the spreading factor and transmission power configuration of each terminal node through the A2C algorithm. The end-to-end latency optimization module determines the service rate benchmark based on the spreading factor and transmission power configuration of each terminal node, through the deep integration of attention mechanisms. The network allocates bandwidth and sends feedback signals to the throughput optimization module to adjust decision weights when there is sustained high latency.

2. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 1, characterized in that, In the end-to-end latency optimization module, the attention weight is determined by a combination of the device priority score, queue length score, and waiting urgency score. node Comprehensive attention score Calculate using the following formula: , in , , These are the weighting coefficients. , , They are nodes Priority score, queue length score, and waiting urgency score; The attention scores are then normalized using softmax to obtain the normalized weights for each node: 。 3. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 2, characterized in that, The depth In the network's action selection, normalized weights are introduced. Value correction: The corrected action score is calculated using the following formula: ， in For state Next action of value, For action Assigned to node bandwidth, For action Total bandwidth For nodes Normalized attention weights This is the penalty coefficient.

4. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 1, characterized in that, The gateway deployment module, throughput optimization module, and end-to-end latency optimization module form a three-level collaborative closed loop through state awareness, parameter transmission, and dynamic feedback. The gateway topology information output by the gateway deployment module serves as the initial channel reference in the state space of the throughput optimization module. The spreading factor and transmission power configuration output by the throughput optimization module are used as input parameters for calculating the service rate benchmark of each node in the end-to-end delay optimization module. When the end-to-end latency optimization module detects persistently high latency, it sends a feedback signal to the throughput optimization module. The throughput optimization module adjusts the decision weights of the spreading factor and transmission power based on the feedback signal, and feeds back the node density information of the latency anomaly area to the gateway deployment module to update the priority of gateway candidate points.

5. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 1, characterized in that, In the throughput optimization module, the state space is defined as a multi-dimensional combination of the current throughput, the load status of each gateway, the channel occupancy rate, and the average signal-to-noise ratio of the nodes. The action space is defined as the combination of the spreading factor and the transmission power; The reward function is determined by throughput.

6. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 5, characterized in that, The A2C algorithm adopts a centralized control architecture, in which the network server collects the status information of each terminal node, makes unified decisions, and issues the optimal spreading factor and transmission power configuration to each terminal node. The algorithm interacts with the environment, selects an action, updates the node configuration, calculates the reward, and selects the optimal parameter configuration based on the reward value.

7. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 1, characterized in that, In the gateway deployment module, the location with the highest node density in each sub-region is selected as the gateway candidate point; All gateway candidate points are grouped into candidate combinations, all combinations are traversed, and combinations that can cover more than 90% of terminal nodes are selected. Calculate the comprehensive evaluation index for each selected combination, and select the gateway combination with the highest index value as the deployment result.

8. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 7, characterized in that, The comprehensive evaluation index is GUE, which is calculated using the following formula: ， in The total area covered by the gateway. For the number of gateways, For intersection, union, and comparison, This is the overlap penalty coefficient.

9. The deployment and performance optimization system for a multi-gateway LoRaWAN network according to claim 8, characterized in that, The sub-regions are divided as follows: ， in Given the side length of the region, divide the region into... Sub-regions; For each sub-region, candidate gateway points are selected. If the terminal node coverage rate in a certain sub-region is less than 90%, the location with the second highest node density in that sub-region is selected as a supplementary candidate point until the overall terminal node coverage rate reaches more than 90%.

10. A method for deploying and optimizing the performance of a multi-gateway LoRaWAN network, characterized in that, A deployment and performance optimization system for a multi-gateway LoRaWAN network as described in any one of claims 1-9 includes the following steps: Determine the gateway deployment location based on the spatial distribution characteristics of terminal nodes, and output gateway topology information; The state space is constructed based on the gateway topology information, and the spreading factor and transmission power configuration of each terminal node are determined by the A2C algorithm. The service rate benchmark is determined based on the spreading factor and transmission power configuration of each terminal node. Bandwidth is allocated through a deep Q network with a fused attention mechanism, and the decision weights for adjusting the spreading factor and transmission power are fed back during periods of sustained high latency.

Images