Artificial intelligence-based internet of things wireless communication path optimization method

Abstract

The application discloses an Internet of Things wireless communication path optimization method based on artificial intelligence, which comprises original data collection, original data optimization, network state prediction model construction, path selection optimization model design and Internet of Things wireless communication path optimization. The original data is obtained through data collection; the improved bidirectional long short-term memory network model is used as the network state prediction model, the implicit multi-source time series data is converted into the explicit demand intensity of different communication paths, the real-time demand intensity of each communication path is accurately identified through the fusion of time series dependence analysis and periodic attention mechanism; the improved reinforcement learning model is used as the path selection optimization model, the dynamic space fully representing the decision environment is constructed, the action constraint network is introduced to ensure the action feasibility, the multi-objective trade-off is realized by using the proximal policy optimization, and thus the personalized path control strategy considering the instant service quality and long-term network efficiency is generated.

Description

Technical Field

[0001] This invention relates to the field of IoT wireless communication path optimization technology, specifically to an IoT wireless communication path optimization method based on artificial intelligence. Background Technology

[0002] IoT wireless communication path optimization refers to the process of intelligently adjusting and optimizing the data transmission path selection strategy of wireless networks with multiple communication modes such as multi-hop, multi-rate, and multi-quality of service levels, using intelligent technology. Through continuous learning of network traffic patterns, node status, and environmental parameters by algorithms, it dynamically plans the combination of parameters such as bandwidth allocation, transmission rate, priority, and routing selection for different data transmission paths, so as to achieve precise matching between communication paths and network requirements and business objectives, and meet the personalized needs of IoT applications for diverse communication services.

[0003] However, traditional IoT wireless communication path optimization methods often rely on preset rules or short-term historical statistics, failing to effectively integrate the dynamic interaction characteristics of network environment, node status, and traffic patterns. This results in the identification of network demand status remaining at a shallow level, and a lack of ability to model long-term traffic patterns. They can only respond to immediate traffic while ignoring potential patterns, leading to rigid strategies that cannot adapt to dynamic network scenarios. Traditional IoT wireless communication path optimization methods often employ heuristic rules or single-objective control, making it difficult to handle multi-dimensional state inputs. This causes strategies to deviate from actual network physical and protocol constraints, leading to network congestion, excessive energy consumption, or service quality degradation. Furthermore, single-objective optimization ignores differences in service types, resulting in unfair resource allocation and overall inefficiency. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an AI-based IoT wireless communication path optimization method. Traditional IoT wireless communication path optimization methods typically rely on preset rules or short-term historical statistics, failing to effectively integrate the dynamic interaction characteristics of network environment, node status, and traffic patterns. This results in superficial identification of network demand states and a lack of ability to model long-term traffic patterns, responding only to immediate traffic while ignoring underlying patterns, leading to rigid strategies that are unsuitable for dynamic network scenarios. This solution creatively employs an improved bidirectional long short-term memory network model as the network state prediction model. It transforms implicit multi-source time-series data, including network environment, traffic patterns, and node status, into explicit demand intensity indicators for different communication paths. By integrating time-series dependency analysis and a periodic attention mechanism, it adaptively learns network traffic patterns and accurately identifies the real-time demand intensity of each communication path. This approach provides interpretable and high-precision demand input for subsequent path selection optimization. Traditional IoT wireless communication path optimization methods often employ heuristic rules or single-objective control, making it difficult to handle multi-dimensional state inputs. This leads to policies deviating from actual network physical and protocol constraints, causing network congestion, excessive energy consumption, or degraded service quality. Furthermore, single-objective optimization ignores differences in service types, resulting in unfair resource allocation and overall inefficiency. This solution creatively adopts an improved reinforcement learning model as the path selection optimization model. Through hierarchical state coding, it integrates network infrastructure conditions, real-time demand intensity, and service context to construct a dynamic space that comprehensively represents the decision-making environment. An action constraint network is introduced, embedding network protocols and physical boundaries as learnable functions into policy generation to ensure action feasibility. Proximal policy optimization achieves adaptive trade-offs among multiple objectives, thereby generating personalized path control strategies that balance immediate service quality and long-term network performance.

[0005] The technical solution adopted by this invention is as follows: The IoT wireless communication path optimization method based on artificial intelligence provided by this invention includes the following steps:

[0006] Step S1: Raw data acquisition;

[0007] Step S2: Optimize the raw data;

[0008] Step S3: Constructing the network state prediction model;

[0009] Step S4: Path selection optimization model design;

[0010] Step S5: IoT wireless communication path optimization.

[0011] Further, in step S1, the raw data acquisition is used to collect the raw data required for IoT wireless communication path optimization. Specifically, by performing data acquisition, a path optimization raw dataset is obtained. The path optimization raw dataset specifically includes a historical path raw dataset and a current path raw dataset. Both the historical path raw dataset and the current path raw dataset include network environment data, traffic characteristic data, and node status data. The historical path raw dataset also includes service quality feedback data.

[0012] Further, in step S2, the raw data optimization is used to optimize the collected raw data, specifically including the following steps:

[0013] Step S21: Data cleaning, used to handle missing values ​​and correct outliers. Specifically, it involves filling gaps in sensors or network elements through time series linear interpolation and identifying and correcting outliers through dynamic thresholding algorithms to improve data quality.

[0014] Step S22: Multi-source alignment processing, used to solve the difference in sampling frequency of heterogeneous data, specifically to achieve temporal alignment of network environment data, traffic feature data and node status data through sliding window aggregation;

[0015] Step S23: Data normalization, used to normalize data to adapt to network load and scene changes. Specifically, the data is processed by using the minimum-maximum normalization method by time slice to eliminate the impact of network fluctuations.

[0016] Step S24: Dataset splitting, used to split the historical dataset into training set and test set, specifically splitting the original dataset of the historical path into path training set and path test set;

[0017] Step S25: Optimize the original data, specifically by optimizing the original dataset of the current path through data cleaning, multi-source alignment processing, and data normalization to obtain the dataset to be processed, and by optimizing the original dataset of the historical path through data cleaning, multi-source alignment processing, data normalization, and dataset segmentation to obtain the path training set and the path test set.

[0018] Further, in step S3, the network state prediction model is constructed to build a model for extracting dynamic temporal features of network traffic and environment and quantifying path demand intensity. Specifically, it is to construct an improved bidirectional long short-term memory network model as the network state prediction model. The improved bidirectional long short-term memory network model specifically includes a multi-source fusion module, a bidirectional feature extraction module, a periodic attention module, and a demand intensity quantification module.

[0019] The construction of the network state prediction model specifically includes the following steps:

[0020] Step S31: Construction of a multi-source fusion module to address the issues of scale differences and feature incompatibility among multi-source data. Specifically, this involves dynamically weighting network environment features, traffic feature data, and node state features using learnable gating weights to obtain multi-source fusion features. The steps include:

[0021] Step S311: Gated weight generation, used for adaptive allocation of data importance weights, specifically by generating weight ratios through linear transformation and Sigmoid activation function to obtain fused gated weights;

[0022] Step S312: Feature transformation, used to improve the expressive power of features, specifically, the network environment features, traffic features and node state features are linearly mapped and then processed by the hyperbolic tangent function to obtain enhanced network environment features, enhanced traffic features and enhanced node state features;

[0023] Step S313: Feature fusion calculation, used to effectively fuse multi-source information, specifically by weighting and summing the enhanced features of each data source through gating weights to obtain multi-source fused features;

[0024] Step S32: Construct a bidirectional feature extraction module to capture the traffic dependencies of long short-term memory networks. Specifically, extract temporal patterns from the forward and backward directions using a bidirectional long short-term memory network to obtain bidirectional temporal features.

[0025] Step S33: Construction of a periodic attention module to enhance the ability to recognize periodic network traffic patterns. Specifically, this involves guiding attention to key periodic nodes through temporal location encoding to obtain periodic enhanced features. The steps include:

[0026] Step S331: Time encoding is generated to inject periodic prior knowledge. Specifically, the 24-hour daily cycle and 168-hour weekly cycle time points are encoded using sine and cosine functions to obtain a periodic encoding vector for attention correction.

[0027] Step S332: Attention energy calculation, used to identify periodic sensitive moments, specifically by fusing bidirectional temporal features and periodic encoding vectors through linear transformation to obtain the attention energy score at each moment;

[0028] Step S333: Obtain periodic enhancement features, which are used to generate periodic enhancement features. Specifically, the attention energy score is transformed into a probability distribution through the Softmax function, and the bidirectional temporal features are weighted and summed to obtain the periodic enhancement features.

[0029] Step S34: Construction of the demand intensity quantification module, used to explicitly quantify implicit path demand. Specifically, this involves learning the demand intensity of each communication path by improving the activation function to obtain path demand intensity features. The steps include:

[0030] Step S341: Feature mapping, used for dimensionality reduction and feature abstraction, specifically by compressing periodic enhancement features through a fully connected layer to obtain compressed periodic enhancement features;

[0031] Step S342: Design an improved activation function to enhance nonlinear representation capability. Specifically, the output range of the Sigmoid activation function is expanded by introducing a learnable slope parameter to obtain the improved activation function. Based on the improved activation function, the compressed periodic enhancement features of each communication path are processed to obtain the demand intensity of each communication path and form a demand intensity vector.

[0032] Step S35: Construct and train the model, specifically by constructing the improved bidirectional long short-term memory network model through the multi-source fusion module, the bidirectional feature extraction module, the periodic attention module, and the demand intensity quantification module. The improved bidirectional long short-term memory network model is constructed by integrating these modules. The model is trained based on the path training set and its performance is verified based on the path test set to obtain the improved bidirectional long short-term memory network model, which serves as the network state prediction model.

[0033] Further, in step S4, the path selection optimization model design is used to design the model required for dynamically generating personalized path selection control strategies. Specifically, it involves constructing an improved reinforcement learning model as the path selection optimization model. The improved reinforcement learning model specifically includes a state encoding module, an action constraint network, and a policy optimization module.

[0034] The path selection optimization model design specifically includes the following steps:

[0035] Step S41: Construction of the state encoding module for structured representation of decision states. Specifically, this involves obtaining a hierarchical state space by separating three layers of information: basic state, contextual state, and temporal state. The basic state represents the network environment and basic node information; the contextual state represents path demand intensity information and service quality feedback information; and the temporal state represents time period information. The construction of the state encoding module includes the following steps:

[0036] Step S411: State hierarchical encoding, used to distinguish the importance of state types, specifically by processing the information of the three states through independent fully connected layers to obtain three sets of state vectors;

[0037] Step S412: State space integration, used to construct a joint state representation, specifically by concatenating the basic state vector, the context state vector, and the time state vector to obtain the integrated decision state space;

[0038] Step S42: Action constraint network design, used to ensure the network feasibility of the path selection strategy, specifically involves learning the mapping from network state to protocol boundary through a neural network to obtain the real-time feasible action range. The steps include:

[0039] Step S421: Constraint boundary prediction, used to model network physical and protocol constraint rules, specifically by processing node state data through a two-layer ReLU network and outputting path selection boundary parameters;

[0040] Step S422: Constraint matrix generation, used to format action boundaries, specifically by scaling the minimum and maximum values ​​of predefined parameters using the Sigmoid activation function to obtain the constraint matrix that the network system can execute;

[0041] Step S43: Design a reward function to balance service quality and network performance. Specifically, a composite reward function for guiding strategy optimization is obtained by weighting communication quality indicators and network resource consumption indicators.

[0042] Step S44: Construction of the policy optimization module, used to stably learn the optimal path selection policy, specifically by using proximal policy optimization and updating the policy under the constraints of the action constraint network;

[0043] Step S45: Construct and train the model. Specifically, this involves constructing an improved reinforcement learning model by integrating the state encoding module, the action constraint network design, the reward function design, and the policy optimization module. The model is then trained based on the path training set and its performance is verified based on the path test set to obtain the improved reinforcement learning model, which serves as the path selection optimization model.

[0044] Further, in step S5, the IoT wireless communication path optimization specifically involves using the dataset to be processed as input to the network state prediction model to obtain a demand intensity reference result, and combining the demand intensity reference result with the dataset to be processed as input to the path selection optimization model to obtain an optimal path selection control strategy. The demand intensity reference result is specifically a demand intensity vector set obtained based on the input.

[0045] The beneficial effects achieved by the present invention using the above solution are as follows:

[0046] (1) Traditional IoT wireless communication path optimization methods typically rely on preset rules or short-term historical statistics, which cannot effectively integrate the dynamic interaction characteristics of network environment, node status and traffic patterns. This results in the identification of network demand status remaining at a shallow level and lacking the ability to model long-term traffic patterns. It can only respond to immediate traffic and ignore potential patterns, leading to rigid strategies that cannot adapt to dynamic network scenarios. This solution creatively adopts an improved bidirectional long short-term memory network model as the network state prediction model. It transforms the implicit multi-source time-series data such as network environment, traffic patterns and node status into explicit demand intensity indicators for different communication paths. By integrating time-series dependency analysis and periodic attention mechanism, it adaptively learns network traffic patterns and accurately identifies the real-time demand intensity of each communication path, providing interpretable and high-precision demand input for subsequent path selection optimization.

[0047] (2) In view of the technical problems of traditional IoT wireless communication path optimization methods, which often use heuristic rules or single-objective control, it is difficult to handle multi-dimensional state inputs, resulting in the policy deviating from the actual network physical and protocol constraints, causing network congestion, excessive energy consumption or service quality degradation. Moreover, single-objective optimization ignores the differences in service types, resulting in unfair resource allocation and low overall efficiency. This solution creatively adopts an improved reinforcement learning model as the path selection optimization model. By integrating network basic conditions, real-time demand intensity and service context through hierarchical state coding, a dynamic space that fully represents the decision-making environment is constructed. An action constraint network is introduced, and network protocols and physical boundaries are embedded as learnable functions in policy generation to ensure action feasibility. Near-end policy optimization is used to achieve adaptive trade-offs between multiple objectives, thereby generating personalized path control strategies that take into account both immediate service quality and long-term network performance. Attached Figure Description

[0048] Figure 1 A flowchart illustrating the AI-based IoT wireless communication path optimization method provided by this invention;

[0049] Figure 2 A flowchart illustrating the optimization of the raw data in step S2;

[0050] Figure 3 A flowchart illustrating the process of building the network state prediction model in step S3;

[0051] Figure 4 A flowchart illustrating the path selection optimization model design for step S4.

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0054] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0055] Example 1, see Figure 1 The present invention provides an AI-based IoT wireless communication path optimization method, which includes the following steps:

[0056] Step S1: Raw data acquisition;

[0057] Step S2: Optimize the raw data;

[0058] Step S3: Constructing the network state prediction model;

[0059] Step S4: Path selection optimization model design;

[0060] Step S5: IoT wireless communication path optimization.

[0061] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the raw data acquisition is used to collect the raw data required for IoT wireless communication path optimization. Specifically, by performing data acquisition, a path optimization raw dataset is obtained. The path optimization raw dataset specifically includes a historical path raw dataset and a current path raw dataset. Both the historical path raw dataset and the current path raw dataset include network environment data, traffic characteristic data, and node status data. The historical path raw dataset also includes service quality feedback data.

[0062] The communication path modes of the Internet of Things (IoT) wireless network specifically include high-bandwidth paths, low-latency paths, and high-reliability paths. The network environment data specifically includes channel quality data, interference intensity data, and link packet loss rate data collected by sensors or network elements. The traffic characteristic data specifically includes data packet timestamps, data stream types, data packet sizes, data stream transmission durations, bandwidth requirements, and latency requirements collected from network traffic monitoring and logs. The node status data specifically includes node remaining energy data, single transmission energy consumption data, node load data, buffer occupancy rate data, wireless module operating status data, and routing table status data collected in real time by node devices. The service quality feedback data specifically includes end-to-end latency data, throughput data, packet loss rate data, and service success rate data.

[0063] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, the raw data optimization is used to optimize the collected raw data, specifically including the following steps:

[0064] Step S21: Data cleaning, used to handle missing values ​​and correct outliers. Specifically, it involves filling gaps in sensors or network elements through time series linear interpolation and identifying and correcting outliers through dynamic thresholding algorithms to improve data quality.

[0065] Step S22: Multi-source alignment processing, used to solve the difference in sampling frequency of heterogeneous data, specifically to achieve temporal alignment of network environment data, traffic feature data and node status data through sliding window aggregation;

[0066] Step S23: Data normalization, used to normalize data to adapt to network load and scene changes. Specifically, the data is processed by using the minimum-maximum normalization method by time slice to eliminate the impact of network fluctuations.

[0067] Step S24: Dataset splitting, used to split the historical dataset into training set and test set, specifically splitting the original dataset of the historical path into path training set and path test set;

[0068] Step S25: Optimize the original data, specifically by optimizing the original dataset of the current path through data cleaning, multi-source alignment processing, and data normalization to obtain the dataset to be processed, and by optimizing the original dataset of the historical path through data cleaning, multi-source alignment processing, data normalization, and dataset segmentation to obtain the path training set and the path test set.

[0069] Example 4, see Figure 1 and Figure 3This embodiment is based on the above embodiment. In step S3, the network state prediction model is constructed to build a model for extracting the dynamic temporal features of network traffic and environment and quantifying the path demand intensity. Specifically, it is to construct an improved bidirectional long short-term memory network model as the network state prediction model. The improved bidirectional long short-term memory network model specifically includes a multi-source fusion module, a bidirectional feature extraction module, a periodic attention module, and a demand intensity quantification module.

[0070] The construction of the network state prediction model specifically includes the following steps:

[0071] Step S31: Construction of a multi-source fusion module to address the issues of scale differences and feature incompatibility among multi-source data. Specifically, this involves dynamically weighting network environment features, traffic feature data, and node state features using learnable gating weights to obtain multi-source fusion features. The steps include:

[0072] Step S311: Gated weight generation, used for adaptive allocation of data importance weights. Specifically, it generates weight ratios through linear transformation and the Sigmoid activation function to obtain fused gated weights. The formula used is as follows:

[0073] ;

[0074] In the formula, This indicates that the gating weights are fused at time t. This represents the Sigmoid activation function. Indicates the gating mapping weights, Indicates the gating mapping bias term. This represents the network environment characteristics at time t. This represents the flow characteristics at time t. This represents the state characteristics of a node at time t;

[0075] Step S312: Feature transformation, used to improve the expressive power of features. Specifically, the network environment features, traffic features, and node state features are linearly mapped and then processed by the hyperbolic tangent function to obtain enhanced network environment features, enhanced traffic features, and enhanced node state features. The formulas used are as follows:

[0076] ;

[0077] In the formula, This indicates the enhanced network environment characteristics at time t. This indicates that the enhanced flow characteristics are at time t. This indicates that the state characteristics of the enhanced node are represented at time t. Represents the hyperbolic tangent function. Indicates the network environment mapping weights. This represents the network environment mapping bias term. Indicates traffic mapping weights, This represents the traffic mapping bias term. Indicates the node state mapping weight. This represents the node state mapping bias term;

[0078] Step S313: Feature fusion calculation, used to effectively fuse multi-source information. Specifically, it involves weighted summation of the enhanced features from each data source using gated weights to obtain multi-source fused features. The formula used is as follows:

[0079] ;

[0080] In the formula, Indicates multi-source fusion characteristics, This represents element-wise multiplication.

[0081] Step S32: Construction of a bidirectional feature extraction module, used to capture the traffic dependencies of long short-term memory networks. Specifically, it extracts temporal patterns from both the forward and backward directions using a bidirectional long short-term memory network to obtain bidirectional temporal features. The formula used is as follows:

[0082] ;

[0083] In the formula, This represents the forward hidden state at time t. This represents the backward hidden state at time t. This represents the forward hidden state at time t-1. This represents the backward hidden state at time t-1. This represents the bidirectional time series characteristics at time t. Represents a long short-term memory network function;

[0084] Step S33: Construction of a periodic attention module to enhance the ability to recognize periodic network traffic patterns. Specifically, this involves guiding attention to key periodic nodes through temporal location encoding to obtain periodic enhanced features. The steps include:

[0085] Step S331: Time encoding generation, used to inject periodic prior knowledge. Specifically, it encodes 24-hour daily and 168-hour weekly time points using sine and cosine functions to obtain a periodic encoding vector for attention correction. The formula used is as follows:

[0086] ;

[0087] In the formula, This represents the periodic encoded vector at time t, where t represents the time index.

[0088] Step S332: Attention energy calculation, used to identify periodic sensitive moments. Specifically, it involves fusing bidirectional temporal features and periodic encoding vectors through linear transformation to obtain the attention energy score at each moment. The formula used is as follows:

[0089] ;

[0090] In the formula, Let represent the attention energy score at time t, and lv represent the learnable energy score calculation vector. Represents the weights for bidirectional time-series feature transformation. This represents the weights of the periodic encoding vector transformation. This represents the energy fraction transformation bias term, and T represents the transpose operation;

[0091] Step S333: Obtain periodic enhancement features, which are used to generate periodic enhancement features. Specifically, the attention energy score is transformed into a probability distribution using the Softmax function, and the bidirectional temporal features are weighted and summed to obtain the periodic enhancement features. The formula used is as follows:

[0092] ;

[0093] In the formula, Cf represents the periodic enhancement feature, and Time represents the total number of time points. express Attention energy score at any given moment;

[0094] Step S34: Construction of the demand intensity quantification module, used to explicitly quantify implicit path demand. Specifically, this involves learning the demand intensity of each communication path by improving the activation function to obtain path demand intensity features. The steps include:

[0095] Step S341: Feature mapping, used for dimensionality reduction and feature abstraction, specifically by compressing periodic enhancement features through a fully connected layer to obtain compressed periodic enhancement features;

[0096] Step S342: Design an improved activation function to enhance nonlinear representation capabilities. Specifically, this involves expanding the output range of the Sigmoid activation function by introducing a learnable slope parameter to obtain the improved activation function. Based on this improved activation function, the compressed periodic enhancement features of each communication path are processed to obtain the demand intensity of each communication path, forming a demand intensity vector. The formula used is as follows:

[0097] ;

[0098] In the formula, This indicates an improved activation function. This represents the compressed periodic enhancement feature of the a-th communication path. The slope parameter represents the learnable slope parameter, and DI represents the demand intensity vector. This indicates the demand intensity for the first communication path. This indicates the demand intensity for the second communication path. This indicates the demand intensity for communication path A. This represents the compressed periodic enhancement feature of the first communication path. This represents the compressed periodic enhancement feature of the second communication path. This represents the compressed periodic enhancement feature of the Ath type of communication path, where A represents the total number of communication path types.

[0099] Step S35: Construct and train the model, specifically by constructing the improved bidirectional long short-term memory network model through the multi-source fusion module, the bidirectional feature extraction module, the periodic attention module, and the demand intensity quantification module. The improved bidirectional long short-term memory network model is constructed by integrating these modules. The model is trained based on the path training set and its performance is verified based on the path test set to obtain the improved bidirectional long short-term memory network model, which serves as the network state prediction model.

[0100] By performing the above operations, this solution addresses the technical problems of traditional IoT wireless communication path optimization methods, which typically rely on preset rules or short-term historical statistics, failing to effectively integrate the dynamic interaction characteristics of network environment, node status, and traffic patterns. This results in the identification of network demand status remaining at a shallow level and lacking the ability to model long-term traffic patterns, only responding to immediate traffic while ignoring potential patterns, leading to rigid strategies that cannot adapt to dynamic network scenarios. This solution creatively adopts an improved bidirectional long short-term memory network model as the network state prediction model. It transforms implicit multi-source time-series data such as network environment, traffic patterns, and node status into explicit demand intensity indicators for different communication paths. By integrating time-series dependency analysis and periodic attention mechanisms, it adaptively learns network traffic patterns and accurately identifies the real-time demand intensity of each communication path, providing interpretable and high-precision demand input for subsequent path selection optimization.

[0101] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the path selection optimization model design is used to design the model required for dynamically generating personalized path selection control strategies. Specifically, it is to construct an improved reinforcement learning model as the path selection optimization model. The improved reinforcement learning model specifically includes a state encoding module, an action constraint network, and a policy optimization module.

[0102] The path selection optimization model design specifically includes the following steps:

[0103] Step S41: Construction of the state encoding module for structured representation of decision states. Specifically, this involves obtaining a hierarchical state space by separating three layers of information: basic state, contextual state, and temporal state. The basic state represents the network environment and basic node information; the contextual state represents path demand intensity information and service quality feedback information; and the temporal state represents time period information. The construction of the state encoding module includes the following steps:

[0104] Step S411: State hierarchical encoding, used to distinguish the importance of state types, specifically by processing the information of the three states separately through independent fully connected layers to obtain three sets of state vectors, using the following formula:

[0105] ;

[0106] In the formula, Represents the basic state vector. Represents the situational state vector. Represents the time state vector. Represents the ReLU activation function. Represents the weights of the basic state mapping. This represents the bias term of the basic state mapping. Represents the context state mapping weights. This represents the context state mapping bias term. Represents the BERT model function. This indicates the input of service quality feedback data. Indicates the weights of the time-state mapping. The time state mapping bias term is represented by Hour, which represents the hour encoding vector of the current time, and DType, which represents the date type encoding vector of the current time, used to introduce the effects of weekdays, weekends, and holidays.

[0107] Step S412: State space integration, used to construct a joint state representation, specifically by concatenating the basic state vector, the context state vector, and the time state vector to obtain the integrated decision state space;

[0108] Step S42: Action constraint network design, used to ensure the network feasibility of the path selection strategy, specifically involves learning the mapping from network state to protocol boundary through a neural network to obtain the real-time feasible action range. The steps include:

[0109] Step S421: Constraint boundary prediction, used to model network physical and protocol constraint rules, specifically by processing node state data through a two-layer ReLU network and outputting six boundary parameters, including bandwidth allocation lower limit parameter, bandwidth allocation upper limit parameter, transmission rate lower limit parameter, transmission rate upper limit parameter, path hop count baseline parameter, and path reliability fluctuation range parameter.

[0110] Step S422: Constraint matrix generation, used to format action boundaries. Specifically, this involves scaling the minimum and maximum values ​​of predefined parameters using the Sigmoid activation function to obtain the constraint matrix executable by the network system. The formula used is as follows:

[0111] ;

[0112] In the formula, PC represents the constraint matrix. This represents the lower limit parameter for bandwidth allocation. This indicates the upper limit parameter for bandwidth allocation. This represents the lower limit parameter of the transmission rate. This parameter represents the upper limit of the transmission rate. Indicates the baseline parameter for path hop count. This parameter represents the range of path reliability fluctuations. Indicates the bandwidth adjustment range. Indicates the range of transmission rate adjustment. Indicates the physical minimum bandwidth. Indicates the physical minimum transmission rate;

[0113] Step S43: Reward function design to balance service quality and network performance. Specifically, this involves obtaining a composite reward function for guiding strategy optimization by weighting communication quality metrics and network resource consumption metrics. The steps include:

[0114] Step S431: Design of communication quality reward, used to quantify the satisfaction of service quality by path selection. Specifically, it calculates the end-to-end latency penalty and the logarithmic reward of throughput to obtain the communication quality reward. The formula used is as follows:

[0115] ;

[0116] In the formula, Indicates a reward for communication quality. Indicates the delayed matching weight. Indicates throughput matching weight, This indicates a delay in expected business operations. Indicates the actual end-to-end delay. Indicates the maximum permissible delay deviation. Indicates actual throughput;

[0117] Step S432: Design of network performance index reward to quantify network resource costs. Specifically, this involves calculating the negative energy consumption value and the network load balancing gain to obtain the network performance index reward. The formula used is as follows:

[0118] ;

[0119] In the formula, Indicates rewards based on network performance metrics. This represents the energy consumption penalty coefficient. This represents the load balancing gain coefficient. This indicates the energy consumption of a single data transmission. Indicates changes in the load of critical nodes;

[0120] Step S433: Calculate the composite reward, specifically by weighting and summing the communication quality reward and the network performance indicator reward to obtain the composite reward;

[0121] Step S44: Construction of the policy optimization module, used to stably learn the optimal path selection policy, specifically adopting near-end policy optimization, updating the policy under the constraints of the action constraint network, and outputting the action dimension including the allocated bandwidth, transmission rate and selected path;

[0122] Step S45: Construct and train the model. Specifically, this involves constructing an improved reinforcement learning model by integrating the state encoding module, the action constraint network design, the reward function design, and the policy optimization module. The model is then trained based on the path training set and its performance is verified based on the path test set to obtain the improved reinforcement learning model, which serves as the path selection optimization model.

[0123] By performing the above operations, this solution addresses the technical problems of traditional IoT wireless communication path optimization methods, which often employ heuristic rules or single-objective control, making it difficult to handle multi-dimensional state inputs. This leads to policies deviating from actual network physical and protocol constraints, causing network congestion, excessive energy consumption, or degraded service quality. Furthermore, single-objective optimization ignores differences in service types, resulting in unfair resource allocation and overall low efficiency. This solution creatively adopts an improved reinforcement learning model as the path selection optimization model. By integrating network infrastructure conditions, real-time demand intensity, and service context through hierarchical state coding, it constructs a dynamic space that comprehensively represents the decision-making environment. It also introduces an action constraint network, embedding network protocols and physical boundaries as learnable functions into policy generation to ensure action feasibility. By utilizing near-end policy optimization, it achieves adaptive trade-offs among multiple objectives, thereby generating personalized path control strategies that balance immediate service quality and long-term network performance.

[0124] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the IoT wireless communication path optimization specifically involves using the dataset to be processed as the input of the network state prediction model to obtain a demand intensity reference result, and combining the demand intensity reference result with the dataset to be processed as the input of the path selection optimization model to obtain an optimal path selection control strategy. The demand intensity reference result is specifically a demand intensity vector set obtained based on the input.

[0125] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 process, method, article, or apparatus.

[0126] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0127] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An AI-based method for optimizing IoT wireless communication paths, characterized by: The method includes the following steps: S1: Raw data acquisition. Through data acquisition, the path optimization raw dataset is obtained. The path optimization raw dataset specifically includes the historical path raw dataset and the current path raw dataset. S2: Raw data optimization. By optimizing the collected raw data, we obtain the dataset to be processed, the path training set, and the path test set. S3: Network state prediction model construction, used to construct the model required to extract the dynamic temporal features of network traffic and environment and quantify the intensity of path demand. Specifically, it constructs an improved bidirectional long short-term memory network model as the network state prediction model. The improved bidirectional long short-term memory network model specifically includes a multi-source fusion module, a bidirectional feature extraction module, a periodic attention module, and a demand intensity quantification module. The multi-source fusion module is used to handle the problems of scale differences and feature incompatibility among multi-source data. Specifically, it obtains multi-source fusion features by dynamically weighting network environment features, traffic feature data, and node state features using learnable gating weights. The steps include: Step S311: Gated weight generation, used for adaptive allocation of data importance weights, specifically by generating weight ratios through linear transformation and Sigmoid activation function to obtain fused gated weights; Step S312: Feature transformation, used to improve the expressive power of features, specifically, the network environment features, traffic features and node state features are linearly mapped and then processed by the hyperbolic tangent function to obtain enhanced network environment features, enhanced traffic features and enhanced node state features; Step S313: Feature fusion calculation, used to effectively fuse multi-source information, specifically by weighting and summing the enhanced features of each data source through gating weights to obtain multi-source fused features; The demand intensity quantification module is used to explicitly quantify implicit path demand. Specifically, it learns the demand intensity of each communication path by improving the activation function to obtain path demand intensity features. The steps include: Step S341: Feature mapping, used for dimensionality reduction and feature abstraction, specifically by compressing periodic enhancement features through a fully connected layer to obtain compressed periodic enhancement features; Step S342: Design an improved activation function to enhance nonlinear representation capability. Specifically, the output range of the Sigmoid activation function is expanded by introducing a learnable slope parameter to obtain the improved activation function. Based on the improved activation function, the compressed periodic enhancement features of each communication path are processed to obtain the demand intensity of each communication path and form a demand intensity vector. S4: Path selection optimization model design, used to design the model required for dynamically generating personalized path selection control strategies, specifically to construct an improved reinforcement learning model as the path selection optimization model. The improved reinforcement learning model specifically includes a state encoding module, an action constraint network, and a policy optimization module. The state encoding module is used to structurally represent the decision state. Specifically, it obtains a hierarchical state space by separating three layers of information: basic state, contextual state, and temporal state. The basic state represents the network environment and basic node information; the contextual state represents path demand intensity information and service quality feedback information; and the temporal state represents time period information. The steps include: Step S411: State hierarchical encoding, used to distinguish the importance of state types, specifically by processing the information of the three states through independent fully connected layers to obtain three sets of state vectors; Step S412: State space integration, used to construct a joint state representation, specifically by concatenating the basic state vector, the context state vector, and the time state vector to obtain the integrated decision state space; S5: IoT wireless communication path optimization, specifically, using the dataset to be processed as input to the network state prediction model to obtain a demand intensity reference result, and combining the demand intensity reference result with the dataset to be processed as input to the path selection optimization model to obtain an optimal path selection control strategy. The demand intensity reference result is specifically a demand intensity vector set obtained based on the input.

2. The method for optimizing IoT wireless communication paths based on artificial intelligence according to claim 1, characterized in that: The construction of the network state prediction model specifically includes the following steps: Step S31: Construction of multi-source fusion module; Step S32: Construction of the bidirectional feature extraction module; Step S33: Construction of the periodic attention module; Step S34: Construction of the demand intensity quantification module; Step S35: Build and train the model.

3. The method for optimizing IoT wireless communication paths based on artificial intelligence according to claim 2, characterized in that: The bidirectional feature extraction module is constructed to capture the long short-term network traffic dependency relationship. Specifically, it extracts the temporal pattern from the forward and backward directions respectively through the bidirectional long short-term memory network to obtain bidirectional temporal features. The aforementioned periodic attention module is constructed to enhance the ability to recognize periodic network traffic patterns. Specifically, it guides attention to key periodic nodes through temporal location encoding to obtain periodic enhanced features. The steps include: Step S331: Time encoding is generated to inject periodic prior knowledge. Specifically, the 24-hour daily cycle and 168-hour weekly cycle time points are encoded using sine and cosine functions to obtain a periodic encoding vector for attention correction. Step S332: Attention energy calculation, used to identify periodic sensitive moments, specifically by fusing bidirectional temporal features and periodic encoding vectors through linear transformation to obtain the attention energy score at each moment; Step S333: Obtain periodic enhancement features, which are used to generate periodic enhancement features. Specifically, the attention energy score is transformed into a probability distribution through the Softmax function, and the bidirectional temporal features are weighted and summed to obtain the periodic enhancement features. The construction and training of the model specifically involves constructing an improved bidirectional long short-term memory network model through the multi-source fusion module, the bidirectional feature extraction module, the periodic attention module, and the demand intensity quantification module. The model is then trained based on the path training set and its performance is verified based on the path test set to obtain the improved bidirectional long short-term memory network model, which serves as the network state prediction model.

4. The method for optimizing IoT wireless communication paths based on artificial intelligence according to claim 1, characterized in that: The path selection optimization model design specifically includes the following steps: Step S41: Construction of the state coding module; Step S42: Design of motion constraint network; Step S43: Design the reward function; Step S44: Construction of the strategy optimization module; Step S45: Build and train the model.

5. The method for optimizing IoT wireless communication paths based on artificial intelligence according to claim 4, characterized in that: The action constraint network design is used to ensure the network feasibility of the path selection strategy. Specifically, it learns the mapping from network state to protocol boundary through a neural network to obtain the real-time actionable range. The steps include: Step S421: Constraint boundary prediction, used to model network physical and protocol constraint rules, specifically by processing node state data through a two-layer ReLU network and outputting path selection boundary parameters; Step S422: Constraint matrix generation, used to format action boundaries, specifically by scaling the minimum and maximum values ​​of predefined parameters using the Sigmoid activation function to obtain the constraint matrix that the network system can execute; The reward function design is used to balance service quality and network performance. Specifically, it obtains a composite reward function for guiding strategy optimization by weighting communication quality indicators and network resource consumption indicators. The strategy optimization module is constructed to stably learn the optimal path selection strategy, specifically by adopting proximal strategy optimization and updating the strategy under the constraints of the action constraint network. The construction and training of the model specifically involves integrating the state encoding module, the action constraint network design, the reward function design, and the policy optimization module to construct an improved reinforcement learning model. The model is then trained based on a path training set and its performance is verified based on a path test set to obtain the improved reinforcement learning model, which serves as a path selection optimization model.

6. The method for optimizing IoT wireless communication paths based on artificial intelligence according to claim 1, characterized in that: Both the historical path original dataset and the current path original dataset include network environment data, traffic characteristic data, and node status data. The historical path original dataset also includes service quality feedback data.

7. The method for optimizing IoT wireless communication paths based on artificial intelligence according to claim 1, characterized in that: The optimization of the original data specifically includes the following steps: Step S21: Data cleaning, used to handle missing values ​​and correct outliers. Specifically, it involves filling gaps in sensors or network elements through time series linear interpolation and identifying and correcting outliers through dynamic thresholding algorithms to improve data quality. Step S22: Multi-source alignment processing, used to solve the difference in sampling frequency of heterogeneous data, specifically to achieve temporal alignment of network environment data, traffic feature data and node status data through sliding window aggregation; Step S23: Data normalization, used to normalize data to adapt to network load and scene changes. Specifically, the data is processed by using the minimum-maximum normalization method by time slice to eliminate the impact of network fluctuations. Step S24: Dataset splitting, used to split the historical dataset into training set and test set, specifically splitting the original dataset of the historical path into path training set and path test set; Step S25: Optimize the original data, specifically by optimizing the original dataset of the current path through data cleaning, multi-source alignment processing, and data normalization to obtain the dataset to be processed, and by optimizing the original dataset of the historical path through data cleaning, multi-source alignment processing, data normalization, and dataset segmentation to obtain the path training set and the path test set.

Images