An internet of things scene sensing and link adaptation optimization method and system

Abstract

This application relates to the interdisciplinary fields of wireless communication and artificial intelligence, and in particular to a method and system for IoT scene perception and link adaptive optimization. The method includes the following steps: acquiring channel state information at the IoT communication receiver; constructing a two-dimensional feature matrix from the real and imaginary components of the channel state information; classifying and identifying the current channel propagation scene to obtain a scene category; constructing a reinforcement learning state vector based on the scene category and channel quality parameters; using a combination of modulation and coding scheme and signal bandwidth as the reinforcement learning action space; inputting the state vector into a deep reinforcement learning agent to infer the optimal link configuration; and transmitting the optimal link configuration to the IoT transmitter via a feedback link to achieve adaptive adjustment of link transmission parameters. The method of this invention possesses a scene-aware communication adaptive mechanism and features diverse optimization dimensions and strong robustness.

Description

Technical Field

[0001] This application relates to the field of wireless communication and artificial intelligence, and in particular to a method and system for IoT scene perception and link adaptive optimization. Background Technology

[0002] With the rapid evolution of sixth-generation wireless communication technology, the Internet of Things (IoT) has become one of the core application areas of next-generation networks. IoT systems, through deep integration with artificial intelligence, achieve efficient and reliable connectivity to massive numbers of devices in complex environments. However, the large-scale deployment of IoT brings significant architectural complexity—in highly mobile scenarios such as connected vehicles, industrial IoT, and drone communications, wireless channels are simultaneously affected by the high-speed movement of terminals and drastic environmental changes, exhibiting high non-stationarity and dynamism.

[0003] To address these challenges, link adaptation (LA) technology is widely used to dynamically adjust the modulation and coding scheme (MCS) based on real-time channel conditions to maintain communication reliability. Traditional automatic modulation and coding (AMC) schemes rely on signal-to-noise ratio (SNR) threshold lookup tables for MCS decisions. However, these methods are inherently susceptible to channel estimation errors and feedback delays. In wireless environments with significant time selectivity or rich multipath propagation characteristics, they struggle to accurately characterize the complex propagation dynamics of the channel, thus weakening their ability to maintain reliable adaptive transmission. Furthermore, relying solely on MCS adjustment while ignoring other link parameters such as signal bandwidth limits the optimization dimensions of traditional link adaptation methods, making it difficult to meet the stringent performance requirements of diverse future network scenarios.

[0004] Deep learning (DL) methods, especially convolutional neural networks (CNNs), with their local receptive field mechanism and parameter-sharing structure, can efficiently extract discriminative features from channel state information (CSI), demonstrating good robustness and generalization ability in scene identification (SI) tasks. However, as network depth and model size increase, the computational complexity and storage overhead of CNNs grow rapidly, posing a serious challenge to their deployment on resource-constrained edge devices.

[0005] Furthermore, deep reinforcement learning (DRL) methods, which combine environment awareness with policy optimization, have been used for link adaptive decision-making. However, existing DRL link adaptive schemes typically only use SNR as the state input, lacking prior awareness of the channel propagation scenario, resulting in insufficient decision-making accuracy and poor convergence stability in heterogeneous multipath fading environments. In addition, existing schemes generally limit the decision space to MCS adjustment, failing to incorporate signal bandwidth into joint optimization, thus restricting further improvements in system throughput.

[0006] In summary, the key issue that urgently needs to be addressed is how to reduce the computational complexity of scene recognition models while ensuring recognition accuracy, and how to combine scene prior information with deep reinforcement learning to achieve joint adaptive optimization of MCS and signal bandwidth in order to cope with the communication challenges in complex and non-stationary IoT environments. Summary of the Invention

[0007] Therefore, it is necessary to provide an IoT link optimization method with an adaptive mechanism that has scene awareness capabilities to address the above-mentioned technical problems. This method has the characteristics of diverse optimization dimensions and strong robustness.

[0008] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows: An IoT scene sensing and link adaptive optimization method includes the following steps: Channel state information is acquired at the communication receiving end of the Internet of Things, and the real and imaginary components of the channel state information are constructed into a two-dimensional feature matrix. Based on the two-dimensional feature matrix, the current channel propagation scenario is classified and identified to obtain the scenario category. Based on the scenario category and channel quality parameters, a reinforcement learning state vector is constructed. The combination of modulation and coding scheme and signal bandwidth is used as the reinforcement learning action space. The state vector is input into a deep reinforcement learning agent to infer the optimal link configuration with the goal of maximizing system throughput. The optimal link configuration is transmitted to the transmitting end of the Internet of Things via a feedback link to achieve adaptive adjustment of the link transmission parameters.

[0009] Preferably, a scene recognition module is used for classification; the scene recognition module includes a group of convolutional feature extraction layers, a dimensionality reduction layer and a classification layer connected in sequence; The convolutional feature extraction layer group includes multiple convolutional blocks to perform multi-level feature extraction on the two-dimensional feature matrix; The dimensionality reduction layer includes a pooling layer that locally downsamples the temporal dimension features, and a global average pooling layer that globally aggregates the full temporal dimension features. The classification layer performs scene classification on the global aggregated features output by the global average pooling layer and outputs the scene category.

[0010] Furthermore, specifically, the scene recognition module sequentially includes a first convolutional block, a second convolutional block, an average pooling layer, a random deactivation layer, a third convolutional block, a fourth convolutional block, a global average pooling layer, a fully connected layer, and a Softmax classification layer; A residual connection is provided between the output of the first convolutional block and the output of the second convolutional block, and cross-layer feature fusion is performed by adding elements one by one; Each of the first to fourth convolutional blocks contains multiple convolutional kernels and a batch normalization layer; wherein the first convolutional block uses ReLU as the activation function, and the second to fourth convolutional blocks use LeakyReLU as the activation function.

[0011] Furthermore, the state vector is composed of scene category, signal-to-noise ratio, and transmission timing index, and is represented as follows:

[0012] in, For time step t The state vector; The scene category is inferred by the scene recognition module; The signal-to-noise ratio is calculated based on the channel state information; This is the sequence number of the current transmission time, used to explicitly encode time evolution information.

[0013] Furthermore, the deep reinforcement learning agent includes an evaluation dual network and a target dual network, both of which contain a shared feature extractor, a state value stream, and an action advantage stream; the Q-value of the evaluation dual network for a state-action pair is estimated using the following formula:

[0014] in, The parameters for evaluating the shared feature extractor in the dual network; These are the parameters for the state-value stream; Parameters for the dominant flow of motion; Value of a state; For action The advantage value; The total number of actions in the action space; This is the traversal variable for each candidate action in the action space.

[0015] Furthermore, the deep reinforcement learning agent is trained by minimizing the following loss function:

[0016] Among them, the target value Calculated using the following formula:

[0017] in, The instant reward at time step t; Discount factor; This represents the state at time step t+1. This refers to the target dual network; These are the parameters of the shared feature extractor, state value stream, and action advantage stream in the target dual network, respectively. The evaluation of the shared feature extractor parameters of the dual network Update according to the following rules. and Follows the same update mechanism:

[0018] in, This is the learning rate.

[0019] Furthermore, the instant reward Calculated using the following piecewise reward function: when = and When >0, =100; when When =0, = 200; Other situations = 100; in, In the state Next action The system throughput obtained; The maximum achievable throughput among all candidate actions is determined by the following formula:

[0020] in, This refers to the action space.

[0021] Furthermore, the deep reinforcement learning agent employs an ε-greedy strategy π when inferring the optimal link configuration. | ): With probability 1 ε choice makes Maximum action arg ; From the action space with probability ε The action is randomly selected from the options; where ε is the exploration rate.

[0022] An IoT scene sensing and link adaptive optimization system includes: The scene recognition module, configured at the communication receiver of the Internet of Things, is used to construct a two-dimensional feature matrix from the real and imaginary components of the channel state information, and to classify and recognize the current channel propagation scene based on the two-dimensional feature matrix to obtain the scene category; The link adaptive decision module is used to construct a reinforcement learning state vector based on the scenario category and channel quality parameters, use the combination of modulation and coding scheme and signal bandwidth as the reinforcement learning action space, input the state vector into the deep reinforcement learning agent, and infer the optimal link configuration with the goal of maximizing system throughput. The parameter configuration module is used to transmit the optimal link configuration to the transmitting end of the Internet of Things through the feedback link, so as to realize the adaptive adjustment of the link transmission parameters.

[0023] Preferably, the link adaptive decision module includes an evaluation dual network, a target dual network, and an experience replay buffer; Both the evaluation dual network and the target dual network include a shared feature extractor, a state value stream, and an action advantage stream, which are used to map the state vector to state value and action advantage value, respectively, and calculate the action value function by adding the state value after centering the advantage mean. The experience replay buffer is used to store experience tuples containing states, actions, immediate rewards, and the next state, and to provide uniformly randomly sampled mini-batch samples to the evaluation dual network to support network training. The evaluation dual network is used to select the optimal action in the next state, and the target dual network is used to calculate the value of the corresponding target action. The two work together to participate in the calculation of the training target value to suppress the overestimation bias of the Q value.

[0024] The beneficial effects of this invention are as follows: This invention constructs a two-dimensional feature matrix from the real and imaginary components of channel state information to identify channel propagation scenarios. The resulting scenario category and channel quality parameters are then combined to construct a reinforcement learning state vector. This expands the basis for link adaptive decision-making from a single dimension of instantaneous channel quality to the dimension of channel propagation scenarios, solving the problems of existing link adaptive mechanisms lacking scenario awareness and adaptability to different propagation environments. Furthermore, this invention uses a combination of modulation and coding schemes and signal bandwidth as the action space. A deep reinforcement learning agent infers the optimal link configuration with the goal of maximizing system throughput, without relying on manually set thresholds or mapping rules. This solves the problems of traditional link adaptive methods having a single optimization dimension and difficulty adapting to time-varying channel conditions, improving the robustness of link transmission configuration in dynamic IoT communication environments. Attached Figure Description

[0025] Figure 1 This is a flowchart of the IoT scene perception and link adaptive optimization method of the present invention in one embodiment.

[0026] Figure 2 This is a specific algorithm diagram of the IoT scene perception and link adaptive optimization method of the present invention in one embodiment.

[0027] Figure 3 This is a CNN network structure diagram of the IoT scene perception and link adaptive optimization method of the present invention in one embodiment.

[0028] Figure 4 This is a framework diagram of the IoT scene perception and link adaptive optimization method of the present invention (Autonomous scenario identification - dueling double deep Q-network, ASI-D3QN) in one embodiment.

[0029] Figure 5 This is a visualization of the channel characteristics of three 3GPP channel models, EPA, TDL-C, and TDL-E, in one embodiment.

[0030] Figure 6 The image shows the training performance results of the IoT scene perception and link adaptive optimization method of the present invention under different learning rates in one embodiment.

[0031] Figure 7 The image shows the training performance results of the IoT scene perception and link adaptive optimization method of the present invention under different regularization parameters in one embodiment.

[0032] Figure 8 This is a training performance graph of the IoT scene perception and link adaptive optimization method of the present invention under different sample datasets in one embodiment.

[0033] Figure 9 This is a graph showing the training accuracy results of different CNN models in one embodiment.

[0034] Figure 10 This is a diagram showing the training results of the chaotic matrix for different CNN models in one embodiment.

[0035] Figure 11 This is a training performance graph of D3QN under different learning rates in one embodiment.

[0036] Figure 12 Different minimum exploration rates in one embodiment The training results of D3QN.

[0037] Figure 13 This is a diagram showing the experimental results of link adaptive optimization based on ASI-D3QN in the TDL-E scenario of one embodiment. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0039] Example 1 like Figure 1 As shown, an IoT scene perception and link adaptive optimization method includes the following steps: Channel state information is acquired at the communication receiving end of the Internet of Things, and the real and imaginary components of the channel state information are constructed into a two-dimensional feature matrix. Based on the two-dimensional feature matrix, the current channel propagation scenario is classified and identified to obtain the scenario category. Based on the scenario category and channel quality parameters, a reinforcement learning state vector is constructed. The combination of modulation and coding scheme and signal bandwidth is used as the reinforcement learning action space. The state vector is input into a deep reinforcement learning agent to infer the optimal link configuration with the goal of maximizing system throughput. The optimal link configuration is transmitted to the transmitting end of the Internet of Things via a feedback link to achieve adaptive adjustment of the link transmission parameters.

[0040] Example 2 In this embodiment, as Figure 2As shown, this embodiment provides an IoT scene perception and link adaptive optimization method, applied to a physical layer communication system using Orthogonal Frequency Division Multiplexing (OFDM) waveforms. The system introduces a scene recognition module (ASI) and a deep reinforcement learning agent based on a dual deep Q-network (D3QN) at the receiver. These two modules collaboratively construct an ASI-D3QN intelligent communication framework, enabling proactive perception of dynamic wireless channels and adaptive optimization of multi-dimensional link parameters.

[0041] At the transmitting end, the information bit stream undergoes convolutional coding and interleaving processes in sequence, then constellation mapping is performed to generate modulation symbols, which are then converted to the time domain by inverse fast Fourier transform (IFFT). A cyclic prefix (CP) is inserted to suppress inter-symbol interference (ISI) caused by multipath propagation, and finally OFDM symbols are formed and transmitted through the wireless channel.

[0042] At the receiver, the short training symbol (STS) in the preamble sequence is used for frame-level time synchronization and coarse carrier frequency offset (CFO) correction. After removing the cyclic prefix, the received signal is reconverted to the frequency domain using a Fast Fourier Transform (FFT). The frequency-domain received signal can be represented as:

[0043] in, For frequency domain signal reception, Transmit symbols in the frequency domain. Let n be the channel coefficient corresponding to the nth subcarrier, where n∈{1,…,53} is the subcarrier index, and n=27 is the DC subcarrier. It is an additive white Gaussian noise (AWGN) with variance σ².

[0044] Subsequently, a long training sequence LTS was used to compensate for the residual fine frequency offset and estimate the channel coefficients. (n). Based on the estimated channel state information (CSI), channel equalization is performed on all subsequent OFDM symbols to compensate for channel distortion during propagation, and finally the original bit stream is recovered through convolutional decoding.

[0045] The acquired complex value of CSI (n) is decomposed into real and imaginary components, constructing a 53×2 two-dimensional feature matrix, which serves as the input to the subsequent scene recognition module:

[0046] Where real[] represents the real component and imag[] represents the imaginary component.

[0047] The CSI matrix mentioned above is obtained directly from the channel estimation module without the need for additional dedicated sensing hardware. This fully utilizes the existing pilot-assisted channel estimation process in the OFDM system, ensuring the engineering feasibility of the proposed method.

[0048] In addition, the receiver calculates the instantaneous signal-to-noise ratio (SNR) based on the estimated CSI, which serves as one of the channel quality parameters for subsequent link adaptive decision-making.

[0049] like Figure 3 As shown, the scene recognition module uses a lightweight deep convolutional neural network (CNN) to classify the two-dimensional CSI feature matrix and identify the propagation scene category of the current channel. The network input tensor has a dimension of 53×2 and passes through the following structure for feature extraction and scene classification.

[0050] The first convolutional block, Conv1, contains 32 convolutional kernels with a kernel size of 9. It uses batch normalization (BN) and rectified linear function (ReLU) as activation functions to perform preliminary local feature extraction on the input feature matrix.

[0051] The second convolutional block, Conv2, contains 32 convolutional kernels of size 9, employing batch normalization and LeakyReLU as activation functions. A residual connection is established between the outputs of the first and second convolutional blocks. Element-wise addition fuses the feature maps from the previous layer with the output of the current layer, promoting cross-layer information transfer and effectively mitigating the vanishing gradient problem in deep networks. LeakyReLU introduces a small non-zero gradient slope in the negative input region, effectively alleviating the "neuron death" problem of traditional ReLU in negative activation regions.

[0052] Average pooling layer: with a kernel size of 4, it performs local downsampling on the time dimension features, thereby compressing the feature dimension and smoothing local fluctuations, and enhancing statistical stability.

[0053] Dropout layer: The dropout ratio ρ=0.1 randomly deactivates some neuron responses during the training phase, reducing the model's over-reliance on local dominant features, alleviating overfitting problems, and improving the model's generalization ability under different scenarios and SNR conditions.

[0054] The third convolutional block Conv3 and the fourth convolutional block Conv4 both contain 32 convolutional kernels with a kernel size of 9. They employ batch normalization and the LeakyReLU activation function to further abstract the features at a deeper level.

[0055] The Global Average Pooling (GAP) layer globally aggregates all temporal features output from the fourth convolutional block, mapping the time-series features into a compact global representation vector. This significantly reduces the parameter size and effectively alleviates overfitting. Compared to traditional fully connected layers, GAP achieves global integration of high-dimensional features without introducing a large number of parameters, making it a key component of the lightweight design of this invention.

[0056] Fully connected layer and Softmax classification layer: The fully connected layer contains 32 neurons and uses the ReLU activation function to perform the final mapping of global features; the Softmax classification layer outputs the probability distribution of each scene category, and takes the category with the highest probability as the scene recognition result at the current time step. Both fully connected layers incorporate L2 regularization, with a weight coefficient η=10. 4 This is to prevent excessive weighting.

[0057] In addition, batch normalization (BN) is applied to each convolutional layer. By normalizing the features within each mini-batch, it effectively suppresses internal covariate shifts, accelerates training convergence, and improves numerical stability.

[0058] The proposed scene recognition module uses cross-entropy (CE) as the loss function, employs the Adam optimizer, and sets an initial learning rate. Adaptive decay strategy: If the verification loss does not decrease for 5 consecutive epochs, the learning rate is halved, with a minimum learning rate limit of 10. 5 We optimized the parameters and trained for 50 epochs.

[0059] In this embodiment, as Figure 4 As shown, based on the current scene category identified by the ASI module, a deep reinforcement learning agent framework based on the dual deep Q-network D3QN is constructed to achieve joint adaptive optimization of the modulation and coding scheme (MCS) and signal bandwidth. D3QN combines the decoupling mechanism of the dual deep Q-network DDQN with the dual-stream structure of the dual network. By decomposing the Q-function into the state value stream and the action advantage stream, it alleviates the overestimation bias of Q-values ​​while improving the sensitivity to differences in action value, thereby achieving more accurate policy optimization.

[0060] In this embodiment, the Markov decision process is modeled as follows: State Space: The state space is designed to provide the agent with a comprehensive observation of the communication environment. The state vector *st* is composed of scene categories. Instantaneous signal-to-noise ratio and transmission timing index Together they constitute:

[0061] in, For time step t The state vector; The scene category is inferred by the scene recognition module; The signal-to-noise ratio is calculated based on the channel state information; This is the sequence number of the current transmission time, used to explicitly encode time evolution information.

[0062] Action space: The action space is constructed as the Cartesian product of the MCS set Q and the signal bandwidth set B, i.e. =Q×B. Where Q includes 9 MCS configurations: BPSK (bit rate 1 / 2), QPSK (bit rate 1 / 2), QPSK (bit rate 3 / 4), 16-QAM (bit rate 1 / 2), 16-QAM (bit rate 3 / 4), 64-QAM (bit rate 2 / 3), 64-QAM (bit rate 3 / 4), 64-QAM (bit rate 5 / 6), and 256-QAM (bit rate 3 / 4); B includes two signal bandwidth configurations: 20MHz and 40MHz. Therefore, the action space contains | =18 candidate actions, covering all combinations of MCS with wide and narrow bandwidth from low-order low bit rate to high-order high bit rate, providing the agent with sufficient decision-making flexibility.

[0063] Reward Function: The reward function is designed to guide the agent to make link configuration decisions with the goal of maximizing system throughput under different communication scenarios and channel quality conditions, while ensuring communication reliability. Define the state. Take action below The obtained system throughput is The maximum achievable throughput is:

[0064] This refers to the throughput value corresponding to the configuration that achieves the highest throughput among all candidate actions. Based on this, a segmented reward function is designed. As follows: When the selected link configuration achieves maximum throughput and the communication link is available ( = and When the value is greater than 0, a positive reward will be given. =+100, to encourage agents to prioritize high-capacity link configurations; when the selected configuration causes communication interruption ( When =0), apply a strong penalty. = 200 is used to explicitly constrain the agent to avoid communication failures; negative rewards are given for other suboptimal cases where the link is available. = 100 promotes rapid convergence of the strategy towards the optimal direction. The above three-stage reward function, by imposing a penalty greater than that for communication interruption, ensures that the agent prioritizes the availability of the communication link while pursuing maximum throughput.

[0065] The D3QN framework comprises two structurally identical dual networks (an evaluation dual network and a target dual network). Each network consists of a shared feature extractor and two parallel value estimation streams: a state-value stream (parameter θ). v Output scalar state value It reflects the overall quality of the current state; the action dominance flow (parameter θ) A Output the advantage value of each action relative to the average level. This reflects the relative merits of different actions. The Q-value of the state-action pair in the dual network is estimated using the following formula:

[0066] Where θ is the parameter for evaluating the shared feature extractor in the dual network; These are the parameters for the state-value stream; Parameters for the dominant flow of motion; Value of a state; The dominant value for the action at; For action space The base number (in this embodiment) =18); a' is the traversal variable for each candidate action in the action space. The advantage function is centered by subtracting the mean of all action advantage values, eliminating redundancy between V and A, and making the estimation of state value and action advantage more accurate.

[0067] In this embodiment, the training process is as follows: At each time step t, the agent observes the state. The Q-values ​​of each action are estimated by evaluating the dual network, and an ε-greedy strategy is used for action selection: with probability (1 / 2)... ε) Select the greedy action arg , with probability ε from the action space The action is randomly selected for exploration. The action is then executed. Then, the environment transitions to the next state. And generate instant rewards experience tuple ={ , , , The data is stored in an experience replay buffer with a capacity of 30,000 records to break temporal dependencies.

[0068] During the optimization phase, mini-batches of size m are uniformly and randomly sampled from the experience replay buffer. To mitigate the overestimation bias of the Q-value, an evaluation dual network is used to select the next state. The optimal action is determined by the target-dual network (parameters). , , Calculate the corresponding value. The training target value is defined as:

[0069] in, Discount factor ( =0.99), This is a target-dual network. The training objective of D3QN is to obtain the optimal network parameters by minimizing the following loss function:

[0070] The dual network parameters θ are updated using stochastic gradient descent. and Following the same update mechanism, the update rules are as follows:

[0071] Where α is the learning rate. Target-dual network parameters ( The parameters of the evaluation dual network are synchronously copied every 1,000 training steps to maintain the stability of the training process.

[0072] During training, the exploration rate ε gradually decreases from the initial value of 1.0 to the minimum value of 0.001, allowing the agent to gradually transition from a fully exploratory mode to a stable execution mode of the optimal policy.

[0073] During system operation, the receiver first uses the ASI module to perform real-time scene identification on the current channel and outputs the scene category. Simultaneously, the instantaneous signal-to-noise ratio is calculated based on the channel estimation results. Combined with transmission timing index Construct the current state vector The input is an evaluation dual network that has completed offline training. The evaluation dual network infers the optimal link configuration based on an ε-greedy policy and transmits it to the transmitter via a dedicated feedback link. The transmitter performs adaptive updates of the link parameters at the beginning of the next transmission frame, thereby achieving real-time matching between the link configuration and the current channel state. This closed-loop mechanism ensures that the link adaptive policy always maintains a high degree of matching with the non-stationary communication environment, effectively maintaining policy optimality and ensuring the stability of communication performance under dynamic conditions.

[0074] Example 3 This embodiment is based on a self-built intelligent communication prototype platform, ICP, for verification. The ICP platform hardware consists of two USRPX310 software-defined radios (one as the transmitter and one as the receiver), and the software environment is implemented using LabVIEW NXG. It features a graphical user interface for real-time monitoring and control of the end-to-end link and integrates an eight-channel channel simulator to reproduce the real propagation environment. The platform operates at a 5.9 GHz carrier frequency and supports two configurable bandwidths: 20 MHz and 40 MHz.

[0075] In this embodiment, as Figure 5 As shown, three representative 3GPP channel models were selected: Extended Pedestrian A (EPA), Tapped Delay Line-C (TDL-C), and Tapped Delay Line-E (TDL-E) to verify the scene recognition accuracy of the ASI module in a heterogeneous environment. CSI samples were collected under the three 3GPP standard channel models: EPA, TDL-C, and TDL-E. The three scenarios showed significant differences in time-frequency characteristics. In the figure, (a) represents EPA, (b) represents TDL-C, and (c) represents TDL-E. The EPA scenario has a smaller channel delay spread and a relatively flat channel response; the TDL-C scenario has rich multipath propagation and exhibits obvious frequency-selective fading; the TDL-E scenario is a line-of-sight (LoS) scenario. Approximately 30,000 CSI samples were collected for each scenario. The corresponding throughput indicators were obtained and labeled based on exhaustive link configuration (18 actions). Offline training and testing were completed in a Python / Keras (TensorFlow backend) environment.

[0076] To determine the optimal hyperparameter configuration, the system evaluated the impact of different learning rates λ, Dropout ratios ρ, and L2 regularization coefficients η on model performance.

[0077] like Figure 6 As shown in Table 1, in the learning rate comparison experiment, when λ=10 -3 The model rapidly achieved an accuracy exceeding 90% during the initial training phase and maintained stable convergence, with an average training accuracy (AveACC) of 0.9952 and an average loss (AveLoss) of only 0.017, significantly outperforming other configurations. λ=10 -1 The worst performing result was observed, with a significantly delayed convergence process. Considering both efficiency and accuracy, λ=10 was ultimately selected. -3 The optimal learning rate is shown in Figure (a), which represents the training accuracy result, and Figure (b), which represents the training loss result.

[0078] Table 1 Training performance at different learning rates

[0079] As shown in Table 2, in the regularization parameter comparison experiment ( Take a value of 0.1 to 0.5. Take 10 -4 Up to 10 -1 ),when When =0.1, regardless of The model exhibits optimal performance regardless of the dropout value, indicating that moderate Dropout is beneficial for feature extraction and generalization; Figure 7 It gives the following: =10 -3 and =10 -1 Under different conditions The training loss curve corresponding to the value. When =10 -4 At that time, the model exhibited superior loss curve characteristics in the initial stage, decreasing rapidly and then converging stably. With... As the initial residual loss increases significantly, the convergence rate slows down considerably, indicating that excessively strong regularization weakens optimization efficiency. This is especially true in... =10 -1 At that point, the initial loss was close to 1.3, indicating a significant early optimization hurdle. When the value is fixed at 0.1, the recognition accuracy decreases monotonically as η increases, indicating that excessive L2 regularization will suppress the model's expressive ability. =0.1、 =10 -4 The combination achieved the highest average accuracy of 0.9954.

[0080] Table 2 differs and Average training under configuration ( =10 -3 )

[0081] like Figure 8 As shown in Table 3, in the dataset size comparison experiment (N is 5,000, 10,000, 20,000, 30,000, 40,000, and 50,000 respectively), the accuracy and loss curves fluctuate significantly when N=5,000, indicating that insufficient data will limit feature learning; the convergence trend stabilizes after N increases; when N exceeds 30,000, the performance improvement gradually approaches saturation, so N=30,000 is selected as the training dataset size. Figure 8 (a) shows the training accuracy result. (b) shows the training loss result.

[0082] Table 3 Training results for different sample datasets N (×10) -4 )

[0083] In summary, the final hyperparameter configuration is: λ=10 -3 ρ=0.1, η=10 -4 N=30,000, training for 50 epochs.

[0084] like Figure 9 As shown in Table 4, under uniform hyperparameter settings, the proposed CNN (a) is compared with CNN (41,795 parameters) (b), DCNN (51,413 parameters) (c), and LSTM network (30,628 parameters) (d). The steady-state recognition accuracy of all methods converges to approximately 99%, but the proposed method shows faster accuracy improvement in the initial training phase, the smallest generalization gap between training and validation curves, the lowest steady-state loss, and the smallest numerical fluctuation, exhibiting the best anti-overfitting ability and numerical stability. Specifically, the proposed method outperforms CNN (AveAcc=0.9932, 41,795 parameters) and DCNN (AveAcc=0.9870, 51,413 parameters) under the conditions of AveAcc=0.9959 and approximately 30,019 parameters.

[0085] Table 4 Training results of different methods

[0086] like Figure 10 As shown, in the confusion matrix analysis, among the CNN (a) proposed in this invention and CNN (41,795 parameters) (b), DCNN (51,413 parameters) (c), and LSTM network (30,628 parameters) (d), the CNN method proposed in this invention achieved the highest discrimination accuracy in all test scenarios, with the vast majority of test samples accurately distributed along the main diagonal. It performed most robustly in the TDL-E scenario, achieving an overall accuracy of 99.75% in the EPA scenario with only a very small number of misclassifications. In contrast, the inter-class confusion of the DCNN and LSTM methods was more pronounced.

[0087] In this embodiment, the model was also systematically trained according to the aforementioned D3QN framework under the key hyperparameter configurations listed in Table 5. Figure 11 The training performance of D3QN under different learning rates is shown, where the horizontal axis represents the training round index. Specifically, Figure 11 (a) describes the action selection accuracy for each round. Figure 11 (b) presents the cumulative reward evolution process starting from the initial training phase.

[0088] Table 5 Key parameters for D3QN training

[0089] Experimental results show that a learning rate of 10-2 With 5×10 -3 This configuration enables more accurate decision-making and faster convergence. With the above configuration, the model reaches a stable optimal policy range within approximately 300 episodes and maintains stable performance in subsequent training. In contrast, when the learning rate is 10... -3 At this rate, the model requires nearly 700 episodes to reach a stable performance range, but ultimately achieves the highest steady-state decision accuracy, with corresponding cumulative rewards superior to other candidate learning rates. Conversely, when the learning rate is below 10... -3 At that time, the learning efficiency dropped significantly; even after 1,500 episodes, the action selection accuracy remained below 50%, and the cumulative reward hardly increased, indicating that there was a significant convergence failure problem in the training process.

[0090] like Figure 12 As shown, minimum exploration rate This has a significant impact on the decision-making performance and policy stability of D3QN agents. Although all configurations exhibit a rapid learning phase within the first 300 episodes, steady-state performance decreases with... Significant differences arise from different values. Higher exploration rates (e.g.) =0.1 and =0.05) Due to continuous random perturbations, the accuracy of the actions is limited to approximately 90% and 95% respectively, forming a performance ceiling. Conversely, when Lowering the threshold to 0.001 effectively suppresses random fluctuations, achieving the maximum number of correct action selections and the most robust convergence curve. This indicates that setting a low exploration lower bound during the convergence phase is crucial for the stable execution of the agent's transition from the environment exploration phase to the optimal link adaptation strategy.

[0091] Example 4 This embodiment performs system verification of the link adaptation performance of the ASI-D3QN framework on the ICP platform, and compares it with the traditional link adaptation scheme based on SNR threshold and the scene awareness scheme that only adjusts MCS based on ASI. Figure 13 As shown in Table 6, the proposed ASI-D3QN algorithm was compared with the traditional method based on SNR threshold under the condition that the channel simulator output power was set to -40dB. Specifically, 500-second independent tests were conducted in EPA, TDL-C and TDL-E scenarios, and a 1-minute sliding window was used to statistically analyze the average throughput and bit error rate (BER).

[0092] Table 6. Statistical Experimental Results of Link Adaptive Optimization Based on ASI-D3QN

[0093] Traditional link adaptive schemes based on SNR thresholds: Under additive white Gaussian noise channel conditions, this method can select the optimal MCS based on SNR. However, in actual multipath fading channels, due to the significant differences in multipath richness and fading characteristics in different scenarios, relying solely on SNR for decision-making often lacks reliability. This leads to a tendency to select higher-order MCSs (such as 256-QAM(3 / 4)) in complex propagation environments, deviating from the actual optimal operation. By using channel characterization information to adjust the scene-aware MCS, a more reasonable MCS selection point can be achieved in different propagation environments.

[0094] Scene-aware MCS adjustment scheme: Utilizing the precision of the ASI module, under a fixed bandwidth of 20MHz, 64-QAM (3 / 4), 64-QAM (3 / 4), and 64-QAM (5 / 6) are selected as the optimal throughput configurations for EPA, TDL-C, and TDL-E scenarios, respectively. However, this scheme is essentially limited to a single MCS parameter dimension and does not jointly consider signal bandwidth optimization, thus limiting potential performance improvements in complex multipath fading environments.

[0095] The ASI-D3QN joint optimization scheme of this invention achieves coordinated optimization of MCS and bandwidth by explicitly incorporating signal bandwidth into the decision variable. In the EPA scenario, 16-QAM (3 / 4) with a 40MHz bandwidth is selected; in the TDL-E scenario, 64-QAM (2 / 3) with a 40MHz bandwidth is selected; and in the TDL-C scenario, the combination of 64-QAM (3 / 4) with 20MHz is maintained. This joint optimization strategy achieves higher system throughput in various propagation environments.

[0096] Experimental results show that ASI-D3QN, while dynamically adjusting link parameters, consistently maintains the bit error rate (BER) within a certain range. Within the threshold, the method achieves continuous throughput maximization while ensuring communication reliability, fully verifying the superior adaptability and significant performance advantages of the proposed method in complex multipath fading environments.

[0097] In summary, the ASI-D3QN framework proposed in this invention effectively solves the robustness problem of link adaptation in complex and dynamic IoT wireless environments by integrating lightweight CNN scene recognition, perception-enhanced state design, joint optimization of MCS and bandwidth, and precise D3QN policy optimization. It outperforms traditional methods in both recognition robustness and link adaptation policy optimization, providing an effective solution for intelligent link optimization in complex and dynamic wireless scenarios.

Claims

1. A method for IoT scene perception and link adaptive optimization, characterized in that, Includes the following steps: Channel state information is acquired at the communication receiving end of the Internet of Things, and the real and imaginary components of the channel state information are constructed into a two-dimensional feature matrix. Based on the two-dimensional feature matrix, the current channel propagation scenario is classified and identified to obtain the scenario category. Based on the scenario category and channel quality parameters, a reinforcement learning state vector is constructed. The combination of modulation and coding scheme and signal bandwidth is used as the reinforcement learning action space. The state vector is input into a deep reinforcement learning agent to infer the optimal link configuration with the goal of maximizing system throughput. The optimal link configuration is transmitted to the transmitting end of the Internet of Things via a feedback link to achieve adaptive adjustment of the link transmission parameters.

2. The method of claim 1, wherein, Specifically, a scene recognition module is used for classification and recognition; the scene recognition module includes a group of convolutional feature extraction layers, a dimensionality reduction layer and a classification layer connected in sequence. The convolutional feature extraction layer group includes multiple convolutional blocks to perform multi-level feature extraction on the two-dimensional feature matrix; The dimensionality reduction layer includes a pooling layer that locally downsamples the temporal dimension features, and a global average pooling layer that globally aggregates the full temporal dimension features. The classification layer performs scene classification on the global aggregated features output by the global average pooling layer and outputs the scene category.

3. The method of claim 2, wherein, Specifically, the scene recognition module sequentially includes a first convolutional block, a second convolutional block, an average pooling layer, a random deactivation layer, a third convolutional block, a fourth convolutional block, a global average pooling layer, a fully connected layer, and a Softmax classification layer; A residual connection is provided between the output of the first convolutional block and the output of the second convolutional block, and cross-layer feature fusion is performed by adding elements one by one; Each of the first to fourth convolutional blocks contains multiple convolutional kernels and a batch normalization layer; wherein the first convolutional block uses ReLU as the activation function, and the second to fourth convolutional blocks use LeakyReLU as the activation function.

4. The method according to claim 2, characterized in that, The state vector is composed of scene category, signal-to-noise ratio, and transmission timing index, and is represented as follows: wherein, is a state vector at time step t ; is a scene class inferred by the scene recognition module; is a signal-to-noise ratio calculated based on the channel state information; is a sequence number at the current transmission time, used for explicitly encoding time evolution information.

5. The method of claim 1, wherein, The deep reinforcement learning agent includes an evaluation dual network and a target dual network, both of which contain a shared feature extractor, a state value stream, and an action advantage stream; the Q-value of the evaluation dual network for a state-action pair is estimated using the following formula: in, The parameters for evaluating the shared feature extractor in the dual network; These are the parameters for the state-value stream; Parameters for the dominant flow of motion; Value of a state; For action The advantage value; The total number of actions in the action space; This is the traversal variable for each candidate action in the action space.

6. The method according to claim 5, characterized in that, The deep reinforcement learning agent is trained by minimizing the following loss function: Among them, the target value Calculated using the following formula: in, The instant reward at time step t; Discount factor; This represents the state at time step t+1. This refers to the target dual network; These are the parameters of the shared feature extractor, state value stream, and action advantage stream in the target dual network, respectively. The evaluation of the shared feature extractor parameters of the dual network Update according to the following rules. and Follows the same update mechanism: in, This is the learning rate.

7. The method according to claim 6, characterized in that, The instant reward Calculated using the following piecewise reward function: when = and When >0, =100; when When =0, = 200; Other situations = 100; in, In the state Next action The system throughput obtained; The maximum achievable throughput among all candidate actions is determined by the following formula: in, This refers to the action space.

8. The method according to claim 5, characterized in that, The deep reinforcement learning agent employs an ε-greedy strategy π when inferring the optimal link configuration. | ): With probability 1 ε choice makes Maximum action arg ; From the action space with probability ε Randomly select actions; Where ε is the exploration rate.

9. An IoT scene perception and link adaptive optimization system, characterized in that, include: The scene recognition module, configured at the communication receiver of the Internet of Things, is used to construct a two-dimensional feature matrix from the real and imaginary components of the channel state information, and to classify and recognize the current channel propagation scene based on the two-dimensional feature matrix to obtain the scene category; The link adaptive decision module is used to construct a reinforcement learning state vector based on the scenario category and channel quality parameters, use the combination of modulation and coding scheme and signal bandwidth as the reinforcement learning action space, input the state vector into the deep reinforcement learning agent, and infer the optimal link configuration with the goal of maximizing system throughput. The parameter configuration module is used to transmit the optimal link configuration to the transmitting end of the Internet of Things through the feedback link, so as to realize the adaptive adjustment of the link transmission parameters.

10. The system according to claim 9, characterized in that, The link adaptive decision module includes an evaluation dual network, a target dual network, and an experience replay buffer; Both the evaluation dual network and the target dual network include a shared feature extractor, a state value stream, and an action advantage stream, which are used to map the state vector to state value and action advantage value, respectively, and calculate the action value function by adding the state value after centering the advantage mean. The experience replay buffer is used to store experience tuples containing states, actions, immediate rewards, and the next state, and to provide uniformly randomly sampled mini-batch samples to the evaluation dual network to support network training. The evaluation dual network is used to select the optimal action in the next state, and the target dual network is used to calculate the value of the corresponding target action. The two work together to participate in the calculation of the training target value to suppress the overestimation bias of the Q value.

Images