Link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks

Abstract

The application discloses a kind of link life prediction and uncertainty perception routing decision method for dynamic wireless network, this method is first according to the historical state sequence of candidate link to generate prior link life prediction result with uncertainty parameter, and prior result is calibrated using a small amount of, on-demand acquisition of active probe information;On this basis, the quality of service demand vector corresponding to service identification SID is combined, the short-term link life, long-term link life and uncertainty related features are adaptively weighted, and then the routing selection for current service demand is completed.The application can improve the reliability of critical link state judgment in the dynamic wireless network environment with limited probe resources, reduce the influence of noise interference on routing selection, thereby improving the reliability and stability of routing decision.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and in particular to a link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks. Background Technology

[0002] With the rapid development of wireless communication networks, wireless ad hoc networks, and multi-hop networks, the link status between network nodes exhibits significant dynamism and uncertainty. Affected by factors such as node mobility, wireless channel fluctuations, interference variations, and node load changes, the availability, reliability, and duration of links constantly change over time, significantly impacting routing decisions and resource scheduling at the network layer. In existing technologies, routing protocols typically rely on instantaneous or short-term statistical indicators to evaluate link status, such as Received Signal Strength Indication (RSSI), Signal-to-Noise Ratio (SNR), Packet Receiver Rate (PRR), or historical packet loss rate. These methods often employ moving averages, exponential weighting, or simple threshold judgments to estimate link quality. However, these methods usually only reflect the current or short-term state of the link and are difficult to effectively predict the future sustainability of the link. When the link changes rapidly or has a potential degradation trend, it can easily lead to frequent route switching or path selection failure. To overcome these problems, some research attempts to introduce time-series-based prediction models to predict link lifetime or future quality using historical link status, such as using recurrent neural networks and Long Short-Term Memory (LSTM) networks to model the link evolution process. While such methods improve prediction accuracy to some extent, existing technologies mostly output only a single predicted value without explicitly modeling the uncertainty of the prediction results. When the distribution of training data changes, environmental interference increases, or observation information is insufficient, the reliability of the prediction results will significantly reduce the overall network performance. Existing technologies often employ fixed periods or simple rules for probing, but lack evaluation of the probing, which may mislead subsequent routing or scheduling decisions. On the other hand, in wireless networks, to improve the accuracy of link state awareness, additional observation information can often be obtained through active probing, such as sending probe packets to measure real-time packet loss rate, latency, or jitter. However, active probing introduces additional communication overhead and energy consumption. If all links are frequently probed, the quantitative evaluation mechanism of the benefits is difficult to achieve optimal allocation of probing resources within a limited probing budget. Furthermore, existing routing decision methods typically assume that predicted values ​​have equal credibility when utilizing prediction results, failing to adaptively weigh link information at different time scales based on prediction uncertainty. In scenarios with multiple services and multiple Quality of Service (QoS) requirements, this approach struggles to balance short-term performance with long-term stability, and is prone to making suboptimal or even incorrect routing choices in highly uncertain environments.

[0003] In summary, existing technologies still have shortcomings in link state prediction, prediction reliability modeling, proactive detection resource allocation, and uncertainty-based routing decisions. A technical solution is still needed that can jointly model link lifetime prediction and its uncertainties, and adaptively optimize perception and routing decisions under the condition of limited detection budget. Summary of the Invention

[0004] Purpose of the Invention: The purpose of this invention is to provide a link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks. This method aims to address the problems of rapid dynamic changes in link status, insufficient reliability of prediction results, and low reliability of routing decisions under conditions of limited probe resources in wireless networks. By analyzing the historical status of links, the invention predicts both short-term and long-term link lifetimes and characterizes their uncertainties. Based on this, it adaptively selects some links for active probes to improve the accuracy of link status awareness. Finally, it combines the prediction results and their uncertainties to make routing decisions, thereby improving the stability of routing selection and the overall network performance.

[0005] Technical Solution: 1. A link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks, comprising the following steps:

[0006] S1, Construct the link input feature vector based on the historical state data of the candidate links;

[0007] S2, the obtained input feature vector is fed into the prediction network, and the predicted mean of short-term and long-term lifetime of the link is output, as well as the uncertainty parameter corresponding to the lifetime prediction.

[0008] S3. Input the obtained prior lifetime prediction results and uncertainty index into the detection scoring network to calculate the detection value score of each candidate link; based on the detection value score, select the links with longer predicted lifetimes and higher uncertainty from the candidate links as active detection target links.

[0009] For the selected target links, an active detection mechanism is triggered to obtain the predicted value and uncertainty after detection. Then, the prior prediction result and the information after active detection are jointly input into the residual update network for joint fusion processing to obtain the updated short-term lifetime, long-term lifetime and corresponding uncertainty, forming the enhanced posterior link state.

[0010] S4, input the QoS requirement vector corresponding to the current service and the enhanced a posteriori link status into the gating network, where the QoS requirements include latency, reliability, throughput or handover stability;

[0011] By using the corresponding weight coefficients output by the gating network, short-term lifetime characteristics, long-term lifetime characteristics, and uncertainty-related characteristics are adaptively weighted.

[0012] Based on the obtained gating weights, the enhanced link status is weighted and fused to form a comprehensive decision feature oriented towards current business needs;

[0013] S5 inputs the comprehensive decision features into the routing decision network and outputs the next-hop selection probability distribution for each candidate neighbor node of the current node; based on the selection probability of the candidate neighbor nodes, the final next-hop node is determined, and the routing forwarding decision is completed.

[0014] Furthermore, the specific steps for constructing the link input feature vector include:

[0015] S11, collect historical status information of multiple consecutive moments according to a preset time interval to form a sliding window sequence; the historical status information includes RSSI, SNR, PRR, packet loss rate, latency, retransmission count, node load and service type.

[0016] S12 maps the raw data from the physical layer, MAC layer and business layer to a fixed time axis to complete the clock alignment between data from different sources.

[0017] S13, perform missing value filling, outlier smoothing and normalization on the aligned data to obtain a standardized temporal input tensor with a uniform structure;

[0018] S14 encodes and concatenates the link physical status features, service identification features, and time attribute features to form a comprehensive input vector corresponding to each time step.

[0019] Furthermore, the prediction network is built on a deep probabilistic learning architecture, and the whole adopts a multi-layer stacked LSTM as the backbone encoder: the bottom LSTM unit is responsible for extracting local temporal change patterns, the higher LSTM further aggregates the evolution patterns over a longer time scale, and Dropout, layer normalization or residual connection are configured between LSTM layers.

[0020] The predictive network backend employs a multi-head probabilistic regression layer to predict and estimate the short-term and long-term link lifetimes, respectively, and outputs the mean lifetime and uncertainty parameters.

[0021] Furthermore, the detection scoring network predicts the prior lifetime value output by the prediction network. With uncertainty indicators Given an input vector, the algorithm passes through several fully connected layers, non-linear activation functions, and normalization processes to output a detection value score for each candidate link. ;

[0022] The residual update network contains several fully connected layers and skip connections, which output the residual correction amount. and This is superimposed on the original prior to generate the enhanced posterior link state.

[0023] Furthermore, a phased decoupled training paradigm is adopted for the detection scoring network and the residual update network:

[0024] The first stage focuses on training the residual update network: by freezing the parameters of the prior prediction network, the network is trained using prior predictions, observations, and the actual lifetime. The network is trained and updated using sample pairs to minimize the mean squared error between the posterior lifetime and the true lifetime.

[0025] The second stage involves optimizing the detection scoring network, with the following specific steps:

[0026] First, by constructing a probe gain label G based on the reduction in prediction error:

[0027] ,

[0028] Among them, if A larger value indicates that the link has been significantly corrected through probing, making it a high-yield probing target; if If the value is close to zero or even negative, it indicates that probing this link is of low value.

[0029] Next, by minimizing the output of the detection scoring network and the detection reward label... The difference is used for training, and the expression is as follows:

[0030] ,

[0031] in, This represents the loss function of the scoring network, used to measure the error between the detection value score output by the scoring network and the actual detection gain. Let represent the mean of the prior link lifetime prediction for the i-th sample; This represents the prior uncertainty parameter of the i-th sample before active detection; This represents a scoring network function used to output a probe value score based on link lifetime prediction results and uncertainties. The label represents the true detection gain for the i-th sample, used to indicate how much the prediction error has been reduced after active detection.

[0032] Furthermore, the gating network employs a lightweight MLP, and the output is... After activation, normalized gating coefficients are formed.

[0033] Furthermore, the routing network adopts a comprehensive decision-making feature-based approach. The probability output layer generates a next-hop selection distribution for candidate neighbor nodes, expressed as follows:

[0034] ,

[0035] in, Indicates candidate actions; Indicates the current routing status; Represents the routing decision network; Let represent the probability of selecting candidate action a under the current routing state s; This represents the comprehensive decision-making characteristics after gating fusion.

[0036] Furthermore, when training the routing decision network, an end-to-end joint training process compatible with deep reinforcement learning is adopted; among which, the end-to-end latency actually generated during network operation is considered. Packet loss rate and link switching frequency Constructing a composite punishment and reward signal :

[0037] ,

[0038] in, , , These are the weighting coefficients for each objective item;

[0039] A proximal strategy optimization algorithm is adopted, and generalized advantage estimation is utilized. Perform closed-loop updates on the strategy parameters.

[0040] Compared with the prior art, the significant advantages of this invention are as follows:

[0041] 1. This invention jointly determines the link lifetime prediction value and uncertainty parameters, prioritizing candidate links with longer predicted lifetimes but insufficient prediction confidence as active detection targets. These links possess the potential to become stable routing links and may significantly alter routing decisions due to additional observations, thus having higher detection value. After acquiring real-time observation information of the target link, the system inputs the prior link lifetime prediction value, uncertainty parameters, and real-time observation information into the update network. The update network outputs residual corrections to calibrate the prior results for posterior state. Through this method, this invention can reduce ineffective detection of low-lifetime, low-return links under limited detection budget conditions, while avoiding instantaneous detection noise directly affecting route selection, thus improving the reliability and stability of critical link state judgment while preserving the stability of historical time-series predictions.

[0042] 2. This invention connects the Service Identifier (SID) with the Quality of Service (QoS) Demand Vector (Q): In the link lifetime prediction phase, the SID is used to distinguish different service types; in the routing decision phase, the SID is further mapped to the QoS Demand Vector (Q), which expresses the different requirements of the current service for latency, reliability, throughput, jitter, and handover stability. The gated network adaptively weights short-term link lifetime, long-term link lifetime, and uncertainty-related characteristics based on the enhanced posterior link state and the QoS Demand Vector (Q), enabling routing decisions to no longer apply the same evaluation criteria to all services, but rather dynamically adjust the link selection criteria according to service requirements. Therefore, this invention can achieve differentiated routing selection oriented towards service objectives in a dynamic wireless network environment, improving the adaptability, stability, and reliability of routing decisions.

[0043] 3. This invention links link lifetime prediction, uncertainty assessment, active probing, state calibration, and routing decision into a continuous closed loop: the prediction network outputs prior link lifetime prediction results and uncertainty parameters; the scoring network calculates the probing value score based on the prediction results; the active probing mechanism acquires real-time observation information of high-value target links; the updating network performs posterior state calibration on the prior results; and the gating network and routing decision network then transform the calibrated link state into the next-hop selection result. This closed-loop process enables link prediction results to guide probing resource allocation, active probing results to improve link state reliability, and the enhanced link state to directly serve routing decisions, thereby reducing information gaps between link measurement, state calibration, and routing execution, and improving transmission stability in dynamic wireless network environments. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall architecture of the present invention;

[0045] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0046] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0047] This invention proposes a link lifetime prediction and routing decision method for dynamic wireless networks. It constructs a two-layer architecture combining link lifetime prediction and uncertainty estimation, such as... Figure 1As shown, this architecture enables differentiated routing selection oriented towards business objectives. This two-layer architecture aims to maintain the reliability and stability of routing decisions under extreme conditions such as drastic topology changes, signal interference fluctuations, random node movement, or load imbalances by deeply quantifying environmental uncertainties. In the overall logical layout, link lifetime prediction is handled by the prediction network, active probing and uncertainty enhancement are performed by the probing scoring network, active probing network, and residual update network, and the final routing output is completed collaboratively by the gating network and the routing decision network. Link lifetime, active probing and uncertainty enhancement, and routing output share a unified link state representation space, ensuring that the prediction results can accurately guide the allocation of probing resources and directly translate into the basis for routing selection, fundamentally eliminating information silos between link measurement, state estimation, and decision execution.

[0048] The flowchart of this invention is as follows Figure 2 As shown, the specific implementation steps are as follows:

[0049] Step 1: Construct the link input feature vector based on the collected historical status information;

[0050] Step 11: Collect historical status data of candidate links;

[0051] For candidate links in the network, historical state information is collected at preset time intervals for multiple consecutive moments to form a sliding window sequence. The historical state information includes RSSI, SNR, PRR, packet loss rate, latency, retransmission count, node load, and service type.

[0052] For any candidate link in the network, a standardized data construction process is first established, using a sliding window to collect the historical state sequence of the link over a consecutive number of time steps as input. The sliding window can be preset based on the network change rate, sampling granularity, and computing resources; for example, it can take the most recent K time steps to form a time series sample of length K. Each time step consists of multi-dimensional observation information of the link, reflecting not only its physical layer communication quality but also its service carrying characteristics and temporal evolution attributes. In this way, instead of relying solely on the instantaneous state at a certain moment, it utilizes the trends, fluctuations, periodicities, and abrupt change patterns hidden in continuous time series to provide sufficient context for link lifetime prediction.

[0053] Step 12: Perform time alignment of multi-source data;

[0054] To eliminate time discrepancies between heterogeneous data from multiple sources, raw data from the physical layer, MAC (Media Access Control Layer), and business layer are uniformly mapped onto a fixed time axis to achieve clock alignment between data from different sources and ensure consistency of features at each moment.

[0055] Step 13: Complete the missing data repair and normalization process;

[0056] The aligned data is then filled with missing values, smoothed for outliers, and normalized to obtain a standardized temporal input tensor with a uniform structure.

[0057] For missing observations caused by inconsistent sampling frequencies, forward filling, linear interpolation, or estimation strategies based on statistical laws can be used to fill them in; for features of different dimensions, standardization or interval normalization methods should be used uniformly to ensure numerical stability during training.

[0058] By performing clock alignment, missing value imputation, outlier smoothing, and normalization, a time-series tensor with a regular structure and consistent dimensions is formed that can be directly input into a neural network.

[0059] Step 14: Construct the link input feature vector;

[0060] The physical status features of the link, the service identification features, and the time attribute features are encoded and concatenated to form a comprehensive input vector corresponding to each time step.

[0061] The historical state sequence of this invention not only covers physical feature tensors characterizing link quality and network environment evolution, but also... This also includes standardized key indicators such as RSSI, SNR, PRR, packet loss rate, queuing delay, transmission delay, retransmission count, and channel occupancy. Furthermore, considering that the core feature extraction layer of the prediction network uses a Long Short-Term Memory (LSTM) network, the feature structure is specifically reconstructed on the input side. Since LSTM excels at capturing long-term dependencies and dynamic trajectories in sequences, this invention discards static network identifier features that contribute little or no direct to temporal evolution, avoiding interference from invalid features in temporal learning. Simultaneously, information reflecting the business context and temporal rhythm is retained to construct the time-series data. The combined input tensor:

[0062] (1)

[0063] in, It represents the business identifier or service type label, used to distinguish the differences in link requirements under different scenarios such as real-time business, best-effort business, periodic collection business or control business. In the subsequent routing decision stage, it is further mapped to the service quality requirement vector Q to characterize the latency, reliability, throughput and switching stability requirements corresponding to different business types. It represents time attributes, which can be absolute timestamps, relative time indexes, periodic time slot numbers, or day-night cycle location codes, etc. The mapping function representing the time attribute can be represented by sine-cosine periodic coding, normalized linear projection, etc., so that the network can perceive potential periodic degradation or regular disturbances. This indicates a splicing operation. This indicates an encoding operation.

[0064] This invention employs a differentiated feature encoding technique: by embedding discrete features through an embedding layer. Labels are mapped to continuous dense vectors, enabling the potential similarities between different business types to be automatically learned in the vector space. Simultaneously, periodic mapping or normalization is applied to the temporal attributes to enhance the model's ability to express temporal context. Subsequently, the transformed temporal features, dense vectors of business requirements, and physical feature tensors are concatenated and fused to form a unified input sequence suitable for sequence learning. This design preserves the authenticity of the original link observations while also endowing the model with the ability to recognize that "the same physical state has different lifetime meanings in different business and temporal contexts."

[0065] Step 2: Feed the input feature vector into the constructed prediction network to extract the link temporal evolution features;

[0066] The obtained time-series input vector is fed into the prediction network, and short-term fluctuation information and long-term evolution trend information of the link in the time dimension are extracted by stacking LSTM.

[0067] The prediction network of this invention is built on a deep probabilistic learning architecture, employing a multi-layered stacked LSTM as the backbone encoder: the bottom-layer LSTM units are responsible for extracting local temporal variation patterns, such as short-term fading, sudden increases in interference, and instantaneous congestion; the higher-layer LSTMs further aggregate evolution patterns over longer time scales, such as node movement trends, cyclical load changes, and slow link degradation. Through this hierarchical temporal design, hidden states can be extracted from complex historical trajectories. This is used to characterize the potential evolution trend of the link at the current moment and its future sustainability. To improve generalization ability and training stability, Dropout, layer normalization, or residual connections can also be configured between LSTM layers to prevent overfitting and alleviate the gradient vanishing problem in deep network training.

[0068] Based on shared time-series representation, the prediction network backend employs a multi-head probabilistic regression layer to predict and estimate short-term and long-term link lifetimes respectively. It can also output the mean lifetime and uncertainty parameters simultaneously under a unified output head. Unlike traditional methods that only output a single lifetime estimate, this invention outputs link lifetime distribution parameters through probabilistic regression, thus explicitly expressing the prediction confidence level. To ensure the output has reasonable physical meaning, the output layer introduces a Softplus activation function to optimize the expected mean lifetime. with standard deviation Set constraints to ensure nonnegative Hengzheng, the expression is as follows:

[0069] (2)

[0070] (3)

[0071] in, , These are the weight matrices for the corresponding regression heads, and... , These are the bias terms for the corresponding regression heads. A very small positive number is added to prevent the denominator from being zero.

[0072] Formulas (2) and (3) above make the output of the prediction network no longer a rigid, fixed value, but can adaptively widen or tighten the prediction interval according to the interpretability of the input state.

[0073] During the pre-training phase, the prediction network utilizes negative log-likelihood loss (NLL) for backpropagation optimization, and its loss function... Defined as:

[0074] (4)

[0075] in, Indicates sample The true link lifetime label, where N represents the number of samples involved in the loss calculation. This loss function has a dual constraint: on the one hand, when the predicted mean... When the deviation from the true lifetime is large, the loss will increase significantly, forcing the learning of a more accurate lifetime estimate; on the other hand, blindly outputting a small prediction standard deviation on highly difficult samples will also lead to problems. Similarly, such errors will be penalized, so the network will automatically learn to output higher uncertainty under unpredictable, highly volatile, and information-scarce conditions. In other words, this loss not only trains "how accurate the predictions are," but also "when to acknowledge uncertainty." This ability is the key foundation for subsequent proactive exploration and robust routing decisions.

[0076] Therefore, the prediction network of this invention mainly performs two functions: first, it generates expected prediction results of future link lifetime based on historical link state sequences; second, it outputs corresponding uncertainty parameters according to the fluctuation, completeness and interpretability of input features, thereby forming a link state characterization that simultaneously includes lifetime prediction values ​​and prediction confidence, providing a basis for subsequent active detection screening, state calibration and risk-aware routing decisions.

[0077] Based on the hidden features extracted by LSTM, a probabilistic regression layer is used to output the predicted mean of the short-term and long-term lifetimes of the link. Along with the predicted lifetime value, an uncertainty parameter corresponding to the lifetime prediction is further output to characterize the reliability of the current prediction result.

[0078] Step 3: Input the obtained prior lifetime prediction results and uncertainty index into the detection scoring network to calculate the detection value score of each candidate link;

[0079] Based on the detection value score, links with longer predicted lifetimes and higher uncertainties are selected from the candidate links as active detection target links. For the selected target links, the active detection mechanism is triggered to obtain the detected predicted value and uncertainty. The prior prediction results and the information after active detection are jointly input into the residual update network for fusion processing.

[0080] Building upon link lifetime prediction, this invention further constructs an active detection strategy based on uncertainty and lifetime characteristics. This strategy aims to improve the state reliability of critical links through a small number of on-demand, targeted active detection actions. This strategy abandons the inefficient methods of traditional network-wide polling, fixed-period detection, or random sampling detection, instead establishing a "targeted on-demand detection" selection principle. That is, detection is only carried out on links that are most likely to affect decision quality and are most likely to be effectively corrected through additional observations, thereby achieving an efficient balance between communication overhead and state reliability.

[0081] Specifically, this invention deploys a detection scoring network based on a lightweight multilayer perceptron. This detection scoring network predicts the prior lifetime value output by the network. With uncertainty indicators The input vector can be further expanded to include auxiliary information such as the current queue state of the candidate link, the number of neighbors, and the number of historical handovers. The detection scoring network performs low-cost mapping of the input through several fully connected layers, nonlinear activation functions, and normalization processing, and outputs a detection value score corresponding to each candidate link. Due to its lightweight structure, it can perform rapid inference even when network node resources are limited, without imposing a significant additional burden on the real-time routing process.

[0082] The core logic of the probe scoring network lies in automatically identifying links with "long predicted lifetimes but high uncertainty." For links with short predicted lifetimes, their approaching failure trend is usually quite clear, and even additional probes are unlikely to change the conclusion that they are unsuitable as primary routes, thus eliminating the need to waste valuable resources on verification. For links with low uncertainty, indicating a high degree of confidence, their predictions can be directly accepted, also requiring no additional probes. Only those links whose predictions show potential for use but lack sufficient certainty are truly worth verifying through real-time observation. This strategy is essentially a value-oriented information gathering mechanism, focusing limited probe budgets on areas "most likely to influence the decision boundary."

[0083] For the high-value target links identified above, an active detection mechanism is immediately triggered to obtain the predicted values ​​and uncertainties after detection; real-time physical status data such as RSSI, PRR, latency, link instantaneous bit error rate, and repeatability are collected. These observations can be obtained through short probe packets, link feedback confirmation information, or implicit measurement results carried in existing service packets, thereby minimizing additional network burden. Unlike traditional approaches, this invention does not simply replace the original predicted values ​​with the observation results. Instead, it uses these physical observations as post-verification evidence, inputting them together with the original prior prediction results into a residual update network employing a residual network architecture for fusion and correction.

[0084] The residual update network contains several fully connected layers and skip connections. Its design goal is not to directly re-predict link lifetime, but rather to learn "how much to correct the original prior results based on the observed information." Specifically, the residual update network outputs the residual correction amount. and Then and Superimposed on the original prior, the enhanced posterior link state is generated:

[0085] (5)

[0086] (6)

[0087] in, , These represent the mean lifetime and uncertainty of the prediction network output, respectively, both derived from the prediction network. , and represent the posterior lifetime mean and posterior uncertainty obtained after fusing real-time observations, respectively.

[0088] Using a residual update network has significant advantages: firstly, prior predictions already provide strong basic estimates, and the update network only needs to focus on bias correction, making the learning process easier; secondly, the residual structure can prevent observation noise from causing drastic fluctuations in the output, improving the stability of the update process. Especially when observations are limited, observation errors are large, or instantaneous state fluctuations are significant, residual update is more robust than complete reestimation.

[0089] This process forms a closed-loop mechanism from "prediction, detection, fusion" to "state calibration" in the inference phase. That is, the prediction network provides a global prior, the detection scoring network determines whether further detection is worthwhile, and the residual update network uses limited physical evidence to perform posterior corrections on critical links, ultimately outputting an enhanced link representation that more closely reflects the true state and is more suitable for routing decisions. In this way, without the need for intensive measurement of all links, the ambiguity of the link states of critical nodes can be significantly reduced with extremely low interference and minimal additional overhead.

[0090] To ensure the effective operation of the aforementioned collaborative mechanism, this invention adopts a phased decoupling training paradigm to reduce the instability caused by gradient interference during joint training. The first phase focuses on training the residual update network: by freezing the parameters of the preceding prediction network, the network uses prior predictions, observations, and the actual lifetime. The network is trained and updated using sample pairs, with the objective of minimizing the mean squared error (MSE) between the posterior lifetime and the true lifetime:

[0091] (7)

[0092] in, This represents the loss function for updating the network, used to train and update the network so that the posterior link lifetime prediction is closer to the true value. This represents the mean posterior link lifetime prediction for the i-th sample after active detection and network update correction.

[0093] The essential goal of this stage is to enable the residual update network to learn the mapping ability of "how to correct prior errors based on finite observations", so that it can make reasonable adjustments under different uncertainty levels, different observation qualities and different link evolution modes.

[0094] The second stage focuses on optimizing the MLP probe scoring network. To enable the probe scoring network to truly learn "which links are worth probing," this invention constructs a probe reward label G based on the reduction in prediction error:

[0095] (8)

[0096] in, This reflects how much the prediction error has decreased relative to the prior error after introducing real-time observations and updating the network. If A larger value indicates that the link has been significantly corrected through probing, making it a high-yield probing target; if A value close to zero or even negative indicates low probe value for that link. The probe scoring network minimizes the difference between its output and the probe reward label. Training based on differences:

[0097] (9)

[0098] in, This represents the loss function of the scoring network, used to measure the error between the detection value score output by the scoring network and the actual detection gain. This represents the mean of the prior link lifetime prediction for the i-th sample, i.e., the mean of the prior link lifetime prediction before active detection. This represents the prior uncertainty parameter of the i-th sample before active detection; This represents a scoring network function used to output a probe value score based on link lifetime prediction results and uncertainties. The label represents the true detection gain for the i-th sample, used to indicate how much the prediction error has been reduced after active detection.

[0099] This training mechanism automatically learns a near-optimal detection value function, ensuring that detection actions are concentrated in critical scenarios where "predictions appear usable but lack credibility and are easily corrected by observation." This avoids ineffective detection of low-value links and reduces problems such as increased bandwidth consumption, energy consumption, and control latency caused by global detection.

[0100] The prior predictions are corrected by the residual update network to obtain the updated short-term lifetime, long-term lifetime and corresponding uncertainty, forming the enhanced posterior link state.

[0101] Step 4: Input the QoS requirement vector corresponding to the current service and the obtained enhanced link status into the gating network. The QoS requirements include requirements such as latency, reliability, throughput or handover stability.

[0102] By utilizing the weight coefficients output by the gating network, short-term lifetime characteristics, long-term lifetime characteristics, and uncertainty-related characteristics are adaptively weighted. Based on the obtained gating weights, the enhanced link states are weighted and fused to form comprehensive decision features oriented towards current business needs.

[0103] To effectively transform the enhanced link-state information at the front end into stable and reliable routing choices, this invention introduces an uncertainty-gated network based on a soft attention mechanism at the decision layer. This gated network receives the updated enhanced state. and the service quality requirements vector of the current business As a joint input. Among them, the enhanced state. It can include posterior lifetime mean, posterior uncertainty, short-term and long-term lifetime components, instantaneous observation correction signal, and necessary auxiliary link attributes; QoS requirement vector. This is used to describe the different preferences of current services for low latency, high reliability, low jitter, low switching frequency, or high throughput. By inputting both link status and service intent into the decision layer, the rigid selection method of "treating all services the same way" can be avoided, enabling differentiated routing based on service objectives.

[0104] Gated networks are essentially lightweight MLPs whose outputs are processed by... After activation, normalized gating coefficients (attention weights) are formed. :

[0105] (10)

[0106] in, This represents the gating weight coefficient, used to adjust the importance of different features in routing decisions; denoted as a multilayer perceptron in a gating network, used to calculate gating scores based on input features; e represents the natural constant. Q represents the enhanced post-link state; Q represents the QoS requirement vector of the current service.

[0107] This gating coefficient This is used to dynamically adjust the relative importance of short-term and long-term lifetime characteristics in decision-making. For example, for ultra-low latency services, more emphasis can be placed on short-term stability to avoid immediate link collapse; for continuous transmission services, more attention is paid to balancing long-term lifetime with handover costs. Simultaneously, gating networks can also integrate uncertainty indicators to suppress high-risk, high-noise links, thus forming a comprehensive characteristic. In other words, gating networks are not simply weighted sums, but rather risk-sensitive re-expressions of augmented states under the constraints of business requirements.

[0108] Step 5: Input the comprehensive decision features into the routing decision network and output the next-hop selection probability distribution for each candidate neighbor node of the current node; determine the final next-hop node based on the selection probability of the candidate neighbor nodes and complete the routing forwarding decision.

[0109] Routing decision networks are based on comprehensive decision characteristics use The probability output layer generates a next-hop selection distribution for candidate neighbor nodes:

[0110] (11)

[0111] in, This represents the candidate action, i.e., the choice of forwarding from the current node to different neighboring nodes; Indicates the current routing status. This represents a routing decision network, also known as a multilayer perceptron. Let represent the probability of selecting candidate action a under the current routing state s; This represents the comprehensive decision-making characteristics after gating fusion.

[0112] The routing decision network outputs a probability distribution rather than a hard decision. This allows it to maintain policy exploration capabilities during training and enables flexible decision-making through maximum probability selection or sampling strategies during actual operation. Compared to routing methods based on fixed rules, manual thresholds, or greedy single-index optimization, the policy network of this invention can automatically learn optimal behavior patterns in complex environments under multi-objective constraints.

[0113] To endow the decision-making module with more accurate adaptive capabilities, this invention further constructs an end-to-end joint training process compatible with Deep Reinforcement Learning (DRL). After treating the routing network as a policy network, the training process is based on the actual end-to-end latency generated during network operation. Packet loss rate and link switching frequency Constructing a composite punishment and reward signal :

[0114] (12)

[0115] in, , , The weighting coefficients for each objective item can be set according to different application scenarios.

[0116] When business prioritizes reliability, the weight of packet loss can be increased; when business is sensitive to switching overhead, the weight of switching frequency can be increased. Through this reward design, the comprehensive requirements of "low latency, low packet loss, and low switching" are uniformly mapped into optimizable policy objectives.

[0117] For the optimization algorithm, the Proximal Policy Optimization (PPO) algorithm is adopted, and generalized advantage estimation is utilized. For strategy parameters Perform closed-loop updates, with the objective function... for:

[0118] (13)

[0119] in, This represents the ratio of the probabilities of the new and old strategies in terms of action probability. These are the trimming parameters. This represents the expected operation on the sampling time step or training samples; This represents the pruning function, used to restrict the policy probability ratio to a certain value. Within a certain range, to avoid excessively large changes in a single strategy update.

[0120] PPO, by limiting the magnitude of each update, avoids drastic policy oscillations, which is beneficial for stable training in complex dynamic routing networks. Since the front-end prediction network, probe scoring network, active probe network, and residual update network already provide the routing decision network with more physically realistic and risk-aware state inputs, the reinforcement learning policy network can converge to a routing policy that balances performance and stability more efficiently, without getting caught in the local optimum problem of frequently switching paths based solely on short-term instantaneous indicators.

[0121] In subsequent operation, steps 2-5 above will be repeated to form a closed-loop update mechanism for link prediction, active detection, state calibration and routing decision, thereby continuously improving transmission stability and routing reliability in dynamic network environments.

[0122] Based on the above steps, this invention achieves continuous closed-loop coordination of prediction, detection, and decision-making. In this closed loop, prediction provides target guidance for detection, giving detection actions a clear direction and cost awareness; detection, in turn, provides physical correction evidence for prediction, improving the reliability of critical link states; the enhanced states are further transformed into robust routing actions oriented towards business needs through gating and policy networks. Ultimately, in dynamic wireless network environments, the frequency of route switching is significantly reduced, the risk of link interruption due to misestimation is decreased, and a better balance is achieved between detection overhead and communication reliability.

Claims

1. A link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks, characterized in that, The steps include the following: S1, Construct the link input feature vector based on the historical state data of the candidate links; S2, the obtained input feature vector is fed into the prediction network, and the predicted mean of short-term and long-term lifetime of the link is output, as well as the uncertainty parameter corresponding to the lifetime prediction. S3. Input the obtained prior lifetime prediction results and uncertainty index into the detection scoring network to calculate the detection value score of each candidate link. Based on the detection value score, links with longer predicted lifetimes and higher uncertainty are selected from the candidate links and used as active detection target links. For the selected target links, an active detection mechanism is triggered to obtain the predicted values ​​and uncertainties after detection; Then, the prior prediction results and the information after active detection are jointly input into the residual update network for joint fusion processing to obtain the updated short-term lifetime, long-term lifetime and corresponding uncertainty, forming the enhanced posterior link state. S4, input the QoS requirement vector corresponding to the current service and the enhanced a posteriori link status into the gating network, where the QoS requirements include latency, reliability, throughput or handover stability; By using the corresponding weight coefficients output by the gating network, short-term lifetime characteristics, long-term lifetime characteristics, and uncertainty-related characteristics are adaptively weighted. Based on the obtained gating weights, the enhanced link status is weighted and fused to form a comprehensive decision feature oriented towards current business needs; S5 inputs the comprehensive decision features into the routing decision network and outputs the next-hop selection probability distribution for each candidate neighbor node of the current node; based on the selection probability of the candidate neighbor nodes, the final next-hop node is determined, and the routing forwarding decision is completed.

2. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 1, characterized in that, The specific steps for constructing the link input feature vector include: S11, collect historical status information of multiple consecutive moments according to a preset time interval to form a sliding window sequence; the historical status information includes RSSI, SNR, PRR, packet loss rate, latency, retransmission count, node load and service type. S12 maps the raw data from the physical layer, MAC layer and business layer to a fixed time axis to complete the clock alignment between data from different sources. S13, perform missing value filling, outlier smoothing and normalization on the aligned data to obtain a standardized temporal input tensor with a uniform structure; S14 encodes and concatenates the link physical status features, service identification features, and time attribute features to form a comprehensive input vector corresponding to each time step.

3. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 1, characterized in that, The prediction network is built on a deep probabilistic learning architecture, and the whole uses a multi-layer stacked LSTM as the backbone encoder: the bottom LSTM unit is responsible for extracting local time change patterns, the higher LSTM further aggregates the evolution patterns over a longer time scale, and Dropout, layer normalization or residual connection are configured between LSTM layers. The predictive network backend employs a multi-head probabilistic regression layer to predict and estimate the short-term and long-term link lifetimes, respectively, and outputs the mean lifetime and uncertainty parameters.

4. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 1, characterized in that, The detection scoring network predicts the prior lifetime value output by the network. With uncertainty indicators Given an input vector, the algorithm passes through several fully connected layers, non-linear activation functions, and normalization processes to output a detection value score for each candidate link. ; The residual update network contains several fully connected layers and skip connections, which output the residual correction amount. and This is superimposed on the original prior to generate the enhanced posterior link state.

5. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 1, characterized in that, A phased decoupled training paradigm is adopted for the detection scoring network and the residual update network: The first stage focuses on training the residual update network: by freezing the parameters of the prior prediction network, the network is trained using prior predictions, observations, and the actual lifetime. The network is trained and updated using sample pairs to minimize the mean squared error between the posterior lifetime and the true lifetime. The second stage involves optimizing the detection scoring network, with the following specific steps: First, by constructing a probe gain label G based on the reduction in prediction error: ， Among them, if A larger value indicates that the link has been significantly corrected through probing, making it a high-yield probing target; if If the value is close to zero or even negative, it indicates that probing this link is of low value. Next, by minimizing the output of the detection scoring network and the detection reward label... The difference is used for training, and the expression is as follows: ， in, This represents the loss function of the scoring network, used to measure the error between the detection value score output by the scoring network and the actual detection gain. Let represent the mean of the prior link lifetime prediction for the i-th sample; This represents the prior uncertainty parameter of the i-th sample before active detection; This represents a scoring network function used to output a probe value score based on link lifetime prediction results and uncertainties. The label represents the true detection gain for the i-th sample, used to indicate how much the prediction error has been reduced after active detection.

6. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 1, characterized in that, The gated network employs a lightweight MLP, and its output is... After activation, normalized gating coefficients are formed.

7. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 1, characterized in that, The routing network adopts based on comprehensive decision-making characteristics. The probability output layer generates a next-hop selection distribution for candidate neighbor nodes, expressed as follows: ， in, Indicates candidate actions; Indicates the current routing status; Represents the routing decision network; Let represent the probability of selecting candidate action a under the current routing state s; This represents the comprehensive decision-making characteristics after gating fusion.

8. The link lifetime prediction and uncertainty-aware routing decision method for dynamic wireless networks according to claim 7, characterized in that, When training the routing decision network, an end-to-end joint training process compatible with deep reinforcement learning is adopted; among which, the actual end-to-end latency generated during network operation is considered. Packet loss rate and link switching frequency Constructing a composite punishment and reward signal : ， in, , , These are the weighting coefficients for each objective item; A proximal strategy optimization algorithm is adopted, and generalized advantage estimation is utilized. Perform closed-loop updates on the strategy parameters.

Images