Routing communication control method and system for dynamically optimizing end-to-end time delay

Abstract

The invention provides a routing communication control method and system for end-to-end delay dynamic optimization. The method comprises the following steps: performing multi-dimensional acquisition and analysis on real-time operation states of candidate paths in a network, performing comprehensive evaluation, and dynamically selecting an optimal path with the minimum end-to-end delay at the current moment for data forwarding; after the optimal path is started, monitoring the time delay change trend of the optimal path, constructing a multi-dimensional dynamic feature, performing similarity and correlation analysis, and extracting a feature input vector; and inputting the vector into a prediction model, generating a short-term prediction result of future time delay, comparing the short-term prediction result with a set threshold, and when the prediction result shows that an over-limit risk exists, triggering a path switching or alarm mechanism. Nonlinear time sequence characteristic analysis and probability prediction judgment are introduced into path evaluation, so that active early warning and path reconstruction before time delay abnormity occurs are realized, and dynamic balance and stable transmission effect of communication time delay can be kept under the condition of link load fluctuation or node performance change.

Description

Technical Field

[0001] This invention belongs to the field of routing communication technology, specifically relating to a routing communication control method and system with dynamic end-to-end latency optimization. Background Technology

[0002] With the rapid development of the Industrial Internet and the Internet of Things, the transmission scale and complexity of data communication networks are constantly increasing, and the end-to-end real-time requirements are becoming increasingly stringent. Especially in scenarios such as smart manufacturing, unmanned logistics, vehicle-to-everything (V2X) and cloud-edge collaborative computing, the transmission latency of data between different nodes has become a key factor affecting system performance and decision-making response speed. Traditional routing algorithms mostly rely on static shortest paths or path selection mechanisms based on fixed cost functions, making it difficult to achieve stable low-latency communication under conditions of dynamic changes in network topology, fluctuations in link load, and differences in node performance.

[0003] Existing latency optimization methods typically only consider static metrics such as path bandwidth, hop count, or node congestion, lacking a dynamic characterization of network state evolution over time. When link utilization or node load changes rapidly, routing selection struggles to reflect these network state updates in a timely manner, leading to problems such as sudden increases in path latency or packet congestion. Furthermore, while some machine learning-based predictive routing methods possess a degree of adaptability, they often overlook the potential nonlinear correlations and temporal dependencies in latency sequences, resulting in poor prediction stability and difficulty in achieving high-confidence latency exceedance warnings. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention provides a routing communication control method and system for end-to-end dynamic delay optimization. This disclosure resolves issues such as delayed route selection response, insufficient path evaluation dimensions, and low prediction accuracy in existing technologies, and is applicable to dynamic delay optimization control in multi-node distributed communication environments.

[0005] This invention provides the following technical solution: According to a first aspect of this disclosure, an end-to-end latency dynamically optimized routing communication method is provided, the method being applicable to lightweight load data transmission or heavy load data transmission, the method comprising the following steps: The system collects and analyzes the real-time operating status of candidate paths in the network from multiple dimensions. It comprehensively evaluates the transmission rate, node load and link utilization of each path and dynamically selects the preferred path that can achieve the minimum end-to-end latency at the current moment from multiple candidate paths for data forwarding. After the path is enabled, the latency change trend of the preferred path is continuously monitored. The timing feature analysis and prediction model are used to determine whether the latency in the future period may exceed the threshold. When the prediction result shows that there is a risk of latency exceeding the limit, the backup path switching mechanism is triggered to maintain the minimum control of end-to-end latency.

[0006] According to a second aspect of this disclosure, an end-to-end delay dynamically optimized routing communication control system is provided, the system comprising: The path awareness and optimization module is used to collect and analyze the real-time operating status of candidate paths in the network from multiple dimensions. Based on the transmission rate, node load and link utilization of each path, it performs a comprehensive evaluation and dynamically selects the optimal path that can achieve the minimum end-to-end latency at the current moment from multiple candidate paths for data forwarding. The latency prediction and switching control module continuously monitors the latency change trend of the preferred path after the path is enabled. It uses time series feature analysis and prediction models to determine whether the latency in the future period may exceed the threshold. When the prediction result indicates that there is a risk of latency exceeding the limit, it triggers the backup path switching mechanism to maintain the minimum control of end-to-end latency.

[0007] This invention achieves a closed-loop control process from real-time perception, trend analysis, risk prediction, and dynamic switching by acquiring and evaluating path status in multiple dimensions, combined with nonlinear temporal feature extraction and predictive modeling based on time windows. By introducing state similarity analysis, reproducible quantitative feature indicators (such as decidability, stagnation, longest diagonal, and similarity entropy), and a probabilistic prediction and judgment mechanism, it can identify trends and assess risks before latency anomalies occur, and proactively execute path switching operations based on the prediction results, thereby maintaining low latency and high reliability of end-to-end communication under different load conditions. Attached Figure Description

[0008] The invention will now be described in more detail with reference to embodiments and the accompanying drawings. Figure 1 A schematic diagram of the structure of an end-to-end delay dynamic optimization routing communication control system 100 according to an embodiment of the present disclosure is shown.

[0009] Figure 2 A schematic diagram of a routing communication control system 110 with dynamic end-to-end delay optimization is shown in some other embodiments.

[0010] Figure 3 A flowchart of a routing communication control method 200 with dynamic end-to-end latency optimization provided according to an embodiment of the present disclosure is shown.

[0011] Figure 4The flowchart of the method 300 for dynamically selecting and forwarding data after collecting and analyzing the real-time operating status of candidate paths in the network according to an embodiment of the present disclosure is shown.

[0012] Figure 5 This diagram illustrates the possible transmission paths when a source forwarding node forwards data to a target node via multiple candidate relay nodes during routing communication.

[0013] Figure 6 The diagram illustrates the parameter settings changes and communication performance of nodes operating under light load scenarios with low arrival rates and low packet loss rates.

[0014] Figure 7 The diagram illustrates the parameter settings changes and communication performance of nodes under heavy load operation scenarios with high arrival rates and high packet loss rates.

[0015] Figure 8 A flowchart of a method 400 for monitoring the latency change trend of a preferred path after the path is enabled and triggering a switch based on a prediction model, according to an embodiment of the present disclosure, is shown.

[0016] Figure 9 The diagram shows the end-to-end time delay sequence and its sampling within a continuous time window.

[0017] Figure 10 The diagram illustrates the phase space reconstruction analysis process of the time delay sequence, along with the state similarity diagram and the thresholded black-and-white similarity diagram.

[0018] Figure 11 A flowchart of a method 500 for similarity and correlation analysis of time-series feature parameters according to an embodiment of the present disclosure is shown.

[0019] Figure 12 A flowchart of a method 600 for calculating multiple comprehensive feature indicators characterizing the stability and trend of time delay data based on the time delay variation law obtained by analysis, according to an embodiment of the present disclosure, is shown.

[0020] Figure 13 A flowchart of a method 700 for making decisions on path switching or issuing alarms based on future latency prediction results according to an embodiment of the present disclosure is shown.

[0021] Figure 14 A comparison chart of the distribution of three latency indicators is shown between the routing communication method with end-to-end latency dynamic optimization according to embodiments of the present disclosure and the moving average method.

[0022] Figure 15 This diagram illustrates a comparison of the initial optimal path and the path after dynamic switching in routing communication under the method of this application and existing methods. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Figure 1 A schematic diagram of an end-to-end delay dynamic optimization routing communication control system 100 according to an embodiment of the present disclosure is shown. The system 100 includes a path awareness and optimization module 102, a delay prediction and switching control module 101, and an execution layer.

[0025] The path awareness and optimization module 102 is used to perform real-time detection and feature acquisition of the operating status of each candidate path in the communication network. The path awareness and optimization module 102 can acquire multi-dimensional operating information including node load, link utilization, physical layer transmission rate, and packet loss rate, and perform statistical analysis and comprehensive evaluation of the collected data to determine the optimal path among multiple candidate paths that achieves the minimum end-to-end latency at the current moment. This module is equivalent to the system's perception layer, used to establish the mapping relationship between network operating status and upper-layer control logic, and continuously output path performance data to the decision-making layer.

[0026] The latency prediction and switching control module 101 is used to continuously monitor and predict the end-to-end latency changes of the preferred path after the path is enabled. Based on the future short latency trend results calculated by the time-series feature extraction and prediction model, module 101 compares the results with a set threshold to determine whether the path may experience latency exceeding the limit. When a latency anomaly or predicted risk is detected, module 101 generates a path switching command and sends it to the execution layer through the dynamic optimization core to perform the switching operation, thereby maintaining the dynamic minimization of end-to-end transmission latency. This module corresponds to the system's decision-making or management layer and is used to implement the closed-loop logic of global latency prediction, path evaluation, and control command generation.

[0027] The execution layer is the physical communication layer of the system of this invention. It includes various network nodes, links, and forwarding units, and is used to execute path selection and switching commands from the upper-layer control module. The execution layer receives the preferred path information provided by the path awareness and optimization module 102, and adjusts the forwarding path of data packets according to the control instructions issued by the delay prediction and switching control module 101, thereby realizing dynamic scheduling between different relay nodes. At the same time, the execution layer returns the real-time link status, path transmission rate, and feedback information to the path awareness and optimization module 102, forming a data acquisition and feedback closed loop.

[0028] In a preferred embodiment, the path awareness and optimization module 102 and the delay prediction and switching control module 101 exchange information through a "prediction and decision" core unit. This core unit is responsible for receiving path status awareness data, node queue status information, and transmission rate indicators from the perception layer, and using these as input parameters for dynamic optimization calculations. The path instructions and switching decisions generated after processing by the prediction and decision unit are synchronously sent to the execution layer for actual control of path scheduling and data flow allocation.

[0029] Furthermore, the perception layer, management layer, and execution layer form a closed-loop control structure through feedback signals. The path feedback information uploaded by the execution layer is parsed by the perception layer and then provided to the management layer for model self-correction, enabling the system to adaptively adjust and enhance robustness under different load conditions. Through this structure, stable latency optimization can be maintained in both light and heavy load scenarios, achieving dynamic optimal control of end-to-end transmission performance.

[0030] Figure 2 A schematic diagram of a routing communication control system 110 with dynamic end-to-end delay optimization is shown in some other embodiments.

[0031] In this embodiment, system 110 includes a path awareness and optimization module 111 and a delay prediction and switching control module 112; The path perception and optimization module 111 includes a path status perception unit 1111 and a path evaluation and selection unit 1112. The delay prediction and switching control module 112 includes a delay monitoring unit 1121, a prediction and judgment unit 1122 and a path switching execution unit 1123. The path status awareness unit 1111 is used to collect and analyze the real-time running status of candidate paths in the network from multiple dimensions. The path evaluation and selection unit 1112 is used to dynamically determine the preferred path that can achieve the minimum end-to-end latency at the current moment based on the comprehensive evaluation results of transmission rate, node load and link utilization obtained by the path state awareness module. The delay monitoring unit 1121 is used to continuously monitor the end-to-end delay changes after the preferred path is enabled, and to generate time-series feature data for trend identification. The prediction and judgment unit 1122 is used to generate a time delay prediction result for a short period of time in the future based on the time series feature data using a prediction model, and to determine whether the path may experience a time delay exceeding the limit in the future period. The path switching execution unit 1123 is used to trigger a switching operation of the backup path or alarm when the output result of the prediction and judgment module indicates that there is a risk of delay exceeding the limit, so as to achieve dynamic minimization control of end-to-end delay.

[0032] Those skilled in the art will understand that Figure 2 The system 110 shown is merely illustrative. In some embodiments, the end-to-end latency dynamically optimized routing communication system provided in this disclosure may contain more or fewer components than system 110.

[0033] Figure 3 A flowchart of a routing communication control method 200 for dynamic optimization of end-to-end latency according to an embodiment of this disclosure is shown. In step 201, after the terminal device generates a communication request, the system performs real-time operation status detection on the selectable paths in the network, collects the transmission rate, load status, and link utilization information of the relay nodes of each path; based on the operation status detection results, the system calculates the comprehensive transmission performance index of the candidate path at the current moment, and performs a comprehensive evaluation based on the transmission capacity, congestion level, and stability of each path to determine the preferred path that can achieve the minimum end-to-end latency; In step 202, during data transmission, the end-to-end latency changes of the currently active path are continuously monitored. A latency observation sequence within a continuous time window is obtained and features are extracted to form time-series feature data for trend identification. A prediction model is constructed based on the time-series feature data to predict the latency change trend of the path in the short term and generate corresponding prediction results. The prediction results are compared with a preset latency threshold. When the prediction results indicate that the path has a latency exceeding the limit risk or a trend of decreasing stability, a backup path switching mechanism is triggered or an early warning signal is issued to achieve dynamic minimization control of end-to-end communication latency.

[0034] This invention achieves a shift from static configuration to dynamic perception and optimization in path selection through multi-dimensional real-time acquisition and comprehensive analysis of candidate paths, enabling adaptive selection of the optimal forwarding path under varying network loads and link conditions. By introducing latency prediction and handover control mechanisms, the system can proactively determine and execute path switching before latency anomalies occur, effectively reducing the average latency and fluctuations in end-to-end communication. Through collaborative feedback between the perception layer, decision layer, and execution layer, this invention achieves closed-loop operation of path status, latency prediction, and handover control, ensuring continuous and stable communication during data transmission and rapid recovery to optimal transmission status when the network environment dynamically changes.

[0035] The above description, in conjunction with the accompanying drawings, outlines an end-to-end delay dynamic optimization routing communication control method 200 and a system 110 that can be used as an end-to-end delay dynamic optimization routing communication control system according to the present invention. However, those skilled in the art will understand that the execution of the steps of method 200 is not limited to the order shown in the figures and described above, but can be performed in any other reasonable order. Furthermore, system 110 does not necessarily include… Figure 2All components shown may include only some of the components necessary to perform the functions described in this invention, and the connection of these components is not limited to the form shown in the figure.

[0036] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0037] In one or more exemplary designs, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. For example, if implemented in software, the functions can be presented as one or more instructions or codes. Stored on a computer-readable medium, or transmitted as one or more instructions or codes on a computer-readable medium.

[0038] The various units of the apparatus disclosed herein can be implemented using discrete hardware components or integrated into a single hardware component, such as a processor. For example, they can be implemented or perform the various exemplary logic blocks, modules, and circuits described herein using a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof for performing the functions described herein.

[0039] Figure 4 The flowchart of the method 300 for dynamically selecting and forwarding data after collecting and analyzing the real-time operating status of candidate paths in the network according to an embodiment of the present disclosure is shown. Figure 5 This diagram illustrates the possible transmission paths when a source forwarding node forwards data to a target node via multiple candidate relay nodes during routing communication. Source forwarding node N... f At a given time, a candidate relay node needs to be selected. Forward the data packet to the target node N. d .

[0040] In step 301, the nth node at time t is calculated. Due to the physical layer data packet transmission rate of transmission contention : ; in, The instantaneous physical layer rate of node n toward the selected next-hop link at time t (determined by modulation coding, beamforming, etc.). Contention intensity is a normalized measure of factors such as concurrent transmission duty cycle and backoff collision rate among neighboring nodes. ; The contention penalty coefficient is determined by the media access control policy (part of the data link layer in the computer network protocol stack, used to determine how multiple nodes can fairly and efficiently share the same wireless channel). n=1,2,…,N; At this point, different candidate relay nodes have different effective bandwidths due to differences in neighbor contention. On each possible forwarding path corresponding to each candidate relay node, the effective physical layer rate of the candidate node is determined through a 301 step. This can effectively reflect the actual bandwidth that the link can provide under interference such as neighbor competition and backoff.

[0041] In step 302, for a data packet of size Pa after data grouping, its path through nodes is calculated. Transmitted single packet service time: ; Once the effective rate is determined, the single-packet service time of each candidate relay node is also obtained. Figure 5 The connection between the candidate node and the target node represents the service time required for data packet transmission.

[0042] In step 303, when the data packet arrives at the node When the process is a non-Poisson general update process, within a sliding window of length L, a set of K interval duration samples of the arrival samples of the next-hop candidate node is statistically analyzed online: And a set of service time samples with a total of U service durations: ;

[0043] Let k be the duration of the k-th interval, where k = 1, 2, ..., K; Let u be the service duration, where u = 1, 2, ..., U; for abbreviation, and then for abbreviation; The results are from statistical analysis, and This is the calculation result for step 302; Further statistical calculations yielded the empirical mean of the interval duration samples. Standard deviation Then calculate the nodes coefficient of variation : ; The empirical mean of the service duration sample was obtained through statistical calculation. Standard deviation Then calculate the nodes service variation coefficient : ; Empirical mean of interval duration samples Standard deviation and the empirical mean of the service duration sample Standard deviation These statistics reflect the traffic fluctuations and service stability of candidate nodes under different time windows, and then the final coefficient of variation is used to... and This will be reflected in the text.

[0044] In step 304, further statistical analysis results from step 303 and the nth node at time t are used. Real-time arrival rate compute nodes Business intensity : ; 1; Business intensity This determines the extent to which candidate nodes approach saturation. Business intensity. An increase indicates significant congestion on the data forwarding path where the relay node resides. The larger the value, the more likely the path will be congested.

[0045] In step 305, based on the arrival variation coefficient and service variation coefficient from step 303, the nth node at time t is obtained. Queuing delay, excluding the time it takes for data packets to be transmitted: ; When multiple data packets arrive at a relay node, a queuing process occurs, where the relay node serves the data packet from the start to the end of the forwarding process. The time from arrival to the start of forwarding is the waiting delay. Since the service performance of each node varies, the waiting delay also depends on the expected service time of a single packet from the start to the end of forwarding at that node. In addition, business intensity The degree of congestion reflects the length of the queue.

[0046] In step 306, consider the nth node at time t. Data packets were successfully transmitted according to the expected service duration. In this case, the total delay from the arrival of the data packet to its successful transmission and the delay before transmission. for: ; in, The node at time t The packet loss rate at the location. Once a packet is lost during the forwarding process of a candidate relay node (i.e., exceeding the forwarding capacity of that relay node), the data packet will be returned from the candidate relay node to the source forwarding node for retransmission (dashed arrow in the figure), which further amplifies the total latency.

[0047] Therefore, through the method 300 of this embodiment, the source forwarding node N f During packet transmission, the path of each candidate relay node needs to undergo effective rate correction, service time calculation, statistical characteristic evaluation, and accumulation of traffic intensity and queuing delay. Finally, the packet loss and retransmission effect is added to obtain the final total latency. The source forwarding node selects the optimal forwarding node based on the latency assessment of these candidate paths. The path that forwards the data packets from the source forwarding node to the target destination node via the optimal forwarding node is the initially determined optimized path with the minimum end-to-end latency.

[0048] In one embodiment of the present invention, Figure 6 This illustrates the parameter settings changes and communication performance of a node operating under light load conditions with low arrival rate and low packet loss rate. In step 301, the effective physical layer rate of the node at time t is calculated. This rate is determined by the instantaneous physical layer rate. 10Mbps and competition intensity and competition penalty coefficient A joint decision. By Figure 6 It can be seen from this that when the competition intensity When fluctuating between 0.20 and 0.80, The value varies accordingly, ranging from approximately 4.40 to 8.60 Mbps, showing a significant decrease in effective rate as contention intensifies. In step 302, based on the packet size... Calculate the service time per package ,Depend on Figure 6 It is evident that, with Changes, The numerical range is approximately 0.116 to 0.227 seconds, meaning that the service time of a single packet increases when the effective rate decreases.

[0049] In step 303, the system performs statistical analysis on the arrival intervals and service duration samples within the sliding window to obtain the arrival variation coefficient. With service variation coefficient These two parameters are used as fixed coefficients in the subsequent calculation of queuing delay. In step 304, based on the arrival rate... With average service time Calculate business intensity ,when When fluctuating between 0.40 and 0.80, The numerical range is approximately 0.047 to 0.166, all within the lightly loaded range well below 1, indicating that the system is in a stable state. In step 305, the queuing delay is calculated using the traffic intensity and the coefficient of variation. ,Depend on Figure 6 As can be seen from the curve, under light load conditions, The numerical variation was small, ranging from approximately 0.004 to 0.032 seconds, with no significant sudden increase, indicating that the queue latency was low in stable operating scenarios.

[0050] In step 306, packet loss rate is taken into account. The impact of this, calculating the total delay ,in, It fluctuates between 0.02 and 0.08. As can be seen from the graph, the total delay... The variation range is approximately 0.13 to 0.27 seconds, generally in the millisecond to sub-second range, indicating that the system can maintain a low total transmission latency under low packet loss rate and light load conditions.

[0051] Therefore, it can be seen that the method 300 in the light load scenario of this disclosure affects the effective rate through the fluctuation of the competition intensity, and the effective rate determines the service time of a single packet; the arrival rate and the service time together determine the service intensity; under light load conditions, the queuing delay is maintained at an extremely low level; after combining the packet loss rate correction, the total delay is maintained between approximately 0.13 and 0.27 seconds, thus demonstrating that the method 300 in this disclosure embodiment can operate stably under light load scenarios.

[0052] In another embodiment, Figure 7 This illustrates the parameter settings changes and communication performance of a node operating under heavy load conditions with high arrival rates and high packet loss rates. The parameter settings in steps 301-303 are the same as under light load conditions. In step 304, the arrival rate is considered... Calculate business intensity based on service time. When arrival rate When it fluctuates between 2.10 and 2.90, The values ​​range from 0.246 to 0.625, a significant improvement compared to lightly loaded scenarios. Approaching 0.6 indicates the system is under high load. In step 305, the queuing delay is calculated. ,because Larger values ​​result in longer waiting times in the queue. The values ​​ranged from 0.028 to 0.273 seconds, representing an improvement of approximately one order of magnitude compared to lightly loaded scenarios. When When the curve approaches 0.6, it shows a sharp upward trend, indicating the amplifying effect of queue congestion. In step 306, this is combined with the packet loss rate. The impact of this, calculating the total delay When the packet loss rate fluctuates between 0.03 and 0.13, the total latency... The calculated latency ranges from approximately 0.16 to 0.52 seconds, significantly higher than the 0.13 to 0.27 seconds in lightly loaded scenarios, and reaches its peak under the combined conditions of high packet loss rate and high traffic intensity. In summary, this embodiment reveals that under heavy load conditions, the combined effect of competition intensity and high arrival rate significantly increases traffic intensity, leading to a rapid increase in queuing latency. This, coupled with the amplifying effect of a higher packet loss rate, results in a significant increase in total latency. In this embodiment, although the node is under heavy load conditions with high arrival rate and high packet loss rate, the service intensity is significantly increased, leading to an increase in queuing latency. However, through the progressive modeling and correction in steps 301-306, the total latency is reduced. The timeout remained within a controllable range (approximately 0.16 to 0.52 seconds), and the overall system operation remained stable. This result demonstrates that the method 300 proposed in this application can guarantee the controllable and stable operation of network services under heavy load scenarios.

[0053] It should be noted that, in order to facilitate comparative analysis of the relative changes between different parameters, Figure 6 and Figure 7 Dimensionless processing was applied to the parameter values, meaning that normalization or relative scaling was used during the calculation process to map different physical quantities (such as rate, arrival rate, service time, and delay) to the same vertical axis scale. This processing method does not involve specific physical units; it only reflects the magnitude trend and interrelationships of the parameters over time. Figure 6 and Figure 7 The left ordinate of each coordinate axis corresponds to the system's input parameters, including the competition intensity. Arrival rate and packet loss rate The right-hand vertical axis corresponds to the system's response parameters, including the effective physical layer rate. Single package service time Business intensity Queue waiting time and total latency By normalizing the parameter curves of the input and response ends, the dynamic coupling relationship of each parameter under different load conditions can be intuitively demonstrated.

[0054] Figure 8 A flowchart illustrating a method 400 for monitoring the latency variation trend of a preferred path after path activation and triggering a switch based on a prediction model, according to an embodiment of this disclosure, is shown. Method 400 primarily focuses on the selected optimal path... The system monitors the latency trend within a continuous window to determine whether the short-term future latency exceeds the limit, obtains real-time historical data on latency changes within that time period, generates a prediction result for the short-term future based on the historical data, and then compares it with the latency threshold. If the latency threshold is exceeded, the node for forwarding data packets is switched, thereby achieving the switching of the data packet forwarding path.

[0055] In step 401, within a continuous time window Internally collect the end-to-end actual delay sequence of the currently enabled path. : Among them, such as Figure 9 As shown, L is the continuous time window. Length, This indicates the starting time of the current observation window indexed by k. The corresponding end-to-end actual delay value. It is the first sample point of this window. Indicates the time at the end of the same window The end-to-end actual delay value is the last sample point in this window. Therefore, the end-to-end actual delay sequence of the currently enabled path. It contains a continuous time-delay observation data sequence of length L, which forms the basis of the subsequent normalization processing and similarity analysis.

[0056] Calculate the normalized delay value indexed by k for each acquired end-to-end actual delay value within the continuous window. , where t= , ,…, ; , These are the end-to-end actual delay sequences. The mean and standard deviation of all sampled values.

[0057] In step 402, a multidimensional temporal feature mapping method is used to map continuous time windows. At each time t, a multidimensional state vector is constructed, that is, the one-dimensional time delay sequence is expanded into a multidimensional state vector composed of the current time delay value and several historical time delay values, thereby forming multiple state trajectory point sequences in the phase space, and obtaining the trajectory point vector corresponding to time t. : , for The transpose of ; where, Taken from the end-to-end actual time delay sequence The minimum self-mutual information of each sampled value is used as the value of each trajectory point in the mapping process. , ,..., Relative to the original sample value The delay parameter (i.e., delay duration), where m is the embedding dimension; That is, the resulting sequence of trajectory points in phase space is an m-dimensional state vector; in some embodiments, m is obtained by the nearest neighbor distance method; m and Both update adaptively with the window.

[0058] In one embodiment of this disclosure, to further reveal the similarity and potential repetition patterns of the time-delay sequence at different time points, the system first preprocesses the time-delay sequence in step 401, using a normalization method to eliminate dimensional differences between different sampling intervals, resulting in a standardized sequence with a mean of zero and a variance of one. ,like Figure 10 As shown in (a), the blue curve represents the normalized result of the end-to-end time delay sequence collected within a continuous time window. This delay value reflects the transmission delay fluctuation of data packets on the currently selected path. Orange, cyan, and magenta background blocks mark three typical burst delay segments occurring within sampling time steps 38–44, 72–77, and 92–100, respectively. Subsequently, in step 402, based on the delay embedding principle of multi-dimensional temporal feature mapping, the standardized delay sequence is reconstructed into state points in phase space. For example, by taking the embedding dimension m=2 and the delay parameter τ=1, the following result is obtained: Figure 10 (b) shows the state vector expanded in a two-dimensional phase space coordinate system. Pairing of adjacent time step trajectory points This refers to the distribution of point clusters formed by paired concentrated regions; when τ is not equal to 1 and m is not equal to 2, for example, when m=3 and τ=2, then the pairing of practice step trajectory points... (will be unfolded in a three-dimensional coordinate system) Figure 10 (b) respectively with Figure 10 (a) The same color is displayed in Figure 10 (a) shows point clusters within the three burst periods, which form distinct local clusters in phase space, characterizing the nonlinear clustering of time-delayed states during these periods. This is achieved by calculating the Euclidean distance between each state vector and applying a quantile threshold. Determining whether they belong to a "similar state" can form a pattern like this. Figure 10 (c) shows the state similarity map. The color intensity in this map represents the distance between state points at different times; warm-colored areas correspond to high similarity (i.e., low distance), and cool-colored areas correspond to low similarity (i.e., high distance). This color map allows for a more intuitive observation of the system's local clustering and state distribution density within burst intervals. (It should be noted that...) Figure 10 (c) The colors in the heatmap are only used to represent the distance magnitude and cannot be directly used to determine the nonlinear temporal state similarity vectors of the diagonal or vertical lines. Blue indicates high similarity and red indicates large difference.

[0059] Therefore, by embedding the delay parameter τ into the normalized delay value and performing spatial mapping in the m-dimensional space, the normalized delay and the m embedded delay values ​​after being delayed by τ time steps from time t can be displayed to form m-dimensional spatial trajectory points. In this way, the m-dimensional state similarity diagram and its binarized black and white state diagram can intuitively present the sudden fluctuations and pattern recurrence characteristics of the end-to-end delay in the time evolution process, providing a basis for subsequent trend prediction and switching judgment.

[0060] For step 403, in some embodiments, the similarity and correlation analysis of the temporal feature parameters can be performed using Euclidean distance. This involves calculating the L2 distance between feature vectors at different times to determine the similarity of path delay patterns. This method is simple to implement and suitable for delay sequence analysis scenarios with relatively balanced feature distributions. In other embodiments, the similarity and correlation analysis can be performed based on Mahalanobis distance or cosine similarity. Mahalanobis distance can effectively eliminate the influence of dimensions when there is correlation between features, thereby improving the accuracy of path state clustering. Cosine similarity can be used to identify the consistency of delay change trends; it is sensitive to changes in feature direction but not to changes in amplitude, making it suitable for network environments with large sudden load changes.

[0061] For step 404, in some embodiments, after calculating the similarity matrix or the two-dimensional spatial point cluster correlation matrix presented in the form of a state similarity graph as described above, the system can employ Dynamic Time Warping (DTW) or Kernel Correlation Analysis (KCCA) methods to capture similar patterns of path delay under nonlinear variation conditions, and maintain the stability of data distribution by adaptively adjusting threshold parameters. In some embodiments, based on the above similarity and correlation analysis results, the system calculates multiple comprehensive feature indicators to characterize the stability and trend of delay, including average volatility, short-time autocorrelation coefficient, coefficient of variation, and gradient of change, to form a feature input vector for subsequent prediction and judgment. In other embodiments, the comprehensive feature indicators may further include path delay recovery time, periodicity intensity, stable interval proportion, or trend direction coefficient, etc., to describe the dynamic stability characteristics of path delay from multiple perspectives, thereby enhancing the representativeness of the prediction input.

[0062] For step 405, in some embodiments, when the feature input vector is introduced into the prediction model, a regression-based modeling approach can be used, such as linear regression, ridge regression, or Bayesian regression models, to obtain the predicted time delay value for the path in the short term and compare it with a preset threshold. In other embodiments, the prediction model can be a lightweight neural network structure, such as a feedforward neural network (FNN), a Bayesian regression model, a one-dimensional convolutional neural network (1D-CNN), or a time-series network based on gated recurrent units (GRUs), used to achieve nonlinear time delay prediction under multi-dimensional feature input conditions. When the prediction result indicates that the path has an over-limit risk, the system automatically triggers backup path switching or issues an early warning signal to achieve dynamic time delay control. It should be noted that a one-dimensional convolutional neural network is a type of lightweight convolutional neural network model, and a gated recurrent unit is a type of recurrent neural network model.

[0063] Figure 11 A flowchart of a method 500 for similarity and correlation analysis of time-series feature parameters according to an embodiment of the present disclosure is shown.

[0064] In step 501, the window is calculated. Vector of the i-th trajectory point With the vector of the j-th trajectory point Similarity index : ,in, , And delayed embedding of the window The number of available state points is N. k ,therefore, , , ; For Heaviside step function, It is a distance threshold used to determine the distance between two trajectory point vectors. and Are they similar enough?

[0065] The calculation result is 0 or 1. ; ;in, Let be the quantile function, q∈(0,1), where q is the probability distribution, representing "the percentile value according to the probability distribution". For a set of samples... (For example, the distance between all pairs of points), if the probability distribution is taken as q=0.2, then q-quantile ( This indicates that 20% of the samples in this data set are not greater than this value. By taking an appropriate quantile as the threshold εk in the state similarity graph, we can ensure that the recurrence points are neither too sparse nor too dense.

[0066] In step 502, according to Further calculate the k-th continuous time window The proportion of trajectory point vector pairs that are judged to be similar among all trajectory point vector pairs. : .

[0067] In step 503, by setting the distribution probability q, the proportion of trajectory point vector pairs judged as similar is thus determined. (Also known as state similarity rate) is within the target similarity ratio range Inside, The proportion of similar trajectory point vector pairs Also known as the state similarity rate, it is used to measure the similarity of system states within a window. The density of repeated occurrences.

[0068] Sliding window Every moment within Corresponding to a multidimensional state vector It consists of the standardized current latency value and several historical latency values. For two different moments within the window... The state points were obtained respectively. ,generally By defining a distance threshold ,when When, the similarity index obtained after being constrained by the step function is... Otherwise, the value is 0. Further calculate the state similarity rate. And by selecting the distance distribution -quantile quantile as a threshold, making Approaching the preset target recurrence rate This method not only ensures the sparsity of the recurrence points but also reveals the state patterns of the time-delay sequence in the phase space.

[0069] Figure 10 In (d), the black dots represent situations where the system states are similar under the two time indices (i.e., White dots indicate significant state differences. The diagonal lines of this black-and-white state diagram reflect the short-term self-similarity of the time series, while the vertical black lines represent the stagnation of the system in a certain state. The short diagonal lines marked in red represent the repetitive patterns of time delay states within a short time range, while the blue vertical lines represent the local stagnation of time delay in a certain state.

[0070] The method 500 according to embodiments of this disclosure achieves adaptive identification of the internal dynamic structure of an end-to-end time delay sequence by introducing a trajectory point similarity calculation and state similarity rate control mechanism based on quantile thresholds in time delay analysis. The distance threshold ε is determined using q-quantile as a function. k This ensures that the density of similar trajectory points under different time windows remains within a reasonable range, avoiding the problem of excessively sparse or dense recurrence points caused by the traditional fixed threshold method. This guarantees that the recognition results of the time delay pattern under different sampling windows are neither overly concentrated nor abnormally dispersed, and can stably reflect the actual pattern of path time delay changes. Furthermore, the state similarity rate (RR) is utilized... k The recurring patterns and stable intervals of path delays are quantitatively evaluated, enabling the system to directly extract intrinsic indicators reflecting network congestion and stability characteristics from the delay sequence without increasing the burden of external measurement. This provides more accurate and stable data for subsequent delay change trend analysis and path switching decisions.

[0071] Figure 12 A flowchart of a method 600 for calculating multiple comprehensive feature indicators characterizing the stability and trend of time delay data based on the time delay variation law obtained by analysis, according to an embodiment of the present disclosure, is shown.

[0072] In step 601, in some embodiments, based on the state similarity rate And the length of the diagonal segment of the black-and-white similarity map after threshold binarization of the state similarity map for similar reaction states, where consecutive state similarity points (i.e., state similarity points at adjacent time points) are arranged along the diagonal. Calculate the decisionability : ; in, Indicates length is The number of consecutive similar points in the diagonal segment; The minimum diagonal length threshold (usually set) ≥2), used to avoid statistically random short-term fluctuations; This indicates the proportion of diagonal structures in the black-and-white similarity diagram, if the analyzed end-to-end time delay sequence If historical data exhibits regularity, then many diagonal line segments running from the upper left to the lower right will appear in the black-and-white similarity image, thus allowing for the determination of degree. The index value will be relatively high. For example, a diagonal line segment with a length of ℓ=3 indicates that the system state evolution trajectory is very similar over three consecutive time points.

[0073] In step 602, the stagnation degree of the black-and-white similarity image is calculated. : ; Indicates length is The number of vertical line segments, The minimum diagonal length threshold (usually set) ≥2); In a black and white similarity diagram, some points are arranged vertically in a continuous manner, indicating that the state stays in a similar position for a relatively long time.

[0074] In step 603, the longest diagonal is calculated: The state similarity graph after thresholding and binarization is traversed and statistically analyzed. All diagonal segments with continuous state similarity points are selected, and the maximum value in the length set of these diagonal segments is taken as the longest diagonal. This metric reflects the longest duration for which a system maintains a similar evolutionary state over a continuous period of time, and is used to measure the predictability and stable persistence of path delay sequences.

[0075] In step 604, the state similarity entropy value is calculated. : ; ; Let be the probability distribution of the length of the diagonal segment. Different from black and white similar images One of all possible diagonal lengths, To calculate the total length of all diagonal segments in the reproduced graph. Used to measure the complexity of the distribution of diagonal lengths. A higher value indicates that the diagonal lengths in the black and white similarity images differ significantly, suggesting a complex and unpredictable distribution of similar points. A lower value indicates that the diagonal lengths of the black-and-white similarity scores are not significantly different, and they are all concentrated around a certain value, resulting in a uniform distribution of similar points. In some embodiments, the feature input vector is a nonlinear temporal state similarity feature vector. The method 600 of this embodiment of the present disclosure, through the calculations of the above steps, finally obtains... Method 600 introduces a nonlinear temporal state similarity feature vector. The computational mechanism enables the system to extract multi-dimensional feature information reflecting dynamic stability and change patterns from end-to-end time delay sequences. Determinability By statistically analyzing the proportion of diagonal structures, the continuity and regularity of system state evolution can be reflected; stagnation degree The residence time of a quantized state in a specific mode; longest diagonal The upper limit reflecting the short-term predictability of a system; state similarity entropy This measures the complexity and uncertainty of latency changes. By simultaneously introducing the above four complementary indicators, the stable range, fluctuation intensity, and abrupt change trends of network operation can be revealed from the latency data itself without relying on external monitoring signals. This provides a more comprehensive and physically meaningful reference for path prediction and handover control, and improves the system's latency identification and control capabilities in complex communication environments.

[0076] Figure 13 A flowchart of a method 700 for making decisions on path switching or issuing alarms based on future delay prediction results according to an embodiment of the present disclosure is shown. In step 701, in some embodiments, a nonlinear temporal state similarity feature vector is used. Input into the regression model and perform a continuous time window indexed by k. The end-to-end short-term delay is predicted at the next h-step, and the short-term predicted delay output is obtained. , Satisfy the mean Standard deviation For a normal distribution: , , These are the mean and standard deviation of multiple end-to-end short-term forecast samples, respectively.

[0077] In step 702, the predicted value is calculated to exceed the maximum tolerable delay. probability : ; in, This is the cumulative distribution function of the standard normal distribution.

[0078] In step 703, when the future Probability of exceeding the prediction limit in step If the probability is not lower than a preset probability threshold β, then within the same time window, there is at least one nonlinear temporal state similarity feature vector. The eigenvalues ​​in the data exceed the safety threshold (i.e.) ,or ,or If the system determines that the current path is high-risk, it should initiate a switchover of the packet forwarding node or activate an early warning mechanism. In some embodiments, the preset probability threshold β ranges from 0.05 to 0.1.

[0079] As a determinability safety threshold, Preferably 0.85; when When the proportion of diagonals in the black-and-white similarity graph of trajectory state point pairs is too high, it indicates that the system evolution pattern is overly regular (it may enter a congested or strongly dependent state), which is a risk signal of excessive data forwarding congestion. Stagnation safety threshold Preferably 0.7; when This indicates that when the proportion of vertical line segments is too high, it means that the system is stagnant for a long time, and data packets may be stuck at this node. The longest diagonal safety threshold, The range is 0.2L to 0.3L, which is 0.2 times the continuous time window. The length L is up to 0.3 times the continuous time window. The length L; when the statistically obtained longest diagonal satisfies This indicates that there is a long sequence of similar evolutionary trajectories in the state similarity graph. In terms of network node behavior, this means that the packet forwarding pattern of the node remains almost unchanged over a long period of time. This means that the node is in a highly regular or singular forwarding state. Although it is highly predictable in the short term, it also reflects that the node may fall into a state of continuous queuing, fixed forwarding paths, or insufficient flexibility, and is prone to congestion or latency anomalies when sudden traffic arrives.

[0080] The method 700 according to an embodiment of the present disclosure uses a nonlinear temporal state feature vector A regression prediction model was introduced to achieve short-term trend prediction and risk assessment of end-to-end path delay, enabling the system to proactively identify potential congestion or abnormal states before actual delay exceeds limits. This is achieved by establishing a relationship between the prediction results and the maximum tolerable delay. The probability comparison relationship, combined with the decision value in the feature vector. Stagnation with the longest diagonal By using multiple state thresholds to jointly determine path risk, a composite judgment mechanism integrating statistical probability and temporal pattern characteristics is constructed according to method 700 of this disclosure. This composite judgment mechanism enables the system to no longer rely on a single threshold or instantaneous measurement results in path switching control, but to simultaneously consider three types of factors: the rate of change of path delay, the degree of state stagnation, and the regularity of temporal structure. It reflects the evolutionary characteristics of network operating status from different dimensions, thereby making a comprehensive judgment and realizing the transformation from passive response to active prediction. This effectively improves the accuracy of delay prediction, congestion identification accuracy, and path switching reliability of the network under complex load fluctuations.

[0081] Comparative Example 1 Experiments were conducted to compare two different prediction mechanisms: the method according to the embodiments of this disclosure and the moving average method. In the experiments, key performance parameters of the network path under the same communication scenario were collected and statistically analyzed simultaneously. To ensure the comparability of the results, all three types of indicators used time delay data of the same dimension (milliseconds, ms). The three types of indicators are: (1) total end-to-end delay, which represents the average time taken for the complete transmission process from the source node to the destination node; (2) queuing delay component, which represents the delay caused by the node waiting for processing in the forwarding queue; and (3) jitter, which is defined as the absolute deviation of the time interval between the arrival of adjacent data packets and is used to characterize short-term fluctuation stability.

[0082] Figure 14 The diagram illustrates a comparison of the distribution of three latency indicators between the routing communication method employing end-to-end latency dynamic optimization according to embodiments of this disclosure and the moving average method. Within the same statistical window, three statistical distribution charts are plotted for each of the three indicators: ① a histogram, used to reflect the distribution density of the sample size across different latency intervals; ② a probability density function distribution chart, used to illustrate the trend of probability density changes in consecutive latency values; and ③ a cumulative distribution function distribution chart, used to measure the cumulative probability level at a given latency threshold.

[0083] Figure 14 As shown in the histogram (a), the end-to-end delay, queuing delay, and jitter of the routing communication controlled by the method according to the embodiments of this disclosure all exhibit a main peak that is significantly to the left and concentrated in a specific distribution, indicating that most data packets are forwarded in a lower delay range; while Figure 14 (d) shows that the histogram distribution of the control method (moving score method) is more dispersed and the tails are more extended, indicating that the probability of large time delay events is higher. Figure 14 (b) In the distribution of the routing communication probability density function controlled by the method according to an embodiment of this disclosure, the peaks of the three index curves are sharper, and the distribution range converges, indicating that the system latency fluctuation is small and the predictability is strong; while Figure 14 (e) shows that the probability density distribution of the control method has a flat peak and a wide distribution range, indicating greater time delay uncertainty. In the cumulative density function distribution, Figure 14 (c) The three curves of the routing communication controlled by the method according to the embodiments of this disclosure all rise rapidly to near 1 within a low latency range, indicating that most data packets are transmitted within a low latency range; while Figure 14 (f) The curve of the control method has a significantly gentler upward slope, and a higher time delay is required to reach the same level of cumulative probability.

[0084] Furthermore, the method described in this application significantly outperforms the control method in all three metrics. Regarding end-to-end latency, Figure 14 (a)- Figure 14 The distribution of the method of this disclosure shown in (c) is concentrated in the 40–60 ms interval, while Figure 14 (d)- Figure 14 (f) shows that the distribution of the control method shifts to the right to the 60–90 ms range, with a higher overall latency level. Regarding queuing latency, the method in this application is stable at around 15–25 ms, while the control method generally exceeds 30 ms, with a significant tail extension. Regarding jitter, the fluctuation amplitude of the method in this application is concentrated in the low-value range, while the jitter distribution of the control method is wider, indicating that its short-term fluctuations are significantly more severe.

[0085] This disclosure employs method 300 to perform standardized preprocessing on the time-delay sequence, method 400 to reconstruct the phase space to obtain a state similarity map, and method 500 to construct a thresholded black-and-white similarity map based on an adaptive threshold. Finally, method 600 is used to extract nonlinear time-series state feature vectors. Method 700 was used to extract... Input the regression prediction model to calculate the future time delay distribution and the probability of exceeding limits. The above steps ensure early identification and dynamic response to potential congestion and anomalies. Combined with... Figure X As can be seen, the method of this application not only demonstrates advantages in terms of more concentrated distribution and less fluctuation in horizontal comparison, but also significantly reduces end-to-end latency, queuing latency, and jitter levels in vertical comparison. Therefore, the end-to-end latency dynamic optimization routing communication method according to the embodiments of this disclosure can achieve a significant reduction in packet forwarding latency and an improvement in the latency stability of the end-to-end transmission process and the reliability of path switching (with fewer data transmission interruptions during dynamic route switching) in complex network communication environments, with overall performance superior to existing technical solutions.

[0086] Comparative Example 2 Figure 15 This diagram illustrates a comparison of the initial optimal path and the path after dynamic switching in routing communication under the method of this application and existing methods. Figure 15 (a) – Figure 15 (f) shows the experimental comparison results of the active delay prediction and threshold self-scheduling path optimization method according to the embodiments of this disclosure and two prior art methods (static shortest path method and passive path switching method) under the same network topology conditions. Figure 15 In each subgraph, Channel 1 and Channel 2 represent independent communication channels that forward two different sets of data packets simultaneously under the same routing environment, and each channel corresponds to an end-to-end transmission path.

[0087] in, Figure 15 (a) and Figure 15 (d) are the initial network topologies of the experiment. The two are in the same routing environment. Nodes N1–N15 represent the communication nodes in the network on a two-dimensional plane. The connection weight is the link propagation delay (unit: ms). Figure 15(b) illustrates the minimum end-to-end latency path initially optimized according to the method of this disclosure, where the forwarding links of channel 1 (blue) are N1→N6→N11→N12, and the forwarding links of channel 2 (green) are N3→N7→N11→N14→N15 (the link propagation delays between nodes are 298ms, 166ms, 335ms, and 85ms, respectively). It can be seen that both paths exhibit a convergent distribution in the network, and the weights between nodes (e.g., 362 for N1→N6 and 143 for N6→N11) are significantly lower than the average link cost, indicating that the system automatically selects a low-latency, low-load link combination.

[0088] Figure 15 (c) shows the path structure after dynamic switching based on the optimized path with the minimum end-to-end latency initially obtained by the method of this application, through dynamic latency monitoring and prediction and judgment based on historical data. Channel 1 (red) after the switching path is N1→N2→N3→N4→N8→N12, and Channel 2 after the switching path is N3→N6→N10→N13→N14→N15 (pink). It can be seen that the switched path, while ensuring end-to-end connectivity, avoids the latency spike node N6 and the high-load links N7–N11, resulting in an overall average path cost reduction of approximately 15% and effectively suppressing latency spikes and transient packet loss.

[0089] In comparison, Figure 15 The comparison method shown in (e) uses the existing static shortest path method (Dijkstra), which determines the path once during the initialization phase (e.g., channel 1, represented by tan, is fixed as N1→N2→N3→N2→N12). However, it does not update the path during communication as the link state changes. This results in the path being unable to be adjusted when the load on node N2 increases, leading to a gradual increase in end-to-end latency. Furthermore, because the static routing algorithm only calculates the shortest path based on the link cost at a single moment, When the link state fluctuates rapidly, repeated round-trip segments (N2→N3→N2) form between nodes N2 and N3, causing local loops in the path and increasing end-to-end latency. In contrast, the dynamic optimization method of this application ( Figure 15 (b) By using real-time delay trend prediction and state similarity analysis, this type of loop phenomenon is effectively avoided.

[0090] Figure 15(f) shows a passive path switching algorithm (AODV source-driven routing protocol) for switching the initial optimal path obtained by the Dijkstra method. This path switching method only triggers switching after detecting link failure or severe congestion. After the path switch, channel 1 is N1→N9→N14→N6→N10→N11→N12, and channel 2 is N3→N5→N11→N15. Although communication can be restored, the switching action is delayed, resulting in significant short-term delay spikes and packet delays in the path.

[0091] A comprehensive comparison shows that the method in this application achieves early identification and proactive detour of potential congested nodes through time-series feature prediction and threshold self-scheduling mechanism, thereby maintaining link continuity and low latency characteristics during the handover process.

[0092] pass Figure 15 (b) and Figure 15 The comparison in (f) shows that the method in this application has a more balanced path weight distribution and a lower main link cost; through Figure 15 (c) and Figure 15 The comparison in (f) shows that the dynamic switching mechanism of this application achieves end-to-end latency reduction and transmission stability improvement without introducing additional routing costs. Therefore, the end-to-end latency dynamic optimization routing communication method of this application can effectively achieve low-latency and highly stable path switching control in complex network environments, and its overall performance is significantly better than existing static or passive switching technologies.

[0093] Those skilled in the art will understand that the method steps described herein are not limited to the order exemplarily shown in the accompanying drawings, but can be performed in any other feasible order. The foregoing description of this disclosure is intended to enable any person of ordinary skill in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those of ordinary skill in the art, and the general principles defined herein can be applied to other variations without departing from the spirit and scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the broadest scope of the principles and novel features disclosed herein.

Claims

1. A method of routing communication control with dynamic optimization of end-to-end latency, said method being applicable to light-weight data transmission or heavy-weight data transmission, characterized in that, The method comprises the following steps: The real-time running state of the candidate paths in the network is collected and analyzed in multiple dimensions, the transmission rate, node load and link utilization of each path are comprehensively evaluated, the preferred path that can realize the minimum end-to-end delay at the current time is dynamically selected from multiple candidate paths for data forwarding; After the path is enabled, the delay change trend of the preferred path is continuously monitored, the time sequence feature analysis and prediction model are used to determine whether the delay in the future period is likely to exceed the threshold, and when the prediction result shows that there is a risk of delay overrun, the standby path switching mechanism is triggered to maintain the minimization control of the end-to-end delay.

2. The method of claim 1, wherein, The dynamic selection of the path for forwarding data after the real-time running state of the candidate paths in the network is collected and analyzed in multiple dimensions comprises: By obtaining the physical layer transmission state information of each candidate relay node at the current time, the actual data transmission rate provided by the node is calculated to reflect the effective bandwidth of different nodes in a competitive environment; Based on the transmission rate, the service time of each candidate node in the single data packet forwarding process is determined to represent the basic processing capacity of the node; The data arrival and transmission process information of the node is collected in a continuous time window, the traffic variation characteristics and service stability indicators are counted to reflect the dynamic fluctuations of the node running state; According to the real-time request amount and average processing time of the node, the business load degree of the node is evaluated to determine the saturation level of the path at the current time; According to the load characteristics and traffic stability of the node, the delay quantitative index caused by queuing and waiting is calculated to describe the congestion degree of the path in the forwarding process; After the waiting delay, processing time and transmission reliability of the node are comprehensively considered, the total communication delay of each candidate path is calculated, and the path with the minimum total delay is selected as the preferred path for end-to-end data forwarding.

3. The method of claim 1, wherein, The monitoring of the delay change trend of the preferred path after the path is enabled and the triggering of switching based on the prediction model comprise: The delay observation data of the currently enabled path is continuously collected in a preset time window, and the observation results are normalized and noise suppressed to form the basic time sequence data that can be used for trend identification; Based on the time sequence data, a multi-dimensional dynamic feature representation is constructed, time sequence feature parameters reflecting the change law of the path delay are extracted, and the running state evolution of the system is described; The similarity and correlation analysis of the time sequence feature parameters is performed to identify the law of the path delay change with time, and the data distribution is kept stable through adaptive adjustment; According to the delay change law obtained by analysis, a plurality of comprehensive feature indicators representing the stability and change trend of the delay data are calculated to form a feature input vector for subsequent prediction judgment; The feature input vector is introduced into the prediction model to generate the delay prediction result of the path in the future short period, and the prediction result is compared with the preset threshold, when the prediction delay exceeds the limit or there is an abnormal trend, the switching of the standby path is triggered or an alarm signal is sent in advance.

4. The method of claim 3, wherein, The similarity and correlation analysis of the time sequence feature parameters comprises: The feature difference degree between any sample pair is calculated based on the multi-dimensional feature samples at each time point in the window, and a similarity relationship is determined according to a preset judgment rule, which is used to identify similar patterns of path delay variation; The overall similarity degree of the system in the time period is obtained by statistically analyzing the similarity proportion of all sample pairs in the continuous time window, which is used to quantify the concentration and repeatability of path delay variation; The similarity determination threshold is adaptively adjusted according to the target distribution range of the overall similarity degree, so that the similarity proportion is kept in a predetermined stable interval, thereby ensuring the representativeness and balance of the trend analysis result.

5. The method of claim 3, wherein, The comprehensive feature indicators representing the stability and variation trend of the delay data include: The regularity degree of the system in the time evolution process is statistically analyzed by analyzing the distribution of the continuous variation segments in the time sequence feature, which is used to reflect the predictability of the path delay; The stability and continuity of the path delay are characterized by evaluating the staying characteristics of the system under a specific running mode by identifying the continuous interval that maintains a similar state between adjacent time points; The longest continuous correlation segment in the time sequence is determined based on the statistical results, to identify the continuous range of the system in a high correlation state; The complexity measurement indicator is calculated according to the length distribution of all feature segments, which is used to describe the diversity and uncertainty of the system state change, thereby forming a comprehensive evaluation vector reflecting the overall dynamic characteristics.

6. The method of claim 3, wherein, The decision or alarm mechanism for path switching based on the future delay prediction result includes: The feature input vector is input into a prediction model to estimate the end-to-end delay variation trend of the current path in the future short period, to generate the corresponding prediction result; According to the prediction result, the probability value of the future delay exceeding the preset delay tolerance is calculated, which is used to evaluate the risk level of the path delay anomaly in the short term; The path state is comprehensively determined by judging whether the probability value of the future delay exceeding the preset delay tolerance reaches or exceeds the preset probability threshold, and combining the judgment result of whether any feature indicator in the feature input vector exceeds its safety boundary parameter, and the data forwarding path is switched or a warning operation is performed when the trigger condition is met.

7. The method of claim 3, wherein, The prediction model for path switching decision or alarm based on the future delay prediction result is any one of a Bayesian regression model, a lightweight convolutional neural network model, a recurrent neural network model or a feedforward neural network model.

8. The method of claim 6, wherein, The value range of the preset probability threshold is 0.05 to 0.1; The safety boundary parameters of the judgment indicators in the feature input vector include: a determinable degree threshold, a stagnation degree threshold and a longest correlation segment threshold; the value range of the determinable degree threshold is 0.81 to 0.86; The value range of the stagnation degree threshold is 0.65 to 0.75; the longest correlation segment threshold is 0.2 to 0.3 times the length of the continuous time window where the multi-dimensional feature samples at each time point are located.

9. A routing communication control system for end-to-end latency dynamic optimization, characterized by, The system includes: The path awareness and selection module is used for multi-dimensional collection and analysis of real-time running states of candidate paths in the network, comprehensive evaluation based on transmission rates, node loads and link utilization of the paths, and dynamic selection of a preferred path capable of realizing minimum end-to-end delay at the current time from the candidate paths for data forwarding. The delay prediction and switching control module is used for continuous monitoring of delay variation trends of the preferred path after the path is enabled, determination of whether the delay in a future period is likely to exceed a threshold by using a timing feature analysis and prediction model, and triggering of a backup path switching mechanism to maintain minimum end-to-end delay when the prediction result indicates that there is a risk of delay overrun.

10. The routing communication control system of claim 9, wherein, The path awareness and selection module includes a path state awareness unit and a path evaluation and selection unit, and the delay prediction and switching control module includes a delay monitoring unit, a prediction and determination unit and a path switching execution unit. The path state awareness unit is used for multi-dimensional collection and analysis of real-time running states of candidate paths in the network. The path evaluation and selection unit is used for dynamic determination of a preferred path capable of realizing minimum end-to-end delay at the current time based on comprehensive evaluation results of transmission rates, node loads and link utilization obtained by the path state awareness module. The delay monitoring unit is used for continuous monitoring of end-to-end delay variation after the preferred path is enabled and formation of timing feature data for trend identification. The prediction and determination unit is used for generation of a delay prediction result in a future short period by using a prediction model based on the timing feature data and determination of whether the path is likely to have delay overrun in the future period. The path switching execution unit is used for triggering of switching operations of a backup path or an alarm to realize dynamic minimum control of end-to-end delay when the prediction and determination module outputs a result indicating that there is a risk of delay overrun.

Images