IoT Gateway Data Routing Method and System for High-Concurrency SMS Sending

By analyzing local sensing data and historical forwarding records, the system predicts alternative forwarding paths for IoT gateways, thus solving the problems of link congestion and failure risks in IoT gateways and improving the stability and efficiency of data routing.

CN123124049APending Publication Date: 2026-11-17SHENZHEN CHENGLIYE TECH DEV CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611427831.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies make it difficult to identify congestion or failure risks in subsequent forwarding of adjacent routing nodes in a timely manner in IoT gateways, resulting in unstable link status, especially when sending SMS messages at high concurrency, which can easily cause link congestion or cascading node failures.

Method used

By acquiring local perception data of the current gateway node, the link congestion risk level of adjacent routing nodes is determined, the response delay time series of historical forwarding decision records is analyzed, and alternative forwarding paths are predicted by combining time series analysis and topology data to generate optimal routing instructions and avoid cascading failures.

Benefits of technology

Effectively identify and predict the risk of link congestion and failure of adjacent routing nodes, reduce the possibility of cascading node failures, and improve the stability and efficiency of IoT gateway data routing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123124049A_ABST
    Figure CN123124049A_ABST
Patent Text Reader

Abstract

This invention relates to the field of IoT communication and data routing technology, and discloses a data routing method and system for IoT gateways carrying high-concurrency SMS transmission. The method determines the link congestion risk level of each adjacent routing node based on the local sensing data of the current gateway node, obtains historical forwarding decision records from risky nodes and extracts response delay time series, and determines the cascading failure probability of the next-hop routing node based on delay fluctuation amplitude and delay decay trend. When the cascading failure probability is not less than a preset cascading failure probability threshold, it predicts alternative forwarding paths using the global view topology data of backup gateway nodes and generates optimal routing instructions for concurrent SMS requests to be forwarded. This invention can reduce the impact of link congestion and node cascading failures on high-concurrency SMS forwarding and improve the stability of IoT gateway data routing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) communication and data routing technology, and in particular to an IoT gateway data routing method and system for carrying high-concurrency SMS sending. Background Technology

[0002] In the device communication network managed by the IoT operating system, the IoT gateway is responsible for receiving, routing and forwarding concurrent SMS requests. When SMS requests arrive in a concentrated manner, the link load between gateway nodes, node processing capacity and response latency may change rapidly.

[0003] In one existing technology, a centralized routing scheduling method is employed, where a central control node aggregates the status of each gateway node and link, and then issues routing decisions. As the number of gateway nodes and SMS concurrency increases, the central control node processes more status data and routing requests, and it also needs to continuously synchronize its status with each gateway node.

[0004] However, state acquisition, transmission, and routing decisions are all time-consuming. There may be a time difference between the network state on which the decision is based and the actual link state. When a gateway node selects a route based solely on the current state of its neighboring routing nodes, this state is difficult to reflect the response delay and failure trend when the neighboring routing nodes forward concurrent SMS requests to the next-hop routing node. When multiple gateway nodes forward to the same neighboring routing node in similar time periods, the node and its next-hop routing node may still experience load concentration, which may lead to link congestion or cascading node failures.

[0005] In summary, existing technologies, with their limited local sensing capabilities and constantly changing link states, struggle to promptly identify congestion or failure risks in subsequent forwarding by adjacent routing nodes, thus impacting the stability of IoT gateway data routing. Summary of the Invention

[0006] This invention provides a data routing method and system for IoT gateways that support high-concurrency SMS sending, in order to solve the technical problem that existing technologies, when local perception is limited and link states are constantly changing, are unable to promptly identify the risk of congestion or failure in subsequent forwarding of adjacent routing nodes, thereby affecting the stability of IoT gateway data routing.

[0007] Firstly, to address the aforementioned technical problems, this invention provides a data routing method for an IoT gateway carrying high-concurrency SMS transmission, comprising: Obtain local perception data of the current gateway node, and determine the link congestion risk level of each adjacent routing node based on the local perception data; The adjacent routing nodes whose link congestion risk level is higher than the preset congestion risk level threshold are identified as risk nodes. Historical forwarding decision records are obtained from the risk nodes, and the response delay time series formed by the risk nodes forwarding historical SMS concurrent requests to the next hop routing node is extracted. Time series analysis is performed on the response delay time series, and the cascading failure probability of the next-hop routing node is determined based on the delay fluctuation amplitude and delay decay trend; When the cascading failure probability is not less than the preset cascading failure probability threshold, the global view topology data of the backup gateway node is obtained, and regression prediction is performed on the global view topology data to obtain the backup forwarding path. Based on the alternative forwarding path, generate the optimal routing instruction for concurrent requests of SMS messages to be forwarded.

[0008] Preferably, determining the link congestion risk level of each adjacent routing node based on the local sensing data includes: Decision tree processing is performed based on the local perception data to determine the information gain of each dimension of features; Dimensional features with information gain greater than a preset gain threshold are identified as splitting features, and decision classification paths are generated based on the splitting features. The congestion probability is obtained by traversing the decision classification path, and the link congestion risk level of the corresponding adjacent routing nodes is determined based on the congestion probability.

[0009] Preferably, dimensional features with information gain not greater than the preset gain threshold are identified as features to be pruned, and these features are deleted from the dimensional features participating in the splitting decision of the current decision tree node.

[0010] Preferably, the congestion probability is obtained by traversing the decision classification path, and the link congestion risk level of the corresponding adjacent routing nodes is determined based on the congestion probability, including: Based on the local sensing data, traverse along the decision classification path to determine the leaf node corresponding to the local sensing data; The initial risk label is determined based on the leaf node, and the congestion probability corresponding to the initial risk label is determined based on the preset mapping relationship between the risk label and the congestion probability. Based on the preset mapping relationship between congestion probability range and link congestion risk level, the link congestion risk level corresponding to the congestion probability is determined.

[0011] Preferably, historical forwarding decision records are obtained from the risk node, and the response delay time sequence formed by the risk node forwarding historical SMS concurrent requests to the next hop routing node is extracted, including: Extract the sending timestamp of each historical SMS concurrent request from the historical forwarding decision record returned by the risk node, and the receiving timestamp of the confirmation response received by the risk node from the next-hop routing node. An initial response delay time series is constructed based on the difference between the sending timestamp and the receiving timestamp. The initial response delay time series is then smoothed and denoised to obtain a smoothed delay time series. The extreme value features are extracted from the smoothed delay time series to determine the response delay time series.

[0012] Preferably, determining the cascading failure probability of the next-hop routing node based on the latency fluctuation amplitude and latency decay trend includes: The delay fluctuation amplitude and delay decay trend are extracted from the response delay time series, and a delay state transition matrix is ​​generated based on the delay fluctuation amplitude and the delay decay trend; The failure evolution path of the next-hop routing node is extracted based on the delay state transition matrix, and the cumulative failure risk value is determined based on the failure evolution path. For a failure evolution path where the cumulative failure risk value is greater than a preset cumulative failure risk threshold, the topology association weight of the next-hop routing node is obtained, and the cumulative failure risk value is weighted according to the topology association weight to obtain the cascading failure probability.

[0013] Preferably, when the cascading failure probability is less than a preset cascading failure probability threshold, the current gateway node determines the target forwarding path based on the local perception data, and generates the optimal routing instruction for the concurrent request of the SMS to be forwarded based on the target forwarding path.

[0014] Preferably, regression prediction is performed on the global view topology data to obtain alternative forwarding paths, including: Node connectivity and link load features are extracted from the global view topology data, and a topology feature matrix is ​​constructed. Support vector regression prediction is performed on the topological feature matrix to predict the path connectivity cost of each candidate backup forwarding path in the preset set of candidate backup forwarding paths. The candidate backup forwarding paths are sorted in ascending order according to the path connectivity cost, and the candidate backup forwarding path with the lowest path connectivity cost is determined as the backup forwarding path.

[0015] Secondly, the present invention provides an IoT gateway data routing system for carrying high-concurrency SMS transmission, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any of the above-mentioned IoT gateway data routing methods for carrying high-concurrency SMS transmission.

[0016] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described IoT gateway data routing methods for carrying out high-concurrency SMS transmission.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention determines the information gain of each dimension feature based on local perception data, uses a preset gain threshold to filter split features and delete features to be pruned, and then determines the link congestion risk level through the stepwise mapping of decision classification path, initial risk label and congestion probability, thereby reducing the participation of dimension features with insufficient information gain in the split decision of the current node, so that the current gateway node can identify the link congestion risk of adjacent routing nodes based on effective risk features.

[0018] (2) The present invention forms a response delay time series from the historical forwarding decision records of risk nodes, generates a delay state transition matrix based on the delay fluctuation amplitude and delay decay trend, and determines the cascading failure probability by combining the failure evolution path, cumulative failure risk value and topology association weight, thereby characterizing the failure evolution risk when adjacent routing nodes continue to forward to the next hop routing node, and reducing the possibility of node cascading failure caused by local routing selection.

[0019] (3) The present invention forms complementary routing branches based on a preset cascading failure probability threshold. When the cascading failure probability is less than the preset cascading failure probability threshold, the current gateway node determines the target forwarding path based on local perception data. When the cascading failure probability is not less than the preset cascading failure probability threshold, the backup forwarding path is predicted using the global view topology data of the backup gateway node. This takes into account both local routing decisions in normal scenarios and global topology prediction in risk scenarios. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the IoT gateway data routing method for carrying high-concurrency SMS sending provided in the first embodiment of the present invention. Detailed Implementation

[0021] 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.

[0022] Reference Figure 1 The first embodiment of the present invention provides a data routing method for an IoT gateway that carries high-concurrency SMS sending, comprising the following steps: S11, Obtain the local perception data of the current gateway node, and determine the link congestion risk level of each adjacent routing node based on the local perception data; S12, the adjacent routing nodes whose link congestion risk level is higher than the preset congestion risk level threshold are identified as risk nodes, historical forwarding decision records are obtained from the risk nodes, and the response delay time sequence formed by the risk nodes forwarding historical SMS concurrent requests to the next hop routing node is extracted. S13, Perform time series analysis on the response delay time series, and determine the cascading failure probability of the next-hop routing node based on the delay fluctuation amplitude and delay decay trend; S14, when the cascading failure probability is not less than the preset cascading failure probability threshold, obtain the global view topology data of the backup gateway node, perform regression prediction on the global view topology data, and obtain the backup forwarding path; S15, Based on the backup forwarding path, generate the optimal routing instruction for concurrent requests of SMS messages to be forwarded.

[0023] In step S11, it is first necessary to acquire the local sensing data of the current gateway node and identify the link congestion risk level of each adjacent routing node based on the local sensing data. The local sensing data can be obtained from the original communication logs stored by the current gateway node. In one implementation, the original communication logs are grouped according to the adjacent routing node identifiers and preset sampling periods. The busy or idle state of the corresponding adjacent routing node is determined based on the request sending records and acknowledgment response records within each sampling period. A state sequence is formed according to the sampling period order, and then the busy state percentage, maximum continuous busy time, and state fluctuation frequency are extracted from the state sequence. The state sequence and corresponding features are only used to illustrate one implementation of the local sensing data.

[0024] In one implementation, determining the link congestion risk level of each adjacent routing node based on the local sensing data includes: Decision tree processing is performed based on the local perception data to determine the information gain of each dimension of features; Dimensional features with information gain greater than a preset gain threshold are identified as splitting features, and decision classification paths are generated based on the splitting features. The congestion probability is obtained by traversing the decision classification path, and the link congestion risk level of the corresponding adjacent routing nodes is determined based on the congestion probability.

[0025] It should be noted that the decision tree is constructed before the real-time routing decision is executed. Historical communication logs are segmented according to a 60-minute sampling window and a 15-minute sliding step, and the segmentation results are grouped according to the identifiers of adjacent routing nodes. For each historical sampling period, the percentage of busy states, the maximum continuous busy time, and the frequency of state fluctuations are statistically analyzed to form a three-dimensional feature of the historical samples. The maximum response duration and maximum allowed response timeout rate specified by the communication protocol are read from the communication service level parameters stored in the current gateway node. Historical SMS concurrent requests whose difference between the received timestamp of the acknowledgment response and the sent timestamp of the corresponding historical SMS concurrent request is not less than the maximum response duration, and historical SMS concurrent requests that do not return an acknowledgment response within the maximum response duration, are identified as response timeout requests. The ratio of the number of response timeout requests to the total number of historical SMS concurrent requests is determined as the response timeout rate. When the response timeout rate is greater than the maximum allowed response timeout rate, the corresponding historical sample is automatically marked as a congested sample; otherwise, the corresponding historical sample is automatically marked as a non-congested sample, thus forming the link congestion result of the historical samples. For example, the maximum permissible response timeout rate is set to five percent according to the communication service level agreement of the device's communication network.

[0026] The decision tree is a classification decision tree based on information gain. Historical samples are divided into training samples and validation samples according to the chronological order of the historical sampling periods, with a ratio of 8:2 between training and validation samples. This chronological division is used to prevent data from later periods from being used in the construction of decision trees for earlier periods. For the training samples, the information gain, represented by the change in information entropy, is calculated, and the information gain is multiplied by 100 to obtain an information gain score. In this embodiment, the information gain is referred to as the information gain score. Candidate gain thresholds are set in a step size of 5 within a score range of 0 to 100. Candidate decision trees are constructed using the training samples, and the classification error rate of each candidate decision tree is calculated using the validation samples. The candidate gain threshold corresponding to the candidate decision tree with the lowest classification error rate is determined as the preset gain threshold. When multiple candidate gain thresholds correspond to the same lowest classification error rate, the candidate gain threshold with the largest value is selected to reduce the number of split nodes in the decision tree. For example, according to the above validation process, the preset gain threshold is set to 30.

[0027] Specifically, firstly, the proportion of busy states and the maximum continuous busy time in the state sequence are statistically analyzed, and the proportion of busy states, the maximum continuous busy time, and the state fluctuation frequency are determined as dimensional features of the local perception data. For the current decision tree node, the training samples in the current decision tree node are sorted in ascending order according to the proportion of busy states, the maximum continuous busy time, and the state fluctuation frequency, and the average of two adjacent feature values ​​with different values ​​is determined as a candidate split threshold. Based on each candidate split threshold, the training samples are divided into two sample subsets, and the information entropy before the split and the weighted information entropy of the two sample subsets after the split are calculated. The difference between the information entropy before the split and the weighted information entropy after the split is multiplied by 100 to obtain the information gain of the corresponding candidate split threshold. For each dimensional feature, the information gain with the largest value and the corresponding candidate split threshold are retained. From the dimensional features with information gain greater than the preset gain threshold, the dimensional feature with the largest information gain is determined as the split feature of the current decision tree node. When the training samples in the current decision tree node have the same link congestion result, the number of training samples is less than 2, or there is no dimensional feature with information gain greater than the preset gain threshold, the splitting of the current decision tree node is stopped, and the current decision tree node is determined as a leaf node.

[0028] It should be noted that the aforementioned busy state percentage, maximum continuous busy time, and state fluctuation frequency are exemplary dimensional features used in this embodiment to illustrate the local perception data processing method, and do not limit the local perception data to only include the above three dimensional features in actual deployment. In actual deployment, local perception data may also include one or more of the following: link bandwidth utilization, node queue length, acknowledgment response timeout rate, packet loss rate, node processing load, number of requests arriving per unit time, or jitter amplitude of acknowledgment responses from adjacent routing nodes. The number of dimensional features participating in the decision tree construction can be configured according to the data types and communication service level parameters that the current gateway node can collect. The total number of dimensional features can be a fixed number or can be dynamically adjusted according to changes in the gateway node's collection capabilities, communication protocol type, or network topology. The decision tree construction process does not rely on a fixed three-dimensional feature structure, but instead generates candidate splitting thresholds and calculates information gain for each of the currently participating dimensional features, selecting the dimensional feature with information gain greater than the preset gain threshold and the largest information gain as the splitting feature of the current decision tree node.

[0029] For example, a certain adjacent routing node has a busy state percentage of 60%, a maximum continuous busy time of 15 seconds, and a state fluctuation frequency of 3 Hz. After processing the above three dimensions of features with a decision tree, each dimension of features obtains a corresponding information gain, where the information gain corresponding to the state fluctuation frequency is 45, and the information gain corresponding to the busy state percentage is 25.

[0030] The information gain of each dimension feature is compared with a preset gain threshold. The dimension feature with information gain greater than the preset gain threshold is determined as the splitting feature of the current decision tree node. The dimension feature with the largest information gain is selected from the candidate splitting features as the splitting feature of the current decision tree node. Two branches are formed according to the candidate splitting threshold corresponding to the splitting feature: one branch is not greater than the candidate splitting threshold and the other branch is greater than the candidate splitting threshold. The splitting process is performed on subsequent decision tree nodes in sequence to generate a decision classification path consisting of a root node, a splitting node, and a leaf node.

[0031] For example, when the preset gain threshold is 30, the information gain 45 corresponding to the state fluctuation frequency is greater than the preset gain threshold 30, and the state fluctuation frequency is determined as a candidate splitting feature. When the information gain of the state fluctuation frequency is greater than the information gain of the other candidate splitting features, the state fluctuation frequency is determined as a splitting feature, and two branches are formed according to the candidate splitting threshold corresponding to the state fluctuation frequency.

[0032] In one implementation, dimensional features whose information gain is not greater than the preset gain threshold are identified as features to be pruned, and the features to be pruned are deleted from the dimensional features participating in the splitting decision of the current decision tree node.

[0033] It should be noted that the deletion targets are the candidate dimensional features participating in the splitting decision of the current decision tree node, not the current decision tree node itself. After deleting the features to be pruned, the remaining dimensional features continue to be used to determine the splitting features and generate the decision classification path.

[0034] For example, when the preset gain threshold is 30, the information gain 25 corresponding to the busy state ratio is not greater than the preset gain threshold 30. The busy state ratio is determined as a feature to be pruned and deleted from the candidate dimension features of the current decision tree node; the state fluctuation frequency continues to participate in the splitting decision of the current decision tree node.

[0035] The decision classification path is traversed and mapped, comparing the features of each dimension of the local sensing data with the branch conditions of each split node in the decision classification path until the corresponding leaf node is reached. For each leaf node, the ratio of the number of congested training samples entering the leaf node to the total number of training samples entering the leaf node is determined as the congestion sample ratio of the leaf node. Leaf nodes with a congestion sample ratio of less than 50% are configured as low-frequency congestion type initial risk labels, leaf nodes with a congestion sample ratio of not less than 50% and less than 80% are configured as medium-frequency congestion type initial risk labels, and leaf nodes with a congestion sample ratio of not less than 80% and not greater than 100% are configured as high-frequency congestion type initial risk labels. The average congestion sample ratio of each leaf node with the same initial risk label is calculated, and the correspondence between the initial risk label and the corresponding average value is determined as the preset risk label and congestion probability mapping relationship.

[0036] The preset mapping relationship between congestion probability ranges and link congestion risk levels includes: a congestion probability of not less than 80% and not greater than 100% corresponds to Level 1 risk; a congestion probability of not less than 50% and less than 80% corresponds to Level 2 risk; and a congestion probability of not less than 0% and less than 50% corresponds to Level 3 risk. Level 1 risk is higher than Level 2 risk, and Level 2 risk is higher than Level 3 risk. This forms a complete conversion relationship from leaf nodes, initial risk labels, congestion probabilities to link congestion risk levels.

[0037] The preset congestion risk level threshold is determined based on the false negative rate of congestion samples and the trigger ratio of risk nodes in the verification samples. Level 1, Level 2, and Level 3 risks are sequentially selected as candidate congestion risk level thresholds. The false negative rate of congestion samples and the trigger ratio of risk nodes corresponding to each candidate congestion risk level threshold are calculated. The maximum allowable false negative rate is read from the communication service level parameters. Among the candidate congestion risk level thresholds where the false negative rate of congestion samples is not greater than the maximum allowable false negative rate, the candidate congestion risk level threshold with the lowest trigger ratio of risk nodes is selected. For example, if the maximum allowable false negative rate is 5%, Level 2 risk is determined as the preset congestion risk level threshold based on the above statistical process.

[0038] For example, after inputting local sensing data containing the percentage of busy states, the maximum continuous busy time, and the frequency of state fluctuations into the decision classification path, the local sensing data sequentially passes through the corresponding branches and reaches leaf node L3; the initial risk label corresponding to leaf node L3 is high-frequency congestion type, and the congestion probability is obtained as 85% according to the preset risk label and congestion probability mapping relationship. Then, according to the preset congestion probability interval and link congestion risk level mapping relationship, the link congestion risk level of the corresponding adjacent routing node is determined as level one risk.

[0039] After the decision tree is constructed, the dimensional features, information gain, split threshold, and child node index corresponding to each decision tree node, the initial risk label corresponding to each leaf node, the preset mapping relationship between risk labels and congestion probability, the preset mapping relationship between congestion probability intervals and link congestion risk levels, and the preset congestion risk level threshold are combined to form decision tree parameter data, which is stored in the current gateway node. When performing real-time routing decisions, the current gateway node calls the decision tree parameter data, inputs the currently acquired local sensing data into the constructed decision tree, and compares it according to the dimensional features and split thresholds corresponding to each decision tree node until the corresponding leaf node is traversed; the decision tree is not reconstructed or the decision tree parameter data is not updated during the real-time routing decision stage.

[0040] In step S12, the link congestion risk level of each adjacent routing node is first compared with the preset congestion risk level threshold. The adjacent routing nodes whose link congestion risk level is higher than the preset congestion risk level threshold are identified as risk nodes. The response delay time series corresponding to the next hop routing node is formed from the historical forwarding decision records of the risk nodes.

[0041] In one implementation, historical forwarding decision records are obtained from the risk node, and the response delay time sequence formed by the risk node forwarding historical concurrent SMS requests to the next hop routing node is extracted, including: Extract the sending timestamp of each historical SMS concurrent request from the historical forwarding decision record returned by the risk node, and the receiving timestamp of the confirmation response received by the risk node from the next-hop routing node. An initial response delay time series is constructed based on the difference between the sending timestamp and the receiving timestamp. The initial response delay time series is then smoothed and denoised to obtain a smoothed delay time series. The extreme value features are extracted from the smoothed delay time series to determine the response delay time series.

[0042] Specifically, the risk node returns historical forwarding decision records from the local forwarding log. The historical forwarding decision records include the request identifier, next-hop routing node identifier, sending timestamp, and receiving timestamp of the corresponding confirmation response for each historical concurrent SMS request. The historical forwarding decision records are grouped according to the next-hop routing node identifier, and the sending timestamp and receiving timestamp are paired using the request identifier in each group.

[0043] For example, in the preset risk level ranking relationship, level 1 risk is higher than level 2 risk. When the link congestion risk level of a certain adjacent routing node is level 1 and the preset congestion risk level threshold is level 2, the adjacent routing node is identified as a risk node, and 120 historical forwarding decision records of the risk node over the past 24 hours are obtained. From one of the historical forwarding decision records, the sending timestamp of the historical SMS concurrent request is extracted as 10:00:00:015 milliseconds, and the receiving timestamp of the risk node receiving the corresponding confirmation response is extracted as 10:00:00:053 milliseconds.

[0044] It should be noted that, considering that transient communication interference in the device communication network may cause abnormal fluctuations in the initial response delay time, a moving average filtering algorithm is used to denoise the initial response delay time series. The difference between each received timestamp and the corresponding sent timestamp is determined as the delay data, and the delay data are arranged in the order of the sent timestamps to construct the initial response delay time series. The moving average method is used to calculate the average value of the delay data within a preset window, and the average value is used to update the delay data at the center of the window to obtain a smooth delay time series.

[0045] The preset window consists of five consecutive delay data points arranged according to their transmission timestamps, with a sliding step of one delay data point. The moving average window uses an odd number of delay data points, ensuring each complete preset window has a unique center position. The five delay data points weaken isolated abrupt changes while preserving adjacent short-term fluctuations, serving as the window size in this embodiment. For each complete preset window, the arithmetic mean of the five delay data points is calculated, and the delay data at the center position of the preset window is updated using this arithmetic mean. The initial delay data points at the beginning and end of the initial response delay time series that cannot form a complete preset window retain their original values.

[0046] For example, the difference between the received timestamp and the sent timestamp is 38 milliseconds. The differences corresponding to each historical concurrent SMS request are arranged in chronological order of the sent timestamps to form an initial response delay time series. For a sliding window consisting of 15 milliseconds, 16 milliseconds, 45 milliseconds, 16 milliseconds, and 13 milliseconds, the average delay data within the window is 21 milliseconds. The 45 milliseconds at the center of the window are updated using 21 milliseconds to obtain the corresponding smoothed delay data.

[0047] According to the preset detection window, the size relationship between each delay data in the smooth delay time series and the adjacent delay data is compared sequentially. The delay data that is greater than the adjacent delay data is determined as the local maximum value, and the delay data that is less than the adjacent delay data is determined as the local minimum value. The delay value and time position corresponding to each local maximum value and each local minimum value are marked in the smooth delay time series to obtain the extreme value feature. The smooth delay time series that has completed the extreme value feature marking is determined as the response delay time series.

[0048] It should be noted that the preset detection window consists of three temporally consecutive smoothed delay data points, with a sliding step of one smoothed delay data point. These three smoothed delay data points simultaneously contain the data to be detected, the preceding adjacent data point, and the following adjacent data point, forming the smallest symmetrical data set for a centrally adjacent comparison relationship. Therefore, the preset detection window is set to three smoothed delay data points. The smoothed delay data point located at the center of the preset detection window is compared with the preceding and following smoothed delay data points respectively. When the smoothed delay data point at the center is strictly greater than the preceding and following smoothed delay data points, it is determined as a local maximum; when the smoothed delay data point at the center is strictly less than the preceding and following smoothed delay data points, it is determined as a local minimum. The first and last delayed data points in the smoothed delay time series are not included in the extreme value determination.

[0049] For example, a smooth delay time series consisting of 18 ms, 20 ms, 24 ms, 35 ms, 29 ms, 23 ms and 20 ms is detected. 35 ms is greater than the adjacent 24 ms and 29 ms. 35 ms is identified as a local maximum and a local maximum mark is added at the time position corresponding to 35 ms. The original time order of each delay data is preserved, and the above smooth delay time series with the local maximum mark is identified as the response delay time series.

[0050] In step S13, time series analysis is first performed on the response delay time series to extract the delay fluctuation amplitude used to characterize the degree of delay change and the delay decay trend used to characterize the delay recovery process. The cascading failure probability of the next-hop routing node is then determined based on the delay fluctuation amplitude and the delay decay trend.

[0051] In one implementation, determining the cascading failure probability of the next-hop routing node based on the latency fluctuation amplitude and latency decay trend includes: The delay fluctuation amplitude and delay decay trend are extracted from the response delay time series, and a delay state transition matrix is ​​generated based on the delay fluctuation amplitude and the delay decay trend; The failure evolution path of the next-hop routing node is extracted based on the delay state transition matrix, and the cumulative failure risk value is determined based on the failure evolution path. For a failure evolution path where the cumulative failure risk value is greater than a preset cumulative failure risk threshold, the topology association weight of the next-hop routing node is obtained, and the cumulative failure risk value is weighted according to the topology association weight to obtain the cascading failure probability.

[0052] Specifically, multiple analysis windows are set for the response delay time series, and the delay fluctuation amplitude and delay decay trend within each analysis window are determined respectively. For example, each analysis window consists of five consecutive sampling points in time, with a sliding step of one sampling point; the first analysis window is set starting from the position in the response delay time series with five preceding sampling points. The largest response delay value within the analysis window is determined as the delay peak value; when the largest response delay value corresponds to multiple sampling points, the last sampling point in time is determined as the sampling point corresponding to the delay peak value. The average response delay of the five consecutive sampling points preceding the analysis window is determined as the delay baseline value, and the larger of the difference between the delay peak value and the delay baseline value and zero is determined as the delay fluctuation amplitude. To ensure that the delay fluctuation amplitude and delay decay trend use the same time scale, five sampling periods are set as a preset decay observation period. Starting from the sampling point corresponding to the delay peak, subsequent sampling points are detected sequentially. When the response delay first falls back to the preset normal delay range, the number of sampling periods from the sampling point corresponding to the delay peak to the first fall is determined as the delay attenuation period number, which represents the delay attenuation trend. When the delay peak is already within the preset normal delay range, the delay attenuation period number is set to zero; when the response delay does not fall back to the preset normal delay range within the preset attenuation observation period, the delay attenuation period number is set to 6; when there are less than 5 sampling periods after the delay peak, the corresponding analysis window does not participate in the delay status statistics.

[0053] The preset normal delay interval, preset delay fluctuation range, and preset delay attenuation trend interval are determined based on historical normal forwarding records before routing decisions are executed. Specifically, from the historical forwarding decision records corresponding to the next-hop routing node, historical forwarding decision records that returned an acknowledgment response within the maximum response time specified by the communication protocol and whose corresponding links were not interrupted are selected to form historical normal forwarding records, which are then arranged according to timestamps. The response delays of each historical SMS concurrent request are extracted from the historical normal forwarding records and arranged in ascending order of response delay values. For an ascending sequence containing n values, the product of n and the corresponding percentile ratio is rounded up to obtain a target sequence number, and the value corresponding to the target sequence number is determined as the corresponding percentile. The smaller value between the 95th percentile response delay value and the maximum response time specified by the communication protocol is determined as the upper limit of normal delay, and the range of values ​​that are not less than zero and not greater than the upper limit of normal delay is determined as the preset normal delay interval.

[0054] Historical latency fluctuation amplitudes are extracted from the historical normal forwarding records using the same analysis window and sliding step size, and then sorted in ascending order of historical latency fluctuation amplitude. The 75th percentile historical latency fluctuation amplitude is determined as the first fluctuation boundary, and the 95th percentile historical latency fluctuation amplitude is determined as the second fluctuation boundary. Latency fluctuation amplitudes not greater than the first fluctuation boundary are considered normal fluctuations, latency fluctuation amplitudes greater than the first fluctuation boundary but not greater than the second fluctuation boundary are considered mild fluctuations, and latency fluctuation amplitudes greater than the second fluctuation boundary are considered severe fluctuations, thus forming the preset latency fluctuation amplitude range.

[0055] The historical normal forwarding records are counted using the same preset attenuation observation period, and then sorted in ascending order of historical delay attenuation period count. The 75th percentile of historical delay attenuation period count is determined as the first attenuation boundary, and the 95th percentile of historical delay attenuation period count is determined as the second attenuation boundary. Delay attenuation period counts not greater than the first attenuation boundary are considered normal attenuation, delay attenuation period counts greater than the first attenuation boundary but not greater than the second attenuation boundary are considered mild attenuation, and delay attenuation period counts greater than the second attenuation boundary are considered severe attenuation, thus forming the preset delay attenuation trend range. The 75th percentile is used to distinguish between the main distribution and the upper quartile distribution in the historical normal forwarding records, and the 95th percentile is used to identify tail delay changes in the historical normal forwarding records.

[0056] When determining the latency state of an analysis window, if the latency fluctuation amplitude of the analysis window is severely fluctuating, the latency decay cycle number is severely decaying, or the latency peak is not less than the maximum response time specified by the communication protocol, the analysis window is determined to be in a severely delayed state. If the analysis window is not determined to be in a severely delayed state, and the latency peak is within the preset normal latency range, the latency fluctuation amplitude is normal, and the latency decay cycle number is normal, the analysis window is determined to be in a normal latency state. Analysis windows other than the normal latency state and the severely delayed state are determined to be in a mild latency state, thus ensuring that each analysis window corresponds to only one latency state.

[0057] The number of transitions from any delayed state to various delayed states between adjacent analysis windows is counted chronologically, including transitions that maintain the same delayed state. The ratio between the corresponding number of state transitions and the total number of state transitions from any delayed state is determined as the state transition probability. When the total number of state transitions from any delayed state is zero, the state transition probability of transitioning from that delayed state to itself is set to 1, and the state transition probability of transitioning from that delayed state to other delayed states is set to 0. The rows and columns of the delayed state transition matrix are arranged in the order of normal delayed state, mild delayed state, and severe delayed state, where the rows of the delayed state transition matrix represent the delayed state before the transition, and the columns of the delayed state transition matrix represent the delayed state after the transition, thus forming a three-row, three-column delayed state transition matrix.

[0058] For example, after determining the delay state of each analysis window, the number of times the normal delay state transitioned to the normal delay state, the slightly delayed state, and the severely delayed state were statistically analyzed: 4 times, 5 times, and 1 time, respectively. The number of times the slightly delayed state transitioned to the normal delay state, the slightly delayed state, and the severely delayed state were analyzed: 3 times, 4 times, and 3 times, respectively. The number of times the severely delayed state transitioned to the normal delay state, the slightly delayed state, and the severely delayed state were analyzed: 3 times, 1 time, and 6 times, respectively. Therefore, the state transition probability from the normal delay state to the slightly delayed state is 50%, and the state transition probability from the slightly delayed state to the severely delayed state is 30%. These state transition probabilities are then written into the delay state transition matrix according to a preset state order.

[0059] The severely delayed state is defined as the preset pre-failure state. Based on the state transition probabilities between each delayed state in the delay state transition matrix, a continuous state sequence is extracted along the gradual deterioration direction from the normal delay state to the mild delay state and then to the severely delayed state, forming the failure evolution path of the next-hop routing node. The product of the state transition probabilities on the failure evolution path is determined as the cumulative failure risk value of the failure evolution path. The cumulative failure risk value is used to characterize the degree to which the response delay gradually deteriorates along the failure evolution path and enters the preset pre-failure state.

[0060] For example, based on the aforementioned delay state transition matrix, a state sequence is extracted that transitions from a normal delay state to a slightly delayed state, and then from a slightly delayed state to a severely delayed state, forming a failure evolution path. Multiplying 50% by 30% yields a cumulative failure risk value of 15%.

[0061] It should be noted that the topology association weight is a dimensionless topology risk amplification coefficient corresponding to the next-hop routing node identifier; the current gateway node reads the topology association weight from the pre-configured node topology association weight table using the next-hop routing node identifier as the query key. For failure evolution paths where the cumulative failure risk value is greater than a preset cumulative failure risk threshold, a multiplicative weighted method is used to determine the cascading failure probability: when the product of the cumulative failure risk value and the topology association weight is not greater than 100%, the product is determined as the cascading failure probability; when the product is greater than 100%, 100% is determined as the cascading failure probability. When the cumulative failure risk value is not greater than the preset cumulative failure risk threshold, the cumulative failure risk value is determined as the cascading failure probability.

[0062] In one implementation, the backup gateway node pre-establishes a node topology association weight table in the control plane based on global view topology data. For any routing node i with at least one active link, the node connectivity d_i of routing node i is calculated, and the maximum node connectivity d_max in the device communication network is obtained. The normalized topology association degree c_i of routing node i is calculated according to c_i=d_i / d_max, and the topology association weight w_i of routing node i is calculated according to w_i=1+3c_i, so that the topology association weight is between 1 and 4. The topology association weight is used to characterize the potential impact range of the surrounding connection relationship after the corresponding routing node becomes unresponsive or the link is interrupted, rather than characterizing the probability of the corresponding routing node itself becoming unresponsive or the link is interrupted. Therefore, the larger the node connectivity, the larger the amplification factor used for the cumulative failure risk value. Routing nodes without active links are not added to the candidate backup forwarding path. The node topology association weight table records the routing node identifier, topology version identifier, and topology association weight. When a routing node joins, leaves, or the link connection changes, the backup gateway node recalculates the topology association weight of the affected routing node, generates a node topology association weight table with the new topology version identifier, and distributes it to the current gateway node. During the real-time routing decision phase, the current gateway node only queries the node topology association weight table based on the next-hop routing node identifier. For example, when the node connectivity of the next-hop routing node equals the maximum node connectivity, the normalized topology association degree is 1, and the topology association weight is 4.

[0063] The preset cumulative failure risk threshold is determined based on historical failure evolution path samples during the offline configuration phase. Historical failure evolution path samples that consistently return acknowledgment responses within the maximum response time specified by the communication protocol and whose corresponding links have not been interrupted during the preset subsequent observation period are identified as reference evolution paths. These reference evolution paths are arranged in ascending order of cumulative failure risk value, and the cumulative failure risk value corresponding to the 95th percentile is determined as the preset cumulative failure risk threshold. This ensures that 95% of the reference evolution paths remain in the calculation branches without topology risk amplification, while the remaining 5% of failure evolution risks are subject to topology risk amplification. When the node topology association weight table or historical sample set is updated, the backup gateway node recalculates the preset cumulative failure risk threshold on the control plane and distributes it to the current gateway node. Only numerical comparisons are performed during the real-time routing decision phase. For example, if the preset subsequent observation period is 5 minutes, and the 95th percentile of the cumulative failure risk value for each reference evolution path is 10%, then the preset cumulative failure risk threshold is set to 10%.

[0064] For example, when the preset cumulative failure risk threshold is 10%, the cumulative failure risk value of the above failure evolution path is 15%, and the topology association weight of the next-hop routing node is four, since 15% is greater than 10%, the 15% cumulative failure risk value is multiplied by the topology association weight four to obtain a 60% cascading failure probability. As another example, when the cumulative failure risk value is 8% and the preset cumulative failure risk threshold is 10%, the cumulative failure risk value is not greater than the preset cumulative failure risk threshold, and 8% is determined as the cascading failure probability.

[0065] In step S14, the cascading failure probability needs to be compared with a preset cascading failure probability threshold. When the cascading failure probability is not less than the preset cascading failure probability threshold, the current gateway node obtains global view topology data from the backup gateway node and predicts a backup forwarding path based on the global view topology data. When the cascading failure probability is less than the preset cascading failure probability threshold, the current gateway node determines the target forwarding path based on local sensing data.

[0066] In one implementation, regression prediction is performed on the global view topology data to obtain alternative forwarding paths, including: Node connectivity and link load features are extracted from the global view topology data, and a topology feature matrix is ​​constructed. Support vector regression prediction is performed on the topological feature matrix to predict the path connectivity cost of each candidate backup forwarding path in the preset set of candidate backup forwarding paths. The candidate backup forwarding paths are sorted in ascending order according to the path connectivity cost, and the candidate backup forwarding path with the lowest path connectivity cost is determined as the backup forwarding path.

[0067] It should be noted that the preset cascading failure probability threshold is determined based on historical cascading failure samples during the offline configuration phase. For each historical cascading failure probability, the response status of the corresponding next-hop routing node is observed within a preset subsequent observation period. When the next-hop routing node experiences a preset number of consecutive acknowledgment response timeouts or the corresponding link is interrupted within the preset subsequent observation period, the corresponding historical sample is marked as a cascading failure sample; otherwise, it is marked as a non-cascading failure sample. Multiple candidate cascading failure probability thresholds are set within a preset candidate probability range according to a preset probability step size. For each candidate cascading failure probability threshold, the missed detection rate of cascading failure samples and the proportion of historical samples entering the global view topology prediction branch are calculated. Among the candidate cascading failure probability thresholds where the missed detection rate of cascading failure samples is not greater than the preset missed detection rate upper limit, the candidate cascading failure probability threshold with the largest value is selected to reduce the number of triggers for global view topology prediction while meeting the cascading failure risk identification requirements. When the historical sample set is updated, the backup gateway node recalculates the preset cascading failure probability threshold in the control plane; only one probability comparison is performed during the real-time routing decision phase. For example, based on the above historical sample statistical process, the preset cascading failure probability threshold is set to ten percent.

[0068] It should be noted that the global view topology data includes active routing nodes in the device communication network, the connection relationships between routing nodes, and the load records of each link. Node connectivity is determined based on the number of neighboring routing nodes currently maintaining active connections with the corresponding routing node. For each active routing node, the maximum bandwidth occupancy ratio of each active link connected to the active routing node within a preset statistical window is determined as the link load characteristic of the active routing node. Node connectivity and link load characteristics are arranged according to the node order in the global view topology data to construct a topology feature matrix.

[0069] For example, the global view topology data contains 40 active routing nodes. For each active routing node, node connectivity and link load characteristics are extracted to construct a 40-row, 2-column topology feature matrix. If one of the active routing nodes currently maintains active connections with 4 adjacent routing nodes, and the maximum bandwidth occupancy ratio among the active links connected to the active routing node is 70%, then the node connectivity of the active routing node is 4 and the link load characteristic is 70%.

[0070] It should be noted that the preset set of candidate backup forwarding paths is pre-generated by the backup gateway node in the control plane. Before the real-time routing decision is executed, the backup gateway node constructs a topology graph of the device communication network, using active routing nodes as graph nodes and active links as graph edges. For each registered target routing node in the device communication network, candidate backup forwarding paths are generated using the K-shortest acyclic path algorithm, starting from the current gateway node and ending at the corresponding target routing node, in ascending order of path hop count. Candidate backup forwarding paths containing offline routing nodes or interrupted links are deleted, and the top K candidate backup forwarding paths are retained to form the preset set of candidate backup forwarding paths. During the offline configuration phase, the candidate K value is increased sequentially, and the data storage volume of the corresponding path and the processing time of one support vector regression prediction are calculated. The maximum candidate K value that does not exceed the current gateway node's reserved path storage volume and the upper limit of the routing decision time specified by the communication service level parameters is determined as K.

[0071] For example, K is set to 25 according to the above statistical process. A preset set of candidate backup forwarding paths is indexed based on the current gateway node identifier and the target routing node identifier and pre-stored in the current gateway node. When performing real-time routing decisions, the SMS target address is extracted from the concurrent SMS requests to be forwarded, the target routing node is determined according to the preset mapping relationship between the SMS target address and the target routing node, and the corresponding preset set of candidate backup forwarding paths is queried based on the current gateway node identifier and the target routing node identifier. The corresponding preset set of candidate backup forwarding paths is regenerated only when a routing node is added, removed, or the link connection relationship changes; when the link load characteristics change, the path is not searched again, but the path characteristics of each candidate backup forwarding path are updated according to the latest global view topology data.

[0072] For each candidate backup forwarding path, the current gateway node extracts the node connectivity and link load characteristics of each routing node traversed by the candidate backup forwarding path from the topology feature matrix. It calculates the average node connectivity, minimum node connectivity, average link load characteristics, and maximum link load characteristics, and counts the number of routing nodes included in the candidate backup forwarding path. The obtained five values ​​are arranged in a fixed order to form a five-dimensional path feature vector.

[0073] Since each candidate backup forwarding path is converted into 5 statistical parameters, the number of routing nodes traversed by the candidate backup forwarding path does not change the dimension of the path feature vector. For each dimension of the five-dimensional path feature vector, minimum-maximum normalization is performed using the minimum and maximum values ​​of the corresponding dimension in the training samples; when the maximum and minimum values ​​of the corresponding dimension are the same, the corresponding normalized feature value is set to zero. The minimum and maximum values ​​of each dimension obtained during the training phase are stored as input feature normalization parameters along with the support vector regression model.

[0074] The support vector regression model is an ε-support vector regression model using a radial basis function kernel, including the support vectors, dual coefficients, and bias parameters retained after training. The backup gateway node acquires at least a preset number of historical path samples within a preset historical training period on the control plane. For each historical path sample, a corresponding five-dimensional path feature vector is extracted as the training input. When a historical path returns an acknowledgment response before the maximum response time specified in the communication protocol, the actual end-to-end response delay corresponding to the historical path is determined as the path connectivity cost label. When a historical path fails to return an acknowledgment response before the maximum response time or experiences a link interruption, the maximum response time is determined as the path connectivity cost label. The training set and validation set are divided into eight to two sets based on the chronological order of the historical path samples.

[0075] Based on the features of each dimension in the training set, input feature normalization parameters are determined, and these parameters are used to normalize both the training and validation inputs. Path connectivity cost labels retain their original values ​​in milliseconds. The support vector regression model employs a radial basis function kernel. A grid search is performed among the parameter combinations consisting of penalty parameter candidate values, insensitive interval parameter candidate values, and radial basis function kernel function parameter candidate values. The mean absolute error (MAE) of the model corresponding to each parameter combination on the validation set is calculated, and the parameter combination with the smallest MAE is determined as the target parameter combination.

[0076] The maximum number of iterations for the Support Vector Regression (SVR) model is set to 600. Model training ends when the model optimization process reaches 600 iterations or the change in the objective function value between two consecutive iterations is no greater than 0.0001. The maximum number of iterations and the threshold for the change in the objective function value are determined based on the upper limit of offline training time and the number of iterations required for the mean absolute error of the validation set to stabilize. After training, the SVR model, input feature normalization parameters, and model version identifier are stored in the current gateway node. During the real-time routing decision phase, only the five-dimensional path feature vectors of each candidate backup forwarding path are input into the SVR model to obtain the path connectivity cost in milliseconds; no model training or parameter updates are performed.

[0077] For example, the preset set of candidate backup forwarding paths contains 25 candidate backup forwarding paths. The corresponding features of the nodes traversed by each candidate backup forwarding path are extracted from the 40-row, 2-column topological feature matrix to form 25 sets of five-dimensional path feature vectors. These vectors are then sequentially input into a pre-configured support vector regression model to obtain the path connectivity cost of each candidate backup forwarding path. The path connectivity costs of candidate backup forwarding paths P1, P2, and P3 are 12 milliseconds, 18 milliseconds, and 26 milliseconds, respectively.

[0078] It should be noted that path connectivity cost is used to characterize the risk of interruption or delay anomalies in candidate backup forwarding paths. The smaller the path connectivity cost, the higher the connectivity stability of the candidate backup forwarding path. The candidate backup forwarding paths are arranged in ascending order of path connectivity cost, and the candidate backup forwarding path that is first and has the smallest path connectivity cost is determined as the backup forwarding path.

[0079] For example, after sorting the candidate backup forwarding paths P1, P2 and P3 with path connectivity costs of 12, 18 and 26 in ascending order, candidate backup forwarding path P1 is in the first position, and candidate backup forwarding path P1 is determined as the backup forwarding path.

[0080] In one implementation, when the cascading failure probability is less than a preset cascading failure probability threshold, the following is included: The current gateway node determines the target forwarding path based on the local sensing data; Based on the target forwarding path, generate the optimal routing instruction for the concurrent requests of the SMS messages to be forwarded.

[0081] Specifically, the current gateway node extracts candidate target forwarding paths that can reach the SMS target address from the local routing table, determines the first adjacent routing node in each candidate target forwarding path, and associates the link congestion risk level and congestion probability of each adjacent routing node obtained in step S11 with the candidate target forwarding path corresponding to the first adjacent routing node. It deletes candidate target forwarding paths whose link congestion risk level of the first adjacent routing node is higher than the preset congestion risk level threshold, and then sorts the remaining candidate target forwarding paths in ascending order according to the congestion probability associated with them. The candidate target forwarding path with the lowest congestion probability is determined as the target forwarding path.

[0082] For example, when the cascading failure probability is 8% and the preset cascading failure probability threshold is 10%, the target forwarding path branch is entered. The congestion probabilities of the first adjacent routing node of candidate target forwarding path A, candidate target forwarding path B, and candidate target forwarding path C in the local routing table are 20%, 35%, and 85%, respectively. The link congestion risk level of the first adjacent routing node of candidate target forwarding path C is higher than the preset congestion risk level threshold, so candidate target forwarding path C is deleted, and then candidate target forwarding path A with the lowest congestion probability is determined as the target forwarding path.

[0083] The current gateway node reads the routing node identifier sequence corresponding to the target forwarding path, determines the next-hop routing node identifier based on the routing node identifier sequence, and writes the routing node identifier sequence and the next-hop routing node identifier into the routing control information for concurrent requests of SMS messages to be forwarded, forming the optimal routing instruction.

[0084] For example, the target forwarding path passes through the current gateway node G0, the adjacent routing node G2 and the target node G5 in sequence. The current gateway node generates an optimal routing instruction that includes the routing node identifier sequence G0-G2-G5 and the next-hop routing node identifier G2 according to the target forwarding path.

[0085] In step S15, an optimal routing instruction is generated for concurrent requests of SMS messages to be forwarded, based on the backup forwarding path.

[0086] For cases where the cascading failure probability is not less than a preset cascading failure probability threshold, the optimal routing instruction for concurrent SMS requests to be forwarded needs to be generated first, based on the alternative forwarding path determined in step S14. The routing instruction generation process when the cascading failure probability is less than the preset cascading failure probability threshold has been completed in another implementation of step S14.

[0087] Specifically, the current gateway node reads concurrent SMS requests to be forwarded from the queue to be forwarded, extracts the request identifier and SMS destination address of the concurrent SMS requests to be forwarded, determines the next-hop routing node according to the routing node identifier sequence corresponding to the backup forwarding path, and writes the request identifier, SMS destination address, routing node identifier sequence and next-hop routing node identifier into the routing instruction to obtain the optimal routing instruction for the concurrent SMS requests to be forwarded.

[0088] For example, if the request identifier of the concurrent SMS request to be forwarded is M001 and the SMS destination address is D09, and the routing node identifier sequence corresponding to the backup forwarding path determined in step S14 is the current gateway node G0, routing node R2 and routing node R5 respectively, then routing node R2 is determined as the next-hop routing node, and an optimal routing instruction containing the request identifier M001, SMS destination address D09, routing node identifier sequence G0-R2-R5 and next-hop routing node identifier R2 is generated.

[0089] In summary, this invention discloses a data routing method for IoT gateways carrying high-concurrency SMS transmission. By generating optimal routing instructions for concurrent SMS requests based on the target forwarding path or backup forwarding path, the method reduces the possibility of selecting forwarding paths with link congestion or cascading failure risks, thereby improving the stability of IoT gateway data routing during high-concurrency SMS transmission.

[0090] The second embodiment of the present invention provides an IoT gateway data routing system for carrying high-concurrency SMS sending, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described above.

[0091] It should be noted that the IoT gateway data routing system for carrying high-concurrency SMS sending provided in this embodiment of the invention is used to execute all the process steps of the IoT gateway data routing method for carrying high-concurrency SMS sending in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0092] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A data routing method for an IoT gateway supporting high-concurrency SMS sending, characterized in that, include: Obtain local perception data of the current gateway node, and determine the link congestion risk level of each adjacent routing node based on the local perception data; The adjacent routing nodes whose link congestion risk level is higher than the preset congestion risk level threshold are identified as risk nodes. Historical forwarding decision records are obtained from the risk nodes, and the response delay time series formed by the risk nodes forwarding historical SMS concurrent requests to the next hop routing node is extracted. Time series analysis is performed on the response delay time series, and the cascading failure probability of the next-hop routing node is determined based on the delay fluctuation amplitude and delay decay trend; When the cascading failure probability is not less than the preset cascading failure probability threshold, the global view topology data of the backup gateway node is obtained, and regression prediction is performed on the global view topology data to obtain the backup forwarding path. Based on the alternative forwarding path, generate the optimal routing instruction for concurrent requests of SMS messages to be forwarded.

2. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 1, characterized in that, The step of determining the link congestion risk level of each adjacent routing node based on the local sensing data includes: Decision tree processing is performed based on the local perception data to determine the information gain of each dimension of features; Dimensional features with information gain greater than a preset gain threshold are identified as splitting features, and decision classification paths are generated based on the splitting features. The congestion probability is obtained by traversing the decision classification path, and the link congestion risk level of the corresponding adjacent routing nodes is determined based on the congestion probability.

3. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 2, characterized in that, The step of determining dimensional features with information gain greater than a preset gain threshold as splitting features also includes determining dimensional features with information gain not greater than the preset gain threshold as pruning features, and deleting the pruning features from each dimensional feature participating in the splitting decision of the current decision tree node.

4. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 2, characterized in that, The process of traversing and mapping along the decision classification path to obtain congestion probabilities, and determining the link congestion risk level of corresponding adjacent routing nodes based on the congestion probabilities, includes: Based on the local sensing data, traverse along the decision classification path to determine the leaf node corresponding to the local sensing data; The initial risk label is determined based on the leaf node, and the congestion probability corresponding to the initial risk label is determined based on the preset mapping relationship between the risk label and the congestion probability. Based on the preset mapping relationship between congestion probability range and link congestion risk level, the link congestion risk level corresponding to the congestion probability is determined.

5. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 1, characterized in that, The step of obtaining historical forwarding decision records from the risk node and extracting the response delay time series formed by the risk node forwarding historical SMS concurrent requests to the next hop routing node includes: Extract the sending timestamp of each historical SMS concurrent request from the historical forwarding decision record returned by the risk node, and the receiving timestamp of the confirmation response received by the risk node from the next-hop routing node. An initial response delay time series is constructed based on the difference between the sending timestamp and the receiving timestamp. The initial response delay time series is then smoothed and denoised to obtain a smoothed delay time series. The extreme value features are extracted from the smoothed delay time series to determine the response delay time series.

6. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 5, characterized in that, The determination of the cascading failure probability of the next-hop routing node based on the latency fluctuation amplitude and latency decay trend includes: The delay fluctuation amplitude and delay decay trend are extracted from the response delay time series, and a delay state transition matrix is ​​generated based on the delay fluctuation amplitude and the delay decay trend; The failure evolution path of the next-hop routing node is extracted based on the delay state transition matrix, and the cumulative failure risk value is determined based on the failure evolution path. For a failure evolution path where the cumulative failure risk value is greater than a preset cumulative failure risk threshold, the topology association weight of the next-hop routing node is obtained, and the cumulative failure risk value is weighted according to the topology association weight to obtain the cascading failure probability.

7. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 1, characterized in that, The step of "when the chain failure probability is not less than a preset chain failure probability threshold" further includes "when the chain failure probability is less than the preset chain failure probability threshold", whereby the current gateway node determines the target forwarding path based on the local perception data, and generates an optimal routing instruction for the concurrent request of the SMS to be forwarded based on the target forwarding path.

8. The IoT gateway data routing method for carrying high-concurrency SMS transmission according to claim 1, characterized in that, The step of performing regression prediction on the global view topology data to obtain alternative forwarding paths includes: Node connectivity and link load features are extracted from the global view topology data, and a topology feature matrix is ​​constructed. Support vector regression prediction is performed on the topological feature matrix to predict the path connectivity cost of each candidate backup forwarding path in the preset set of candidate backup forwarding paths. The candidate backup forwarding paths are sorted in ascending order according to the path connectivity cost, and the candidate backup forwarding path with the lowest path connectivity cost is determined as the backup forwarding path.

9. An IoT gateway data routing system for supporting high-concurrency SMS sending, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.