A method, system, device and medium for optimizing stable transmission of Internet of Things data

Abstract

The present invention discloses a method, system, device, and medium for optimizing the stable transmission of Internet of Things (IoT) data, relating to the field of IoT. The method comprises generating multiple test scenarios through a multi-network simulation model; inputting the multiple test scenarios into a target IoT to obtain IoT test data; inputting the IoT test data into a transmission performance prediction model to obtain performance prediction data; inputting the performance prediction data into a data mining and analysis model to obtain data mining information; generating a data stability analysis report based on the data mining information; and identifying anomalies based on the data stability analysis report to form a detection and maintenance plan. The present invention effectively improves the stability, reliability, and operation and maintenance efficiency of IoT systems, and has broad application prospects and significant engineering practical value.

Description

Technical Field

[0001] This specification relates to the field of the Internet of Things, and more specifically, the present invention relates to a method, system, device, and medium for optimizing stable transmission of Internet of Things data. Background Art

[0002] With the rapid development of the Internet of Things (IoT) technology, large-scale heterogeneous terminal devices exchange data through wireless access networks, core networks, and the Internet. The data transmission stability of IoT systems is increasingly becoming a key factor affecting system reliability, real-time performance, and availability.

[0003] In the existing technology, static scenario simulation or limited scenario testing is usually used to evaluate the data transmission performance of the Internet of Things. However, due to the single test environment and the lack of dynamic and changeable network condition simulation, it is difficult to truly reflect the actual transmission behavior of the Internet of Things system in a complex environment, resulting in insufficient performance evaluation coverage, delayed anomaly detection, and untimely system maintenance response.

[0004] In addition, traditional methods mostly rely on rule settings or static thresholds to judge anomalies, which cannot effectively adapt to the dynamic changes in network status and the interaction of multiple factors, and lack intelligent and systematic stability analysis and fault tracing capabilities.

[0005] Therefore, there is an urgent need for a method that can achieve comprehensive evaluation of the end-to-end transmission performance and stability optimization of the IoT system based on intelligent learning and deep analysis mechanisms in a changing network environment. Summary of the Invention

[0006] The Summary of the Invention introduces a series of simplified concepts that will be further described in the Detailed Description of the Invention. The Summary of the Invention is not intended to limit the key features and essential features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0007] In a first aspect, the present invention provides a method for optimizing stable transmission of IoT data, the method comprising:

[0008] Generate multiple test scenarios through a multi-network simulation model, where the multiple test scenarios include wireless access network test scenarios, core network test scenarios, and Internet test scenarios;

[0009] Input the above test scenarios into the target IoT to obtain IoT test data;

[0010] Inputting the IoT test data into a transmission performance prediction model to obtain performance prediction data, wherein the performance prediction data includes transmission delay, packet loss rate, and throughput;

[0011] Inputting the performance prediction data into a data mining and analysis model to obtain data mining information;

[0012] Generate a data stability analysis report based on the above data mining information;

[0013] Determine abnormal points based on the above data stability analysis report to form a detection and maintenance plan.

[0014] In a feasible implementation, the above-mentioned multiple test scenarios are generated by the multi-network simulation model, including:

[0015] Record the current network status, where the network status includes node distribution information, link connection information, and link parameter information;

[0016] Based on the current network state, a target action is selected from the action space, wherein the target action includes adding a node, deleting a node, modifying a link delay, adjusting a link bandwidth, introducing a link failure, and changing a node mobility model;

[0017] After executing the target action, the network state is updated based on the impact of the action, and the complexity, realism, and testability of the generated scenario are evaluated based on the preset reward function.

[0018] Using reinforcement learning algorithms to train strategies, we can select the optimal action under different network states to generate test scenarios that meet the target complexity requirements.

[0019] The multiple test scenarios finally generated are output when the termination conditions are met, wherein the termination conditions include reaching a preset number of nodes, number of links, complexity level and maximum number of generation steps.

[0020] In a feasible implementation, the above-mentioned selecting a target action from the action space includes:

[0021] At each decision moment, a random number r between 0 and 1 is generated;

[0022] When the random number r is greater than the preset exploration rate ε, the action with the largest action value in the current network state is selected;

[0023] When the random number r is less than or equal to the exploration rate ε, an action is randomly selected from the action space, where the exploration rate ε is used to control the balance between exploration and exploitation and is dynamically decayed according to the training process.

[0024] In one feasible implementation, the aforementioned reinforcement learning algorithm training strategy enables selection of optimal actions under different network states to generate test scenarios that meet target complexity requirements, including:

[0025] Sense the current network status in each decision cycle, including the number of nodes, link connectivity, and link parameter information;

[0026] Based on the current network state, a target action is selected according to a preset action selection strategy that prioritizes both exploration and exploitation.

[0027] Based on the results of the action execution, the immediate reward is evaluated according to the complexity of the above scenario, the above realism and the above testability;

[0028] The policy parameters are updated using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes one or more of Q-Learning, Deep Q Network, and Proximal Policy Optimization.

[0029] In a feasible implementation, the transmission performance prediction model includes a multimodal feature encoding layer, a feature fusion layer, a joint modeling layer, and a performance indicator prediction layer;

[0030] The multimodal feature encoding layer is used to perform feature encoding on the network test data of the Internet of Things system. The multimodal feature encoding layer includes:

[0031] Network topology encoding module, used to extract topology structure features based on node distribution information and link connection information;

[0032] Link dynamic feature encoding module, used to extract timing dynamic features based on the link's time-varying delay, packet loss rate, and bandwidth information;

[0033] Traffic statistics feature encoding module, used to extract static features based on traffic statistics characteristics of nodes and links;

[0034] The feature fusion layer is used to perform weighted fusion of the multimodal features based on the self-attention mechanism to generate a unified comprehensive feature representation;

[0035] The joint modeling layer is used to model spatial relationships using a graph convolutional network and time series using a bidirectional long short-term memory network based on the comprehensive feature representation.

[0036] The above performance indicator prediction header is used to predict the transmission delay, packet loss rate and throughput performance indicators respectively according to the joint modeling results.

[0037] In a feasible implementation, the training process of the above data mining and analysis model specifically includes:

[0038] Performing data cleaning and standardization on the standard performance prediction data to extract characteristic information, wherein the characteristic information includes characteristic information of the current value, rate of change, hysteresis characteristics, and local fluctuation characteristics of the performance indicators;

[0039] Based on the above feature information, a dynamic Bayesian network is used to learn the causal structure and construct a dynamic causal relationship diagram between performance indicators that evolves over time.

[0040] Based on the above dynamic causal relationship diagram, identify abnormal nodes and abnormal indicator changes, and trace the abnormal propagation path to infer the abnormal cause chain;

[0041] Introducing a contrastive learning mechanism, using real abnormal links and random links for training optimization;

[0042] An incremental update strategy is adopted to update the dynamic causal relationship graph and anomaly detection model in real time based on the newly added performance prediction data to obtain the above-mentioned data mining and analysis model.

[0043] In a feasible implementation, based on the above feature information, a dynamic Bayesian network is used to perform causal structure learning to construct a dynamic causal relationship graph between performance indicators that evolves over time, including:

[0044] Dividing the performance characteristic information into a plurality of time slices in chronological order, wherein each of the time slices corresponds to a time step;

[0045] In each of the above time slices, the synchronization dependency between different performance indicators in the same time step is determined based on the conditional independence test, and a synchronization dependency structure is generated;

[0046] Between the adjacent time slices, based on the characteristics of performance indicators changing over time, the temporal transfer dependency between time steps is learned to generate a transfer structure across time steps.

[0047] The above-mentioned synchronization dependency structure and the above-mentioned cross-time-step transfer structure are combined, and based on the Bayesian estimation method, the conditional probability distribution parameters between the nodes in the dynamic causal relationship graph are learned to form a dynamic causal relationship graph that describes the temporal evolution characteristics of the performance indicators.

[0048] In a second aspect, the present invention proposes an IoT data stable transmission optimization system, comprising:

[0049] A first generating unit is configured to generate multiple test scenarios through a multi-network simulation model, wherein the multiple test scenarios include a wireless access network test scenario, a core network test scenario, and an Internet test scenario;

[0050] A first acquisition unit is configured to input the aforementioned multiple test scenarios into a target IoT to acquire IoT test data;

[0051] A second acquisition unit is configured to input the IoT test data into a transmission performance prediction model to obtain performance prediction data, wherein the performance prediction data includes transmission delay, packet loss rate, and throughput;

[0052] a third acquisition unit, configured to input the performance prediction data into a data mining and analysis model to obtain data mining information;

[0053] A second generating unit is used to generate a data stability analysis report based on the above data mining information;

[0054] The determination unit is used to determine abnormal points based on the above data stability analysis report to form a detection and maintenance plan.

[0055] In a third aspect, the present invention proposes an electronic device comprising: a memory and a processor, characterized in that the processor is used to implement the steps of the method for optimizing stable transmission of Internet of Things data as described in any one of the first aspects when executing the computer program stored in the memory.

[0056] In a fourth aspect, the present invention proposes a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the method for optimizing stable transmission of Internet of Things data as described in any one of the first aspects are implemented.

[0057] In summary, the method for optimizing the stable transmission of IoT data provided by the present invention can systematically construct a test scenario with multiple levels, multiple variations, and multiple instability factors for wireless access networks, core networks, and the Internet by introducing a multi-network simulation model. Compared with traditional static single-scenario testing, it greatly improves the authenticity and diversity of the test environment, thereby significantly enhancing the ability of IoT systems to evaluate data transmission stability in different complex environments. In the performance prediction stage, the present invention uses multimodal feature coding, graph neural networks, and temporal neural networks to jointly model, comprehensively considers multi-dimensional performance indicators such as transmission delay, packet loss rate, and throughput, and realizes a comprehensive prediction of end-to-end transmission performance, overcoming the information loss problem caused by the local evaluation of a single indicator in the existing technology, and effectively improving the accuracy and comprehensiveness of the overall performance perception of the system. Furthermore, in the process of data mining and analysis, the present invention uses a dynamic Bayesian network to model the causal relationship between performance indicators that evolves over time, and combines a comparative learning mechanism to optimize and infer the cause chain of anomalies. It can accurately identify abnormal nodes and their propagation links, infer the causes of anomalies, and greatly improve the intelligence, interpretability, and positioning accuracy of anomaly detection, solving the problems of lagging anomaly detection and inability to trace causes in existing methods. At the same time, the present invention introduces an incremental update strategy, dynamically correcting the causal relationship model and anomaly detection model based on newly added performance data, ensuring that the system can adapt to changes in network status and performance evolution trends in real time. In the face of dynamic changes in network structure and frequent device access and exit, it can still maintain efficient and accurate data mining and stability analysis capabilities. Finally, the present invention generates a data stability analysis report based on data mining information, and determines anomalies based on this to form a detection and maintenance plan. It constructs an integrated closed-loop optimization mechanism from performance evaluation, anomaly detection to maintenance decision-making, effectively improving the stability, reliability and operation and maintenance efficiency of the Internet of Things system, and has broad application prospects and significant engineering practice value.

[0058] The method for optimizing stable transmission of IoT data proposed in the present invention, and other advantages, objectives, and features of the present invention will be partially reflected in the following description, and will also be understood by those skilled in the art through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present description. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0060] Figure 1 A schematic diagram of the process of optimizing a method for stable data transmission in the Internet of Things provided by an embodiment of the present invention;

[0061] Figure 2 A schematic diagram of the structure of an IoT data stable transmission optimization system provided by an embodiment of the present invention;

[0062] Figure 3 A structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments.

[0064] See also Figure 1 , which is a flow chart of a method for optimizing stable data transmission in the Internet of Things provided by an embodiment of the present invention, which may specifically include:

[0065] S110. Generate multiple test scenarios using a multi-network simulation model, where the multiple test scenarios include a wireless access network test scenario, a core network test scenario, and an Internet test scenario;

[0066] For example, a multi-network simulation model was first used to generate diverse test scenarios, covering wireless access networks, core networks, and the Internet. By simulating different network layers and transmission environments, a variety of typical application scenarios, such as access points, transit points, and public network transmission, were effectively constructed, providing a rich set of test conditions for subsequent IoT system stability assessments.

[0067] S120: Input the above-mentioned multiple test scenarios into the target Internet of Things to obtain Internet of Things test data;

[0068] Exemplarily, after generating the test scenarios, the multiple test scenarios are input into the target IoT system to be tested for simulation or actual deployment. By running IoT terminal devices under different network conditions, end-to-end data transmission information is collected and recorded to form IoT test data, providing basic data support for performance prediction and anomaly analysis.

[0069] S130: Input the IoT test data into a transmission performance prediction model to obtain performance prediction data, wherein the performance prediction data includes transmission delay, packet loss rate, and throughput;

[0070] Exemplarily, the acquired IoT test data is then input into a transmission performance prediction model. The transmission performance prediction model outputs multi-dimensional performance prediction data including transmission delay, packet loss rate and throughput through feature encoding, feature fusion and joint modeling processes, thereby judging the transmission quality performance of the IoT system under different test scenarios in advance.

[0071] S140, inputting the performance prediction data into a data mining and analysis model to obtain data mining information;

[0072] Exemplarily, the performance prediction data is then input into a data mining and analysis model, and through a series of data analysis methods such as causal structure learning, anomaly detection and anomaly cause chain reasoning, potential abnormal patterns, performance degradation trends and system stability risk points are identified to generate preliminary data mining information.

[0073] S150, generating a data stability analysis report based on the above data mining information;

[0074] For example, based on the data mining information obtained above, a data stability analysis report of the Internet of Things system is further generated. The report includes the overall stability score of the system, the identification of major abnormal nodes and links, the inference results of the causes of abnormalities, the risk level assessment, etc., to provide a basis for subsequent maintenance decisions and optimization adjustments.

[0075] S160. Determine abnormal points based on the data stability analysis report to form a detection and maintenance plan.

[0076] For example, finally, based on the abnormal nodes, abnormal links and corresponding abnormal causes marked in the data stability analysis report, the abnormal points that need maintenance are determined, and targeted detection and maintenance plans are formulated, including but not limited to node replacement, link reconstruction, parameter optimization and other operational strategies, so as to improve the stability and reliability of the Internet of Things system in a complex and changeable network environment.

[0077] In summary, the method for optimizing the stable transmission of IoT data provided by the present invention can systematically construct a test scenario with multiple levels, multiple variations, and multiple instability factors for wireless access networks, core networks, and the Internet by introducing a multi-network simulation model. Compared with traditional static single-scenario testing, it greatly improves the authenticity and diversity of the test environment, thereby significantly enhancing the ability of IoT systems to evaluate data transmission stability in different complex environments. In the performance prediction stage, the present invention uses multimodal feature coding, graph neural networks, and temporal neural networks to jointly model, comprehensively considers multi-dimensional performance indicators such as transmission delay, packet loss rate, and throughput, and realizes a comprehensive prediction of end-to-end transmission performance, overcoming the information loss problem caused by the local evaluation of a single indicator in the existing technology, and effectively improving the accuracy and comprehensiveness of the overall performance perception of the system. Furthermore, in the process of data mining and analysis, the present invention uses a dynamic Bayesian network to model the causal relationship between performance indicators that evolves over time, and combines a comparative learning mechanism to optimize and infer the cause chain of anomalies. It can accurately identify abnormal nodes and their propagation links, infer the causes of anomalies, and greatly improve the intelligence, interpretability, and positioning accuracy of anomaly detection, solving the problems of lagging anomaly detection and inability to trace causes in existing methods. At the same time, the present invention introduces an incremental update strategy, dynamically correcting the causal relationship model and anomaly detection model based on newly added performance data, ensuring that the system can adapt to changes in network status and performance evolution trends in real time. In the face of dynamic changes in network structure and frequent device access and exit, it can still maintain efficient and accurate data mining and stability analysis capabilities. Finally, the present invention generates a data stability analysis report based on data mining information, and determines anomalies based on this to form a detection and maintenance plan. It constructs an integrated closed-loop optimization mechanism from performance evaluation, anomaly detection to maintenance decision-making, effectively improving the stability, reliability and operation and maintenance efficiency of the Internet of Things system, and has broad application prospects and significant engineering practice value.

[0078] In a feasible implementation, the above-mentioned multiple test scenarios are generated by the multi-network simulation model, including:

[0079] Record the current network status, where the network status includes node distribution information, link connection information, and link parameter information;

[0080] Based on the current network state, a target action is selected from the action space, wherein the target action includes adding a node, deleting a node, modifying a link delay, adjusting a link bandwidth, introducing a link failure, and changing a node mobility model;

[0081] After executing the target action, the network state is updated based on the impact of the action, and the complexity, realism, and testability of the generated scenario are evaluated based on the preset reward function.

[0082] Using reinforcement learning algorithms to train strategies, we can select the optimal action under different network states to generate test scenarios that meet the target complexity requirements.

[0083] The multiple test scenarios finally generated are output when the termination conditions are met, wherein the termination conditions include reaching a preset number of nodes, number of links, complexity level and maximum number of generation steps.

[0084] For example, the system first records the current IoT system's network status information in real time or periodically. This network status includes at least: node distribution information (such as the number of nodes, node location coordinates, and node connection density); link connection information (such as link topology and link presence flags); and link parameter information (such as link bandwidth, link latency, link packet loss rate, and link stability indicators). By recording the current network status, an initial environmental model of the IoT system is established, providing basic data support for subsequent action decisions and test scenario evolution.

[0085] Based on the recorded current network status, the system selects a target action from a preset action space. The action space includes, but is not limited to: add node; remove node; modify link delay; adjust link bandwidth; inject link failure; and change node mobility model.

[0086] After executing the selected target action, the system immediately updates the network state based on the impact of the action. The update process can define the corresponding state transition function (StateTransitionFunction) according to the action type. The network state update can be expressed as:

[0087]

[0088] in, represents the network state at time step t; represents the target action selected at time step t; Indicates the action The new network state arrived after ; f(·) is the state transition function, which is determined by the action type and the network response mechanism.

[0089] Reward function R( , ) can be constructed as follows:

[0090]

[0091] in, To perform an action at time step t The reward value after Score the complexity of the generated scene; Score the authenticity of the generated scene; Test Score the testability of generated scenarios; is the historical mean and standard deviation of the complexity score; is the historical mean and standard deviation of the authenticity score; is the historical mean and standard deviation of the testability score; is an S-type normalization function (such as the Sigmoid function), defined as: Used to normalize the input features to intervals to improve the comparability of different scoring items; The weight coefficient corresponding to each scoring item, ,and ; is the penalty coefficient; As the indicator function, when the generated network status When the constraints are violated (such as node isolation, link disconnection, and topology break), the output is 1, otherwise the output is 0, which serves to punish illegal states.

[0092] The reward function is standardized to prevent different metrics from affecting the reward calculation balance due to different numerical scales. Sigmoid nonlinear normalization is used to make reward changes smooth and avoid oscillations during training. Constrained penalty terms are introduced to force the network evolution process to avoid unusable or invalid test scenarios.

[0093] Based on the above, reinforcement learning algorithms (such as Q-Learning, DQN, PPO, etc.) are used to train action selection strategies, enabling the system to adaptively select the optimal action under different network conditions, guiding the scenario to evolve towards the target complexity, authenticity, and testability requirements. The strategy training goal is to maximize the cumulative reward:

[0094]

[0095] Among them, π* is the optimal strategy; π is the action selection strategy; Discount factor , controls the importance of long-term rewards; T is the maximum number of steps in the generation process.

[0096] Through continuous training, reinforcement learning agents can optimize action decision paths and generate high-quality test scenarios that meet the target complexity requirements.

[0097] During the test scenario generation process, real-time detection is performed to determine whether the preset termination conditions are met. These termination conditions include, but are not limited to: the number of nodes reaching the target range; the number of links meeting the preset density; the scenario complexity level reaching the set threshold; and the maximum number of generation steps being reached.

[0098] When any termination condition is met, the scenario evolution is stopped and the final generated multiple test scenarios are output for subsequent data transmission testing and performance evaluation of the IoT system.

[0099] Through the above specific implementation, the present invention can flexibly select scenario evolution actions based on the current network status, and dynamically generate multiple test scenarios that meet the target complexity, authenticity and testability requirements under the guidance of the reinforcement learning algorithm. Compared with the traditional static configuration or rule-driven test scenario generation method,

[0100] This invention not only significantly improves the richness and adaptability of the testing environment, but also significantly enhances the intelligence of the testing process, effectively improving the comprehensiveness and accuracy of data transmission performance evaluation of IoT systems under different network conditions. During the action selection process, a reinforcement learning strategy can be used to balance exploration and optimality.

[0101] In a feasible implementation, the above-mentioned selecting a target action from the action space includes:

[0102] At each decision moment, a random number r between 0 and 1 is generated;

[0103] When the random number r is greater than the preset exploration rate ε, the action with the largest action value in the current network state is selected;

[0104] When the random number r is less than or equal to the exploration rate ε, an action is randomly selected from the action space, where the exploration rate ε is used to control the balance between exploration and exploitation and is dynamically decayed according to the training process.

[0105] For example, at each decision time t, the system generates a random number r uniformly distributed between 0 and 1 to determine the current action selection method. This random number can be generated based on a standard pseudo-random number generator, ensuring a certain degree of uncertainty is introduced into the decision process to promote exploration behavior.

[0106] Based on the comparison between the generated random number r and the current preset exploration rate ε, the following two different action selection strategies are executed:

[0107] (1) When the random number r is greater than the exploration rate ε, the system selects the action A* with the largest action value (Q value) under the current network state, which is:

[0108]

[0109] in, To indicate the current state The expected cumulative reward of executing action A under the given set is ∑A=1,A,where A is the current action space. At this time, the system tends to choose the action that is currently considered optimal, maximizing both immediate and future returns.

[0110] (2) When the random number r is less than or equal to the exploration rate ε, the system randomly selects an action A' from the action space A, that is:

[0111]

[0112] By selecting random actions, the agent is guided to explore new state transition paths, discover potentially better action sequences, and avoid falling into local optimality.

[0113] In order to achieve the natural evolution of the training process from exploration to utilization, the system sets the exploration rate ε to decay dynamically with the training process.

[0114] The exponential decay formula can be used to update the ε value:

[0115]

[0116] in, is the initial exploration rate, which is generally set to a higher value (such as 0.9); decay is the decay factor (such as 0.99), which controls the speed at which the exploration rate decreases at each step; step is the current number of training steps; It is the lower limit of the minimum exploration rate, ensuring that there is still a small amount of exploration even in the later stages.

[0117] Through the above dynamic adjustment mechanism, extensive exploration is encouraged in the early stages of training, while gradually focusing on the optimal action path in the later stages of training, thereby improving the stability and convergence of the strategy.

[0118] The action selection strategy adopted in this implementation effectively balances exploration and utilization during IoT test scenario generation. This ensures a thorough search of the action space, avoiding early regression into local optima, while also enabling focused selection of high-value actions in the later stages of training, accelerating strategy convergence and significantly improving the efficiency and quality of test scenario evolution. Furthermore, the dynamic decay mechanism of the exploration rate further enhances the system's adaptability, flexibly adjusting action selection preferences at different training stages and improving the intelligence and robustness of the test scenario generation process. This provides solid support for optimizing data stability in IoT systems operating in complex and changing network environments.

[0119] In one feasible implementation, the aforementioned reinforcement learning algorithm training strategy enables selection of optimal actions under different network states to generate test scenarios that meet target complexity requirements, including:

[0120] Sense the current network status in each decision cycle, including the number of nodes, link connectivity, and link parameter information;

[0121] Based on the current network state, a target action is selected according to a preset action selection strategy that prioritizes both exploration and exploitation.

[0122] Based on the results of the action execution, the immediate reward is evaluated according to the complexity of the above scenario, the above realism and the above testability;

[0123] The policy parameters are updated using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes one or more of Q-Learning, Deep Q Network, and Proximal Policy Optimization.

[0124] For example, within each decision cycle, the system perceives and extracts real-time network status information about the current IoT environment. This network status includes, but is not limited to, the number of nodes (e.g., the total number of online nodes and node density distribution); link connectivity (e.g., inter-node link topology and link connectivity); and link parameter information (e.g., link bandwidth, link latency, link stability indicators, and link packet loss rate). By perceiving these multi-dimensional network status characteristics, the system provides comprehensive and accurate environmental input for action selection, ensuring that action decisions are appropriately tailored to the current network situation.

[0125] After acquiring the current network state, the system selects target actions according to a preset exploration-exploitation balance strategy. Specifically, this includes: using an ε-greedy strategy or other action selection mechanisms that combine action value (such as Q-value) with random exploration; prioritizing actions with the highest current action value (Q-value) during the exploitation phase to maximize expected returns; and randomly selecting actions during the exploration phase to guide the strategy in discovering potentially optimal action sequences. The exploration rate ε dynamically decays based on the number of training steps, prioritizing exploration in the early stages and exploitation in the later stages. This action selection strategy allows the system to effectively balance the exploration of new paths with the optimal action exploitation during training.

[0126] After executing the target action, the system comprehensively evaluates the changes in the scenario in three dimensions: complexity, realism, and testability, based on the updated network status, and calculates the immediate reward value.

[0127] The reward function can be exemplified as follows:

[0128]

[0129] in, are the complexity, authenticity and testability scoring functions respectively; is the weighting coefficient of each score item, ,and ; is the penalty coefficient for illegal status; It is an indicator function that outputs a penalty when the network state violates constraints such as structural connectivity or node validity.

[0130] Complexity score function It is used to measure the complexity of the current network scenario and reflect whether the network structure is complex and diverse enough. It can be constructed based on indicators such as the number of nodes, number of links, average number of layers, and topology value of the network, such as:

[0131]

[0132] Among them, Nnodes is the current number of network nodes; Nlinks is the current number of network links; is the preset maximum number of nodes and maximum number of links (normalized benchmark); Entropy (G) is the network topology value used to measure the complexity of the network structure (for example, based on the degree distribution value); : Weighted coefficients of each indicator, ,and .

[0133] Authenticity score function Measuring the consistency of the generated scenario with the characteristics of real IoT deployments reflects the authenticity of the generated scenario. This can be based on the closeness of link parameters (latency, cloud packet rate, bandwidth, etc.) to real statistical data, such as:

[0134]

[0135] Among them, MSE (Delay) is the mean square error between the current link delay and the actual IoT link delay distribution; MSE (Loss) is the mean square error of the packet loss rate; MSE (Bandwidth) is the mean square error of the bandwidth utilization; is the weighting coefficient of each feature item, ,and .

[0136] Testability score function It is used to measure the support of the current network scenario for the test task, reflecting its testability and analysis value. It can be calculated based on indicators such as connectivity, coverage, and path diversity, such as:

[0137]

[0138] Among them, ConnRatio(G) is the network connectivity ratio (number of connected components / number of nodes), the higher the connectivity ratio, the better; PathDiversity(G) is the path diversity index between node pairs (such as the average number of reachable paths between node pairs); IsolatedNodes(G) is the number of isolated nodes; : weighting coefficient, ,and A high connectivity rate and good path diversity indicate that the network is more conducive to multi-path testing; a large number of isolated nodes reduces testability and is therefore used as a negative penalty.

[0139] The policy parameters are updated using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes one or more of Q-Learning, Deep Q Network, and Proximal Policy Optimization.

[0140] Through this implementation, the present invention is able to adopt an action selection strategy that prioritizes both exploration and exploitation based on the dynamic perception of network status. This strategy, combined with a multi-dimensional reward evaluation mechanism based on complexity, authenticity, and testability, utilizes a reinforcement learning algorithm to continuously optimize the action decision-making process, thereby intelligently generating high-quality, diverse test scenarios under different network environments. Compared to traditional rule-setting or static strategy methods, this invention significantly improves the intelligence, adaptability, and optimization of test scenario evolution, effectively enhancing the accuracy and coverage of IoT system transmission performance evaluation, and possesses extremely high practical application value and promotion prospects.

[0141] In a feasible implementation, the transmission performance prediction model includes a multimodal feature encoding layer, a feature fusion layer, a joint modeling layer, and a performance indicator prediction layer;

[0142] The multimodal feature encoding layer is used to perform feature encoding on the network test data of the Internet of Things system. The multimodal feature encoding layer includes:

[0143] Network topology encoding module, used to extract topology structure features based on node distribution information and link connection information;

[0144] Link dynamic feature encoding module, used to extract timing dynamic features based on the link's time-varying delay, packet loss rate, and bandwidth information;

[0145] Traffic statistics feature encoding module, used to extract static features based on traffic statistics characteristics of nodes and links;

[0146] The feature fusion layer is used to perform weighted fusion of the multimodal features based on the self-attention mechanism to generate a unified comprehensive feature representation;

[0147] The joint modeling layer is used to model spatial relationships using a graph convolutional network and time series using a bidirectional long short-term memory network based on the comprehensive feature representation.

[0148] The above performance indicator prediction header is used to predict the transmission delay, packet loss rate and throughput performance indicators respectively according to the joint modeling results.

[0149] Exemplarily, first, the system extracts and encodes features from the collected network test data through a multimodal feature encoding layer, which includes the following submodules:

[0150] (1) Network topology coding module: used to extract topological structure features based on node distribution information and link connection information. Using features such as node position relationship, link connectivity matrix, node degree, connectivity rate, etc.

[0151] The network structure is encoded through graph convolution (GCN) or node embedding methods (such as GraphSAGE) to obtain a vector representation of each node and the overall topology.

[0152] (2) Link dynamic feature encoding module: This module is used to extract temporal dynamic features based on the performance indicators of the link that change over time (such as link delay, packet loss rate, and bandwidth change). A bidirectional long short-term memory network (Bi-LSTM) is used to capture the temporal evolution pattern and periodic change of the link state, and obtain the link-level dynamic behavior feature representation.

[0153] (3) Traffic Statistical Feature Coding Module: This module is used to extract static features based on the traffic statistical characteristics of nodes and links (such as average traffic, peak traffic, and traffic standard deviation). A multi-layer perceptron (MLP) is used to map and enhance the extracted statistics to form a static feature vector that describes traffic behavior.

[0154] Through the feature encoding of the above three different sources, the system can comprehensively characterize the network structure, link dynamic changes and traffic characteristics, laying the foundation for subsequent feature fusion and predictive modeling.

[0155] After completing the multimodal feature extraction, the system performs weighted fusion on the multimodal features through the feature fusion layer to generate a unified comprehensive feature representation. The present invention adopts the self-attention mechanism, which is based on the importance of each modal feature to the target task (transmission performance prediction).

[0156] Dynamically assigning weights to different features enables information interaction and fusion between features, avoiding information redundancy or conflict caused by simple splicing. The integrated features after fusion not only contain network structure information, but also integrate dynamic link behavior and traffic statistics patterns, with stronger discriminability and adaptability.

[0157] After feature fusion is completed, the system enters the joint modeling stage to further explore the deep correlation between features. The joint modeling layer includes:

[0158] (1) Spatial relationship modeling: Based on the fusion feature representation, the graph convolutional network (GCN) is used to model the spatial dependency relationship between nodes and links, capturing the local neighborhood structure characteristics and long-range dependency characteristics.

[0159] (2) Time series modeling: Based on spatial modeling, we further use bidirectional long short-term memory networks (Bi-LSTM) or temporal convolutional networks (TCN) to model dynamic changes in the time dimension.

[0160] Capture trends and fluctuation patterns in performance indicators over time.

[0161] Through joint space-time modeling, the system can fully understand the complex and dynamically changing network characteristics in the Internet of Things environment.

[0162] Based on the comprehensive feature representation output by the joint modeling layer, the performance metric prediction head includes independent performance metric prediction submodules (Multi-Head Prediction), which are used to predict the following performance metrics: transmission delay (Delay), packet loss rate (Packet Loss Rate), and throughput (Throughput). Each prediction head uses an independent multi-layer perceptron (MLP) for regression prediction, enabling specialized optimization and customized modeling for different performance metrics, improving prediction accuracy and robustness.

[0163] Through the transmission performance prediction model designed in this embodiment, the present invention realizes unified modeling based on multimodal features, fully integrating network structure features, link dynamic behavior features and traffic statistics features.

[0164] By deeply exploring data correlations through joint spatial-temporal modeling, the prediction system can accurately and efficiently predict key performance indicators such as transmission delay, packet loss rate, and throughput in complex and changing IoT environments. Compared to traditional single-feature input or single-layer modeling methods, this invention significantly improves the accuracy, stability, and adaptability of performance predictions, providing strong support for data stability analysis and optimization in IoT systems, and has important engineering application value and promotion prospects.

[0165] In a feasible implementation, the training process of the above data mining and analysis model specifically includes:

[0166] Performing data cleaning and standardization on the standard performance prediction data to extract characteristic information, wherein the characteristic information includes characteristic information of the current value, rate of change, hysteresis characteristics, and local fluctuation characteristics of the performance indicators;

[0167] Based on the above feature information, a dynamic Bayesian network is used to learn the causal structure and construct a dynamic causal relationship diagram between performance indicators that evolves over time.

[0168] Based on the above dynamic causal relationship diagram, identify abnormal nodes and abnormal indicator changes, and trace the abnormal propagation path to infer the abnormal cause chain;

[0169] Introducing a contrastive learning mechanism, using real abnormal links and random links for training optimization;

[0170] An incremental update strategy is adopted to update the dynamic causal relationship graph and anomaly detection model in real time based on the newly added performance prediction data to obtain the above-mentioned data mining and analysis model.

[0171] For example, first, the system performs data preprocessing on the collected standard performance prediction data.

[0172] It mainly includes: data cleaning: eliminating invalid, abnormal, and seriously missing data records to ensure the quality of input data; data standardization: normalizing or standardizing the standard deviation of different performance indicators (such as delay, packet loss rate, and throughput) to eliminate the influence between different dimensions and improve training stability; feature extraction: based on the cleaned and standardized data, extracting the following four types of feature information: current values ​​of performance indicators (such as the delay value, packet loss rate value, and throughput value at the current moment); rate of change characteristics (such as the delay change rate, packet loss rate change rate, and calculation of first-order differences); lag characteristics (such as the historical performance indicator values ​​of the previous n time steps to capture timing dependencies); local fluctuation characteristics (such as the sliding window mean and sliding window standard deviation to measure the degree of indicator fluctuation).

[0173] Through the above processing, a set of feature vectors with unified format and consistent dimensions is formed, laying the foundation for causal structure modeling and anomaly detection.

[0174] After feature extraction is completed, the system performs causal structure learning based on the Dynamic Bayesian Network (DBN). The causal relationship graph not only reveals the static correlation between performance indicators,

[0175] It can also characterize the causal propagation characteristics that change over time.

[0176] After completing causal relationship modeling, the system performs anomaly detection and cause analysis based on the dynamic causal relationship graph. Specifically, this includes detecting unusual fluctuations in performance indicators within the causal relationship graph (e.g., exceeding the predicted range or abnormally amplified rates of change); marking abnormal nodes and tracing their possible root causes and propagation paths along the causal chain; and forming a root cause chain to clarify the order of anomaly occurrence and the scope of impact. Causal chain inference distinguishes directly abnormal nodes from indirectly affected nodes, improving the accuracy of fault location and anomaly interpretation.

[0177] To further improve the accuracy and robustness of anomaly causal chain inference, this paper introduces a contrastive learning mechanism. Specifically, this mechanism uses real anomaly links manually annotated or inferred from historical data as positive samples; randomly generates pseudo-links that do not conform to causal logic as negative samples; trains the algorithm to minimize the difference between positive and negative links in feature space; and optimizes the anomaly chain inference model for better discrimination in complex scenarios. This approach not only improves the accuracy of anomaly chain inference but also enhances the model's adaptability to changes in anomaly type and network structure.

[0178] Considering the dynamic and frequently changing nature of IoT system environments, this paper employs an incremental learning strategy. Specifically, this strategy involves: feeding newly arriving performance prediction data into the training process in real time; dynamically revising the causal graph structure and conditional probability parameters to adapt to indicator evolution trends; updating the anomaly detection model to maintain sensitivity to new anomaly patterns; and avoiding model aging and failure associated with traditional static training methods. This incremental update strategy enables the system to continuously learn online, continuously improving the timeliness and accuracy of anomaly detection and causal inference.

[0179] Through this implementation, the present invention not only implements dynamic causal relationship modeling based on multi-feature information during data stability analysis in IoT systems, but also combines anomaly detection with causal chain tracing, further introduces a comparative learning mechanism to optimize inference accuracy, and maintains the model's long-term adaptability and effectiveness through an incremental learning mechanism. Compared to traditional methods based on static features or rule-based settings, this invention significantly improves the intelligence, interpretability, and real-time nature of anomaly detection, effectively enhancing the IoT system's ability to ensure stability in complex dynamic environments, and possesses extremely high practical engineering application value and promotional potential.

[0180] In a feasible implementation, based on the above feature information, a dynamic Bayesian network is used to perform causal structure learning to construct a dynamic causal relationship graph between performance indicators that evolves over time, including:

[0181] Dividing the performance characteristic information into a plurality of time slices in chronological order, wherein each of the time slices corresponds to a time step;

[0182] In each of the above time slices, the synchronization dependency between different performance indicators in the same time step is determined based on the conditional independence test, and a synchronization dependency structure is generated;

[0183] Between the adjacent time slices, based on the characteristics of performance indicators changing over time, the temporal transfer dependency between time steps is learned to generate a transfer structure across time steps.

[0184] The above-mentioned synchronization dependency structure and the above-mentioned cross-time-step transfer structure are combined, and based on the Bayesian estimation method, the conditional probability distribution parameters between the nodes in the dynamic causal relationship graph are learned to form a dynamic causal relationship graph that describes the temporal evolution characteristics of the performance indicators.

[0185] For example, the system first divides the extracted performance characteristic information into a series of continuous time slices (Time Slices), each of which corresponds to the performance status of the IoT system at a fixed time step. The length of the time slice can be set to seconds, minutes, or hours according to application requirements; each slice contains the performance indicator feature vectors of all nodes or links at that time step, such as latency, packet loss rate, throughput, and their derived characteristics (rate of change, volatility, etc.). Time slicing allows for the structuring of complex and continuous time series data, facilitating the subsequent independent modeling of synchronization dependencies and time transfer dependencies.

[0186] Within each time slice, the system uses a conditional independence test to learn the synchronous dependencies between different performance indicators at the same time step (intra-slice dependencies). The specific steps include: selecting any two performance indicators (such as latency and packet loss rate); testing their conditional independence at the current time step, given other performance indicators as a conditional set; if the two indicators still have a significant correlation after controlling for other indicators, establishing a synchronous causal edge between them; repeating this process for all pairs of indicators, ultimately generating the synchronous dependency structure for the current time step.

[0187] Conditional independence tests can use standard statistical methods, such as the Pearson chi-square test, mutual information test, or Z-test based on the Gaussian assumption, with flexible selection based on the data distribution characteristics. Synchronous dependency learning can reveal direct interactions between performance indicators within a single time step, providing a foundation for local anomaly detection and chain inference.

[0188] After synchronous dependency learning is complete, the system further learns inter-slice dependencies between performance indicators in adjacent time slices. The specific method includes: for any performance indicator A, taking values ​​at time step t and time step t+1, establishing a transition test; analyzing the impact of the state of A at time step t on other performance indicators (such as B) at time step t+1; if A at time t has significant predictive power or causal influence on B at time t+1, establishing an inter-time-step dependency edge between the two; repeating this process for all indicator pairs, forming a temporal transfer dependency structure.

[0189] By learning temporal transfer dependencies, the system can capture the potential causal propagation chain of performance indicators evolving over time, and support the modeling and prediction of abnormal trends and indicator contagion effects.

[0190] After learning the synchronization dependency structure and the time-shift dependency structure separately, the system combines them to form a complete dynamic causal graph. Synchronization dependency edges describe the immediate causal relationships within the same time step, while time-shift edges describe the evolving causal relationships between adjacent time steps. The whole structure forms a directed acyclic graph (DBN) model.

[0191] Then, based on the existing data samples, the Bayesian estimation method is used to learn the conditional probability distribution (CPD) of each causal edge in the dynamic causal relationship graph.

[0192] Specifically, for each node, the conditional probability distribution is learned under a given parent node state. If the performance indicator is a continuous value, a Gaussian distribution can be used to fit the conditional probability. If the indicator is a discrete value, a multinomial distribution can be used for modeling. After completing parameter learning through Bayesian estimation, the dynamic causal relationship diagram not only reflects the structural dependencies between indicators but also provides a quantitative estimate of causal strength, providing solid support for subsequent anomaly detection, causal chain inference, and system behavior prediction.

[0193] Through this implementation, the present invention can dynamically construct a causal relationship diagram between performance indicators in the IoT system that evolves over time based on performance characteristic information. This not only achieves efficient modeling of synchronous causal relationships within the same time step, but also captures the evolutionary causal path across time steps. Combined with the Bayesian estimation method to accurately model the conditional probability relationship between indicators, it significantly improves the accuracy and explainability of system abnormality pattern recognition, abnormality tracing, and behavior prediction. Compared with traditional static analysis or simple time series analysis methods, the present invention can more intelligently and adaptively respond to the complex and changeable performance evolution process in the IoT environment.

[0194] It has broad application value and engineering practice significance.

[0195] like Figure 2 As shown, the present invention also provides an Internet of Things data stable transmission optimization system, including:

[0196] A first generating unit 101 is configured to generate multiple test scenarios using a multi-network simulation model, wherein the multiple test scenarios include a wireless access network test scenario, a core network test scenario, and an Internet test scenario;

[0197] A first acquisition unit 102 is configured to input the aforementioned multiple test scenarios into a target IoT to acquire IoT test data;

[0198] A second acquiring unit 103 is configured to input the IoT test data into a transmission performance prediction model to acquire performance prediction data, wherein the performance prediction data includes transmission delay, packet loss rate, and throughput;

[0199] The third acquisition unit 104 is used to input the performance prediction data into the data mining and analysis model to obtain data mining information;

[0200] The second generating unit 105 is used to generate a data stability analysis report based on the above data mining information;

[0201] The determination unit 106 is configured to determine abnormal points according to the data stability analysis report to form a detection and maintenance plan.

[0202] The above system can also perform the following steps:

[0203] In a feasible implementation, the above-mentioned multiple test scenarios are generated by the multi-network simulation model, including:

[0204] Record the current network status, where the network status includes node distribution information, link connection information, and link parameter information;

[0205] Based on the current network state, a target action is selected from the action space, wherein the target action includes adding a node, deleting a node, modifying a link delay, adjusting a link bandwidth, introducing a link failure, and changing a node mobility model;

[0206] After executing the target action, the network state is updated based on the impact of the action, and the complexity, realism, and testability of the generated scenario are evaluated based on the preset reward function.

[0207] Using reinforcement learning algorithms to train strategies, we can select the optimal action under different network states to generate test scenarios that meet the target complexity requirements.

[0208] The multiple test scenarios finally generated are output when the termination conditions are met, wherein the termination conditions include reaching a preset number of nodes, number of links, complexity level and maximum number of generation steps.

[0209] In a feasible implementation, the above-mentioned selecting a target action from the action space includes:

[0210] At each decision moment, a random number r between 0 and 1 is generated;

[0211] When the random number r is greater than the preset exploration rate ε, the action with the largest action value in the current network state is selected;

[0212] When the random number r is less than or equal to the exploration rate ε, an action is randomly selected from the action space, where the exploration rate ε is used to control the balance between exploration and exploitation and is dynamically decayed according to the training process.

[0213] In one feasible implementation, the aforementioned reinforcement learning algorithm training strategy enables selection of optimal actions under different network states to generate test scenarios that meet target complexity requirements, including:

[0214] Sense the current network status in each decision cycle, including the number of nodes, link connectivity, and link parameter information;

[0215] Based on the current network state, a target action is selected according to a preset action selection strategy that prioritizes both exploration and exploitation.

[0216] Based on the results of the action execution, the immediate reward is evaluated according to the complexity of the above scenario, the above realism and the above testability;

[0217] The policy parameters are updated using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes one or more of Q-Learning, Deep Q Network, and Proximal Policy Optimization.

[0218] In a feasible implementation, the transmission performance prediction model includes a multimodal feature encoding layer, a feature fusion layer, a joint modeling layer, and a performance indicator prediction layer;

[0219] The multimodal feature encoding layer is used to perform feature encoding on the network test data of the Internet of Things system. The multimodal feature encoding layer includes:

[0220] Network topology encoding module, used to extract topology structure features based on node distribution information and link connection information;

[0221] Link dynamic feature encoding module, used to extract timing dynamic features based on the link's time-varying delay, packet loss rate, and bandwidth information;

[0222] Traffic statistics feature encoding module, used to extract static features based on traffic statistics characteristics of nodes and links;

[0223] The feature fusion layer is used to perform weighted fusion of the multimodal features based on the self-attention mechanism to generate a unified comprehensive feature representation;

[0224] The joint modeling layer is used to model spatial relationships using a graph convolutional network and time series using a bidirectional long short-term memory network based on the comprehensive feature representation.

[0225] The above performance indicator prediction header is used to predict the transmission delay, packet loss rate and throughput performance indicators respectively according to the joint modeling results.

[0226] In a feasible implementation, the training process of the above data mining and analysis model specifically includes:

[0227] Performing data cleaning and standardization on the standard performance prediction data to extract characteristic information, wherein the characteristic information includes characteristic information of the current value, rate of change, hysteresis characteristics, and local fluctuation characteristics of the performance indicators;

[0228] Based on the above feature information, a dynamic Bayesian network is used to learn the causal structure and construct a dynamic causal relationship diagram between performance indicators that evolves over time.

[0229] Based on the above dynamic causal relationship diagram, identify abnormal nodes and abnormal indicator changes, and trace the abnormal propagation path to infer the abnormal cause chain;

[0230] Introducing a contrastive learning mechanism, using real abnormal links and random links for training optimization;

[0231] An incremental update strategy is adopted to update the dynamic causal relationship graph and anomaly detection model in real time based on the newly added performance prediction data to obtain the above-mentioned data mining and analysis model.

[0232] In a feasible implementation, based on the above feature information, a dynamic Bayesian network is used to perform causal structure learning to construct a dynamic causal relationship graph between performance indicators that evolves over time, including:

[0233] Dividing the performance characteristic information into a plurality of time slices in chronological order, wherein each of the time slices corresponds to a time step;

[0234] In each of the above time slices, the synchronization dependency between different performance indicators in the same time step is determined based on the conditional independence test, and a synchronization dependency structure is generated;

[0235] Between the adjacent time slices, based on the characteristics of performance indicators changing over time, the temporal transfer dependency between time steps is learned to generate a transfer structure across time steps.

[0236] The above-mentioned synchronization dependency structure and the above-mentioned cross-time-step transfer structure are combined, and based on the Bayesian estimation method, the conditional probability distribution parameters between the nodes in the dynamic causal relationship graph are learned to form a dynamic causal relationship graph that describes the temporal evolution characteristics of the performance indicators.

[0237] like Figure 3 As shown, the present invention also provides an electronic device 30, including a memory 310, a processor 320 and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, any step of the above-mentioned method for optimizing stable transmission of IoT data is implemented.

[0238] The present invention also provides a computer program product, comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program or computer-executable instructions from the computer-readable storage medium and executes the computer program or computer-executable instructions, causing the electronic device to perform any step of the method for optimizing stable IoT data transmission described above.

[0239] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An AI-based method for optimizing stable transmission of IoT data, characterized in that: include: Generate multiple test scenarios through a multi-network simulation model, wherein the multiple test scenarios include a wireless access network test scenario, a core network test scenario, and an Internet test scenario; Inputting the multiple test scenarios into a target IoT to obtain IoT test data; Inputting the IoT test data into a transmission performance prediction model to obtain performance prediction data, wherein the performance prediction data includes transmission delay, packet loss rate, and throughput; inputting the performance prediction data into a data mining and analysis model to obtain data mining information; generating a data stability analysis report based on the data mining information; Determine abnormal points based on the data stability analysis report to form a detection and maintenance plan; The multi-network simulation model generates a variety of test scenarios, including: Record the current network status, wherein the network status includes node distribution information, link connection information, and link parameter information; Based on the current network state, selecting a target action from an action space, wherein the target action includes adding a node, deleting a node, modifying a link delay, adjusting a link bandwidth, introducing a link failure, and changing a node mobility model; After executing the target action, the network state is updated based on the impact of the action, and the complexity, realism, and testability of the generated scenario are evaluated based on a preset reward function. Using reinforcement learning algorithms to train strategies, we can select the optimal action under different network states to generate test scenarios that meet the target complexity requirements. Outputting the multiple test scenarios finally generated when a termination condition is met, wherein the termination condition includes reaching a preset number of nodes, number of links, complexity level, and maximum number of generation steps; The selecting a target action from the action space includes: At each decision moment, a random number r between 0 and 1 is generated; When the random number r is greater than the preset exploration rate ε, the action with the largest action value in the current network state is selected; When the random number r is less than or equal to the exploration rate ε, an action is randomly selected from the action space, where the exploration rate ε is used to control the balance between exploration and exploitation and is dynamically decayed according to the training process.

2. The AI-based IoT data stable transmission optimization method according to claim 1 is characterized in that: The reinforcement learning algorithm training strategy is used to select the optimal action under different network states to generate test scenarios that meet the target complexity requirements, including: Sense the current network status in each decision cycle, including the number of nodes, link connectivity, and link parameter information; Based on the current network state, selecting a target action according to a preset action selection strategy that emphasizes both exploration and exploitation; Based on the result of the action execution, evaluating the immediate reward according to the complexity of the scenario, the realism, and the testability; The policy parameters are updated using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes one or more of Q-Learning, deep Q network, and proximal policy optimization.

3. The AI-based IoT data stable transmission optimization method according to claim 1 is characterized in that: The transmission performance prediction model includes a multimodal feature encoding layer, a feature fusion layer, a joint modeling layer, and a performance indicator prediction layer; The multimodal feature coding layer is used to perform feature coding on the network test data of the Internet of Things system. The multimodal feature coding layer includes: Network topology encoding module, used to extract topology structure features based on node distribution information and link connection information; Link dynamic feature encoding module, used to extract timing dynamic features based on the link's time-varying delay, packet loss rate, and bandwidth information; Traffic statistics feature encoding module, used to extract static features based on traffic statistics characteristics of nodes and links; The feature fusion layer is used to perform weighted fusion of the above multimodal features based on the self-attention mechanism to generate a unified comprehensive feature representation; The joint modeling layer is used to perform spatial relationship modeling using a graph convolutional network and time series modeling using a bidirectional long short-term memory network based on the comprehensive feature representation; The performance indicator prediction layer is used to predict transmission delay, packet loss rate and throughput performance indicators respectively according to the joint modeling results.

4. The AI-based IoT data stable transmission optimization method according to claim 1 is characterized in that: The training process of the data mining and analysis model specifically includes: Performing data cleaning and standardization on the standard performance prediction data to extract characteristic information, wherein the characteristic information includes characteristic information of the current value, rate of change, hysteresis characteristics, and local fluctuation characteristics of the performance indicator; Based on the feature information, a dynamic Bayesian network is used to learn the causal structure and construct a dynamic causal relationship graph between performance indicators that evolves over time. Based on the dynamic causal relationship graph, identify abnormal nodes and abnormal indicator changes, and trace the abnormal propagation path to infer the abnormal cause chain; Introducing a contrastive learning mechanism, using real abnormal links and random links for training optimization; An incremental update strategy is adopted to update the dynamic causal relationship graph and the anomaly detection model in real time based on the newly added performance prediction data to obtain the data mining and analysis model.

5. The AI-based IoT data stable transmission optimization method according to claim 4 is characterized in that: Based on the feature information, a dynamic Bayesian network is used to perform causal structure learning to construct a dynamic causal relationship graph between performance indicators that evolves over time, including: Dividing the performance characteristic information into a plurality of time slices in chronological order, wherein each time slice corresponds to a time step; In each of the time slices, determining the synchronization dependency between different performance indicators in the same time step based on a conditional independence test, and generating a synchronization dependency structure; Between adjacent time slices, based on the characteristics of performance indicators changing over time, learning the time transfer dependency between time steps and generating a transfer structure across time steps; The synchronization dependency structure and the cross-time-step transfer structure are combined, and based on the Bayesian estimation method, the conditional probability distribution parameters between the nodes in the dynamic causal relationship graph are learned to form a dynamic causal relationship graph that describes the temporal evolution characteristics of the performance indicators.

6. An AI-based IoT data stable transmission optimization system, characterized by: include: A first generating unit is configured to generate multiple test scenarios through a multi-network simulation model, wherein the multiple test scenarios include a wireless access network test scenario, a core network test scenario, and an Internet test scenario; A first acquisition unit is configured to input the plurality of test scenarios into a target Internet of Things to acquire Internet of Things test data; a second acquiring unit, configured to input the IoT test data into a transmission performance prediction model to acquire performance prediction data, wherein the performance prediction data includes transmission delay, packet loss rate, and throughput; a third acquisition unit, configured to input the performance prediction data into a data mining and analysis model to obtain data mining information; A second generating unit, configured to generate a data stability analysis report based on the data mining information; a determination unit, configured to determine abnormal points based on the data stability analysis report to form a detection and maintenance plan; The multi-network simulation model generates a variety of test scenarios, including: Record the current network status, wherein the network status includes node distribution information, link connection information, and link parameter information; Based on the current network state, selecting a target action from an action space, wherein the target action includes adding a node, deleting a node, modifying a link delay, adjusting a link bandwidth, introducing a link failure, and changing a node mobility model; After executing the target action, the network state is updated based on the impact of the action, and the complexity, realism, and testability of the generated scenario are evaluated based on a preset reward function. Using reinforcement learning algorithms to train strategies, we can select the optimal action under different network states to generate test scenarios that meet the target complexity requirements. Outputting the multiple test scenarios finally generated when a termination condition is met, wherein the termination condition includes reaching a preset number of nodes, number of links, complexity level, and maximum number of generation steps; The selecting a target action from the action space includes: At each decision moment, a random number r between 0 and 1 is generated; When the random number r is greater than the preset exploration rate ε, the action with the largest action value in the current network state is selected; When the random number r is less than or equal to the exploration rate ε, an action is randomly selected from the action space, where the exploration rate ε is used to control the balance between exploration and exploitation and is dynamically decayed according to the training process.

7. An electronic device comprising: A memory and a processor, characterized in that the processor is used to implement the steps of the AI-based IoT data stable transmission optimization method as described in any one of claims 1 to 5 when executing the computer program stored in the memory.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the AI-based IoT data stable transmission optimization method according to any one of claims 1 to 5 are implemented.

Images