Network switch maintenance decision support system based on AI

Abstract

The invention discloses an AI-based network switch maintenance decision support system, and relates to the technical field of intelligent operation and maintenance of network switches, the system obtains multiple layers of operation parameters such as port flow, packet loss rate, CPU and memory occupation, temperature, fan speed and energy consumption in real time through a data acquisition module; the data processing module fuses the time sequence and the topological features, and constructs a multi-dimensional operation parameter matrix and a weighted topological feature map; the intelligent state evaluation module calculates a load balancing index and automatically adjusts the load; establishing an AI dynamic diagnosis model, evaluating a health state based on a convolutional neural network, and generating a speed reduction and heat dissipation strategy; the abnormal trend monitoring module predicts future temperature and load by using LSTM, and triggers an adjustment strategy; and the structure health assessment module calculates a topology health coefficient, executes port self-inspection and maintenance log records, and realizes real-time monitoring, intelligent diagnosis and maintenance decision support of the switch state.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for network switches, specifically to an AI-based network switch maintenance decision support system. Background Technology

[0002] As data centers continue to expand and network architectures become more complex, the operational stability of switches, as core devices for data transmission and control, directly impacts the communication efficiency and service continuity of the entire system. Traditional switch status monitoring primarily relies on threshold judgments for single parameters such as port speed, power consumption, and CPU utilization, which struggles to reflect potential abnormal trends under the coupling of multiple parameters. Especially in high-concurrency, multi-service scenarios, there is a significant nonlinear correlation and time-series dependency between port traffic fluctuations, temperature changes, and CPU load; relying solely on fixed thresholds or empirical rules is insufficient to promptly identify abnormal signs.

[0003] Existing anomaly detection systems mostly rely on static thresholds or linear models for status assessment, which cannot effectively cope with dynamically changing operating environments and complex coupling relationships between multidimensional parameters. When network load changes abruptly or equipment overheats locally, the system often only triggers alarms after the fault occurs, resulting in delayed warnings and untimely responses. Furthermore, existing methods generally lack a comprehensive evaluation mechanism for port topology, failing to utilize topological characteristics to reflect state transmission and health differences between ports, thus limiting the accurate assessment of the overall system health status. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based network switch maintenance decision support system to solve the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based network switch maintenance decision support system, comprising:

[0006] The data acquisition module is used to comprehensively monitor the operating status of the switch in a multi-layer structure, and collect port traffic rate Rp, packet loss rate Lp, CPU utilization rate Uc, memory utilization rate Mm, internal device temperature Ts, fan speed Nf, power supply power Pe and energy consumption rate Es in real time.

[0007] The data processing module is used to fuse the time and topology features of the collected data: by aligning the sampling time of multiple nodes, normalizing parameters, and constructing feature matrices, a unified multidimensional operating parameter matrix Xt is formed; by combining port connection relationships and traffic coupling strength, an adjacency matrix and weighted edge modeling technology are used to construct a weighted topology feature map that reflects the port association characteristics of the switch.

[0008] The intelligent status assessment module is used to extract the port traffic rate Rp from the unified multidimensional operating parameter matrix Xt, and combine it with the corresponding window mean. and port traffic standard deviation The load balancing index LBI is calculated and compared with the load balancing threshold Lth to determine whether the load between switch ports is balanced. If the load is unbalanced, a load migration strategy is automatically generated.

[0009] The AI ​​dynamic diagnostic model building module is used to build and optimize the intelligent diagnostic model of the switch's operating status based on the convolutional neural network. It integrates multi-dimensional operating parameters and topological features for training to form the AI ​​dynamic diagnostic model. By extracting the multi-dimensional operating parameter matrix Xt at different time steps t in the training sample set, the health status index HSI is calculated and compared with the health status critical threshold Hth to determine whether the overall operating health status of the switch is qualified. If it is not qualified, the module generates speed reduction, heat dissipation compensation and energy saving strategies.

[0010] The abnormal trend monitoring module is used to perform time series prediction of key switch parameters using a Long Short-Term Memory (LSTM) network to obtain future temperature change predictions. CPU load growth rate prediction Combined with port traffic standard deviation The composite anomaly prediction index (CAPI) is calculated and compared with the composite anomaly threshold (Cth) to determine whether there is an abnormal trend in the predicted values ​​of key parameters in the future. If there is, strategies such as fan speed adjustment, local flow rate limiting, and process resource release are implemented.

[0011] The structural health assessment module is used to extract port node state vectors based on weighted topology feature graphs, generate a comprehensive topology state vector Vcur through graph pooling, calculate the structural health coefficient JG, and compare it with the structural health threshold Jth to determine whether the topology health status of the switch port is qualified. If it is not qualified, it automatically performs port self-test and reconnection operations, marks abnormal ports, and generates maintenance logs.

[0012] Preferably, the data acquisition module includes an operating performance acquisition unit, a heat dissipation control status acquisition unit, and an energy consumption characteristic acquisition unit;

[0013] The operational performance acquisition unit is used to deploy traffic monitoring probes and packet loss detection sensors at the port layer of enterprise networks, industrial control networks, and large data centers to monitor the data transmission and reception process of each port in real time; by installing high-precision traffic sensors and packet loss statistics modules at the switch port interfaces, the port traffic rate Rp and packet loss rate Lp are collected in real time to establish the first data group.

[0014] The control and heat dissipation status acquisition unit is used to deploy a microprocessor monitoring module and an integrated temperature and speed sensor on the control processing layer and internal heat dissipation structure of the switch. Through the embedded sampling circuit, it monitors the CPU utilization rate Uc, memory utilization rate Mm, internal temperature Ts and fan speed Nf in real time and establishes a second data group.

[0015] The energy consumption feature acquisition unit is used to install power sensors and energy consumption metering units at key nodes of the switch power module and power supply circuit to collect power Pe and energy consumption rate Es in real time, and record the power fluctuation of the equipment in combination with the sampling timestamp to establish a third data group.

[0016] Preferably, the data processing module includes a time-series feature extraction unit and a topology load feature unit;

[0017] The time-series feature extraction unit is used to preprocess the first, second, and third data groups; to perform multi-node sampling time alignment using NTP and PTP technologies; to address the sampling frequency differences, piecewise linear interpolation and spline interpolation algorithms are used for equal-time interval reconstruction and signal interpolation alignment; parameters of different dimensions are standardized to the interval [0, 1], and a feature alignment matrix is ​​constructed according to the time series index for normalization and feature alignment; the three data groups are merged into a unified multi-dimensional operating parameter matrix Xt through matrix splicing and time-series registration techniques.

[0018] The topology load feature unit is used to extract window mean, variance and trend time series features based on a unified multidimensional operating parameter matrix Xt, and then using sliding window segmentation and wavelet threshold smoothing technology. It also uses port connection status and traffic information to treat each port of the switch as a graph node, port connection relationship as an edge, and traffic as an edge weight. Adjacency matrix modeling technology is used, and the traffic coupling strength between ports is combined as a weighted edge feature to construct a weighted topology feature graph.

[0019] Preferably, the intelligent state assessment module includes a first computing unit and a first analysis unit;

[0020] The first calculation unit is used to extract the port traffic rate Rp from the unified multidimensional operating parameter matrix Xt, and combine it with the corresponding window mean. and port traffic standard deviation After dimensionless processing, the load balancing index LBI is calculated and obtained.

[0021] The first analysis unit is used to obtain a first evaluation result by comparing the load balancing index LBI with the load balancing critical threshold Lth through a preset load balancing critical threshold Lth:

[0022] When the load balancing index LBI is greater than or equal to the load balancing threshold Lth, it indicates that the load is balanced among the switch ports. No adjustment is made, and continuous monitoring is performed.

[0023] When the Load Balancing Index (LBI) is less than the Load Balancing Threshold (Lth), it indicates an imbalance in the load between switch ports, triggering the first warning instruction and generating the first strategy: automatically generating a load migration instruction to redistribute data flows between ports, achieving local adaptive balancing and restoring the overall balance of port traffic distribution; then updating the Load Balancing Index (LBI) value until it is greater than or equal to the Load Balancing Threshold (Lth).

[0024] Preferably, the AI ​​dynamic diagnostic model building module includes a model building unit, a second calculation unit, and a second analysis unit;

[0025] The model building unit is used to construct an initial AI diagnostic model using a convolutional neural network (CNN). It trains and tests the CNN initial model using a unified multidimensional operating parameter matrix Xt and a weighted topological feature map as input data. During the training phase, port layer operating features, control layer heat dissipation features, and energy consumption layer power features are loaded at the model input. Supervised learning is used for multiple rounds of iterative training to optimize the convolutional layer parameters and activation function responses. During the testing phase, cross-validation is used to verify the model's generalization ability, outputting the initial health status identification results. The trained CNN initial model serves as the basic AI intelligent diagnostic model. Simultaneously, the intermediate layer outputs of the multidimensional operating parameter matrix Xt and the weighted topological feature map in the CNN are used as feature vectors to characterize the comprehensive state characteristics of the network switch at the port layer, control layer, and energy consumption layer. Based on the acquired intermediate layer feature information, the AI ​​intelligent diagnostic basic model undergoes secondary training and optimization. Historical operating data and abnormal sample data are used as the training set, and the model weights are dynamically adjusted to form a dynamic AI diagnostic model. The trained AI dynamic diagnostic model serves as a health assessment and trend prediction tool for the operating status, enabling real-time identification and intelligent judgment of the status of key nodes in the network switch.

[0026] Preferably, the second computing unit is used to obtain the input feature function by extracting a multi-dimensional running parameter matrix Xt at different time steps t in the training sample set based on the AI ​​dynamic diagnostic model, and calculating the feature mapping result through convolutional layers and fully connected layers. The model loss function is updated using the backpropagation algorithm, and continuously optimized in each iteration t to obtain the dynamic weight parameters of the corresponding features. After dimensionless processing, the Health Status Index (HSI) is calculated and obtained.

[0027] Preferably, the second analysis unit is used to obtain a second evaluation result by comparing the Health Status Index (HSI) with the Health Status Critical Thres ...

[0028] When the Health Status Index (HSI) is greater than or equal to the critical health status threshold (Hth), it indicates that the overall health status of the switch is qualified and should be continuously monitored.

[0029] When the Health Status Index (HSI) is less than the critical health status threshold (Hth), it indicates that the overall health status of the switch is unqualified, triggering a second warning instruction and generating a second strategy: reducing the upper limit of port speed to achieve thermal balance of the load ports; performing fan speed compensation to improve local heat dissipation capacity; performing energy throttling to reduce power supply load; after execution, the HSI value is recalculated, and when the Health Status Index (HSI) recovers to ≥Hth, the strategy execution is stopped and the correction process parameters are recorded.

[0030] Preferably, the abnormal trend monitoring module includes a third calculation unit and a third analysis unit;

[0031] The third computing unit is used to perform time-series prediction of key parameters for future periods based on a unified multidimensional operating parameter matrix Xt and a Long Short-Term Memory (LSTM) network. It trains the LSTM model by inputting port flow rate Rp, CPU utilization Uc, and internal device temperature Ts to capture parameter change trends and long-term dependencies. During the training phase, supervised learning and backpropagation algorithms are used for multiple rounds of iterative optimization. In the prediction phase, the port flow rate Rp, CPU utilization Uc, and internal device temperature Ts from the unified multidimensional operating parameter matrix Xt are input into the trained LSTM model in time series, and forward propagation is used to obtain the predicted temperature changes for future periods. CPU load growth rate prediction Combined with port traffic standard deviation After dimensionless processing, the composite anomaly prediction index (CAPI) is calculated and obtained.

[0032] Preferably, the third analysis unit is used to obtain a third evaluation result by using a preset composite anomaly threshold Cth and comparing the composite anomaly prediction index CAPI with the composite anomaly threshold Cth:

[0033] When the composite anomaly prediction index CAPI < composite anomaly threshold Cth, it indicates that the predicted values ​​of key parameters are within the normal range and should be continuously monitored.

[0034] When the composite anomaly prediction index CAPI is greater than or equal to the composite anomaly threshold Cth, it indicates that there is an abnormal trend in the key parameters in the future, triggering the third early warning instruction and generating the third strategy: adjusting the cooling fan speed in advance according to the predicted temperature change to improve heat dissipation efficiency; triggering local traffic rate limiting to implement traffic rate limiting on key ports according to the predicted port load fluctuations to alleviate port load pressure; and releasing process resources to release non-critical process resources for control layer nodes with predicted CPU load growth rate to reduce load pressure.

[0035] Preferably, the structural health assessment module includes a fourth calculation unit and a fourth analysis unit;

[0036] The fourth computing unit is used to extract the state vector set of each port node of the switch based on the weighted topology feature map of the switch port, and to fuse the node vector set to generate a comprehensive topology state vector Vcur by graph pooling method. It is then compared with the reference port topology state vector Vref and the cosine similarity method is used to calculate and obtain the structural health coefficient JG.

[0037] The fourth analysis unit is used to preset the structural health threshold Jth, and compare the structural health coefficient JG with the structural health threshold Jth to obtain the fourth evaluation result, including:

[0038] When the structural health coefficient JG ≥ structural health threshold Jth, it indicates that the topology health status of the switch port is qualified and should be continuously monitored.

[0039] When the structural health coefficient JG < structural health threshold Jth, it indicates that the health status of the switch port topology is unqualified, triggering the fourth warning instruction and generating the fourth strategy: perform port self-test and reconnection operations, check abnormal ports and reconnect them, check port availability; mark unrecoverable ports as abnormal and prompt manual maintenance; and record maintenance logs.

[0040] This invention provides an AI-based network switch maintenance decision support system. It has the following beneficial effects:

[0041] (1) The AI-based network switch maintenance decision support system collects parameters such as traffic rate, packet loss rate, CPU utilization, temperature and power in real time by deploying multiple types of sensors at the port layer, control layer and energy consumption layer. After time alignment, normalization and feature matrix construction, a unified multi-dimensional operating parameter matrix Xt is formed. Combined with the port connection relationship, a weighted topology feature map is constructed to realize the collaborative expression of time sequence features and topology features, effectively supporting the accurate modeling and comprehensive diagnosis of the state of complex network switches.

[0042] (2) This AI-based network switch maintenance decision support system uses a convolutional neural network (CNN) to construct an AI dynamic diagnostic model. It performs multiple rounds of supervised training and dynamic weight correction by inputting a multi-dimensional operating parameter matrix Xt and a weighted topology feature map, and calculates the Health Status Index (HSI). Combined with the critical health status threshold Hth, it achieves intelligent assessment of the overall health status of the switch. When an anomaly is detected, the system can automatically generate speed reduction, heat dissipation compensation, and energy-saving strategies, realizing a model-driven dynamic maintenance closed loop.

[0043] (3) The AI-based network switch maintenance decision support system captures the time-series change trends of key parameters such as port traffic, CPU utilization and internal temperature through the Long Short-Term Memory (LSTM) network, predicts future temperature changes and CPU load growth rate, and calculates the Composite Anomaly Prediction Index (CAPI) by combining the standard deviation of port traffic. When CAPI exceeds the threshold Cth, the system can execute strategies such as fan speed adjustment, traffic speed limit and resource release in advance to realize the proactive protection mechanism of "early warning-intervention-recovery".

[0044] (4) This AI-based network switch maintenance decision support system extracts port node state vectors based on weighted topology feature graphs and calculates the structural health coefficient JG using graph pooling and cosine similarity. When JG is lower than the threshold Jth, the system automatically triggers port self-checks and reconnection operations, marks abnormal ports, and generates maintenance logs. This method realizes health judgment from the perspective of structural topology, improving the intelligence and autonomy of network switch fault identification and repair. Attached Figure Description

[0045] Figure 1 This is a flowchart of the AI-based network switch maintenance decision support system of the present invention. Detailed Implementation

[0046] 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.

[0047] Example 1

[0048] Please see Figure 1 This invention provides an AI-based network switch maintenance decision support system, comprising:

[0049] The data acquisition module is used to comprehensively monitor the operating status of the switch in a multi-layer structure, and collect port traffic rate Rp, packet loss rate Lp, CPU utilization rate Uc, memory utilization rate Mm, internal device temperature Ts, fan speed Nf, power supply power Pe and energy consumption rate Es in real time.

[0050] The data processing module is used to fuse the time and topology features of the collected data: by aligning the sampling time of multiple nodes, normalizing parameters, and constructing feature matrices, a unified multidimensional operating parameter matrix Xt is formed; by combining port connection relationships and traffic coupling strength, an adjacency matrix and weighted edge modeling technology are used to construct a weighted topology feature map that reflects the port association characteristics of the switch.

[0051] The intelligent status assessment module is used to extract the port traffic rate Rp from the unified multidimensional operating parameter matrix Xt, and combine it with the corresponding window mean. and port traffic standard deviation The load balancing index LBI is calculated and compared with the load balancing threshold Lth to determine whether the load between switch ports is balanced. If the load is unbalanced, a load migration strategy is automatically generated.

[0052] The AI ​​dynamic diagnostic model building module is used to build and optimize the intelligent diagnostic model of the switch's operating status based on the convolutional neural network. It integrates multi-dimensional operating parameters and topological features for training to form the AI ​​dynamic diagnostic model. By extracting the multi-dimensional operating parameter matrix Xt at different time steps t in the training sample set, the health status index HSI is calculated and compared with the health status critical threshold Hth to determine whether the overall operating health status of the switch is qualified. If it is not qualified, the module generates speed reduction, heat dissipation compensation and energy saving strategies.

[0053] The abnormal trend monitoring module is used to perform time series prediction of key switch parameters using a Long Short-Term Memory (LSTM) network to obtain future temperature change predictions. CPU load growth rate prediction Combined with port traffic standard deviation The composite anomaly prediction index (CAPI) is calculated and compared with the composite anomaly threshold (Cth) to determine whether there is an abnormal trend in the predicted values ​​of key parameters in the future. If there is, strategies such as fan speed adjustment, local flow rate limiting, and process resource release are implemented.

[0054] The structural health assessment module is used to extract port node state vectors based on weighted topology feature graphs, generate a comprehensive topology state vector Vcur through graph pooling, calculate the structural health coefficient JG, and compare it with the structural health threshold Jth to determine whether the topology health status of the switch port is qualified. If it is not qualified, it automatically performs port self-test and reconnection operations, marks abnormal ports, and generates maintenance logs.

[0055] In this embodiment, a network switch intelligent decision-making system with full-cycle monitoring, dynamic learning, and adaptive maintenance capabilities is constructed by integrating multi-layer operating parameter acquisition, topology feature modeling, AI dynamic diagnosis, anomaly trend prediction, and structural health assessment. This system not only enables multi-dimensional perception and unified analysis of port traffic, control performance, and energy consumption status, but also achieves real-time diagnosis of health status and prediction of future anomalies through joint modeling of convolutional neural networks and long short-term memory networks. When operational anomalies or structural imbalances are detected, the system can automatically generate corresponding maintenance strategies, including load migration, heat dissipation compensation, energy throttling, fan speed adjustment, and port self-test reconnection, thereby achieving autonomous diagnosis, proactive control, and intelligent repair of the switch, significantly improving equipment operational reliability and maintenance efficiency.

[0056] Example 2

[0057] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the data acquisition module includes an operating performance acquisition unit, a control and heat dissipation status acquisition unit, and an energy consumption characteristic acquisition unit;

[0058] The operational performance acquisition unit is used to deploy traffic monitoring probes and packet loss detection sensors at the port layer of enterprise networks, industrial control networks, and large data centers to monitor the data transmission and reception process of each port in real time; by installing high-precision traffic sensors and packet loss statistics modules at the switch port interfaces, the port traffic rate Rp and packet loss rate Lp are collected in real time to establish the first data group.

[0059] The control and heat dissipation status acquisition unit is used to deploy a microprocessor monitoring module and an integrated temperature and speed sensor on the control processing layer and internal heat dissipation structure of the switch. Through the embedded sampling circuit, it monitors the CPU utilization rate Uc, memory utilization rate Mm, internal temperature Ts and fan speed Nf in real time and establishes a second data group.

[0060] The energy consumption feature acquisition unit is used to install power sensors and energy consumption metering units at key nodes of the switch power module and power supply circuit to collect power Pe and energy consumption rate Es in real time, and record the power fluctuation of the equipment in combination with the sampling timestamp to establish a third data group.

[0061] In this embodiment, by collecting multi-dimensional parameters of operating performance, heat dissipation control, and energy consumption characteristics in real time, a complete data foundation is formed, enabling comprehensive and multi-level monitoring of switch port traffic, packet loss rate, CPU / memory usage, internal temperature, fan speed, and power consumption. This data acquisition method not only accurately reflects the dynamic operating status of the switch under different network environments and load conditions, but also provides reliable basic information for subsequent data processing, intelligent status assessment, and AI diagnostics, thereby improving the visualization, intelligence, and precision of switch operation management.

[0062] Example 3

[0063] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the data processing module includes a time-series feature extraction unit and a topology load feature unit;

[0064] The time-series feature extraction unit is used to preprocess the first, second, and third data groups; to perform multi-node sampling time alignment using NTP and PTP technologies; to address the sampling frequency differences, piecewise linear interpolation and spline interpolation algorithms are used for equal-time interval reconstruction and signal interpolation alignment; parameters of different dimensions are standardized to the interval [0, 1], and a feature alignment matrix is ​​constructed according to the time series index for normalization and feature alignment; the three data groups are merged into a unified multi-dimensional operating parameter matrix Xt through matrix splicing and time-series registration techniques.

[0065] The topology load feature unit is used to extract window mean, variance and trend time series features based on a unified multidimensional operating parameter matrix Xt, and then using sliding window segmentation and wavelet threshold smoothing technology. It also uses port connection status and traffic information to treat each port of the switch as a graph node, port connection relationship as an edge, and traffic as an edge weight. Adjacency matrix modeling technology is used, and the traffic coupling strength between ports is combined as a weighted edge feature to construct a weighted topology feature graph.

[0066] In this embodiment, by performing time-series alignment, interpolation reconstruction, and normalization on multi-source data, and by modeling port connection relationships and traffic coupling strength as a weighted topology feature map, a fusion expression of multi-dimensional operating parameters and topology association features is achieved. This data processing method can accurately reflect the changes in the operating status of the switch in time and structure, providing high-quality, structured input data for intelligent status assessment, AI diagnosis, and anomaly trend prediction, thereby significantly improving the system's status perception accuracy and decision reliability.

[0067] Example 4

[0068] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1Specifically, the intelligent state assessment module includes a first computing unit and a first analysis unit;

[0069] The first calculation unit is used to extract the port traffic rate Rp from the unified multidimensional operating parameter matrix Xt, and combine it with the corresponding window mean. and port traffic standard deviation After dimensionless processing, the load balancing index (LBI) is calculated using the following formula:

[0070]

[0071] The first analysis unit is used to obtain a first evaluation result by comparing the load balancing index LBI with the load balancing critical threshold Lth through a preset load balancing critical threshold Lth:

[0072] When the load balancing index LBI is greater than or equal to the load balancing threshold Lth, it indicates that the load is balanced among the switch ports. No adjustment is made, and continuous monitoring is performed.

[0073] When the Load Balancing Index (LBI) is less than the Load Balancing Threshold (Lth), it indicates an imbalance in the load between switch ports, triggering the first warning instruction and generating the first strategy: automatically generating a load migration instruction to redistribute data flows between ports, achieving local adaptive balancing and restoring the overall balance of port traffic distribution; then updating the Load Balancing Index (LBI) value until it is greater than or equal to the Load Balancing Threshold (Lth).

[0074] The load balancing threshold Lth is determined by: statistical analysis of extensive network switch port traffic monitoring data, extracting the load balancing index distribution range under normal load distribution conditions, and the imbalance index range under port congestion or local overload conditions. Combining this with network administrator experience and switch design specifications, a reasonable load balancing threshold is determined. Referencing port load management standards and service quality requirements for enterprise networks, data centers, and industrial control networks, these standards typically define the maximum permissible fluctuation range of port load. This threshold is used to accurately determine whether the switch port load is balanced, ensuring data transmission stability and throughput efficiency.

[0075] In this embodiment, the load balancing index (LBI) of port traffic is calculated and compared with a preset threshold (Lth) to achieve real-time intelligent assessment of the switch port load status. When load imbalance is detected, the system can automatically generate a load migration strategy to dynamically adjust the distribution of data flow among ports, thereby achieving local adaptive balancing, effectively improving the overall balance of port traffic distribution, and enhancing the stability and reliability of network switch operation.

[0076] Example 5

[0077] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the AI ​​dynamic diagnostic model building module includes a model building unit, a second computing unit, and a second analysis unit;

[0078] The model building unit is used to construct an initial AI diagnostic model using a convolutional neural network (CNN). It trains and tests the CNN initial model using a unified multidimensional operating parameter matrix Xt and a weighted topological feature map as input data. During the training phase, port layer operating features, control layer heat dissipation features, and energy consumption layer power features are loaded at the model input. Supervised learning is used for multiple rounds of iterative training to optimize the convolutional layer parameters and activation function responses. During the testing phase, cross-validation is used to verify the model's generalization ability, outputting the initial health status identification results. The trained CNN initial model serves as the basic AI intelligent diagnostic model. Simultaneously, the intermediate layer outputs of the multidimensional operating parameter matrix Xt and the weighted topological feature map in the CNN are used as feature vectors to characterize the comprehensive state characteristics of the network switch at the port layer, control layer, and energy consumption layer. Based on the acquired intermediate layer feature information, the AI ​​intelligent diagnostic basic model undergoes secondary training and optimization. Historical operating data and abnormal sample data are used as the training set, and the model weights are dynamically adjusted to form a dynamic AI diagnostic model. The trained AI dynamic diagnostic model serves as a health assessment and trend prediction tool for the operating status, enabling real-time identification and intelligent judgment of the status of key nodes in the network switch.

[0079] In this embodiment, an AI dynamic diagnostic model based on convolutional neural networks is constructed, integrating multi-dimensional operating parameters and port topology features to achieve intelligent health assessment and trend prediction of the switch's operating status. This module can dynamically adjust model weights, fully utilize historical data and abnormal samples, and achieve real-time identification and intelligent judgment of the status of key nodes, significantly improving the accuracy of network switch fault prediction and the timeliness of maintenance decisions.

[0080] Example 6

[0081] This embodiment is an explanation based on Embodiment 5. Please refer to it. Figure 1 Specifically, the second computing unit is used to obtain the input feature function by extracting the multi-dimensional running parameter matrix Xt at different time steps t in the training sample set based on the AI ​​dynamic diagnostic model, and calculating the feature mapping results through convolutional layers and fully connected layers. The model loss function is updated using the backpropagation algorithm, and continuously optimized in each iteration t to obtain the dynamic weight parameters of the corresponding features. After dimensionless processing, the Health Status Index (HSI) is calculated using the following formula:

[0082]

[0083] In the formula, n represents the number of feature branches participating in the fusion calculation.

[0084] In this embodiment, the Health Status Index (HSI) is calculated by the second computing unit based on an AI dynamic diagnostic model, enabling a quantitative assessment of the switch's operational status. This unit can dynamically optimize the weight parameter Wit of each feature branch, allowing the HSI to reflect the comprehensive impact of multi-dimensional operating parameters in real time. This improves the accuracy and reliability of the overall health status assessment of the switch, providing a precise basis for intelligent maintenance strategies.

[0085] Example 7

[0086] This embodiment is an explanation based on Embodiment 5. Please refer to it. Figure 1 Specifically, the second analysis unit is used to obtain a second evaluation result by comparing the Health Status Index (HSI) with the preset health status critical threshold Hth, and performing comparative analysis.

[0087] When the Health Status Index (HSI) is greater than or equal to the critical health status threshold (Hth), it indicates that the overall health status of the switch is qualified and should be continuously monitored.

[0088] When the Health Status Index (HSI) is less than the critical health status threshold (Hth), it indicates that the overall health status of the switch is unqualified, triggering a second warning instruction and generating a second strategy: reducing the upper limit of port speed to achieve thermal balance of the load ports; performing fan speed compensation to improve local heat dissipation capacity; performing energy throttling to reduce power supply load; after execution, the HSI value is recalculated, and when the Health Status Index (HSI) recovers to ≥Hth, the strategy execution is stopped and the correction process parameters are recorded.

[0089] The health status threshold Hth is obtained by statistically analyzing the distribution of the switch's health status index under normal operation and abnormal fault conditions, based on historical operational data and abnormal state samples. This distribution, combined with operational experience and equipment design requirements, determines a reasonable health status threshold. Reference is made to network equipment health assessment specifications, manufacturer maintenance guidelines, and operational security standards, which typically specify the permissible ranges for key parameters and overall health indicator thresholds. This threshold is used to accurately assess the overall operational health status of the switch, supporting fault prevention and dynamic maintenance strategy generation.

[0090] In this embodiment, the second analysis unit compares the Health Status Index (HSI) with the preset threshold Hth, which can determine the overall health status of the switch in real time. When an anomaly occurs, it can automatically trigger an early warning and execute dynamic maintenance strategies, such as port speed adjustment, fan speed compensation, and energy throttling, to ensure the thermal balance, energy consumption control, and load optimization of the switch during operation. At the same time, it realizes closed-loop feedback and recording of strategy execution, which improves the system's adaptive maintenance capability and operational reliability.

[0091] Example 8

[0092] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the abnormal trend monitoring module includes a third calculation unit and a third analysis unit;

[0093] The third computing unit is used to perform time-series prediction of key parameters for future periods based on a unified multidimensional operating parameter matrix Xt and a Long Short-Term Memory (LSTM) network. It trains the LSTM model by inputting port flow rate Rp, CPU utilization Uc, and internal device temperature Ts to capture parameter change trends and long-term dependencies. During the training phase, supervised learning and backpropagation algorithms are used for multiple rounds of iterative optimization. In the prediction phase, the port flow rate Rp, CPU utilization Uc, and internal device temperature Ts from the unified multidimensional operating parameter matrix Xt are input into the trained LSTM model in time series, and forward propagation is used to obtain the predicted temperature changes for future periods. CPU load growth rate prediction Combined with port traffic standard deviation After dimensionless processing, the Composite Anomaly Prediction Index (CAPI) is calculated and obtained, as shown in the following formula:

[0094]

[0095] In the formula, w1, w2 and w3 represent weighting coefficients.

[0096] Characterizing future temperature changes in forecasting The impact on the Composite Anomaly Index (CAPI) is the highest weighted indicator and is a key indicator that directly reflects the contribution of temperature changes in switch equipment to heat dissipation load and potential overheating risk.

[0097] : Characterizes the standard deviation of port traffic The impact on the Composite Anomaly Index (CAPI) is the second highest weight, reflecting the sensitivity of port traffic fluctuations to local congestion and load anomaly trends.

[0098] Characterizing CPU load growth rate prediction The impact on the Composite Anomaly Index (CAPI) is of medium weight, reflecting the contribution of changes in the load of control layer nodes to the overall operating pressure and anomaly trends.

[0099] By constructing a system based on future temperature change predictions CPU load growth rate prediction Combined with port traffic standard deviation The weighted composite anomaly prediction index (CAPI) can quantify the potential abnormal trends of key switch parameters in the future, providing a scientific basis for adjusting fan speed, traffic limiting, and resource release strategies in advance.

[0100] In this embodiment, the third computing unit uses LSTM to perform time series prediction of key switch parameters and calculates the Composite Anomaly Prediction Index (CAPI) by combining the port traffic standard deviation. This enables the early identification of abnormal trends in key parameters, providing a basis for predictive maintenance of the system. It also enables proactive monitoring and early warning of temperature fluctuations, CPU load changes, and port traffic anomalies, thereby improving the stability and reliability of the switch operation.

[0101] Example 9

[0102] This embodiment is an explanation based on Embodiment 8. Please refer to it. Figure 1 Specifically, the third analysis unit is used to obtain a third evaluation result by comparing the composite anomaly prediction index CAPI with the composite anomaly threshold Cth, based on a preset composite anomaly threshold Cth:

[0103] When the composite anomaly prediction index CAPI < composite anomaly threshold Cth, it indicates that the predicted values ​​of key parameters are within the normal range and should be continuously monitored.

[0104] When the composite anomaly prediction index CAPI is greater than or equal to the composite anomaly threshold Cth, it indicates that there is an abnormal trend in the key parameters in the future, triggering the third early warning instruction and generating the third strategy: adjusting the cooling fan speed in advance according to the predicted temperature change to improve heat dissipation efficiency; triggering local traffic rate limiting to implement traffic rate limiting on key ports according to the predicted port load fluctuations to alleviate port load pressure; and releasing process resources to release non-critical process resources for control layer nodes with predicted CPU load growth rate to reduce load pressure.

[0105] The composite anomaly threshold Cth is obtained by analyzing historical time-series data of key parameters of multi-layer network switches, examining the composite exponential distribution of parameters such as CPU utilization, port traffic fluctuations, and device temperature under normal fluctuations and abnormal trends. This, combined with network security strategies and performance design requirements, determines a reasonable composite anomaly judgment threshold. Reference is made to data center management specifications, switch load tolerance standards, and predictive maintenance guidelines, which typically define trigger thresholds for critical parameter anomalies. This threshold is used to identify potential future abnormal trends in critical parameters in advance, ensuring timely response of control strategies.

[0106] In this embodiment, the third analysis unit compares and analyzes the composite anomaly prediction index CAPI with the preset threshold Cth. When an abnormal trend is detected in the future key parameters, an early warning can be automatically triggered and control strategies can be executed, including fan speed adjustment, traffic rate limiting of key ports, and release of non-critical process resources. This enables proactive intervention and risk mitigation, effectively reducing the risk of switch failures caused by abnormal temperature or load, and improving the safety and reliability of system operation.

[0107] Example 10

[0108] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the structural health assessment module includes a fourth calculation unit and a fourth analysis unit;

[0109] The fourth computing unit is used to extract the state vector set of each port node of the switch based on the weighted topology feature map of the switch ports, and to fuse the node vector set to generate a comprehensive topology state vector Vcur using a graph pooling method. This vector is then compared with the reference port topology state vector Vref, and the structural health coefficient JG is calculated using the cosine similarity method, as shown in the following formula:

[0110]

[0111] In the formula, Vref represents the reference port topology state vector, which is obtained from the ideal port topology state vector provided by the manufacturer;

[0112] The fourth analysis unit is used to preset the structural health threshold Jth, and compare the structural health coefficient JG with the structural health threshold Jth to obtain the fourth evaluation result, including:

[0113] When the structural health coefficient JG ≥ structural health threshold Jth, it indicates that the topology health status of the switch port is qualified and should be continuously monitored.

[0114] When the structural health coefficient JG < structural health threshold Jth, it indicates that the health status of the switch port topology is unqualified, triggering the fourth warning instruction and generating the fourth strategy: perform port self-test and reconnection operations, check abnormal ports and reconnect them, check port availability; mark unrecoverable ports as abnormal and prompt manual maintenance; and record maintenance logs.

[0115] The structural health threshold Jth is obtained by statistically analyzing the distribution range of structural health coefficients under normal and abnormal port connection states, based on weighted feature map data of the switch port topology. This, combined with network design specifications and operational experience, determines a reasonable structural health assessment threshold. Referencing switch topology management standards, redundant port design specifications, and fault tolerance guidelines, these standards typically specify the minimum acceptable similarity or health coefficient for port topologies. This threshold is used to accurately determine the health status of the port topology, ensuring network reliability and maintainability.

[0116] In this embodiment, the structural health assessment module analyzes the weighted topology characteristics of the switch ports, calculates the structural health coefficient JG, and compares it with a preset threshold Jth to determine the health status of the port topology in real time. When a port anomaly is detected, the system can automatically perform port self-tests and reconnection operations, mark the abnormal port and prompt manual maintenance, and record maintenance logs, thereby achieving rapid location and handling of port faults and improving the overall reliability and maintenance efficiency of the switch.

[0117] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value, it is acceptable.

[0118] The above formulas are all derived from software simulation using a large amount of data and are selected to be close to the actual values. The coefficients in the formulas are set by those skilled in the art according to the actual situation. The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any equivalent substitutions or changes made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the protection scope of the present invention.

Claims

1. An AI-based network switch maintenance decision support system, characterized by, Comprise: Data acquisition module for comprehensive monitoring of the running state of the switch under the multi-layer structure, real-time acquisition of port flow rate Rp, packet loss rate Lp, CPU occupancy rate Uc, memory occupancy rate Mm, device internal temperature Ts, fan speed Nf, power supply power Pe and energy consumption rate Es; Data processing module for time and topology feature fusion of collected data: through multi-node sampling time alignment, parameter normalization and feature matrix construction, a unified multi-dimensional running parameter matrix Xt is formed; combined with port connection relationship and flow coupling strength, adjacency matrix and weighted edge modeling technology are used to construct a weighted topology feature graph reflecting the port association characteristics of the switch; An intelligent state evaluation module is used to extract the port flow rate Rp in the unified multi-dimensional running parameter matrix Xt, combine the corresponding window mean and the port flow standard deviation , calculate and obtain the load balancing index LBI, and compare and analyze the load balancing critical threshold Lth to determine whether the loads among the switch ports are balanced, and if the loads are unbalanced, automatically generate a load migration strategy. AI dynamic diagnosis model establishment module for constructing and optimizing an intelligent diagnosis model of the switch running state based on a convolutional neural network, training multi-dimensional running parameters and topology features to form an AI dynamic diagnosis model; by extracting multi-dimensional running parameter matrix Xt at different time steps t in the training sample set, calculating the health status index HSI, and comparing it with the health status critical threshold Hth, the overall running health status of the switch is judged, if not qualified, speed reduction, heat dissipation compensation and energy consumption throttling strategies are generated; An abnormal trend monitoring module is configured to use a long short-term memory network (LSTM) to perform time series prediction on the key parameters of the switch, to obtain a temperature change prediction for a future period and CPU load growth rate prediction ; combined with the port flow standard deviation , a composite abnormal prediction index (CAPI) is calculated and compared with a composite abnormal threshold (Cth) to determine whether the predicted value of the future key parameter has an abnormal trend, and if so, the strategies of fan speed regulation, local flow speed limiting and process resource release are executed. The structure health assessment module is used for extracting port node state vector based on the weighted topology feature graph, generating a comprehensive topology state vector Vcur through graph pooling, calculating the structure health coefficient JG, and comparing it with the structure health threshold Jth to judge whether the switch port topology structure health status is qualified, if not qualified, automatically execute port self-checking and reconnection operation, mark abnormal port and generate maintenance log. 2.The AI-based network switch maintenance decision support system according to claim 1, wherein, The data acquisition module comprises a running performance acquisition unit, a control heat dissipation state acquisition unit and an energy consumption feature acquisition unit; The running performance acquisition unit is used for laying flow monitoring probes and packet loss detection sensors at the port layer of enterprise network, industrial control network and large data center to monitor the data transmission process of each port in real time; by installing high-precision flow sensors and packet loss statistical modules at the switch port interface, the port flow rate Rp and packet loss rate Lp are collected in real time to establish the first data set; The control heat dissipation state acquisition unit is used for laying microprocessor monitoring modules and temperature and speed integrated sensors on the control processing layer and internal heat dissipation structure of the switch, and real-time monitoring of CPU occupancy rate Uc, memory occupancy rate Mm, device internal temperature Ts and fan speed Nf is realized through embedded sampling circuit to establish the second data set; The energy consumption feature acquisition unit is used for installing power sensors and energy consumption measurement units at the key nodes of the switch power module and power supply circuit to collect power supply power Pe and energy consumption rate Es in real time, and record the device power fluctuation combined with the sampling time stamp to establish the third data set. 3.The AI-based network switch maintenance decision support system according to claim 1, wherein, The data processing module comprises a time sequence feature extraction unit and a topology load feature unit; The time sequence feature extraction unit is used for data preprocessing of the first data set, the second data set and the third data set; NTP and PTP technology is used for multi-node sampling time alignment; For the difference of sampling frequency, piecewise linear interpolation and spline interpolation algorithm are used for equal time interval reconstruction, signal interpolation alignment; Different dimensional parameters are standardized to the interval [0, 1], and a feature alignment matrix is constructed according to the time sequence index, and normalization and feature alignment are performed; Through matrix splicing and time sequence registration technology, three types of data groups are fused to form a unified multi-dimensional running parameter matrix Xt; The topology load feature unit is used to extract window mean, variance and trend time sequence features based on the unified multi-dimensional running parameter matrix Xt, and the switch ports are used as graph nodes, the port connection relationship is used as edge, and the flow is used as edge weight, the adjacency matrix modeling technology is used, and the flow coupling strength between ports is used as the weighted edge feature, and the weighted topology feature graph is constructed. 4.The AI-based network switch maintenance decision support system according to claim 1, wherein, The intelligent state evaluation module includes a first calculation unit and a first analysis unit; The first calculation unit is configured to extract a port flow rate Rp in a unified multi-dimensional operation parameter matrix Xt, combine a corresponding window mean and a port flow standard deviation After dimensionless processing, a load balancing index LBI is calculated and obtained. The first analysis unit is used to compare and analyze the load balancing index LBI and the load balancing critical threshold Lth through the preset load balancing critical threshold Lth, and obtain the first evaluation result, including: When the load balancing index LBI is greater than or equal to the load balancing critical threshold Lth, it indicates that the load balancing between the switch ports is balanced, and no adjustment is made, and continuous monitoring is performed; When the load balancing index LBI is less than the load balancing critical threshold Lth, it indicates that the load balancing between the switch ports is unbalanced, and a first warning instruction is triggered to generate a first strategy: automatically generating a load migration instruction to redistribute data flow between ports to locally adaptively balance and restore overall balance of port flow distribution; then update the load balancing index LBI value until it is greater than or equal to the load balancing critical threshold Lth. 5.The AI-based network switch maintenance decision support system according to claim 1, wherein, The AI dynamic diagnosis model establishment module includes a model establishment unit, a second calculation unit and a second analysis unit; The model establishment unit is used to construct an AI diagnosis initial model using a convolutional neural network, and to train and test the convolutional neural network initial model with the unified multi-dimensional running parameter matrix Xt and the weighted topology feature graph as input data; The training stage runs features of the port layer, heat dissipation features of the control layer and power features of the energy consumption layer by loading the port layer at the input end of the model, adopts a supervised learning mode to perform multi-round iterative training, and optimizes the convolution layer parameters and the activation function response; the test stage verifies the generalization ability of the model by the cross-validation method, outputs the initial health state recognition result, and takes the trained convolutional neural network initial model as the AI intelligent diagnosis basic model; meanwhile, the multi-dimensional running parameter matrix Xt and the weighted topological feature map in the intermediate layer of the convolutional neural network are output as feature vectors, which are used to represent the comprehensive state features of the network switch at the port layer, the control layer and the energy consumption layer; based on the obtained intermediate layer feature information, the AI intelligent diagnosis basic model is trained and optimized twice, the historical running data and the abnormal sample data are used as the training set to dynamically correct the model weight, and the AI dynamic diagnosis model is formed; the trained AI dynamic diagnosis model is used as the health evaluation and trend prediction of the running state, and the real-time recognition and intelligent judgment of the key node state of the network switch.

6. The AI-based network switch maintenance decision support system according to claim 5, wherein, The second computing unit is configured to extract a multi-dimensional operation parameter matrix Xt at different time steps t in a training sample set based on an AI dynamic diagnosis model, calculate a feature mapping result through a convolution layer and a full connection layer, and obtain an input feature function The model loss function is updated by a gradient through a back propagation algorithm, and the dynamic weight parameters of the corresponding features are obtained by continuously optimizing in each iteration round t After dimensionless processing, a health state index HSI is calculated and obtained. 7.The AI-based network switch maintenance decision support system according to claim 5, wherein, The second analysis unit is used to compare and analyze the health state index HSI and the health state critical threshold Hth by the preset health state critical threshold Hth, and obtain a second evaluation result, including: When the health state index HSI is greater than or equal to the health state critical threshold Hth, it indicates that the overall operation health state of the switch is qualified, and continuous monitoring is performed; When the health state index HSI is less than the health state critical threshold Hth, it indicates that the overall operation health state of the switch is unqualified, a second warning instruction is triggered, and a second strategy is generated: reducing the upper limit of the port rate to balance the load of the port; compensating the fan speed to improve the local heat dissipation capacity; throttling the energy consumption to reduce the power load; after execution, the HSI value is recalculated, and when the health state index HSI returns to Hth, the strategy execution is stopped and the correction process parameters are recorded. 8.The AI-based network switch maintenance decision support system of claim 1, wherein, The abnormal trend monitoring module includes a third calculation unit and a third analysis unit; The third calculation unit is used to perform time series prediction on the key parameters of the future period based on the unified multi-dimensional running parameter matrix Xt by using the long short-term memory network LSTM, train the LSTM model by inputting the port flow rate Rp, the CPU occupancy rate Uc and the device internal temperature Ts, and capture the parameter change trend and long-term dependence; In the training stage, the supervised learning and back propagation algorithm are used for multi-round iteration optimization; in the prediction stage, the port flow rate Rp, CPU occupancy rate Uc and device internal temperature Ts in the unified multi-dimensional running parameter matrix Xt are input into the trained LSTM model in time sequence, the model forward propagation calculation is adopted to obtain the temperature change prediction of the future period and CPU load growth rate prediction ; combined with the port flow standard deviation , the dimensionless treatment is carried out to obtain the composite abnormal prediction index CAPI. 9.The AI-based network switch maintenance decision support system according to claim 8, wherein, The third analysis unit is used to compare and analyze the composite abnormal prediction index CAPI and the composite abnormal threshold Cth by the preset composite abnormal threshold Cth, and obtain a third evaluation result, including: When the composite abnormal prediction index CAPI is less than the composite abnormal threshold Cth, it indicates that the predicted value of the future key parameter is in the normal range, and continuous monitoring is performed; When the composite anomaly prediction index CAPI is greater than or equal to the composite anomaly threshold Cth, it indicates that there is an abnormal trend in the future key parameters, a third early warning instruction is triggered, and a third strategy is generated: the cooling fan speed is adjusted in advance, the fan speed is adjusted according to the predicted temperature change, and the heat dissipation efficiency is improved; local flow speed limiting is triggered, flow speed limiting is implemented on the key port according to the predicted port load fluctuation, and the port load pressure is relieved; process resource release is performed, and non-key process resources are released for the control layer nodes with predicted CPU load growth rate, so as to reduce the load pressure. 10.The AI-based network switch maintenance decision support system according to claim 1, wherein, The structure health assessment module comprises a fourth computing unit and a fourth analysis unit. The fourth computing unit is configured to extract a state vector set of each port node of the switch based on a weighted topology feature map of the switch port, fuse the node vector set to generate a comprehensive topology state vector Vcur through a graph pooling method, and compare the comprehensive topology state vector Vcur with a reference port topology state vector Vref, calculate and obtain a structure health coefficient JG by using a cosine similarity method. The fourth analysis unit is configured to preset a structure health threshold Jth, compare the structure health coefficient JG with the structure health threshold Jth, and obtain a fourth assessment result, which comprises: When the structure health coefficient JG is greater than or equal to the structure health threshold Jth, it indicates that the health state of the switch port topology is qualified, and the monitoring is continued. When the structure health coefficient JG is less than the structure health threshold Jth, it indicates that the health state of the switch port topology is unqualified, a fourth early warning instruction is triggered, and a fourth strategy is generated: a port self-checking and reconnection operation is started, the abnormal port is checked and reconnected, and the availability of the port is checked; the port that cannot be recovered is marked as abnormal, and manual maintenance is prompted; and a maintenance log is recorded.

Images