A method and system for visualizing server component operation and maintenance decisions based on BMCs

Abstract

The application discloses a kind of based on BMC's server component operation and maintenance decision visualization method and system, it is related to server intelligent operation and maintenance technical field.The method includes: through BMC out-of-band channel acquisition and standardization processing server all component operating state time series data, obtain multidimensional state feature;Multi-dimensional state feature is input topological learning network and executes online unsupervised learning, generates and dynamically updates inter-component dependent correlation topology;Based on dynamic dependent correlation topology, the abnormal signal of component is executed multi-step time series propagation reasoning, generates abnormal propagation spatio-temporal evolution prediction result;Accordingly generate operation and maintenance decision guide visualization content and corresponding disposal guide information;Collect operation execution effect data and update model parameters reversely, complete closed loop iteration.The application can realize the dynamic modeling of component real dependent relationship and abnormal propagation spatio-temporal prediction, improve abnormal analysis accuracy and fault disposal efficiency, guarantee server full life cycle operation and maintenance performance stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent server operation and maintenance technology, specifically to a visualization method and system for server component operation and maintenance decisions based on BMC. Background Technology

[0002] With the rapid development of the cloud computing and big data industries, the scale of data center servers continues to expand, and the complexity of operation and maintenance has increased significantly. As the out-of-band management core of the server, the Baseboard Management Controller (BMC) can independently collect real-time time-series data on the operating status of all components such as CPU, memory, hard disk, power supply, and fans, and has become the core data source of modern server operation and maintenance systems.

[0003] Currently, BMC-based server operation and maintenance systems generally use pre-defined static dependency topologies as the basic architecture for anomaly analysis and root cause localization. Under this architecture, the dependencies between components are usually pre-configured by the manufacturer according to the hardware design drawings when the server leaves the factory, and remain fixed throughout the server's entire service life.

[0004] However, this technical architecture based on static dependency topology has inherent technical limitations: the actual coupling dependencies between server components are not static, but dynamically evolve as the server's operating state changes. Throughout the server's lifecycle, factors such as hardware aging, decreased heat dissipation efficiency due to dust accumulation, changes in electrical coupling caused by load fluctuations, component replacement, and firmware upgrades will all cause the actual dependencies between components to continuously deviate from the static topology at the time of manufacture.

[0005] Because existing technologies cannot automatically learn and dynamically update changes in these real dependencies, anomaly analysis based on static topology cannot accurately reflect the actual operating status of the server, resulting in a large number of false alarms and missed alarms. The accuracy of root cause location continues to decline with the service life of the server. At the same time, it is impossible to predict the anomaly propagation trend based on real dependencies, and can only issue single-point alarms after a failure occurs. Operations personnel can only see scattered status data and cannot intuitively understand the scope of impact and development trend of the failure, resulting in low efficiency in failure handling. In addition, due to the lack of an effective feedback mechanism, the system cannot correct topology deviations based on actual operation and maintenance results, causing operation and maintenance performance to continue to deteriorate with the service life of the server.

[0006] Therefore, there is an urgent need in this field for a server component operation and maintenance decision visualization method and system that can automatically learn and dynamically update the real dependencies between server components based on the full time series data of BMC, realize the prediction and decision guidance visualization of the spatiotemporal evolution of anomaly propagation, and have closed-loop self-iterative capability. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a visualization method for server component operation and maintenance decisions based on BMC, comprising the following steps: S1: Collect the operating status time-series data of all server components through the BMC out-of-band channel, perform standardization processing on the operating status time-series data, and obtain the multi-dimensional status features corresponding to each component; S2: Input the multidimensional state features into the topology learning network, perform online unsupervised learning, and generate and dynamically update the dependency association topology between server components; S3: Based on the dynamic dependency association topology, perform multi-step temporal propagation inference on the component abnormal signal to generate the spatiotemporal evolution prediction result of abnormal propagation; S4: Based on the spatiotemporal evolution prediction results, generate operation and maintenance decision guidance visualization content and corresponding handling guidance information; S5: Collect the execution effect data of the operation corresponding to the disposal guidance information, and update the model parameters of the topology learning network and the temporal propagation inference based on the execution effect data to complete the closed-loop iteration of the model.

[0008] Preferably, the step of obtaining the multidimensional state features corresponding to each component in step S1 includes: S11: Calculate the physical coupling-derived characteristics between components within the same chassis for the standardized runtime time-series data. S12: The physical coupling-derived features are added to the basic state features of the corresponding components to obtain the multi-dimensional state features of each component for topology learning.

[0009] Preferably, the step of generating and dynamically updating the dependency association topology between server components in step S2 includes: S21: Initialize a minimum-scale topology learning network, map the multidimensional state features to the topology learning network, and perform unsupervised learning; S22: When the cumulative quantization error of the topology learning network exceeds a preset threshold, perform incremental network growth at the position corresponding to the largest error to obtain an updated topology learning network. S23: Map the multidimensional state features to the updated topology learning network, perform unsupervised learning, and generate a dynamic dependency association topology between server components based on the learned mapping results.

[0010] Preferably, the step of generating the dynamic dependency association topology between server components in step S23 includes: S231: Statistically analyze the mapping positional relationship of the multidimensional state features of different components in the updated topology learning network within multiple consecutive time windows; S232: Calculate the dependency strength weights between components based on the co-occurrence characteristics of mapping positions within a continuous time window; S233: Construct and update a dynamic dependency association topology online using components as nodes and dependency strength weights as weighted edges.

[0011] Preferably, the step S3, which involves performing multi-step temporal propagation inference on the component anomaly signal to generate a spatiotemporal evolution prediction result for the anomaly propagation, includes: S31: Extract the component anomaly degree from the multidimensional state features, use the component anomaly degree as a node feature, and combine it with the dynamic dependency association topology to construct graph structure data for anomaly propagation reasoning; S32: Based on the physical location distance of components, perform modulation processing on the edge weights in the graph structure data. The closer the physical location of the components, the higher the enhancement coefficient is assigned to the edge weights, thus obtaining the modulated graph structure data. S33: Based on the modulated graph structure data, perform multi-step temporal propagation inference to generate spatiotemporal evolution prediction results for anomalous propagation.

[0012] Preferably, after step S33, the following step is further included: S34: Perform a confidence assessment on the generated abnormal propagation path to obtain the confidence assessment result; S35: When the highest confidence level in the confidence assessment result is lower than the preset threshold, the adaptive update of the dynamic dependency association topology in step S2 is triggered, and the updated dynamic dependency association topology is reused for multi-step temporal propagation inference in step S3.

[0013] Preferably, the step of generating visualization content to guide operation and maintenance decisions in step S4 includes: S41: Map the dynamic dependency association topology to a three-dimensional virtual operation and maintenance space, and use a force-directed layout algorithm to calculate the spatial position of each component node; S42: Generate a three-dimensional basic view of operation and maintenance by encoding the dependency strength between components through node spatial distance, encoding the risk of abnormal propagation of components through node color temperature, and encoding the urgency of component operation and maintenance through node pulsation frequency.

[0014] Preferably, the step of generating the operation and maintenance decision guidance visualization content and corresponding handling guidance information in step S4 further includes: S43: Based on the spatiotemporal evolution prediction results of the anomaly propagation, the temporal process of anomaly propagation is rendered in the three-dimensional operation and maintenance basic view using dynamic particle flow to obtain a three-dimensional operation and maintenance view with anomaly propagation path. S44: Generate an operation and maintenance sequence that matches the anomaly propagation trend, predict the component status after the operation is executed based on the operation and maintenance sequence, overlay and render the status prediction effect in the three-dimensional operation and maintenance view with the anomaly propagation path, and generate operation and maintenance decision guidance visualization content and corresponding handling guidance information.

[0015] Preferably, the step S5, which involves updating the model parameters of the topology learning network and the temporal propagation inference based on the execution effect data, includes: S51: Based on the component status data collected after the operation and maintenance operation is executed, evaluate the actual recovery effect of the operation and obtain the recovery effect evaluation result; S52: Integrate the full data of this abnormal case with the recovery effect evaluation results to construct a training sample, and update the mapping weights of the topology learning network and the edge weight parameters of the temporal propagation inference based on the training sample.

[0016] Furthermore, to achieve the above objectives, the present invention also provides a BMC-based server component operation and maintenance decision visualization system, including a memory, a processor, and a BMC-based server component operation and maintenance decision visualization program stored on the memory and executable on the processor, wherein the program is configured to implement the steps of the BMC-based server component operation and maintenance decision visualization method.

[0017] The beneficial effects of this invention are: 1. Achieving dynamic and accurate modeling of component dependencies. By integrating three types of physical coupling-derived features—temperature gradient, airflow transfer coefficient, and voltage fluctuation transfer coefficient—and employing an incremental unsupervised learning network adapted to the low-computing-power environment of BMC, a dependency association topology with components as nodes and actual coupling strength as weighted edges is generated and dynamically updated. This overcomes the shortcomings of existing static dependency topologies that cannot adapt to coupling drift caused by factors such as hardware aging, changes in heat dissipation efficiency, and component replacement, reducing spurious dependencies and improving the accuracy of anomaly analysis and root cause localization.

[0018] 2. Achieve spatiotemporal joint prediction of anomaly propagation. Based on dynamic dependency-related topology, and combined with the physical location of components, edge weights are modulated according to physical laws. Multi-step temporal propagation inference is performed to generate anomaly propagation spatiotemporal evolution prediction results that include spatial paths, time nodes, impact intensity, and range. This overcomes the shortcomings of existing technologies that can only issue single-point alarms after a fault occurs, extending the operation and maintenance mode from passive handling to proactive prediction, reserving a window period for operation and maintenance, and helping to reduce the risk of fault escalation.

[0019] 3. Construct a decision-guided visualization and end-to-end closed-loop self-iterative system. Core decision dimensions such as dependency strength, anomaly risk, and operational urgency are encoded as 3D spatial visual variables. Combined with dynamic particle flow rendering and operational operation simulation, decision-guided visualization content is generated. Simultaneously, operational operation execution effect data is collected, and the parameters of the topology learning and propagation inference model are updated in reverse, constructing an end-to-end closed-loop self-iterative mechanism. This overcomes the shortcomings of existing operational systems that can only display scattered status data and lack effective feedback mechanisms, reducing the cognitive load on operational personnel, improving fault handling efficiency, and ensuring the continuous stability of operational performance throughout the server's entire lifecycle. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the steps of the server component operation and maintenance decision visualization method based on BMC as described in an embodiment of the present invention; Figure 2 This is a structural diagram of the server component operation and maintenance decision visualization system based on BMC as described in an embodiment of the present invention.

[0021] Figure Labels 10. Memory; 20. Processor. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described content is only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] The features and effects of the present invention will be further described in detail below with reference to embodiments.

[0024] like Figure 1 As shown, this invention provides a visualization method for server component operation and maintenance decisions based on BMC, including the following steps: S1: Collect the operating status time-series data of all server components through the BMC out-of-band channel, perform standardization processing on the operating status time-series data, and obtain the multi-dimensional status features corresponding to each component; S2: Input the multidimensional state features into the topology learning network, perform online unsupervised learning, and generate and dynamically update the dependency association topology between server components; S3: Based on the dynamic dependency association topology, perform multi-step temporal propagation inference on the component abnormal signal to generate the spatiotemporal evolution prediction result of abnormal propagation; S4: Based on the spatiotemporal evolution prediction results, generate operation and maintenance decision guidance visualization content and corresponding handling guidance information; S5: Collect the execution effect data of the operation corresponding to the disposal guidance information, and update the model parameters of the topology learning network and the temporal propagation inference based on the execution effect data to complete the closed-loop iteration of the model.

[0025] In this embodiment, the dynamic dependency association topology refers to a dynamic graph structure with server hardware components as nodes and the actual coupling dependency strength between components as weighted edges, which can be updated in real time with changes in server operating status; the spatiotemporal evolution prediction result refers to multi-dimensional prediction data including the spatial path, time node, influence intensity, and influence range of anomaly propagation. This method can run entirely in the BMC embedded environment, or the data collected by the BMC can be uploaded to the cluster management platform for execution, and the core logic of the two deployment methods is the same; step S1 runs continuously without interruption, and the sampling period of the BMC's operating status data can be adjusted within the range of 1-10 seconds according to computing resources, with a default value of 5 seconds; steps S2 to S5 perform background updates at a preset period when there is no anomaly trigger, and the preset period can be adjusted within the range of 1-24 hours according to the BMC's computing resources, with a default value of 6 hours, and is executed immediately when an anomaly triggers. When running on the cluster management platform, the sampling period can be appropriately reduced to 1 second, the initial size of the topology learning network can be increased to 5×5, and the continuous time window can be extended to 10-20 sampling periods to obtain more accurate topology modeling results.

[0026] The statistical period for standardization uses a sliding time window of 100 sampling periods, with a sliding step of 1 sampling period. This means that the statistical value within the current window is recalculated every time a new sample is collected. The core formula for standardization is: In the formula: This represents the k-th dimension of the original running state time series data at time t; , Let be the mean and standard deviation of the k-th dimension data at time t within the sliding window s; To prevent division by zero, when the calculated standard deviation is less than 10... 6 At that time, take 10 6 As a denominator, this ensures that the standardized results can adapt to the slow drift of server operating status in real time.

[0027] This embodiment overcomes the inherent shortcomings of existing static topology, post-event alarm, and data-display-based operations and maintenance (O&M) by constructing a full-process intelligent O&M pipeline encompassing "data acquisition - dynamic topology modeling - anomaly propagation reasoning - decision visualization - closed-loop feedback." It automatically learns the true dynamic dependencies of components using full-volume BMC time-series data, solving the problem of root cause location accuracy decaying over service time. Through dynamic topology, it achieves spatiotemporal joint prediction of anomaly propagation, upgrading the O&M mode from passive handling to proactive prediction. It elevates visualization from data presentation to decision guidance, bridging the cognitive gap between fragmented data and O&M decision-making. Through full-link closed-loop feedback, it achieves automated accumulation of O&M experience and self-iteration of model capabilities, ensuring continuous and stable O&M performance throughout the server's entire lifecycle.

[0028] In one feasible implementation, for a single-server local BMC deployment scenario, the step of obtaining the multi-dimensional state characteristics corresponding to each component in step S1 includes: S11: Calculate the physical coupling-derived characteristics between components within the same chassis for the standardized runtime timing data; S12: The physical coupling-derived features are added to the basic state features of the corresponding components to obtain the multi-dimensional state features of each component for topology learning.

[0029] In this embodiment, the basic state characteristics refer to the standardized BMC raw acquisition data, including but not limited to: CPU temperature, memory temperature, hard disk temperature, power supply temperature, motherboard chipset temperature, CPU voltage, memory voltage, power supply output voltage, fan speed, hard disk read / write speed, CPU utilization, and memory utilization.

[0030] Physical coupling-derived characteristics refer to features that reflect the strength of direct interactions between components, calculated based on the physical laws of server hardware. The specific calculation method is as follows: Temperature gradient: In the formula: , The real-time temperatures of components i and j. The distance between the physical centers of the two components is the three-dimensional Euclidean distance.

[0031] Single fan airflow transfer coefficient: In the formula: Let n be the real-time rotational speed of the nth fan; Let V be the combined equivalent projected area of ​​components i and j along the normal direction of the nth fan outlet, which can be calculated based on the server's factory hardware layout diagram; V is the total internal volume of the chassis. When the components are completely obscured... Set to 0; when partially obstructed, calculate based on the proportion of the unobstructed projected area; components not in the direct airflow path. Set to 0.

[0032] Total air volume transfer coefficient: In the formula, N is the total number of fans in the chassis; Let be the single-fan airflow transfer coefficient of the nth fan to components i and j.

[0033] Voltage fluctuation transfer factor: In the formula: Let be the Pearson correlation coefficient of the voltage signals of components i and j over the most recent 10 sampling periods. This coefficient measures the degree of linear correlation between two variables, ranging from -1 to 1. A value closer to 1 indicates a stronger correlation. The standard formula for its calculation is: Where E is the expected value calculated based on the sampling interval. , These are the average values ​​of the voltage signals of components i and j, respectively. , These represent the standard deviations of the corresponding voltage signals.

[0034] The multi-voltage signal component takes the maximum absolute value of the correlation coefficients of all voltage signals as its value. When the sampling periods are not synchronized, the correlation coefficient is calculated after resampling and aligning the time axis using linear interpolation.

[0035] The multidimensional state features of component i are its own basic state features, which are then combined with the physical coupling derivative features of this component and all other components in the chassis. The feature dimensions of all components are unified as "basic feature dimension + (total number of components in the chassis - 1) × 3" to ensure consistent input dimensions.

[0036] This embodiment injects prior constraints of hardware physics laws into subsequent topology learning by specifically designing three types of physically coupled derived features: temperature gradient, airflow transfer coefficient, and voltage fluctuation transfer coefficient. This fundamentally avoids the generation of false dependencies that do not conform to the actual hardware connection, heat dissipation, and power supply logic by unsupervised learning. By using unified feature dimension splicing rules, it ensures that servers with different numbers of components and different hardware configurations can be adapted to this method without additional adjustments to the model input structure.

[0037] In one feasible implementation, for a single-server local BMC deployment scenario, the step S2 of generating and dynamically updating the dependency association topology between server components includes: S21: Initialize a minimum-scale topology learning network, map the multidimensional state features to the topology learning network, and perform unsupervised learning; S22: When the cumulative quantization error of the topology learning network exceeds a preset threshold, perform incremental network growth at the position corresponding to the largest error to obtain an updated topology learning network. S23: Map the multidimensional state features to the updated topology learning network, perform unsupervised learning, and generate a dynamic dependency association topology between server components based on the learned mapping results.

[0038] The step of generating the dynamic dependency association topology between server components as described in step S23 includes: S231: Statistically analyze the mapping positional relationship of the multidimensional state features of different components in the updated topology learning network within multiple consecutive time windows; S232: Calculate the dependency strength weights between components based on the co-occurrence characteristics of mapping positions within a continuous time window; S233: Construct and update a dynamic dependency association topology online using components as nodes and dependency strength weights as weighted edges.

[0039] In this embodiment, the topology learning network is an incremental growing self-organizing map network (GSOM) designed to adapt to the low computing power environment of BMC. During initialization, a minimum scale neuron grid of 2×2 or 3×3 is used, and the initial weight vector of each neuron is the average value of the first 100 input samples.

[0040] To avoid model oscillations and wasted computing power caused by a full topology reconstruction, the core growth rule is as follows: when the accumulated quantization error of the model exceeds a preset threshold, the neuron with the largest error is found, the Euclidean distance of the weight vectors of this neuron and all its neighboring neurons (Manhattan distance is 1) is calculated, the neighboring neuron with the largest Euclidean distance is selected, and a new neuron is inserted between these two neurons instead of a full topology reconstruction; the initial weight of the newly inserted neuron is the average of the weights of its two neighboring neurons, to avoid model oscillations caused by topological abrupt changes.

[0041] In this embodiment, the weight update rule for the GSOM unsupervised learning neurons adapted to the incremental growth characteristics is as follows: For each input multidimensional state feature vector, find its best matching neuron (BMU, i.e., the neuron with the smallest Euclidean distance to the input vector), and perform local weight updates only on the BMU and neurons within its topological neighborhood. The update formula is: In the formula: Let be the weight vector of neuron j at time t; The learning rate decays exponentially with the number of iterations, with an initial value of 0.1 and a minimum decay to 0.001. This is the current input multidimensional state feature vector.

[0042] Neuron j belongs to the current topological neighborhood of the BMU only if Weights are updated only when the condition is met; otherwise, the weights remain unchanged. (Topological neighborhood) The neighborhood radius is defined as the set of neurons whose distance from the BMU Manhattan distance does not exceed the current neighborhood radius. The neighborhood radius decreases linearly with the number of iterations, with an initial value of 1 and a minimum shrinkage to 0 (only updating the BMU itself). The single-round iteration termination condition for unsupervised learning is: the cumulative quantization error change rate of the single-round input samples (the absolute value of the difference between the cumulative quantization error of this round and the cumulative quantization error of the previous round divided by the cumulative quantization error of the previous round) is less than 1%, or the maximum number of iterations of 100 is reached to avoid unnecessary computational power consumption. For newly inserted neurons in incremental growth, a weight synchronous update rule is executed: after the new neuron is inserted, based on its initial weight, the same update rule as above is used to fine-tune the weights based on the most recent 20 input samples, ensuring that the new neuron quickly adapts to the input data distribution without affecting the weight stability of the original neurons.

[0043] In this embodiment, the "adjacency relationship" of neurons is defined as neurons with a Manhattan distance of 1 in the topological grid, rather than Euclidean distance; the co-activation mode analysis only counts stable co-occurrence within a continuous time window, and single or intermittent co-occurrence is not included in the dependency strength calculation, thereby filtering out spurious correlations caused by instantaneous load fluctuations; Cumulative quantization error refers to the average Euclidean distance between all input vectors and their best-matching neuron weight vectors. The preset threshold can be adjusted within the range of 0.1-0.3 based on BMC computing power resources. Here, the continuous time window refers to a sliding window with a length of 5-10 sampling periods, used to filter instantaneous noise. The formula for calculating the dependency strength weight is: In the formula: C is the number of times the feature vectors of components i and j are mapped to the same or adjacent neurons within N consecutive time windows; The online update of edge weights uses an exponential moving average method, and the update formula is as follows: In the formula: α is the attenuation factor, which can be adjusted in the range of 0.05-0.4 according to the hardware drift rate; The edge weights between components i and j at the previous time step are given.

[0044] Exponential moving average is a method for smoothing time series data. By assigning higher weights to recent data, it enables the model to adapt to data changes more quickly. When the edge weight is below 0.05, the edge is automatically deleted to avoid invalid associations consuming computational resources.

[0045] This embodiment introduces physical coupling-derived features to provide prior constraints for topology learning, effectively avoiding the learning of spurious associations that do not conform to the physical laws of the hardware; through the incremental GSOM growth mechanism, the computational power consumption of the model is reduced, and it can run stably online at the BMC; through the analysis of neuron co-activation patterns in multiple continuous time windows, temporally misaligned spurious correlations are filtered out, and only substantial physical coupling dependencies with continuous co-occurrence characteristics are retained; through online updating of edge weights using an exponential moving average method, coupling relationship drift caused by factors such as server hardware aging, firmware upgrades, and component replacements can be smoothly adapted.

[0046] In another optional implementation, for deployment scenarios where the number of server components increases or decreases dynamically, the topology learning network is implemented using an incremental growing neural gas network (GNG) adapted to the low computing power environment of BMC.

[0047] Initially, a minimal network of 2 neurons is used, with initial connection weights between neurons set to 0. During unsupervised learning, for each input sample, the first and second best-matching neurons are found, the weight vectors of the two neurons are updated, and a connection is established between them. Simultaneously, the age of all connections (the number of input sample iterations since the connection was established) is updated. When the accumulated quantization error exceeds a preset threshold, a new neuron is inserted between the neuron with the largest error and its connected neighboring neuron with the largest error, completing incremental network growth. When the age of a connection exceeds a preset threshold (50-100 input sample iterations), isolated neurons with that connection and no connection are deleted, achieving dynamic topology pruning. This implementation does not require a pre-defined fixed mesh structure, better adapting to scenarios where the number of nodes changes due to dynamic expansion and replacement of server components. The remaining feature mapping rules, dependency strength calculation methods, and online edge weight update rules remain consistent with the aforementioned GSOM topology learning implementation method deployed on a single server's local BMC.

[0048] In one feasible implementation, for a scenario where a single server is deployed locally at the BMC end, the step S3, which involves performing multi-step temporal propagation inference on the component anomaly signal to generate a spatiotemporal evolution prediction result of the anomaly propagation, includes: S31: Extract component anomaly degree from the multidimensional state features, use the component anomaly degree as node feature, and combine it with the dynamic dependency association topology to construct graph structure data for anomaly propagation reasoning; S32: Based on the physical location distance of components, perform modulation processing on the edge weights in the graph structure data. The closer the physical location of the components, the higher the enhancement coefficient is assigned to the edge weights, thus obtaining the modulated graph structure data. S33: Based on the modulated graph structure data, perform multi-step temporal propagation inference to generate spatiotemporal evolution prediction results for anomaly propagation.

[0049] In this embodiment, the component anomaly degree is calculated using the Z-score normalization method, and the formula is: In the formula: This represents the current value of the k-th dimension state feature of the component; , This represents the average and standard deviation of this feature under normal operating conditions over the past 7 days.

[0050] The normal operating baseline state uses an automatic identification method to exclude periods with abnormal alarms, component replacements, firmware upgrades, and extreme load periods with CPU utilization exceeding 90% or falling below 10% within the past 7 days. Newly launched servers use the common baseline state parameters of the same model of server for the first 7 days, and automatically switch to their own baseline state after accumulating 7 days of valid data.

[0051] When any feature When the absolute value exceeds 3, the component is determined to be abnormal; the abnormality degree is the sum of all abnormal characteristics. The absolute maximum value. If there are no abnormal characteristics, the component is judged to be normal, and the abnormality degree is set to 0 by default.

[0052] To incorporate the physical propagation laws of heat transfer, vibration transmission, and electrical coupling within the server chassis into the model, this invention specifically designs an edge weight modulation process. Specifically, based on the physical distance between components, the original edge weights in the dynamically dependent topology are enhanced. Edge weights corresponding to components with closer physical locations receive higher enhancement coefficients. The modulation formula is as follows: In the formula: is the original edge weight between components i and j in the dynamically dependent topology; k is the physical proximity enhancement coefficient, which can be adjusted in the range of 0.2-0.6 according to the server chassis heat dissipation design; The three-dimensional Euclidean distance between the two components; This represents the maximum distance between any two components within the chassis.

[0053] In this embodiment, the physical location distance is based on the three-dimensional Euclidean distance of the server's factory hardware layout diagram, rather than the two-dimensional projection distance; the edge weight modulation only applies to the anomaly propagation inference process and does not modify the edge weights of the original dynamically dependent topology, thereby ensuring the independence and accuracy of topology learning. Multi-step temporal propagation inference is implemented using a temporal propagation model based on graph attention mechanism. The number of inference steps is set to 3-5 steps, corresponding to a prediction duration of 15-30 minutes in the future. Each inference step corresponds to a prediction duration of 5-6 minutes, that is, 3 inference steps correspond to the next 15 minutes, and 5 inference steps correspond to the next 30 minutes. This can be adjusted according to the server sampling period and the typical fault propagation speed. The reasoning process begins with the initial anomalous state, and the initial risk value is defined as the initial risk value. The abnormality level is defined as the degree of abnormality of the abnormal component, while the initial risk value for the normal component is 0.

[0054] In the m-th step of reasoning, the formula for calculating the affected risk value of node j is: In the formula: Let j be the set of neighboring nodes; The attention coefficients calculated for the graph attention mechanism range from 0 to 1.

[0055] The standard graph attention mechanism is used for calculation, which involves calculating the similarity of the anomaly features of nodes i and j through a shared single-layer feedforward neural network, and then performing softmax normalization. Softmax normalization is a method that transforms multiple values ​​into a probability distribution so that the sum of all output values ​​is 1. The standard calculation formula is as follows: This embodiment constructs graph structure data by combining component anomaly degree with dynamic dependency topology, realizing joint modeling of spatial propagation and temporal evolution of abnormal signals; by fusing component physical location information to modulate edge weights, physical laws such as heat transfer, vibration transfer, and electrical coupling within the server chassis are injected into the model, effectively eliminating false propagation paths; through multi-step temporal reasoning, the fault propagation trend can be predicted in advance, reserving sufficient window period for operation and maintenance.

[0056] In another optional implementation, for deployment scenarios where BMC computing resources are extremely limited, the multi-step temporal propagation inference is implemented using a multi-step temporal propagation model based on Markov chains.

[0057] Each component is treated as a state node in a Markov chain. The outlier risk value of a node is the core influencing factor on the state transition probability, and the modulated edge weights are... First, row normalization is performed to satisfy the constraint that the row sum of the transition matrix is ​​1, which is then used as the initial state transition probability matrix between nodes. During multi-step inference, the state transition probability matrix at step m is the m-th power of the initial transition probability matrix, and the affected risk value of node j at step m is the product of the initial abnormal state vector and the state transition probability matrix at step m. This implementation does not require training a neural network model, resulting in lower computational power and memory overhead. It can further adapt to the operating environment of low-configuration and older BMC models. The corresponding rules for the number of inference steps and prediction duration, as well as the application method of risk value output, are consistent with the aforementioned implementation of graph attention temporal propagation inference deployed on a single server local BMC.

[0058] In a preferred embodiment, for a single-server local BMC deployment scenario, the following steps are included after step S33: S34: Perform a confidence assessment on the generated abnormal propagation path to obtain the confidence assessment result; S35: When the highest confidence level in the confidence assessment result is lower than the preset threshold, the adaptive update of the dynamic dependency association topology in step S2 is triggered, and the updated dynamic dependency association topology is reused for multi-step temporal propagation inference in step S3.

[0059] In this embodiment, propagation path confidence assessment and topology adaptive triggering are performed after step S33: For all possible propagation paths obtained through inference, the posterior confidence is calculated using a Bayesian inference framework. The confidence calculation formula is as follows: P = 0.5P₁ + 0.3P₂ + 0.2P₃ In the formula: P1 is the normalized value of the product of the edge dependency strengths of each edge on the path. The normalization method is to sum the products of the edge dependency strengths of all possible propagation paths obtained in this inference, and divide the product of a single path by the sum to obtain the value of P1. P2 represents the frequency of this path in historical anomaly cases; P3 represents the actual verification accuracy of this path in historical cases.

[0060] For newly launched servers with no historical anomaly cases, a confidence cold start rule is implemented: the initial values ​​of P2 and P3 are both set to 0.5. After the device accumulates 10 valid anomaly cases, it will automatically switch to the actual value based on its own historical data statistics, and the initial value of 0.5 will be completely removed from the statistical sample. For server clusters of the same model, it is supported to import historical anomaly case data from other devices in the same cluster as the initial value for cold start, which accelerates the convergence of confidence assessment.

[0061] Historical anomaly cases are stored in a graph database, and propagation paths are stored as node sequences. A subgraph matching algorithm is used to retrieve identical or similar paths, and their frequency and verification accuracy are statistically analyzed. The subgraph matching algorithm is used to find subgraphs in a large graph that have the same or similar structure as a given subgraph. In this embodiment, an approximate subgraph matching method based on node labels and edge weights is used. The implementation rules for the approximate subgraph matching are as follows: paths with completely matching node labels and an absolute difference in edge weights of less than 0.2 are considered similar paths. During matching retrieval, paths with completely identical node sequences are prioritized. If no completely matching path is found, similar paths are then retrieved. The matching results are sorted from high to low similarity, and the P2 and P3 values ​​of the top 5 paths are calculated. The propagation paths are then sorted from high to low confidence, and the top three propagation path candidate sets are output. When the highest confidence level is lower than the preset threshold (0.7-0.85), the incremental GSOM model's weight adaptive update is immediately triggered. The update method is as follows: extract the multi-dimensional state features of the nodes and their direct neighbors in the most recent 100 sampling periods, perform incremental unsupervised learning only on the neurons corresponding to these nodes (without updating the weights of other neurons), and then recalculate and update the corresponding edge weights according to the method in step S23, without performing training and full topology reconstruction of all neurons; after the update is completed, the multi-step temporal propagation inference in step S33 is re-executed.

[0062] This embodiment forms an internal closed loop between the inference layer and the modeling layer through propagation path confidence assessment and local topology adaptive update mechanism, ensuring the continuous accuracy of anomaly propagation prediction while guaranteeing model stability; by weighted fusion of multiple confidence factors, it effectively reduces confidence misjudgment caused by a single factor.

[0063] In one feasible implementation, for a single server local operation and maintenance terminal deployment scenario, the step of generating operation and maintenance decision guidance visualization content in step S4 includes: S41: Map the dynamic dependency association topology to a three-dimensional virtual operation and maintenance space, and use a force-directed layout algorithm to calculate the spatial position of each component node; S42: Generate a three-dimensional basic view of operation and maintenance by encoding the dependency strength between components through node spatial distance, encoding the risk of abnormal propagation of components through node color temperature, and encoding the urgency of component operation and maintenance through node pulsation frequency.

[0064] The steps in step S4, which involve generating visualized content for operation and maintenance decision guidance and corresponding handling guidance information, further include: S43: Based on the spatiotemporal evolution prediction results of the anomaly propagation, the temporal process of anomaly propagation is rendered in the three-dimensional operation and maintenance basic view using dynamic particle flow to obtain a three-dimensional operation and maintenance view with anomaly propagation path. S44: Generate an operation and maintenance sequence that matches the anomaly propagation trend, predict the component status after the operation is executed based on the operation and maintenance sequence, overlay and render the status prediction effect in the three-dimensional operation and maintenance view with the anomaly propagation path, and generate operation and maintenance decision guidance visualization content and corresponding handling guidance information.

[0065] In this embodiment, The three-dimensional virtual operation and maintenance space adopts a right-handed Cartesian coordinate system, with the lower left rear corner of the server chassis as the origin (0,0,0). The width direction of the chassis is the X-axis, the depth direction is the Y-axis, and the height direction is the Z-axis. The coordinate system range is mapped 1:1 to the actual physical size of the chassis. The initial position anchoring rule of each component node is as follows: based on the actual physical three-dimensional coordinates of the component within the chassis, it is mapped to the corresponding initial position in the three-dimensional virtual operation and maintenance space. Force-guided layout iteration is only executed within the limited range of the initial position, and the maximum offset does not exceed 1 / 2 of the corresponding component's physical size, ensuring the consistency between the virtual layout and the actual physical layout of the server and reducing the cognitive load of operation and maintenance personnel.

[0066] The force-guided layout algorithm determines the final spatial position of each component node through the combined action of three independent force fields. The iterative solution formula is as follows: Dependent on gravity In the formula: This is the gravitational constant (default value 1.0), used to adjust the weight of the influence of the inter-component dependency strength on the node distance. For components with higher dependency strength, the weight can be increased accordingly. This value makes nodes closer together in virtual space; , Let be the position vectors of nodes i and j.

[0067] Risk repulsion In the formula: This is the repulsion constant (default value 0.6), used to adjust the spatial isolation of high-risk components. For scenarios with higher anomaly risks, it can be increased accordingly. This value makes high-risk nodes more noticeable in virtual space; , Let be the abnormal propagation risk value for nodes i and j.

[0068] Urgency Centripetal Force In the formula: This is the centripetal force constant (default value 0.4), used to adjust the degree to which high-urgency components converge towards the center of the view. For scenarios with higher operational urgency, this value can be increased accordingly. This value concentrates the core processing nodes in the central area of ​​the view; The urgency of maintenance for node i.

[0069] Node position update In the formula: The learning rate is 0.1, which decreases linearly with the number of iterations.

[0070] The iteration termination condition of the force-guided layout algorithm is that the average change in node position is less than 0.1 pixels, or the maximum number of iterations of 500 is reached.

[0071] In this embodiment, the urgency of component maintenance is calculated jointly by the component's importance and the risk value of anomaly propagation, using the following formula: In the formula: Let i be the risk value for abnormal propagation of component i; The importance weight of component i, with a value ranging from 0 to 1, represents highly important components. =1.0, Medium Importance =0.6, low importance =0.3. The node visual variable mapping rules are as follows: Node color temperature: Calculated using a linear mapping of the HSV color space, with the following formula: In the formula: The hue value of node i in the HSV color space; Let be the abnormal propagation risk value of node i, ranging from 0 to 1. That is, a risk value of 0 corresponds to blue (240°), and a risk value of 1 corresponds to red (0°).

[0072] Node pulsation frequency: In the formula: Let be the pulsation frequency of node i; Let represent the urgency of node i, with a value ranging from 0 to 1. That is, an urgency of 0 corresponds to no pulse, and an urgency of 1 corresponds to 2 pulses per second.

[0073] Node size: The diameter is positively correlated with the importance of the component. The importance of a component can be predefined based on the impact of its failure on the server's business. For example, CPU and memory are high importance (30 pixels in diameter), power supply and system hard drive are medium importance (20 pixels in diameter), and fans and temperature sensors are low importance (10 pixels in diameter). The default diameter is 10-30 pixels.

[0074] The dynamic particle flow rendering rules are as follows: The particle emission sources are candidate components for the root cause of the anomaly, with each source emitting 20-50 particles per second; The particle flow direction points towards the affected component along the dependency relationship edge, and the particle velocity calculation formula is: In the formula: The particle velocity representing the propagation of the abnormal signal from component i to component j; The predicted propagation time of the anomaly is in seconds.

[0075] Particle density is positively correlated with propagation probability. A propagation probability above 0.8 corresponds to high density, 0.5-0.8 corresponds to medium density, and below 0.5 corresponds to low density. The particle color is time-dependent: 0-5 minutes corresponds to red, 5-15 minutes to orange, and 15-30 minutes to yellow. The particle lifespan is 2-5 seconds, and the transparency decreases linearly over time.

[0076] The generation of operation and maintenance (O&M) operation sequences is based on O&M knowledge graph matching, prioritizing the matching of the most effective operations in the past for the same anomaly pattern and the same propagation path. The O&M knowledge graph contains four core entities: anomaly pattern, propagation path, O&M operation, and operation effect, as well as three core relationships: "anomaly pattern - corresponding propagation path", "propagation path - corresponding O&M operation", and "O&M operation - historical effect".

[0077] Matching similar anomaly patterns uses a rule where the cosine similarity of the anomaly degree vector is greater than 0.8. Matching similar propagation paths uses the aforementioned approximate subgraph matching rule based on node labels and edge weights. The matching priority is: paths with completely identical node sequences > similar paths > similar anomaly patterns. When there are no matching historical cases, a general operation and maintenance sequence based on component importance is output, prioritizing the inspection and handling of high-importance components.

[0078] During the operation simulation, the changes in the state of the components corresponding to the operation (such as setting the abnormality to 0 after replacing the power module, and reducing the temperature by 30% after adjusting the fan speed) are input into the anomaly propagation model. The future risk values ​​of all components are recalculated, and the risk reduction area is displayed in the 3D view with a green semi-transparent overlay, and the potential secondary risk area is displayed with a yellow semi-transparent overlay.

[0079] The component status changes corresponding to operations are determined based on historical operation effect data stored in the operation and maintenance knowledge graph, taking the average effect value of similar operations under similar abnormal modes. For new operations without historical data, simplified models based on physical laws are used for prediction, such as the linear relationship model between fan speed and temperature change, and the linear relationship model between power supply output power and voltage fluctuation. Status change parameters can be manually fine-tuned according to server model and actual operating environment.

[0080] This embodiment reduces the cognitive load on maintenance personnel by directly encoding the core decision-making dimensions into a force-oriented layout of three-dimensional spatial visual variables; it transforms the abstract anomaly propagation process into an intuitive visual effect through dynamic particle flow temporal rendering; and it enables maintenance personnel to assess the operation effect in advance and avoid secondary failures caused by blind operation through simulation and effect prediction of the operation operation sequence.

[0081] In another optional implementation, for server cluster scenarios with massive historical operation and maintenance cases, the operation and maintenance operation sequence generation is implemented using an operation and maintenance decision model based on deep reinforcement learning. A deep Q-network (DQN) is used to construct the reinforcement learning model, with the anomaly vector of server components and dynamic dependency topology as the model input state space, standard operation and maintenance instructions as the model output action space, and the fault recovery rate after operation execution as the reward function. The model is pre-trained based on massive historical anomaly cases of the same model of server, and outputs the operation and maintenance operation sequence with the maximum cumulative reward during online inference. Simultaneously, it supports online updating of model parameters based on real-time operation results to adapt to the personalized operating conditions of different servers.

[0082] For example, the deep Q-network adopts a 3-layer fully connected neural network structure. The input layer dimension is "number of components × (anomaly degree dimension + topology connection dimension)", the hidden layer dimension is 256, and the output layer dimension is the total number of standard operation and maintenance instructions. The state space encoding method is as follows: the anomaly degree value of each component is concatenated with the weights of all adjacent edges of that component to form a one-dimensional vector, and then the vectors of all components are concatenated in a fixed order to form a global state vector. The action space is defined as a discrete set of 12 standard operation and maintenance operations, including: checking fan speed, adjusting fan speed, checking power supply output voltage, replacing power supply module, checking hard drive S... MART information, hard drive replacement, BMC restart, operating system restart, PCIe device isolation, PCIe device replacement, memory temperature check, memory replacement; The reward function is defined as: fault recovery rate within 10 minutes after operation execution × 100 - operation time (minutes) × 2 - business interruption duration (minutes) × 10. If the operation causes secondary faults, an additional penalty of -200 is applied; The model training adopts an experience replay mechanism with a replay buffer size of 10000, a batch size of 32, a target network update frequency of once every 100 steps, a discount factor γ = 0.95, and a learning rate of 0.0001.

[0083] The remaining visualization rendering rules and operation simulation pre-play logic are consistent with the aforementioned knowledge graph operation and maintenance sequence generation implementation method deployed on a single server local operation and maintenance terminal.

[0084] In one feasible implementation, for a single-server local closed-loop deployment scenario, the step S5, which involves reversing the model parameters of the topology learning network and temporal propagation inference based on the execution effect data, includes: S51: Based on the component status data collected after the operation and maintenance operation is executed, evaluate the actual recovery effect of the operation and obtain the recovery effect evaluation result; S52: Integrate the full data of this abnormal case with the recovery effect evaluation results to construct a training sample, and update the mapping weights of the topology learning network and the edge weight parameters of the temporal propagation inference based on the training sample.

[0085] In this embodiment, the S5 closed-loop feedback update has a higher priority than the regular periodic update of S2: when the S5 feedback update is triggered, the currently executing S2 periodic update is immediately paused, and the S5 feedback update is completed first; the next periodic update of S2 will be executed based on the latest edge weights after the S5 feedback, and the recovery effect assessment is based on the component status data within 10 minutes after the operation and maintenance is performed to calculate the recovery rate. In the formula: Z init Z is the initial value of the component's anomaly level before the operation and maintenance is performed. final This represents the final anomaly level of the component after the maintenance operation is performed.

[0086] Based on the recovery rate, the operational effectiveness is divided into three levels: Fully effective: recovery rate ≥80%; Partially effective: 50% ≤ recovery rate <80%; Ineffective: Recovery rate <50%.

[0087] When the initial anomaly degree Z of the component init When the value is 0, it indicates that the component is in a completely healthy state and there is no clear source of abnormal signal. This operation is not a valid case of abnormal handling. Under this boundary condition, the recovery rate is not calculated, and the sample data does not participate in the subsequent update of the parameters of the topology learning network and the time-series propagation inference model in step S52. It is directly discarded to avoid invalid samples from disturbing the model iteration.

[0088] The full set of anomaly case data includes: anomaly occurrence time, root cause component, initial anomaly strength, propagation path and confidence level obtained through inference, sequence of executed operation and maintenance operations, operation execution time, sequence of component state changes after operation execution, and final recovery effect.

[0089] The feedback update rule of the topology learning network is specifically designed for this invention: For a propagation path corresponding to a fully valid operation, increase the weight of all edges on the path. , where P is the confidence level of the path; For the propagation paths corresponding to some effective operations, the edge weight adjustment is linearly related to the recovery rate: =(Recovery rate - 0.5) 0.05(1+P), the edge weight is still limited to the range of 0-1 after the update; For the propagation path corresponding to an invalid operation, reduce the weight of all edges on the path. =0.05+0.05 (1-P); After the edge weights are updated, they are limited to the range of 0-1 to avoid overflow. In this embodiment, the feedback update only adjusts the edge weight size without changing the topology, thereby ensuring the stability of the model. Abnormal case data is stored incrementally, retaining data from the most recent 12 months and periodically deleting expired data. When the same propagation path is verified as effective in 5 consecutive abnormal cases, its edge weight update step size is reduced to 50% of the original to prevent overfitting.

[0090] The time-series propagation inference model is updated using mini-batch gradient descent, with the loss function being the mean squared error loss, as shown in the formula: In the formula: The risk value predicted by the model. The actual observed risk value; the learning rate for model training is set to 0.001, and the batch size is set to 8.

[0091] Mini-batch gradient descent is a commonly used neural network training method. Each iteration uses a small subset of training samples to calculate gradients and update model parameters, balancing training speed and convergence stability. The training sample input for the time-series propagation inference model consists of the dynamic dependency topology and initial anomaly vector at the time of anomaly occurrence. Labels are the actual anomaly sequence observed in each component within 15-30 minutes after the anomaly occurs. The model update cycle can be set to once a week, or an update can be performed when the number of accumulated new anomaly cases reaches 100.

[0092] This embodiment achieves the standardization and data-driven accumulation of operational experience by quantitatively evaluating the actual recovery effect of operation and maintenance operations; it solves the problem of the continuous degradation of the performance of traditional fixed models as the server service time increases by feeding back abnormal cases and recovery effects to update the topology model and inference model; and it forms a continuous improvement cycle of "practice-learning-optimization-re-practice" through a full-link closed-loop feedback mechanism, so that the system's operation and maintenance capabilities are continuously improved with the accumulation of operation and maintenance experience.

[0093] In another optional implementation, for collaborative deployment scenarios involving multi-node server clusters of the same model, the topology learning network updates the model parameters inversely with the time-series propagation inference, employing a cross-device model update mechanism based on incremental federated learning. Each node calculates the model parameter update gradient based on local anomaly cases, uploading only the encrypted gradient information to the federated learning center node. The center node only performs gradient aggregation operations and does not store any original operational data, ensuring data privacy and security. The center node performs weighted aggregation on the gradient information of all nodes to generate globally updated model parameters, which are then distributed to each node to complete the local model update.

[0094] For example, the incremental federated learning uses the FedAvg aggregation algorithm, and the gradient encryption uses the Paillier additive homomorphic encryption algorithm. After each node accumulates 10 valid anomaly cases locally, it calculates the update gradient of the local model parameters, encrypts the gradient using the Paillier algorithm, and uploads it to the central node. The central node performs a weighted average aggregation of the encrypted gradients of all nodes, with the weight being the number of local training samples of each node. After aggregation, the encrypted global gradient is distributed to each node, and each node decrypts it and updates its local model parameters. The model synchronization cycle is once a week, or synchronization is triggered when more than 30% of the nodes in the cluster have completed local updates. For heterogeneous BMC devices, a hierarchical update mechanism for model parameters is adopted, which only updates the fully connected layer parameters of the topology learning network and the propagation inference model, while sharing the underlying feature extraction layer parameters.

[0095] This implementation method can achieve cross-device sharing of operation and maintenance experience and collaborative optimization of models while protecting the privacy of operation and maintenance data and the security of business data on each server. It solves the problem of insufficient model updates and slow convergence caused by insufficient abnormal case samples on a single server. The other update rules and effect evaluation criteria are consistent with the aforementioned implementation method of updating model parameters in a local closed-loop deployment on a single server.

[0096] like Figure 2 As shown, a server component operation and maintenance decision visualization system based on BMC is disclosed. The system includes a memory, a processor, and a server component operation and maintenance decision visualization program based on BMC stored in the memory and executable on the processor. The program is configured to implement the steps of the server component operation and maintenance decision visualization method based on BMC.

[0097] The server component operation and maintenance decision visualization system based on BMC provided by this invention adopts the server component operation and maintenance decision visualization method based on BMC in the above embodiments. It can construct a full-process intelligent operation and maintenance pipeline of "data acquisition - dynamic topology modeling - anomaly propagation inference - decision visualization - closed-loop feedback," deeply integrating incremental topology learning driven by full-volume time-series BMC data with anomaly propagation spatiotemporal evolution inference based on physical priors. Combined with multi-dimensional visual encoding for operation and maintenance decision guidance visualization, it achieves an intuitive presentation of the impact of faults, realizing a fundamental shift in server operation and maintenance mode from passive post-event handling to proactive prediction and early warning. Furthermore, the full-link closed-loop feedback mechanism ensures the continuous stability of the system's operation and maintenance performance throughout the server's entire lifecycle. Compared with existing technologies, the beneficial effects of the server component operation and maintenance decision visualization system based on BMC provided by this invention are the same as those of the server component operation and maintenance decision visualization method based on BMC provided in the above embodiments. Other technical features of the server component operation and maintenance decision visualization system based on BMC are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0098] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A visualization method for server component operation and maintenance decisions based on BMC, characterized in that, Includes the following steps: S1: Collect the operating status time-series data of all server components through the BMC out-of-band channel, perform standardization processing on the operating status time-series data, and obtain the multi-dimensional status features corresponding to each component; S2: Input the multidimensional state features into the topology learning network, perform online unsupervised learning, and generate and dynamically update the dependency association topology between server components; S3: Based on the dynamic dependency association topology, perform multi-step temporal propagation inference on the component abnormal signal to generate the spatiotemporal evolution prediction result of abnormal propagation; S4: Based on the spatiotemporal evolution prediction results, generate operation and maintenance decision guidance visualization content and corresponding handling guidance information; S5: Collect the execution effect data of the operation corresponding to the disposal guidance information, and update the model parameters of the topology learning network and the temporal propagation inference based on the execution effect data to complete the closed-loop iteration of the model.

2. The server component operation and maintenance decision visualization method based on BMC according to claim 1, characterized in that, The step of obtaining the multidimensional state features corresponding to each component in step S1 includes: S11: Calculate the physical coupling-derived characteristics between components within the same chassis for the standardized runtime time-series data. S12: The physical coupling-derived features are added to the basic state features of the corresponding components to obtain the multi-dimensional state features of each component for topology learning.

3. The server component operation and maintenance decision visualization method based on BMC according to claim 1, characterized in that, The step of generating and dynamically updating the dependency association topology between server components as described in step S2 includes: S21: Initialize a minimum-scale topology learning network, map the multidimensional state features to the topology learning network, and perform unsupervised learning; S22: When the cumulative quantization error of the topology learning network exceeds a preset threshold, perform incremental network growth at the position corresponding to the largest error to obtain an updated topology learning network. S23: Map the multidimensional state features to the updated topology learning network, perform unsupervised learning, and generate a dynamic dependency association topology between server components based on the learned mapping results.

4. The server component operation and maintenance decision visualization method based on BMC according to claim 3, characterized in that, The step of generating the dynamic dependency association topology between server components as described in step S23 includes: S231: Statistically analyze the mapping positional relationship of the multidimensional state features of different components in the updated topology learning network within multiple consecutive time windows; S232: Calculate the dependency strength weights between components based on the co-occurrence characteristics of mapping positions within a continuous time window; S233: Construct and update a dynamic dependency association topology online using components as nodes and dependency strength weights as weighted edges.

5. The server component operation and maintenance decision visualization method based on BMC according to claim 1, characterized in that, The step S3, which involves performing multi-step temporal propagation inference on the component anomaly signal to generate a spatiotemporal evolution prediction result for anomaly propagation, includes: S31: Extract the component anomaly degree from the multidimensional state features, use the component anomaly degree as a node feature, and combine it with the dynamic dependency association topology to construct graph structure data for anomaly propagation reasoning; S32: Based on the physical location distance of components, perform modulation processing on the edge weights in the graph structure data. The closer the physical location of the components, the higher the enhancement coefficient is assigned to the edge weights, thus obtaining the modulated graph structure data. S33: Based on the modulated graph structure data, perform multi-step temporal propagation inference to generate spatiotemporal evolution prediction results for anomalous propagation.

6. The server component operation and maintenance decision visualization method based on BMC according to claim 5, characterized in that, Step S33 is followed by the following steps: S34: Perform a confidence assessment on the generated abnormal propagation path to obtain the confidence assessment result; S35: When the highest confidence level in the confidence assessment result is lower than the preset threshold, the adaptive update of the dynamic dependency association topology in step S2 is triggered, and the updated dynamic dependency association topology is reused for multi-step temporal propagation inference in step S3.

7. The server component operation and maintenance decision visualization method based on BMC according to claim 1, characterized in that, The steps in step S4 for generating visualization content to guide operation and maintenance decisions include: S41: Map the dynamic dependency association topology to a three-dimensional virtual operation and maintenance space, and use a force-directed layout algorithm to calculate the spatial position of each component node; S42: Generate a three-dimensional basic view of operation and maintenance by encoding the dependency strength between components through node spatial distance, encoding the risk of abnormal propagation of components through node color temperature, and encoding the urgency of component operation and maintenance through node pulsation frequency.

8. The server component operation and maintenance decision visualization method based on BMC according to claim 7, characterized in that, The steps in step S4, which involve generating visualized content for operation and maintenance decision guidance and corresponding handling guidance information, further include: S43: Based on the spatiotemporal evolution prediction results of the anomaly propagation, the temporal process of anomaly propagation is rendered in the three-dimensional operation and maintenance basic view using dynamic particle flow to obtain a three-dimensional operation and maintenance view with anomaly propagation path. S44: Generate an operation and maintenance sequence that matches the anomaly propagation trend, predict the component status after the operation is executed based on the operation and maintenance sequence, overlay and render the status prediction effect in the three-dimensional operation and maintenance view with the anomaly propagation path, and generate operation and maintenance decision guidance visualization content and corresponding handling guidance information.

9. The server component operation and maintenance decision visualization method based on BMC according to claim 1, characterized in that, The step S5, which involves updating the model parameters of the topology learning network and temporal propagation inference based on the execution effect data, includes: S51: Based on the component status data collected after the operation and maintenance operation is executed, evaluate the actual recovery effect of the operation and obtain the recovery effect evaluation result; S52: Integrate the full data of this abnormal case with the recovery effect evaluation results to construct a training sample, and update the mapping weights of the topology learning network and the edge weight parameters of the temporal propagation inference based on the training sample.

10. A server component operation and maintenance decision visualization system based on BMC, characterized in that, The system includes a memory, a processor, and a BMC-based server component operation and maintenance decision visualization program stored in the memory and executable on the processor, the program being configured to implement the steps of the BMC-based server component operation and maintenance decision visualization method as described in any one of claims 1 to 9.

Images