Data center distributed monitoring management method and system based on edge computing

Abstract

The application belongs to the technical field of distributed monitoring, and discloses a data center distributed monitoring management method and system based on edge computing. The running state data of a target edge node is collected in real time, and the cooperative working state between the node and the associated node is monitored. Based on these data, multi-dimensional features are extracted, a weighted matrix and a Bayesian network inference model are constructed, and the self-healing ability of the target edge node is evaluated. When the target node has an abnormal trend, the self-healing ability score is compared with a preset threshold value. If the score is lower than the threshold value, the target node is determined as a fault node and intervention measures are triggered. The self-healing ability can be intelligently evaluated, excessive intervention can be avoided, fault recovery efficiency can be improved, and system resource waste can be reduced. The application is suitable for data centers under edge computing architecture, especially in large-scale distributed computing environments, and can effectively improve the efficiency of fault management and resource scheduling.

Description

Technical Field

[0001] This invention relates to the field of distributed monitoring technology, specifically to a data center distributed monitoring and management method and system based on edge computing. Background Technology

[0002] With the widespread deployment of edge computing in various applications, data centers need to handle a large number of edge nodes distributed across different geographical locations. These nodes typically have limited computing resources, low bandwidth, and unstable network connections, making them susceptible to various factors such as hardware failures, network problems, and high loads. Due to the close interconnectivity and collaborative work between edge nodes, node failures not only affect the operation of a single node but may also lead to a decline in the performance of adjacent nodes or the entire system.

[0003] Many existing systems determine node intervention needs solely based on resource consumption. This means they judge a node's health status by its individual resource consumption (such as CPU, memory, and bandwidth) and intervene when thresholds are exceeded. However, they fail to consider that edge nodes can recover from failures through their own dynamic adjustments or collaboration with other nodes. For example, a node might experience performance degradation due to momentary load fluctuations or resource contention, but this doesn't mean the node is incapable of self-repairing through load adjustments, resource reallocation, or short-term recovery mechanisms. Existing static monitoring methods ignore this, leading to premature intervention, wasting system resources, and failing to maximize the utilization of nodes' inherent resilience and redundant resources. Summary of the Invention

[0004] In view of the above-mentioned shortcomings mentioned in the background art, the purpose of this invention is to provide a data center distributed monitoring and management method and system based on edge computing.

[0005] The first aspect of this invention provides a distributed monitoring and management method for data centers based on edge computing, applied to data centers, the method comprising the following steps: S1 collects real-time operational status data of the target edge node, as well as the collaborative working status between the target edge node and its associated edge nodes; S2, Multi-dimensional feature data extracted from the running status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes; S3. Based on the multi-dimensional feature data, construct a weighted matrix and a Bayesian network inference model, evaluate the self-healing ability of the target edge node based on the weighted matrix and the Bayesian network inference model, and obtain a self-healing ability score. S4. When a preset type of abnormal trend appears at the target edge node, the self-healing ability score of the target node is compared with a preset threshold. If the self-healing ability score is lower than the preset threshold, the target node is identified as a faulty node and intervention is triggered on the target edge node.

[0006] A second aspect of the present invention provides a data center distributed monitoring and management system based on edge computing, applied in a data center, the system comprising: The data acquisition module is used to collect real-time operational status data of the target edge nodes, as well as the collaborative working status between the target edge nodes and their associated edge nodes. The feature extraction module is used to extract multi-dimensional feature data from the running status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes. The self-healing capability assessment module is used to construct a weighted matrix and a Bayesian network inference model based on the multi-dimensional feature data, and to assess the self-healing capability of the target edge node based on the weighted matrix and the Bayesian network inference model to obtain a self-healing capability score. The intervention decision module is used to compare the self-healing ability score of the target node with a preset threshold when a preset type of abnormal trend appears at the target edge node. If the self-healing ability score is lower than the preset threshold, the target node is identified as a faulty node and intervention is triggered on the target edge node.

[0007] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in any of the preceding claims.

[0008] A fourth aspect of the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the method as described in any of the preceding claims.

[0009] Compared with existing technologies, this invention collects and analyzes the operational and collaborative working status of target edge nodes and associated nodes in real time. Combining multi-dimensional feature data, it constructs a weighted matrix and a Bayesian network inference model to accurately assess the self-healing capabilities of nodes, optimizes the fault recovery process, and maximizes resource utilization efficiency. This invention can effectively improve the operational stability of data centers, enhance the intelligence level of resource management, reduce system resource waste, and significantly improve the efficiency and accuracy of fault repair. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall process of a data center distributed monitoring and management method based on edge computing disclosed in an embodiment of the present invention; Figure 2This is a schematic diagram of the system architecture of the technical solution of the present invention; Figure 3 This is a schematic diagram of the structure of a data center distributed monitoring and management system based on edge computing disclosed in an embodiment of the present invention. Detailed Implementation

[0011] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0013] This invention discloses a distributed monitoring and management method for data centers based on edge computing, comprising the following steps S1 to S4. The solution of this invention is applied to a management server or distributed monitoring system in a data center. This system is deployed in a cloud computing center or core network node to perform unified monitoring, health status assessment, and fault management of numerous edge nodes distributed in different geographical locations. These edge nodes typically possess limited computing resources, low bandwidth, and unstable network connections, and collaborate to complete complex computing and data processing tasks.

[0014] Please see Figure 1 This invention provides a distributed monitoring and management method for data centers based on edge computing, applicable to data centers. The method includes the following steps: S1 collects real-time operational status data of the target edge node, as well as the collaborative working status between the target edge node and its associated edge nodes; In this embodiment, for the target edge node, a real-time data acquisition process is initiated, and the acquired data includes the following two types: The target edge node's own operational status data: This type of data reflects the node's internal health status and resource usage, including but not limited to: CPU utilization (e.g., real-time utilization as a percentage), memory usage (e.g., used memory size, memory usage percentage, swap partition utilization), bandwidth utilization (e.g., inbound and outbound traffic rates of network interfaces, in Mbps or Gbps), disk I / O read / write rates, number of currently running processes, system average load (e.g., average load values ​​over the past 1 minute, 5 minutes, and 15 minutes), and hardware health indicators such as node temperature and power consumption.

[0015] The collaborative working status between the target edge node and its associated edge nodes: This type of data reflects the interaction relationships and collaborative health of nodes in a distributed system. Associated edge nodes refer to other edge nodes that have task dependencies, data exchanges, or redundant backup relationships with the target edge node. The collaborative working status specifically includes: communication latency (e.g., network round-trip time in milliseconds, RTT) between the target node and each associated node, link bandwidth utilization, data transmission rate, packet loss rate, connection stability (e.g., number and duration of connection interruptions), synchronization status of task collaborative execution, and mutual heartbeat signals. This data can be obtained in real time by deploying probes on network devices, analyzing traffic mirroring, or extracting it from the node's log system.

[0016] In some embodiments, the target edge node is a node that exhibits a preset type of abnormal trend during the monitoring process. The preset type of abnormal trend includes, but is not limited to, excessively high resource utilization, abnormal load fluctuations, increased network latency, and abnormal collaborative work between nodes.

[0017] Specifically, the basic health indicators of all edge nodes are continuously monitored, and dynamic or static thresholds are set to identify abnormal trends. When a node exhibits certain abnormal signs, it is marked as a target edge node, triggering subsequent steps S2 to S4 for in-depth analysis.

[0018] The preset types of abnormal trends are quantitative descriptions of the precursory characteristics of potential node failures. For example, excessively high resource utilization can be defined as: CPU utilization exceeding 85% for 5 consecutive minutes, or memory occupancy consistently exceeding 90%. Abnormal load fluctuations can be defined as: the variance of the system's average load exceeding three times the normal fluctuation range within a short period (e.g., within 1 minute), or the load value suddenly dropping to near zero (potentially indicating process freeze or crash). Increased network latency can be defined as: communication latency with critical associated nodes exceeding twice the historical average latency, or exceeding the maximum latency threshold tolerated by the business (e.g., 100ms). Abnormal inter-node collaboration can manifest as: heartbeat signal interruptions exceeding 3 times, or a data synchronization failure rate exceeding 5% during task collaboration.

[0019] When a node exhibits any one or more of the above-mentioned abnormal trends, it is considered to be on the verge of performance degradation or failure, but its self-healing capability is uncertain. Therefore, it is designated as the target node and proceeds to the self-healing capability assessment process described below.

[0020] S2, Multi-dimensional feature data extracted from the running status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes; In this embodiment, while the raw data is abundant, directly using it to determine a node's self-healing potential presents two problems: first, the data volume is enormous and its dimensions are complex, leading to low processing efficiency; second, a single indicator is insufficient to reflect the interaction relationships between nodes. For example, a spike in a node's CPU utilization cannot distinguish whether it's due to a problem with its own program or a task backlog caused by a failure of a related node. Therefore, it is necessary to extract multi-dimensional features from the raw data that can characterize collaborative relationships.

[0021] As an example, the multi-dimensional feature data extracted from the operational status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes includes: Based on the CPU utilization, memory usage, bandwidth utilization, and load of the target edge node, assess the resource sharing capability between the target edge node and its associated edge nodes. This dimension aims to quantify the ability of a target edge node to obtain resource support from associated nodes when facing resource constraints.

[0022] In practice, the first step is to read the CPU utilization, memory usage, bandwidth utilization, and current load of the target edge node. For example, suppose target node A has a CPU utilization of 92%, a memory utilization of 95%, a bandwidth utilization of 80%, and a system average load (1 minute) of 8.5 (while the number of CPU cores is 8). These indicators all suggest that node A is severely overloaded. Next, all associated nodes that have task dependencies or data exchange relationships with A (such as nodes B, C, and D) are scanned to obtain their idle resource status. For example, node B currently has 40% CPU idle and 30% memory idle; node C has 10% CPU idle and 15% memory idle; and node D has 5% CPU idle and 5% memory idle. Through weighted calculations (e.g., CPU idle rate weighted at 0.5, memory idle rate weighted at 0.3, and bandwidth idle rate weighted at 0.2), the potential resource support capability that each associated node can provide to A is evaluated. Finally, the resource sharing capability characteristic value can be quantified into a comprehensive score, for example, a weighted average based on the support capabilities of associated nodes B, C, and D. A higher score indicates that A is more likely to receive assistance from surrounding nodes when resources are scarce. In this example, since nodes C and D are also quite busy, only node B has a lot of free resources, so the resource sharing capability score is probably at a medium level.

[0023] Based on the communication latency, bandwidth usage, and data transmission rate between the target edge node and its associated edge nodes, the data transmission capability between the target edge node and its associated edge nodes is evaluated. This dimension aims to quantify the ability of target nodes and associated nodes to exchange data efficiently and stably.

[0024] In practice, calculations are performed based on the communication latency, bandwidth usage, and data transmission rate collected in step S1. For example, for target node A and its key associated node B, the average communication latency was monitored to be 15ms, far below the 100ms threshold tolerated by the business; the link bandwidth utilization rate was 60%, with a 40% margin remaining; and the actual data transmission rate was stable at around 80Mbps, close to the theoretical maximum rate of 100Mbps. However, for nodes A and C, the communication latency fluctuated significantly, drastically jittering between 10ms and 200ms, and the packet loss rate reached 3%. Through statistical analysis, a data transmission stability index between A and each associated node was calculated, which could be defined as a negative correlation function between the reciprocal of the latency and the packet loss rate. Ultimately, the data transmission capability characteristic value can be a weighted average of the transmission capabilities of A and all associated nodes, with the weights set according to business importance (e.g., the transmission capability of node B, which undertakes the main task distribution, has a higher weight). The higher this characteristic value, the smoother the data exchange channel between the target node and the cluster, and the more conducive it is to collaborative recovery.

[0025] Based on the load conditions and task allocation of the target edge node and associated edge nodes, evaluate the load balancing capability between the target edge node and associated edge nodes; This dimension aims to quantify the dynamic adjustment capability of a distributed system to migrate load from overloaded nodes (target nodes) to idle nodes.

[0026] In practice, the current load of the target edge node and its associated edge nodes, as well as historical task allocation records, are analyzed. For example, it is found that target node A currently has extremely high load, while its associated nodes B and C have relatively low load. Simultaneously, reviewing the task allocation logs from the past 5 minutes reveals that the task scheduler attempted to migrate some tasks from A to B and C, but the migration success rate was only 20%, and the load increase on B and C after migration was slow, indicating that the task scheduling strategy or the application itself does not support fast and effective load migration. Conversely, in another node cluster, when the load of a node exceeds a threshold, the container orchestration system (such as Kubernetes) can smoothly migrate the Pods on it to other nodes within 30 seconds, demonstrating extremely strong load balancing capabilities. Based on this, the load balancing capability characteristic value of the cluster where the target node resides is evaluated, which comprehensively reflects the cluster's dynamic scheduling efficiency and task portability.

[0027] Based on the redundancy task allocation, fault recovery strategy and resource scheduling capabilities of the target edge node and associated edge nodes, the collaborative repair capability between the target edge node and associated edge nodes is evaluated. Collaborative recovery capability is the most critical dimension for evaluating a node's self-healing ability. It aims to quantify whether the collaborative cluster to which the target node resides can compensate for its functionality through redundancy mechanisms, task reallocation, or policy adjustments when the target node experiences a failure or performance degradation, thereby creating conditions for recovery. The evaluation is based on a comprehensive assessment of the redundant task allocation between the target edge node and its associated edge nodes, pre-defined fault recovery strategies, and resource scheduling capabilities. For example, for a critical database service running on target node A, a hot standby instance is detected on associated node B (redundant task allocation), and an automatic master-slave failover strategy is configured (fault recovery strategy). When A's heartbeat signal is interrupted, the resource scheduler (such as the distributed coordination service ZooKeeper) can migrate the virtual IP to B within 5 seconds, completing service takeover (resource scheduling capability). For another non-critical data cleaning task, only an automatic restart strategy after task failure is configured, with no redundant instance. This information is combined to calculate a collaborative recovery capability characteristic value for each target node. For example, if critical services on node A have redundant backups and a robust failover strategy, its collaborative recovery capability score will be very high; conversely, if all services operate as single points of failure, the score will be extremely low.

[0028] The resource sharing capability, the data transmission capability, the load balancing capability, and the collaborative repair capability are used as the multi-dimensional feature data.

[0029] The aforementioned feature data is stored in the form of numerical vectors, such as [resource sharing capability: 0.75, data transmission capability: 0.82, load balancing capability: 0.60, collaborative repair capability: 0.95], which together constitute a multi-dimensional feature dataset describing the collaborative working relationship between the target edge node and its associated edge nodes.

[0030] S3. Construct a collaborative weighted matrix and a Bayesian network inference model based on the multi-dimensional feature data, evaluate the self-healing ability of the target edge node based on the collaborative weighted matrix and the Bayesian network inference model, and obtain a self-healing ability score. In this embodiment, the feature data extracted in step S2 is static and discrete. The interactions between them and how they collectively affect the self-healing ability of the nodes require a reasoning framework capable of handling uncertainty and dependencies. Therefore, this step introduces a collaborative weighted matrix and a Bayesian network reasoning model to transform the multi-dimensional feature data into a comprehensive self-healing ability score for subsequent intervention decisions.

[0031] Specifically, firstly, a collaborative weighting matrix is ​​constructed based on multi-dimensional feature data to quantify the support strength of each associated edge node for the self-healing capability of the target edge node. Then, using the target node and its associated nodes as nodes in a Bayesian network, the dependencies between nodes are determined based on the collaborative weighting matrix, thus constructing the Bayesian network structure. Next, historical sample data is used to learn the parameters of the Bayesian network, obtaining a conditional probability table for each node, forming a Bayesian network inference model. Finally, multi-dimensional feature data from the current and historical moments are input into this model, and a dynamic Bayesian network is constructed using the collaborative weighting matrix. Probabilistic inference is then used to obtain the predicted state probability sequence of the target edge node within a future time window, and the self-healing capability score is calculated accordingly.

[0032] As an example, a collaborative weighted matrix and Bayesian network inference model are constructed based on the aforementioned multi-dimensional feature data, including: S31, based on the multi-dimensional feature data, calculate the collaborative influence factor between the target edge node and each associated edge node, and construct a collaborative weighted matrix based on the collaborative influence factor; wherein, the collaborative influence factor is used to characterize the degree of support or influence intensity of the associated edge node on the self-healing ability of the target edge node; In practice, the four-dimensional feature data (resource sharing capability, data transmission capability, load balancing capability, and collaborative repair capability) calculated in step S2 are used to calculate the collaborative impact factor between the target node and each associated node. This factor is a comprehensive quantitative indicator designed to measure the extent to which a specific associated node can provide effective support when the target node experiences an anomaly.

[0033] Let the target node be A, and its associated node set be... For each associated node Based on the resource sharing capabilities between A and X Data transmission capability Load balancing capability Collaborative repair capabilities These four characteristic values ​​are used to calculate the synergistic influence factor. The calculation formula can use a weighted summation method:

[0034] in, This is a weighting factor that can be dynamically adjusted based on the business scenario. For example, for computationally intensive tasks, resource sharing capabilities are more important, so it can be set... For communication-intensive tasks, the weight of data transmission capability can be appropriately increased, such as... The weighting coefficients can be configured according to actual business needs, or they can be automatically learned from historical data through machine learning methods. Details will not be elaborated here.

[0035] After calculating the synergistic influence factors for all associated nodes, these factors are constructed into a synergistic weighted matrix W. The rows and columns of the matrix correspond to all relevant edge nodes (including the target node itself; typically, the self-influence factor can be set to 1 or 0). Elements in the matrix... This represents the collaborative influence factor of node i on node j. For target node A, the corresponding rows or columns in the collaborative weighted matrix reflect its interaction with the entire associated cluster. For example, if the associated nodes are only B, C, and D, the matrix would look like this (only the part related to A is shown):

[0036] In this context, the main diagonal elements are typically set to 1, representing the node's complete influence over itself. Off-diagonal elements... This indicates the strength of support that associated node X provides for target node A.

[0037] S32, using the target edge node and its associated edge nodes as nodes of the Bayesian network, the dependency relationship between nodes is determined according to the collaborative weighting matrix, and the Bayesian network is constructed. In practice, the target node and all its associated nodes are abstracted as random variable nodes in a Bayesian network. The state of each node can be discretized into several categories, for example, three states: healthy (indicating that the node is operating normally and all indicators are good), sub-healthy (indicating that the node's performance has declined, but it has not completely failed and may recover through self-adjustment), and faulty (indicating that the node can no longer work normally and requires external intervention). The state division can be refined according to actual needs, for example, into five levels.

[0038] The structure of the Bayesian network, i.e., the directed edges between nodes, is determined based on the aforementioned collaborative weighting matrix. The determination rule is: if the collaborative influence factor... Greater than a certain preset threshold If the threshold is 0.6, then the state of node i is considered to directly affect the state of node j, so a directed edge is added from node i to node j in the network. This threshold can be set based on historical experience or optimized from the training data through cross-validation.

[0039] Continuing the previous example, suppose target node A has associated nodes B, C, and D. The calculated synergistic influence factors are respectively , , If the threshold If only the influence factor of B on A exceeds the threshold, then there will be a directed edge from node B to node A in the Bayesian network, indicating that the state of node B affects the state of node A. The influences of C and D on A are insignificant, so no edges are added. Furthermore, the mutual influence between nodes can be considered; for example, if there are edges from A to B simultaneously (i.e., ... If the value is greater than the threshold, then a bidirectional dependency is formed. However, in most cases, the influence relationship is unidirectional.

[0040] In this way, previously isolated nodes establish logical causal dependencies through a collaborative weighted matrix, forming a Bayesian network structure capable of describing fault propagation and collaborative recovery mechanisms. For example, if a storage node has a high collaborative influence factor on a compute node, then there will be a directed edge in the network pointing from the storage node to the compute node, indicating that the health status of the storage node affects the self-repair success rate of the compute node.

[0041] S33. Using historical sample data, parameters of the Bayesian network are learned to obtain the conditional probability table of each node, thus forming a Bayesian network inference model.

[0042] In practice, after determining the structure of the Bayesian network, each node in the network needs to be equipped with a conditional probability table to quantify the strength of dependencies between nodes. The conditional probability table describes the probability that a node is in a certain state given the state of its parent node.

[0043] Parameter learning is performed using a large amount of sample data stored in a historical database. These historical samples contain state records of various nodes under different scenarios (such as normal operation, peak load, partial failure, network jitter, etc.). Each sample is a snapshot of the state of all relevant nodes at a given point in time, along with corresponding multi-dimensional feature data. By statistically analyzing this historical data, a conditional probability table for each node can be calculated.

[0044] Taking node A as an example, assume its parent node in the Bayesian network is B (according to step S32, B points to A). Both nodes A and B are categorized into three states: "healthy," "sub-healthy," and "faulty." Therefore, the conditional probability table for node A needs to include the probability of A being in any of the three states for each state of B. For example, based on historical data: when B is "healthy," the probability of A being "healthy" is 0.95, "sub-healthy" is 0.04, and "faulty" is 0.01; when B is "sub-healthy," the probability of A being "healthy" is 0.70, "sub-healthy" is 0.25, and "faulty" is 0.05; when B is "faulty," the probability of A being "healthy" is 0.10, "sub-healthy" is 0.30, and "faulty" is 0.60. These statistical results constitute the conditional probability table for node A.

[0045] Parameter learning can employ methods such as maximum likelihood estimation or Bayesian estimation. For scenarios with a large amount of historical data, maximum likelihood estimation is simple and effective; for sparse data, a prior distribution can be introduced for smoothing. After learning is complete, each node obtains a conditional probability table, and the Bayesian network at this point possesses inference capabilities, becoming a mature Bayesian network inference model.

[0046] As an example, the self-healing ability of target edge nodes is evaluated based on a weighted matrix and a Bayesian network inference model, resulting in a self-healing ability score, including: S34, arrange the multi-dimensional feature data of the current moment and multiple consecutive historical moments in chronological order to form a time-series feature dataset, and input the time-series feature dataset into the Bayesian network inference model to obtain the initial state probability distribution of the target edge node at each moment through probabilistic inference. In practical implementation, to achieve more accurate evaluation, this invention further introduces a time dimension, considering the dynamic characteristics of node state evolution over time. Multi-dimensional feature data of the target node, sampled at fixed time intervals (e.g., every 30 seconds) over a past period (e.g., the past 10 minutes), is extracted from the database. This data is then arranged in chronological order. (in Arrange them (at the current time) to form a time-series feature dataset.

[0047] The feature data from each time step in this dataset is sequentially input into a pre-trained static Bayesian network inference model for probabilistic inference. For each time step... The model calculates the posterior probability distribution of the target node's current state based on the observed features (i.e., multi-dimensional feature data) and the states of other nodes in the Bayesian network (this requires utilizing the state information of associated nodes, but in this step, only the feature observations of the target node itself can be considered, while the states of other nodes can be obtained through historical data or synchronous inference). This is achieved using the Bayesian formula. Here, it is assumed that at the inference time, the states of associated nodes can be obtained by synchronously inputting their own historical features into the model, or by using an approximation method, i.e., using the state probabilities from historical moments as input to the current moment.

[0048] For example, at time t1, given the multi-dimensional feature data of that time, the model outputs the state probability distribution of target node A as: healthy 0.92, sub-healthy 0.06, faulty 0.02. At time t2, given the feature data of t2, the output probability distribution is: healthy 0.88, sub-healthy 0.10, faulty 0.02. This continues until the current time tn, where the probability distribution is: healthy 0.45, sub-healthy 0.40, faulty 0.15. Thus, a sequence of state probability changes from the past to the present is obtained, denoted as... Each of them It is a three-dimensional vector (corresponding to the probabilities of the three states).

[0049] S35, determine the collaborative influence weight between the target edge node and the associated edge node according to the collaborative weighting matrix, construct the time transition probability matrix in the dynamic Bayesian network based on the collaborative influence weight, and combine the initial state probability distribution to obtain the state probability prediction sequence of the target edge node in the future preset time window using dynamic Bayesian network inference. In practice, the static Bayesian network is extended into a dynamic Bayesian network (DBN). The dynamic Bayesian network can model the evolution of the system state over time and usually consists of two parts: the initial state distribution (obtained from step S34) and the state transition model.

[0050] First, we need to define the time transition probability, which is the probability that a node will transition from one state to the next. This transition probability is not fixed but is influenced by the cooperative relationships between nodes. Based on the cooperative weighted matrix constructed above, we determine the time transition probability matrix of the target node A.

[0051] Let the state of target node A at time t be... Let the state space be {healthy, sub-healthy, faulty}. Then the state transition probability is... This depends on the current state s and the collaborative influence factors of associated nodes on A. One modeling approach is to use the corresponding factors in the collaborative weighting matrix as adjustment parameters to modify the base transition probabilities.

[0052] Specifically, a basic state transition matrix can be predefined. , representing the natural transition probability of a node's state in the absence of cooperative influence. For example, the basic transition matrix is:

[0053] Then, based on the collaborative influence factors of all associated nodes on A in the collaborative weighting matrix, a collaborative correction coefficient is calculated. For example, the average of the collaborative influence factors of all associated nodes on A can be used as... ,Right now .if A larger value indicates stronger support for A from the associated cluster. Therefore, A's self-recovery capability should be enhanced; that is, the probability of transitioning from a sub-healthy state to a healthy state should increase, while the probability of transitioning from a healthy state to a failure state should decrease. Thus, a function can be designed to adjust the basic transition matrix to obtain the actual time transition probability matrix. For example, the actual probability of transitioning from a sub-healthy state to a healthy state can be set as follows: , where β is the adjustment coefficient, and ensures that all transition probabilities are normalized.

[0054] Assume the average collaborative influence factor of target node A If the probability of transitioning from sub-health to health in the basic transition matrix is ​​0.20, then the adjusted probability might become 0.20×(1+0.5×0.75)=0.20×1.375=0.275 (assuming β=0.5). Simultaneously, the probability of transitioning from health to failure might be appropriately reduced.

[0055] After obtaining the time transition probability matrix, combine it with the current time obtained in step S34. Initial state probability distribution The future time step is calculated iteratively using inference algorithms (such as the forward algorithm) of dynamic Bayesian networks. The state probability distribution is given. Here, m is the prediction step number, corresponding to the length of a preset future time window. For example, if the next 5 minutes are predicted, and each step is taken every 30 seconds, then m = 10. The prediction process is as follows:

[0056]

[0057] And so on, to obtain the probability prediction sequence of future states. .

[0058] S36, calculate the self-healing ability score of the target edge node based on the state probability prediction sequence. The self-healing ability score is positively correlated with the probability that the node will maintain a normal state or recover to normal within a future time window.

[0059] In practice, the final self-healing ability score is calculated based on the state probability prediction sequence within the future time window obtained in step S35. The scoring logic is: the more likely a node is to remain in a "healthy" state in the future, or the more likely it is to quickly recover from a "sub-healthy" state to a "healthy" state, the higher its self-healing ability score will be.

[0060] For example, the probability of being in a "healthy" state at each future time point is taken, and its average value is calculated. For example, if the probabilities of being in a healthy state in the next 5 minutes (10 time points) are 0.60, 0.55, 0.50, 0.45, 0.40, 0.38, 0.35, 0.33, 0.30, and 0.28 respectively, then the average health probability is approximately 0.414, and the self-healing ability score can be set at 41.4 points (out of 100).

[0061] Additionally, the potential for recovery can be considered. For example, a self-healing ability score can be defined as... ,in As a discount factor, the more recent the time period, the higher the weight, reflecting a greater focus on the possibility of near-term recovery. If Therefore, the weighted sum of the above sequences is: 0.60 + 0.9 × 0.55 + 0.9 2 The score is obtained by multiplying by 0.50 and then by adding ...

[0062] Alternatively, it can be defined by combining the probability of transitioning from a sub-healthy state to a healthy state. For example, the score can be defined as the cumulative probability of a node transitioning from an unhealthy state (sub-healthy or faulty) to a healthy state within a future time window.

[0063] Assuming the target node is currently in a sub-healthy state, if a dynamic Bayesian network predicts that it has an 80% probability of recovering to a healthy state within the next 5 minutes, and once recovered, has a 90% probability of maintaining a stable healthy state, then its self-healing ability score will be high, for example, 85 points. Conversely, if it is predicted that its future state will continue to deteriorate, with a recovery probability of less than 10%, then its self-healing ability score will be less than 20 points.

[0064] Through the steps described above, the complex inter-node collaborative relationships and multi-dimensional feature data are transformed into an intuitive self-healing capability score. This score comprehensively reflects the self-recovery potential of a node under its own and the cluster's collaborative effects.

[0065] S4. When a preset type of abnormal trend appears at the target edge node, the self-healing ability score of the target node is compared with a preset threshold. If the self-healing ability score is lower than the preset threshold, the target node is identified as a faulty node and intervention is triggered on the target edge node.

[0066] In this embodiment, the abnormal trends detected in step S1 are combined with the self-healing ability score calculated in step S3 to make the final intervention decision. That is, not all anomalies are intervened immediately, but only those anomalies that cannot be recovered by the node itself and its surroundings are identified as real faults and triggered for processing.

[0067] Specifically, when the monitoring mechanism detects a pre-defined abnormal trend in a target edge node (such as excessively high resource utilization, abnormal load fluctuations, increased network latency, or abnormal inter-node collaboration), it does not immediately take any intervention measures against the node. Instead, it compares the calculated self-healing capability score with a pre-defined threshold. This threshold is a pre-configured value used to define the boundary between "self-healing" and "requiring intervention." The threshold setting can be dynamically adjusted based on various factors, including but not limited to: Business Importance Level: For edge nodes carrying core business functions (such as payment transactions and real-time control), the tolerance for business interruption is extremely low, so the threshold should be set high. For example, it can be set to 80 points, meaning that a node is only allowed to recover on its own if the probability of self-healing is higher than 80%; otherwise, immediate intervention is required. For nodes carrying non-core business functions (such as log collection and data backup), a longer period of self-recovery attempts can be tolerated, and the threshold can be set lower, such as 30 points.

[0068] System redundancy: If the cluster containing the target node has high redundancy (e.g., using an N+M backup mechanism), even if a single node fails, the service can be seamlessly taken over by other nodes. In this case, a lower intervention threshold can be set, giving the node more time to recover. Conversely, if the cluster redundancy is low, node failure may cause service impairment. Therefore, a higher intervention threshold should be set to allow for earlier intervention.

[0069] Historical Failure Modes: This feature learns the failure modes of different nodes in historical data. For nodes historically prone to permanent failures, the intervention threshold can be dynamically increased; for nodes historically with multiple successful self-healing events, the threshold can be appropriately decreased.

[0070] Timeframe: During peak business periods, to ensure system stability, the intervention threshold can be temporarily lowered (i.e., intervention is more easily triggered) to prevent nodes from being in a sub-optimal state for extended periods, thus affecting overall performance. During off-peak business periods, the intervention threshold can be raised to give nodes more opportunities to self-repair.

[0071] For example, suppose the system sets a default intervention threshold of 50 points for a video rendering cluster. When target node A is performing large-scale graphics rendering, its GPU utilization exceeds 95% for five consecutive minutes (an abnormal trend of "excessive resource utilization"), triggering an evaluation. Considering the collaborative relationship between node A and surrounding nodes B, C, and D, a self-healing capability score of 65 points is calculated. Since 65 points is higher than the threshold of 50 points, it is considered that although node A is currently overloaded, its surrounding cluster has sufficient idle GPU resources available for scheduling, and its task migration strategy is sound. Node A has a high probability of recovering itself through task migration. Therefore, it is not identified as a faulty node; only a log entry of "Node A overloaded, self-healing capability score 65, higher than the threshold, continue observation" is recorded, and its status changes are continuously monitored.

[0072] For example, suppose another target node D detects more than 5 interruptions in its heartbeat signal to its critical associated node E (belonging to the category of "abnormal inter-node collaboration"), triggering an evaluation process. Its self-healing capability score is calculated to be only 18 points. Model reasoning reveals that the core services on node D lack redundant backups, and network partitioning causes it to lose connection with the cluster, making it unable to recover on its own. Since 18 points is far below the threshold of 50 points, node D is officially identified as a "faulty node," immediately triggering intervention actions for that node.

[0073] Intervention actions can be a pre-configured series of automated scripts or operational processes, including but not limited to the following types: Alarm Notification: High-priority alarm information will be sent to on-duty maintenance personnel through multiple channels (such as SMS, email, WeChat Work, DingTalk, and telephone voice alarms). The alarm content should include: fault node identifier (such as IP address, hostname, and data center location), fault trigger time, anomaly trend type, self-healing capability score, and preliminary fault diagnosis information (such as anomaly indicators collected in S1). For example: "[High-Risk Alarm] Edge node D (IP: 192.168.1.100) was identified as a fault node at 14:23:05 on March 1, 2026. Anomaly type: Heartbeat interruption with associated nodes. Self-healing capability score: 18 points. Please handle immediately."

[0074] Node isolation: Automatically isolate the failed node from the service cluster, stopping the allocation of new computing tasks or service requests to that node. In container orchestration environments such as Kubernetes, isolation can be achieved by marking the node as "unschedulable." On a load balancer, the backend servers of the node can be marked as offline. This prevents the failed node from further impacting the overall service quality and avoids the spread of the fault.

[0075] Automatic restart: Attempt to perform a cold restart (i.e., restart the physical machine) or service restart (i.e., restart critical processes) on the failed node via a remote management interface (such as IPMI, BMC, SSH). The restart operation can be progressive; for example, first attempt a graceful restart of the service process, if that fails, perform a system restart, and if that still fails, attempt a hardware restart. Record the number of restart attempts and the results.

[0076] Service Migration and Reconstruction: For critical service instances running on nodes, resource scheduling interfaces are invoked to quickly rebuild these services on other healthy edge nodes. For example, in a containerized environment, Pods can be rescheduled to other available nodes; in a virtual machine environment, pre-prepared virtual machine images can be launched; and in a serverless architecture, functions can be triggered to execute in other regions. During service migration, data consistency and state synchronization issues need to be considered. Seamless service switching can be ensured through shared storage or data replication technologies.

[0077] Create maintenance work orders: Automatically create fault handling work orders to IT service management platforms (such as ServiceNow, Jira), submit all relevant information of the fault node (historical monitoring data, self-healing capability assessment report, current status, etc.) as attachments, and assign them to the corresponding maintenance team for manual investigation and handling.

[0078] Trigger standby node takeover: If the cluster adopts the active-standby mode, the standby node can be automatically triggered to become the primary node, switching the virtual IP or service traffic to the standby node, achieving business recovery in seconds.

[0079] Intervention actions can be a single operation or a combination of the above operations, executed sequentially according to a preset strategy. For example, an automatic restart can be attempted first. If the node returns to normal after the restart, isolation is lifted and its service role is restored; if the restart fails or the node remains abnormal after restarting, service migration is triggered and an alarm notification is sent.

[0080] Through the above mechanism, this invention achieves precise and intelligent management of edge nodes, significantly reducing the frequency of invalid alarms and manual intervention. Simultaneously, by assessing self-healing capabilities, it maximizes the utilization of the edge nodes' own recovery capabilities and the cluster's collaborative redundancy resources, improving resource utilization and operational efficiency while ensuring system stability.

[0081] Please see Figure 3 This invention also provides a data center distributed monitoring and management system 100 based on edge computing, applied to a data center, the system comprising: The data acquisition module 10 is used to collect the running status data of the target edge node in real time, as well as the collaborative working status between the target edge node and its associated edge nodes. Feature extraction module 20 is used to extract multi-dimensional feature data from the running status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes. The self-healing capability assessment module 30 is used to construct a weighted matrix and a Bayesian network inference model based on the multi-dimensional feature data, and to assess the self-healing capability of the target edge node based on the weighted matrix and the Bayesian network inference model to obtain a self-healing capability score. The intervention decision module 40 is used to compare the self-healing ability score of the target node with a preset threshold when a preset type of abnormal trend appears at the target edge node. If the self-healing ability score is lower than the preset threshold, the target node is identified as a faulty node and intervention is triggered on the target edge node.

[0082] As an example, the target edge node is a node that exhibits a preset type of abnormal trend during the monitoring process. The preset type of abnormal trend includes, but is not limited to, excessively high resource utilization, abnormal load fluctuations, increased network latency, and abnormal collaboration between nodes.

[0083] As an example, the feature extraction module 20 is configured to: Based on the CPU utilization, memory usage, bandwidth utilization, and load of the target edge node, assess the resource sharing capability between the target edge node and its associated edge nodes. Based on the communication latency, bandwidth usage, and data transmission rate between the target edge node and its associated edge nodes, the data transmission capability between the target edge node and its associated edge nodes is evaluated. Based on the load conditions and task allocation of the target edge node and associated edge nodes, evaluate the load balancing capability between the target edge node and associated edge nodes; Based on the redundancy task allocation, fault recovery strategy and resource scheduling capabilities of the target edge node and associated edge nodes, the collaborative repair capability between the target edge node and associated edge nodes is evaluated. The resource sharing capability, data transmission capability, load balancing capability, and collaborative repair capability are used as the multi-dimensional feature data.

[0084] As an example, the self-healing ability assessment module 30 is configured to: Based on the multi-dimensional feature data, the collaborative influence factor between the target edge node and each associated edge node is calculated, and a collaborative weighted matrix is ​​constructed based on the collaborative influence factor; wherein, the collaborative influence factor is used to characterize the degree of support or influence intensity of the associated edge node on the self-healing ability of the target edge node; Using the target edge node and its associated edge nodes as nodes of the Bayesian network, the dependencies between nodes are determined according to the collaborative weighting matrix, and a Bayesian network is constructed. By using historical sample data to learn the parameters of the Bayesian network, a conditional probability table for each node is obtained, forming a Bayesian network inference model.

[0085] As an example, the self-healing ability assessment module 30 is configured to: The multi-dimensional feature data of the current moment and multiple consecutive historical moments are arranged in chronological order to form a time-series feature dataset. The time-series feature dataset is then input into the Bayesian network inference model, and the probability distribution of the initial state of the target edge node at each moment is obtained through probabilistic inference. The collaborative influence weight between the target edge node and the associated edge node is determined based on the collaborative weighting matrix. The time transition probability matrix in the dynamic Bayesian network is constructed based on the collaborative influence weight. Combined with the initial state probability distribution, the state probability prediction sequence of the target edge node in the future preset time window is obtained by reasoning using the dynamic Bayesian network. The self-healing ability score of the target edge node is calculated based on the state probability prediction sequence. The self-healing ability score is positively correlated with the probability that the node will remain in a normal state or recover to normal within a future time window.

[0086] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding embodiments.

[0087] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A distributed monitoring and management method for data centers based on edge computing, applied to data centers, characterized in that, The method includes the following steps: S1 collects real-time operational status data of the target edge node, as well as the collaborative working status between the target edge node and its associated edge nodes; S2, Multi-dimensional feature data extracted from the running status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes; S3. Based on the multi-dimensional feature data, construct a weighted matrix and a Bayesian network inference model, evaluate the self-healing ability of the target edge node based on the weighted matrix and the Bayesian network inference model, and obtain a self-healing ability score. S4. When a preset type of abnormal trend appears at the target edge node, the self-healing ability score of the target node is compared with a preset threshold. If the self-healing ability score is lower than the preset threshold, the target node is identified as a faulty node and intervention is triggered on the target edge node. Based on the aforementioned multi-dimensional feature data, a collaborative weighted matrix and Bayesian network inference model are constructed, including: S31, based on the multi-dimensional feature data, calculate the collaborative influence factor between the target edge node and each associated edge node, and construct a collaborative weighted matrix based on the collaborative influence factor; wherein, the collaborative influence factor is used to characterize the degree of support or influence intensity of the associated edge node on the self-healing ability of the target edge node; S32, using the target edge node and its associated edge nodes as nodes of the Bayesian network, the dependency relationship between nodes is determined according to the collaborative weighting matrix, and the Bayesian network is constructed. S33. Using historical sample data, parameters of the Bayesian network are learned to obtain the conditional probability table of each node, thus forming a Bayesian network inference model.

2. The data center distributed monitoring and management method based on edge computing according to claim 1, characterized in that: The target edge node is a node that exhibits a preset type of abnormal trend during the monitoring process. The preset type of abnormal trend includes, but is not limited to, excessively high resource utilization, abnormal load fluctuations, increased network latency, and abnormal collaboration between nodes.

3. The data center distributed monitoring and management method based on edge computing according to claim 1, characterized in that: Multi-dimensional feature data extracted from the operational status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes includes: Based on the CPU utilization, memory usage, bandwidth utilization, and load of the target edge node, assess the resource sharing capability between the target edge node and its associated edge nodes. Based on the communication latency, bandwidth usage, and data transmission rate between the target edge node and its associated edge nodes, the data transmission capability between the target edge node and its associated edge nodes is evaluated. Based on the load conditions and task allocation of the target edge node and associated edge nodes, evaluate the load balancing capability between the target edge node and associated edge nodes; Based on the redundancy task allocation, fault recovery strategy and resource scheduling capabilities of the target edge node and associated edge nodes, the collaborative repair capability between the target edge node and associated edge nodes is evaluated. The resource sharing capability, the data transmission capability, the load balancing capability, and the collaborative repair capability are used as the multi-dimensional feature data.

4. The data center distributed monitoring and management method based on edge computing according to claim 1, characterized in that: The self-healing capability of target edge nodes is evaluated based on a weighted matrix and a Bayesian network inference model, resulting in a self-healing capability score, including: S34, arrange the multi-dimensional feature data of the current moment and multiple consecutive historical moments in chronological order to form a time-series feature dataset, and input the time-series feature dataset into the Bayesian network inference model to obtain the initial state probability distribution of the target edge node at each moment through probabilistic inference. S35, determine the collaborative influence weight between the target edge node and the associated edge node according to the collaborative weighting matrix, construct the time transition probability matrix in the dynamic Bayesian network based on the collaborative influence weight, and combine the initial state probability distribution to obtain the state probability prediction sequence of the target edge node in the future preset time window using dynamic Bayesian network inference. S36, calculate the self-healing ability score of the target edge node based on the state probability prediction sequence. The self-healing ability score is positively correlated with the probability that the node will maintain a normal state or recover to normal within a future time window.

5. A data center distributed monitoring and management system based on edge computing, applied to a data center, characterized in that: The system includes: The data acquisition module is used to collect real-time operational status data of the target edge nodes, as well as the collaborative working status between the target edge nodes and their associated edge nodes. The feature extraction module is used to extract multi-dimensional feature data from the running status data and the collaborative working status to characterize the collaborative working relationship between the target edge node and its associated edge nodes. The self-healing capability assessment module is used to construct a weighted matrix and a Bayesian network inference model based on the multi-dimensional feature data, and to assess the self-healing capability of the target edge node based on the weighted matrix and the Bayesian network inference model to obtain a self-healing capability score. The intervention decision module is used to compare the self-healing ability score of the target node with a preset threshold when a preset type of abnormal trend appears at the target edge node. If the self-healing ability score is lower than the preset threshold, the target node is identified as a faulty node and intervention is triggered on the target edge node. The self-healing ability assessment module is configured to: Based on the multi-dimensional feature data, the collaborative influence factor between the target edge node and each associated edge node is calculated, and a collaborative weighted matrix is ​​constructed based on the collaborative influence factor; wherein, the collaborative influence factor is used to characterize the degree of support or influence intensity of the associated edge node on the self-healing ability of the target edge node; Using the target edge node and its associated edge nodes as nodes of the Bayesian network, the dependencies between nodes are determined according to the collaborative weighting matrix, and a Bayesian network is constructed. By using historical sample data to learn the parameters of the Bayesian network, a conditional probability table for each node is obtained, forming a Bayesian network inference model.

6. The data center distributed monitoring and management system based on edge computing according to claim 5, characterized in that: The target edge node is a node that exhibits a preset type of abnormal trend during the monitoring process. The preset type of abnormal trend includes, but is not limited to, excessively high resource utilization, abnormal load fluctuations, increased network latency, and abnormal collaboration between nodes.

7. A data center distributed monitoring and management system based on edge computing according to claim 5, characterized in that: The feature extraction module is configured to: Based on the CPU utilization, memory usage, bandwidth utilization, and load of the target edge node, assess the resource sharing capability between the target edge node and its associated edge nodes. Based on the communication latency, bandwidth usage, and data transmission rate between the target edge node and its associated edge nodes, the data transmission capability between the target edge node and its associated edge nodes is evaluated. Based on the load conditions and task allocation of the target edge node and associated edge nodes, evaluate the load balancing capability between the target edge node and associated edge nodes; Based on the redundancy task allocation, fault recovery strategy and resource scheduling capabilities of the target edge node and associated edge nodes, the collaborative repair capability between the target edge node and associated edge nodes is evaluated. The resource sharing capability, the data transmission capability, the load balancing capability, and the collaborative repair capability are used as the multi-dimensional feature data.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-4.

9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-4.

Images