Server failure prediction method and device, electronic equipment and storage medium

Abstract

The application discloses a server fault prediction method and device, electronic equipment and storage medium, and relates to the technical field of servers, comprising: by constructing a component dependency graph, the dependency relationship between components is clearly pointed out, so that the fault prediction process can more accurately identify the path of fault propagation. And based on the running state sequence of the server component to determine the state feature vector, the running state change of each component in the target period can be captured, which helps to improve the identification ability of complex fault modes. In addition, by dynamically adjusting the connection weight between nodes in the component dependency graph to form a target dependency graph model, finally, based on the fault prediction result determined by the target dependency graph model, the abnormal components that may exist in the server can be clearly pointed out, fully considering the dependency relationship between components, improving the accuracy of the obtained fault prediction result, and thus solving the problem of low accuracy of the fault prediction result of the server.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, electronic device and storage medium for predicting server failures. Background Technology

[0002] In the field of server operation and maintenance management, fault prediction technology is of great significance for ensuring the stable operation of systems and improving operational efficiency. Most server fault prediction methods in related technologies are based on a single data source, such as performance monitoring data or system logs, using threshold detection or simple machine learning models to identify potential fault risks. However, these methods fail to fully consider the interdependencies between the various hardware components within the server, resulting in low accuracy in predicting server faults. Summary of the Invention

[0003] This application provides a method, apparatus, electronic device, and storage medium for predicting server failures, in order to at least address the problem of low accuracy in predicting server failures in related technologies.

[0004] This application provides a server fault prediction method, comprising: constructing a component dependency graph based on the hardware topology of a target server, wherein the component dependency graph includes multiple interconnected nodes, the nodes indicating a server component in the target server, and the connection relationships between the nodes indicating the dependency relationships between the multiple server components; determining a state feature vector for each of at least one server component based on its respective operating state sequence, wherein the operating state sequence indicates the change in the operating state of the server component within a target period; determining the connection weights between the multiple nodes in the component dependency graph based on the at least one state feature vector to obtain a target component dependency graph, wherein the connection weights indicate the path of state propagation of fault states among the multiple server components; and determining a fault prediction result based on the target component dependency graph, wherein the fault prediction result indicates at least one abnormal component among the multiple server components.

[0005] This application also provides a server fault prediction device, comprising: a construction unit for constructing a component dependency graph based on the hardware topology of a target server, wherein the component dependency graph includes multiple interconnected nodes, the nodes indicating a server component in the target server, and the connection relationships between the nodes indicating the dependency relationships between the multiple server components; a first determination unit for determining a state feature vector of at least one server component based on the respective operating state sequence of at least one server component in the target server, wherein the operating state sequence indicates the change of the operating state of the server component within a target period; a second determination unit for determining the connection weights between the multiple nodes in the component dependency graph based on at least one state feature vector to obtain a target component dependency graph, wherein the connection weights indicate the path of state propagation of fault states among the multiple server components; and a third determination unit for determining a fault prediction result based on the target component dependency graph, wherein the fault prediction result indicates at least one abnormal component among the multiple server components.

[0006] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described server fault prediction methods.

[0007] This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of any of the above-described server fault prediction methods.

[0008] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described server fault prediction methods.

[0009] This application constructs a component dependency graph that not only includes representation nodes of server components but also explicitly indicates the dependencies between components. This enables the fault prediction process to more accurately identify fault propagation paths. Furthermore, determining state feature vectors based on the operational state sequences of server components captures the operational state changes of each component within a target period, which helps improve the ability to identify complex fault modes. In addition, by dynamically adjusting the connection weights between nodes in the component dependency graph to form a target dependency graph model, it can better adapt to changes in the server's operating environment, identify actual fault propagation paths, and thus improve the accuracy of prediction results. Finally, the fault prediction results determined based on the target dependency graph model can clearly identify potentially abnormal components in the server, fully considering the dependencies between components, improving the accuracy of the obtained fault prediction results, and thus solving the problem of low accuracy in server fault prediction. Attached Figure Description

[0010] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A hardware structure block diagram of a server device for a server fault prediction method provided in an embodiment of this application;

[0012] Figure 2 This is a schematic diagram of one of the optional server fault prediction methods according to an embodiment of this application;

[0013] Figure 3 This is a second schematic diagram of an optional server fault prediction method according to an embodiment of this application;

[0014] Figure 4 This is a third schematic diagram of an optional server fault prediction method according to an embodiment of this application;

[0015] Figure 5 This is a structural block diagram of a server fault prediction device according to an embodiment of this application;

[0016] Figure 6 This is a schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0018] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0019] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The methods and embodiments provided in this application can be executed on a server device or a similar computing device. Taking running on a server device as an example, Figure 1 This is a hardware structure block diagram of a computer device for a server fault prediction method according to an embodiment of this application. Figure 1 As shown, the server device may include one or more ( Figure 1 Only one is shown in the image. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The server device may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the server equipment described above. For example, the server equipment may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0021] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the server fault prediction method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thus implementing the above-described method. The memory 104 may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the server device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0022] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by a communication provider for the server device. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0023] This embodiment provides a method for predicting server failures. Figure 2 This is a flowchart of a server fault prediction method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps:

[0024] S202, Build a component dependency graph based on the hardware topology of the target server. The component dependency graph includes multiple interconnected nodes. The nodes are used to indicate a server component in the target server, and the connection relationship between the nodes is used to indicate the dependency relationship between multiple server components.

[0025] Optionally, in this embodiment, the target server may refer to, but is not limited to, a specific server device for which fault prediction is to be performed. For example, a rack server in a company's data center used to run the core business system, whose hardware configuration includes 2 central processing units, 64GB of memory, 10 enterprise-grade solid-state drives, and 2 redundant power supplies, can be identified as the target server.

[0026] Optionally, in this embodiment, the hardware topology may refer to, but is not limited to, the physical connection methods and logical relationships between various hardware components within the target server. It reflects the relational architecture between components, serves as the basis for constructing the component dependency graph, and ensures that the dependency graph conforms to the actual hardware layout of the server. For example, in the target server, the CPU is connected to the memory module via a processor socket on the motherboard, the memory module is connected to the motherboard chipset via a memory slot, the hard drive is connected to the storage controller on the motherboard via a SATA interface, and the power supply unit supplies power to components such as the CPU, memory, and hard drive via power lines. These physical connections collectively constitute the hardware topology of the server.

[0027] Optionally, in this embodiment, the component dependency graph can be, but is not limited to, a graph-based data structure, consisting of nodes representing server components and edges representing dependencies between components, used to visually represent the relationships between components. For example, for the target server mentioned above, a component dependency graph can be constructed, where the nodes are CPU1, CPU2, hard disk, power supply, etc., and the edges are set according to the hardware topology.

[0028] Optionally, in this embodiment, the dependency relationship may be, but is not limited to, an objectively existing association logic between components formed based on the server hardware topology. This includes physical connection dependencies, such as circuit connections and power supply line connections between components, as well as signal transmission dependencies, such as data interaction and control signal transmission paths between components.

[0029] Optionally, in this embodiment, server components may refer to, but are not limited to, various hardware units that constitute the target server. For example, the central processing unit, memory module, solid-state drive, power supply unit, cooling fan, etc., in the target server are all server components, and the operating status of each component directly affects the overall performance and stability of the server.

[0030] S204, Based on the respective operating state sequences of at least one server component in the target server, determine the respective state feature vectors of at least one server component, wherein the operating state sequence is used to indicate the changes in the operating state of the server component within the target period.

[0031] Optionally, in this embodiment, the running state sequence may refer to, but is not limited to, a set of server component running state data recorded in chronological order within the target period, which can reflect the temporal change pattern of the component running state.

[0032] Optionally, in this embodiment, the state feature vector can be, but is not limited to, a vector extracted from the server component running state sequence that can characterize the key features of the component running state. It can transform time series data into a feature form that is easy to calculate and analyze, providing data support for determining the connection weights of the component dependency graph.

[0033] To illustrate further, for the utilization operation status sequence of the target CPU in the target server within 1 hour, the vector obtained by calculating indicators such as maximum value, minimum value, average value, variance, and trend slope can be determined as the state feature vector of the target CPU.

[0034] S206, determine the connection weights between multiple nodes in the component dependency graph based on at least one state feature vector to obtain the target component dependency graph, wherein the connection weights are used to indicate the path of state propagation of fault states between multiple server components;

[0035] Optionally, in this embodiment, the connection weight may refer to, but is not limited to, the weight value of the connection edge between nodes in the component dependency graph. It is used to quantify the probability of a fault state propagating between different server components and is a core parameter for constructing the target component dependency graph, which can accurately identify the fault propagation path.

[0036] Optionally, in this embodiment, the target component dependency graph can be, but is not limited to, a graph structure formed by assigning connection weights to the connection edges between nodes based on the original component dependency graph, which can reflect the propagation characteristics of faults between components.

[0037] S208, Based on the target component dependency graph, determine the fault prediction result, wherein the fault prediction result is used to indicate at least one abnormal component among multiple server components.

[0038] Optionally, in this embodiment, the fault prediction result may refer to, but is not limited to, the component information indicating anomalies in the server obtained through target component dependency graph analysis. This is the final output of the fault prediction method, providing maintenance personnel with clear fault location information and facilitating timely fault handling. For example, by analyzing the target component dependency graph of the target server, it is found that the state feature vector of memory module A is abnormal, and its connection weight with the target CPU is high. It is determined that memory module A has a fault risk and may affect the operation of the target CPU. The output fault prediction result may include, but is not limited to, the abnormal component being memory module A and the potentially affected component being the target CPU.

[0039] Optionally, in this embodiment, abnormal components may refer to, but are not limited to, server components that are determined to be in abnormal operating status or have a risk of failure in the fault prediction results. These are the key targets of operation and maintenance personnel and can help them quickly locate the source of the fault.

[0040] Optionally, in this embodiment, a component dependency graph consisting of nodes and connecting edges is constructed based on the actual hardware connection method and logical relationship of the target server. Nodes correspond to server components, and connecting edges correspond to the communication connection relationship between components.

[0041] Next, for one or more server components in the target server, the running status data of each component recorded in chronological order within the target period is obtained as a running status sequence. Then, through data processing and feature extraction, the state feature vector corresponding to each component is obtained.

[0042] Then, using the state feature vectors of one or more server components that have been acquired, the weight values ​​of the connection edges between nodes in the component dependency graph are calculated through a specific algorithm, and the original component dependency graph is upgraded into a target component dependency graph with connection weights. The magnitude of the connection weights can reflect the possibility of a fault propagating between the corresponding components.

[0043] It is best to utilize the target component dependency graph, combined with anomaly analysis of component state feature vectors, and through fault propagation path analysis, risk assessment, and other methods, to identify abnormal components in the server and form fault prediction results.

[0044] It should be noted that by constructing a component dependency graph that closely matches the actual layout of the server hardware, a structured foundation is provided for fault propagation analysis, ensuring that the fault analysis conforms to the hardware association logic. By extracting the state feature vectors of the component running state sequences, time-series data is transformed into a quantifiable and analyzable feature form, providing data support for assessing the probability of fault propagation. By calculating the connection weights between nodes in the component dependency graph, the target component dependency graph is obtained, accurately quantifying the probability of fault propagation between components and clarifying the fault propagation path. By identifying abnormal components through the target component dependency graph and outputting fault prediction results, precise fault location information is provided to maintenance personnel. This achieves the goals of improving the accuracy of server fault prediction, clarifying fault propagation paths, shortening fault troubleshooting time, and improving server maintenance efficiency and reliability.

[0045] Through the embodiments of this application, by constructing a component dependency graph, which not only includes representation nodes of server components but also explicitly indicates the dependencies between components, the fault prediction process can more accurately identify fault propagation paths. Furthermore, determining state feature vectors based on the operational state sequences of server components can capture the operational state changes of each component within a target period, including short-term anomalies and long-term degradation, which helps improve the ability to identify complex fault modes. In addition, by dynamically adjusting the connection weights between nodes in the component dependency graph to form a target dependency graph model, it can better adapt to changes in the server operating environment, identify actual fault propagation paths, and thus improve the accuracy of prediction results. Finally, the fault prediction results determined based on the target dependency graph model can clearly identify potentially abnormal components in the server, fully considering the dependencies between components and improving the accuracy of the obtained fault prediction results.

[0046] As an optional approach, based on the target component dependency graph, the fault prediction results are determined, including:

[0047] Obtain the failure probability coefficient for each node in the target component dependency graph. The failure probability coefficient indicates the probability that the server component corresponding to the node will fail.

[0048] Traverse each node in the target component dependency graph, and if the failure probability coefficient of the current node satisfies the probability condition, determine the current node as the target node and the server component corresponding to the current node as the abnormal component.

[0049] Optionally, in this embodiment, the failure probability coefficient may be, but is not limited to, a value calculated based on the state feature vector of the server component and the propagation state of the nodes in the target component dependency graph. It is used to quantify the likelihood of the corresponding server component failing and is a quantitative indicator for judging whether the component is abnormal.

[0050] Optionally, in this embodiment, the probability condition may be, but is not limited to, a threshold condition set based on server operation and maintenance requirements, the degree of impact of component failure, and historical failure data, used to determine whether the failure probability coefficient reaches the component abnormality standard. This condition serves as the basis for distinguishing between normal and abnormal components, ensuring the objectivity and consistency of abnormal component identification.

[0051] Optionally, in this embodiment, the target node may be, but is not limited to, the node whose corresponding failure probability coefficient satisfies a preset probability condition during the traversal of the target component dependency graph nodes.

[0052] Optionally, in this embodiment, based on the running state sequence of the server components corresponding to each node in the target component dependency graph, the second-level fluctuation and long-term trend features are extracted through a dual-channel time series analysis architecture and fused into a state feature vector. Then, combined with the state of the nodes after multiple rounds of propagation in the graph attention mechanism, a nonlinear transformation is used to calculate the failure probability coefficient corresponding to each node, which quantifies the possibility of component failure.

[0053] Next, all nodes in the target component dependency graph are accessed in a preset order, and the failure probability coefficient of each node is compared with the preset probability condition of the component corresponding to that node. If the failure probability coefficient reaches or exceeds the probability condition threshold, the node is marked as the target node, and the server hardware component associated with the target node is identified as an abnormal component with a failure risk.

[0054] The embodiments provided in this application first calculate the failure probability coefficient of each node in the target component dependency graph, transforming the component failure risk into a quantifiable numerical indicator. Then, the nodes are traversed and the target nodes and abnormal components are filtered according to the probability conditions, ensuring the quantification, comprehensiveness and objectivity of abnormal component identification. This achieves the technical effect of accurately locating high-risk abnormal components of the server, avoiding the omission of potential faults, reducing human subjective judgment errors, and providing clear fault handling targets for operation and maintenance personnel.

[0055] As an optional approach, obtain the failure probability coefficient corresponding to each node in the target component dependency graph, including:

[0056] Obtain the state vector corresponding to the first node in the target component dependency graph, where the state vector is used to indicate the running state of the first node;

[0057] By performing a nonlinear transformation on the state vector, the failure probability coefficient corresponding to the first node is obtained.

[0058] Optionally, in this embodiment, the first node may be, but is not limited to, any node in the target component dependency graph for which the failure probability coefficient is to be calculated, and is the specific associated object of the state vector and the failure probability coefficient.

[0059] Optionally, in this embodiment, the state vector may be, but is not limited to, a vector used to comprehensively quantify and reflect the operating state of the component at different time scales.

[0060] Optionally, in this embodiment, the nonlinear transformation can be, but is not limited to, a process of performing mathematical operations on the node state vector through a preset learnable weight vector and activation function, which can map the state vector to a probability value in the [0,1] interval, thereby realizing the transformation from component operating state characteristics to fault probability.

[0061] Optionally, in this embodiment, firstly, the first node to be analyzed is located in the target component dependency graph, and the state vector of the first node is obtained; secondly, a preset nonlinear transformation formula is used to transform the state vector into a fault probability coefficient in the [0,1] interval, thereby quantifying the fault risk of the component corresponding to the first node.

[0062] The embodiments provided in this application achieve the technical effect of accurately quantifying the failure risk of a single server component by first obtaining the state vector of the first node in the target component dependency graph and then performing a nonlinear transformation on the state vector.

[0063] As an optional approach, obtain the state vector corresponding to the first node in the target component dependency graph, including:

[0064] The state feature vector corresponding to the first node is determined as the state vector of the first node;

[0065] Perform at least one state propagation on the target component dependency graph, and in each state propagation, perform the following operations:

[0066] Calculate the first attention score for each of the first node and at least one second node, where the second node is dependent on the first node, and the first attention score is used to indicate the degree of influence of the first node on the second node;

[0067] Based on at least one first attention score, the state vectors of at least one second node are weighted and summed to obtain the neighbor state vectors.

[0068] The state vector of the first node is updated based on the state vectors of its neighbors.

[0069] Optionally, in this embodiment, the state vector may be, but is not limited to, a vector used to comprehensively characterize the operating state of the component corresponding to the first node. It is initially determined by the state feature vector and, after multiple state propagation updates, incorporates the influence of the second node, thus more accurately reflecting the actual operating state of the component.

[0070] Optionally, in this embodiment, state propagation can be, but is not limited to, in the target component dependency graph, based on the inter-node dependency relationship and attention score, the process of passing the state vector information of the second node to the first node and updating the state vector of the first node. This is the core mechanism for integrating the mutual influence between components, so that the state vector can reflect the changes in the running state brought about by the component association.

[0071] Optionally, in this embodiment, the second node may be, but is not limited to, a node in the target component dependency graph that has a dependency relationship with the first node, corresponding to other hardware components of the server, and a neighboring node in the state propagation process of the first node, providing state information of associated components for the update of the state vector of the first node, and ensuring the mutual influence between the state vector fusion components.

[0072] Optionally, in this embodiment, the first attention score may be, but is not limited to, a value calculated by a graph attention mechanism algorithm based on the state vectors of the first node and the second node, and is used to quantify the degree of influence of the first node on the second node.

[0073] Optionally, in this embodiment, the neighbor state vector may be, but is not limited to, a vector obtained by weighted summation of the state vectors of all second nodes of the first node based on the first attention score.

[0074] Optionally, in this embodiment, firstly, the state feature vector generated by multi-source data processing and multi-scale feature extraction of the component corresponding to the first node is directly determined as the initial state vector of the first node; then, at least one state propagation is performed on the target component dependency graph. In each propagation, the first attention score between the first node and the associated second node is calculated first, and then the state vector of the second node is weighted and summed based on the score to obtain the neighbor state vector. Finally, the state vector of the first node is updated using the neighbor state vector, and finally the state vector of the first node is obtained by fusing its own features and the influence of associated components.

[0075] Through the embodiments provided in this application, by first determining the state feature vector of the first node as the initial state vector, and then through at least one state propagation, the influence of the second node on the first node is integrated, thereby achieving the technical effect of enabling the state vector of the first node to fully reflect the component's own operating characteristics and the influence of related components, improving the accuracy of state vector representation, and providing accurate input for the calculation of fault probability coefficients.

[0076] As an optional approach, determining fault prediction results based on the target component dependency graph also includes:

[0077] Calculate the second attention score corresponding to the third node and at least one fourth node in the target component dependency graph, wherein there is a dependency relationship between the server component corresponding to the third node and the server component corresponding to the fourth node, and the second attention score is used to indicate the degree of influence of the third node on the fourth node;

[0078] The edge between the fourth node corresponding to the second attention score that satisfies the attention condition and the third node is determined as the fault propagation path, where the fault prediction result includes the fault propagation path.

[0079] Optionally, in this embodiment, the second attention score may be, but is not limited to, a value calculated by a graph attention mechanism algorithm based on the state vectors of the third node and the fourth node, used to quantify the degree of influence of the component corresponding to the third node on the component corresponding to the fourth node. The larger the value, the higher the probability that the failure of the third node will cause the failure of the fourth node.

[0080] Optionally, in this embodiment, the attention condition may be, but is not limited to, a threshold condition set based on server operation and maintenance experience, historical fault data and component fault impact level, used to determine whether the second attention score reaches the standard of high fault propagation probability, and is the basis for distinguishing between ordinary dependent connection edges and fault propagation paths.

[0081] Optionally, in this embodiment, the fault propagation path may be, but is not limited to, the connection edge between the third node and the fourth node that satisfies the attention condition in the target component dependency graph, which is a high-probability path for the fault to propagate from the component corresponding to the third node to the component corresponding to the fourth node.

[0082] Optionally, in this embodiment, firstly, in the target component dependency graph, based on the dependency relationship between the third node and the fourth node, the state vectors of the two are obtained, and the second attention score of the third node and each fourth node is calculated by the graph attention mechanism algorithm to quantify the degree of influence of the third node on the fourth node; secondly, the second attention score is compared with the preset attention conditions, the fourth node that meets the conditions is selected, the connection edge between the fourth node and the third node is determined as the fault propagation path, and the path is integrated with the abnormal component and the fault probability coefficient to form a complete fault prediction result, and finally the diagnostic information containing the fault propagation path is output.

[0083] The embodiments provided in this application first calculate the second attention scores of the third and fourth nodes based on the graph attention mechanism, then filter and determine the fault propagation path according to the attention conditions, and incorporate the fault propagation path into the fault prediction result, thereby achieving the technical effect of revealing the potential path of server fault propagation from abnormal components to related components.

[0084] As an optional approach, before determining the state feature vectors of at least one server component based on the respective running state sequences of at least one server component in the target server, the method further includes:

[0085] Acquire multi-source sensor data generated by the target server during operation using at least one sensor;

[0086] According to the network time protocol, the timestamps of multi-source sensor data are aligned.

[0087] Resampling is performed on data with different sampling frequencies from multi-source sensor data to obtain a multi-dimensional state vector sequence;

[0088] The multidimensional state vector corresponding to each time step in the multidimensional state vector sequence is assigned to a node in the component dependency graph to obtain the running state sequence.

[0089] Optionally, in this embodiment, the multi-source sensor data can be, but is not limited to, a set of operational data collected by at least one sensor from different hardware components or different monitoring dimensions of the target server, covering performance indicators, health parameters, and environmental variables, etc. It is the original data source for generating multi-dimensional state vector sequences and operational state sequences, and can comprehensively reflect the operational status of server components.

[0090] Optionally, in this embodiment, the Network Time Protocol (NTP) may be, but is not limited to, the Internet Protocol used to synchronize computer system clocks, which can calibrate the timestamps of data collected by different sensors to a unified time reference, ensuring the consistency of data in the time dimension.

[0091] Optionally, in this embodiment, timestamp alignment can be, but is not limited to, a process based on network time protocols to adjust the acquisition timestamps of multi-source sensor data to the same time scale and precision, ensuring that sensor data from different sources and of different types can correspond in the time dimension, laying a time consistency foundation for subsequent resampling and vector allocation.

[0092] Optionally, in this embodiment, the sampling frequency may be, but is not limited to, the time interval between sensor data acquisition, i.e., the number of times data is acquired per unit time. Different types of sensors have different sampling frequencies due to monitoring requirements and hardware characteristics, which is the core basis for performing resampling operations.

[0093] Optionally, in this embodiment, the resampling operation may be, but is not limited to, a process of adjusting data streams with different sampling frequencies in multi-source sensor data to a uniform sampling frequency through interpolation or downsampling, to ensure that all data have the same time step and provide data in a uniform format for generating multi-dimensional state vector sequences.

[0094] Optionally, in this embodiment, the multidimensional state vector sequence may be, but is not limited to, a set of vectors arranged in chronological order after timestamp alignment and resampling, each time step containing multidimensional sensor data, which can comprehensively reflect the operating status of multiple components or multiple monitoring dimensions of the target server at each time point, and is a direct data carrier allocated to the component dependency graph nodes.

[0095] Optionally, in this embodiment, the running state sequence may be, but is not limited to, a time-series state data set exclusive to each node formed by allocating the multidimensional state vector of each time step in the multidimensional state vector sequence to the corresponding node according to the node division of the component dependency graph. It can reflect the running state changes of a single component at different time steps and is the core data for subsequent extraction of state feature vectors and calculation of fault probability coefficients.

[0096] Optionally, in this embodiment, firstly, multi-source sensor data generated during the operation of the target server is collected by at least one sensor; secondly, the multi-source data is timestamped based on the network time protocol to eliminate time inconsistencies caused by sensor clock deviations; next, resampling is performed on data with different sampling frequencies to unify all data to the same sampling frequency and generate a multi-dimensional state vector sequence; finally, according to the node division of the component dependency graph, the multi-dimensional state vector of each time step is assigned to the corresponding node to form a unique running state sequence for each node.

[0097] The embodiments provided in this application first utilize sensors to collect multi-source data, then use NTP to achieve timestamp alignment and resampling to generate a unified format of multi-dimensional state vector sequence, and finally distribute the data to the component dependency graph nodes to obtain the running state sequence, thereby achieving the technical effect of providing targeted and time-series running state data for each node in the component dependency graph.

[0098] As an optional approach, based on the respective operating state sequences of at least one server component in the target server, the state feature vectors of at least one server component are determined, including:

[0099] The running state sequence is input into a temporal convolutional network to obtain a short-time feature vector output by the temporal convolutional network. The short-time feature vector is used to predict the possible failure of at least one server component. The state feature vector includes the short-time feature vector.

[0100] The running state sequence is input into the sequence modeling network to obtain the long-term feature vector output by the sequence modeling network. The long-term feature vector is used to predict the remaining lifetime of at least one server component. The state feature vector includes the long-term feature vector.

[0101] Obtain the first weight value corresponding to the short-term feature vector, where the first weight value is used to indicate the importance of the short-term feature vector;

[0102] Obtain the second weight value corresponding to the long-term feature vector, where the second weight value is used to indicate the importance of the long-term feature vector;

[0103] Based on the first and second weight values, the short-term feature vector and the long-term feature vector are weighted and fused to obtain the state feature vector.

[0104] Optionally, in this embodiment, the Temporal Convolutional Network (TCN) can be, but is not limited to, a deep learning model for processing time-series data. Through a multi-layer causal dilated convolutional structure, it can effectively capture short-term local dependencies and transient abnormal patterns in time-series data, ensuring accurate identification of sudden abnormal features during component operation.

[0105] Optionally, in this embodiment, the short-term feature vector may be, but is not limited to, a vector output by a temporal convolutional network after processing the running state sequence, containing information such as the instantaneous fluctuation characteristics and sudden abnormal patterns of the component within a short time scale, and is mainly used to predict the short-term failures that may occur in the server component.

[0106] Optionally, in this embodiment, the sequence modeling network can be, but is not limited to, a network model that is good at capturing long-term dependencies in time series data. Through memory units or historical state accumulation mechanisms, it can extract long-term trends and gradual change patterns in time series data. It is the core model for extracting long-term feature vectors and provides long-term state basis for predicting the remaining lifetime of components.

[0107] Optionally, in this embodiment, the long-term feature vector may be, but is not limited to, a vector output by the sequence modeling network after processing the running state sequence, containing information such as the long-term trend characteristics and gradual degradation patterns of the component over a long time scale.

[0108] Optionally, in this embodiment, the first weight value may be, but is not limited to, a value used to quantify the importance of short-term feature vectors in state feature vectors; alternatively, in this embodiment, the second weight value may be, but is not limited to, a value used to quantify the importance of long-term feature vectors in state feature vectors.

[0109] Optionally, in this embodiment, firstly, the running state sequence of the server component is input into the temporal convolutional network and the sequence modeling network respectively. The TCN outputs a short-term feature vector containing short-term transient anomaly features, and the sequence modeling network outputs a long-term feature vector containing long-term trend features. Secondly, according to the component type, running scenario, and historical data, the first weight value of the short-term feature vector and the second weight value of the long-term feature vector are obtained respectively. Finally, the short-term and long-term feature vectors are fused by weighted summation to generate a state feature vector that simultaneously covers short-term anomalies and long-term trends.

[0110] To further illustrate, suppose the CPU1 component of the target server is running under high load conditions for core business operations. Historical data shows that 80% of failures in this scenario are caused by short-term performance spikes or instantaneous temperature anomalies. Based on this, the first weight value of the short-term feature vector of CPU1 is 0.65, indicating that in the state feature vector, the importance of short-term abnormal features is higher than that of long-term trend features. If CPU1 is running under low load conditions, the first weight value is adjusted to 0.35 to reduce the impact of short-term fluctuations.

[0111] The embodiments provided in this application first use a temporal convolutional network to extract short-term feature vectors and a sequence modeling network to extract long-term feature vectors. Then, the weight values ​​of the two types of feature vectors are dynamically determined according to the needs of the scenario. Finally, the weighted fusion is used to obtain the state feature vector, thereby achieving the technical effect of making the state feature vector simultaneously cover the short-term anomaly and long-term degradation information of the component, taking into account the needs of fault prediction and remaining life assessment, and improving the accuracy of the comprehensive characterization of the component's operating status.

[0112] As an optional approach, after determining the fault prediction results based on the target component dependency graph, the method also includes:

[0113] The fault prediction results are matched with historical fault repair data to obtain the target repair operation. The historical fault repair data is used to indicate the repair operations performed on the target server. The target repair operation is used to prevent possible anomalies that may occur during the operation of the target server.

[0114] Optionally, in this embodiment, the historical fault repair data may be, but is not limited to, a dataset containing fault scenario information and corresponding repair operations when the target server has previously experienced a fault. The fault scenario information includes historical abnormal components, historical fault types, historical fault propagation paths, etc., and the repair operations include specific troubleshooting steps, component replacement plans, parameter adjustment strategies, etc.

[0115] Optionally, in this embodiment, the target repair operation may be, but is not limited to, a specific operation plan formed by matching the current fault prediction result with historical fault repair data to select the repair operation that best matches the current prediction scenario in the historical data and adaptively adjusting it. This plan is used to prevent the current prediction anomaly of the target server from occurring. It is a key output connecting fault prediction and actual operation and maintenance execution, ensuring that effective measures can be taken in advance to avoid faults.

[0116] Optionally, in this embodiment, based on the fault prediction result as input, historical fault repair data of the target server is retrieved, and a similarity matching algorithm is used to find the historical repair scheme that best matches the current prediction scenario. Adaptive adjustments are made based on the specific differences of the current fault to ultimately form the target repair operation. Among them, historical fault repair data is the basis for ensuring the feasibility of the target repair operation, and the core value of the target repair operation is to prevent the current predicted server anomaly from actually occurring by implementing intervention measures in advance.

[0117] The embodiments provided in this application match the current fault prediction results with the historical fault repair data of the target server, and migrate the effective repair solutions verified by past operation and maintenance practices to the current prediction scenario. This achieves the goal of transforming fault prediction results into executable preventive repair operations, thereby reducing the actual failure rate and operation and maintenance costs of the server.

[0118] As an optional approach, the fault prediction results are matched with historical fault repair data to obtain the target repair operation, including at least one of the following:

[0119] The fault prediction results are matched with historical fault repair data to obtain the first repair operation, wherein the first repair operation is used to refresh the memory cache of the target server, and the target repair operation includes the first repair operation.

[0120] The fault prediction results are matched with historical fault repair data to obtain the second repair operation, wherein the second repair operation is used to reallocate the computing resources of the target server, and the target repair operation includes the second repair operation.

[0121] The fault prediction results are matched with historical fault repair data to obtain the third repair operation. The third repair operation is used to modify the firewall rules of the target server. The target repair operation includes the third repair operation.

[0122] Optionally, in this embodiment, the first repair operation may be, but is not limited to, a specific operation for refreshing the target server's memory cache obtained by matching the fault prediction result with historical fault repair data. Its core function is to clear the abnormal data accumulated in the memory cache, repair the cache read-write link deviation, and prevent server lag, data read-write errors and other faults caused by memory cache abnormalities. It is a preventive operation for potential memory-related faults.

[0123] Optionally, in this embodiment, the second repair operation may be, but is not limited to, an operation obtained through matching for reallocating the target server's computing resources, which aims to alleviate the resource pressure on high-load components and prevent component failures caused by insufficient resources or imbalanced allocation by adjusting the resource allocation ratio.

[0124] Optionally, in this embodiment, the third repair operation may be, but is not limited to, an operation obtained through matching to modify the firewall rules of the target server. The core is to adjust the firewall's filtering policy for network packets to prevent abnormal network connections or component failures caused by network attacks due to unreasonable firewall rules.

[0125] Optionally, in this embodiment, based on the components that may fail in the fault prediction results, historical fault repair data of the target server is retrieved, and historical repair operations that are compatible with the current faulty component and fault scenario are selected. For different fault types, three specific repair operations are generated, such as a first repair operation to resolve memory cache anomalies, a second repair operation to alleviate computing resource imbalance, and a third repair operation to optimize firewall rules. The target repair operation includes at least one of the above three types of operations, and ultimately, the failure of the target server is prevented by executing these operations.

[0126] The embodiments provided in this application accurately match the components that may fail in the fault prediction results with historical fault repair data, and generate targeted repair operations such as refreshing memory cache, reallocating computing resources, and modifying firewall rules. This transforms the fault prediction results into specific and executable preventive operation and maintenance actions, relies on historical operation and maintenance experience to ensure the effectiveness of the repair operations, avoids the actual occurrence of potential faults in memory, computing resources, and networks, and improves the operational stability and availability of the target server.

[0127] As an optional solution, in order to better understand the process of the above-mentioned server fault prediction method, the following describes the execution flow of the above-mentioned server fault prediction method in conjunction with optional embodiments, but it is not intended to limit the technical solution of the embodiments of this application.

[0128] Optionally, in this embodiment, as Figure 3 As shown, this embodiment illustrates a server fault prediction method integrating AI diagnostic functions, comprising the following steps:

[0129] S302, Obtain the hardware topology of the target server, and construct a dependency graph model containing multiple hardware components and their connection relationships based on the hardware topology. The dependency graph model is used to represent the physical dependencies and signal transmission paths between the components.

[0130] Furthermore, the hardware configuration information of the target server is read through the server baseboard management controller, and the central processing unit, memory module, storage device, power supply unit, cooling fan and motherboard power supply domain are identified as graph nodes.

[0131] Based on the hardware design documents and circuit connection logic, determine the electrical dependencies between each node, establish directed connection edges, and form an initial set of connection relationships;

[0132] An initial propagation weight is configured for each directed connection edge. The initial propagation weight is set according to the strength of the physical influence between components and is used to characterize the possibility of fault signals being transmitted from the source component to the associated component in subsequent graph calculations.

[0133] It should be noted that the construction of the graph model not only reflects the physical connection relationship between hardware components, but also quantifies the potential influence of different connection paths in fault propagation through the initial propagation weight. This provides a priori knowledge basis for the subsequent graph attention mechanism, ensuring that the modeling of fault propagation paths conforms to the actual circuit and system architecture logic, and avoiding unreasonable reasoning paths that may be generated by purely data-driven methods.

[0134] S304: Collect multi-source sensor data generated by the target server during operation, perform time synchronization and sampling frequency unification processing on the sensor data, and generate a time-series state sequence corresponding to each component in the dependency graph model.

[0135] Furthermore, the performance monitoring module continuously collects CPU utilization, memory usage, and disk I / O latency as performance indicators, while the health monitoring module collects hard drive operating temperature, number of bad sectors, power supply output voltage, and fan speed as health parameters.

[0136] The network time protocol is used to align the timestamps of all sensor data at the millisecond level, eliminating clock skew between data acquisition devices.

[0137] Resampling is performed on data streams with different sampling frequencies, and the health parameters acquired at low frequencies are extended to the same sampling granularity as the performance indicators through linear interpolation, forming a multidimensional state vector sequence with a uniform time step.

[0138] The multidimensional state vector of each time step is assigned to the corresponding hardware component node in the dependency graph model to form a temporal state sequence aligned with the graph structure.

[0139] It should be noted that the time synchronization and sampling frequency unification process ensures the consistency and comparability of multi-source heterogeneous sensor data in the time dimension, eliminates state misalignment caused by clock deviation of acquisition equipment or difference in sampling period, and ensures that the generated time-series state sequence can truly reflect the collaborative operation status of each hardware component in the same time segment, providing a reliable data foundation for subsequent cross-component correlation analysis.

[0140] S306, based on time-series state sequences, extracts second-level fluctuation features and long-term trend features through a dual-channel time-series analysis architecture, and integrates the second-level fluctuation features and long-term trend features into a comprehensive health characterization to reflect the changes in the operating status of components at different time scales;

[0141] Furthermore, the performance metrics of high-frequency sampling in the temporal state sequence are input into the temporal convolutional network, and the transient anomaly patterns are extracted through multi-layer causal dilated convolution, outputting the transient anomaly feature vector;

[0142] Low-frequency sampled health parameters are input into the sequence modeling unit, and the component aging index is calculated by combining the historical state offset accumulation mechanism, and the slow degradation feature vector is output.

[0143] A gating fusion mechanism is introduced to weight and integrate instantaneous abnormal feature vectors and slow degradation feature vectors to generate a comprehensive health characterization that simultaneously reflects sudden disturbances and gradual deterioration.

[0144] The calculation process of the gating fusion mechanism is shown in formulas (1) and (2):

[0145] ;

[0146] ;

[0147] in, Represents the instantaneous anomaly feature vector. This represents a slowly degrading feature vector. The learnable gated weight matrix, It is the Sigmoid activation function. This is the fused comprehensive health representation vector;

[0148] It should be noted that the gating fusion mechanism automatically adjusts the contribution ratio of instantaneous abnormal features and slow degradation features through learnable parameters, avoiding subjective bias caused by manually setting weights. This enables the model to dynamically adjust its focus according to different workloads and operating stages, thereby improving the comprehensive health representation's ability to express complex failure modes and its generalization performance.

[0149] S308 maps the comprehensive health representation to the corresponding component nodes in the dependency graph model, and combines the connection relationship of the dependency graph model to use the graph attention mechanism to carry out multiple rounds of state propagation between nodes, dynamically adjust the influence weight of each connection path in fault propagation, and identify high-probability fault propagation paths.

[0150] Furthermore, the comprehensive health representation vector of each hardware component is used as the initial state vector of the corresponding graph node;

[0151] In each propagation round, the attention score between adjacent nodes is calculated. The attention score reflects the degree of influence of the source node's state change on the target node.

[0152] The state information of neighboring nodes is weighted and aggregated based on the attention score, and the hidden state of the current node is updated.

[0153] The calculation process of the attention score is shown in formulas (3) and (4):

[0154]

[0155] ;

[0156] in, Indicates the source node The state vector, Represents the target node The state vector, To share the transformation matrix, For learnable vectors in the attention mechanism, For nodes To the node Normalized attention weights Represents a node The set of neighboring nodes;

[0157] It should be noted that the graph attention mechanism dynamically calculates the influence weights between nodes in each round of state propagation. It not only considers the topological connection relationship, but also adaptively adjusts the information transmission intensity according to the actual operating state changes of each component, thereby identifying the propagation path that plays a dominant role in the specific fault evolution process and enhancing the accuracy and timeliness of fault causal reasoning.

[0158] S310: Determine the risk significance of each component based on the node status after propagation is completed, generate server-level fault risk assessment results based on risk significance, and output a warning signal when the fault risk assessment results meet the warning triggering conditions.

[0159] Furthermore, a nonlinear transformation is performed on the final node state vector of each component after multiple rounds of propagation to generate a local risk score, which represents the probability of the component malfunctioning.

[0160] The local risk scores of all components are merged using a maximum pooling method to generate a global failure risk score that represents the overall health status of the server.

[0161] The calculation process for the local risk score is shown in formula (5):

[0162] (5);

[0163] in, Represents a node In the The final state vector after round propagation, A learnable rating weight vector. It is the Sigmoid activation function. For components Local risk score;

[0164] The global fault risk score is compared with the dynamic early warning threshold. When the global fault risk score exceeds the threshold continuously for a preset time, a fault early warning signal is generated and pushed to the operation and maintenance management platform.

[0165] It should be noted that the global fault risk score uses a maximum pooling method to aggregate local risk scores, which can effectively capture the state deterioration trend of the weakest link in the system and ensure that high-risk faults are not masked by overall averaging. At the same time, combined with the continuous over-threshold duration judgment mechanism, it reduces false alarms caused by occasional fluctuations and improves the stability and credibility of early warning decisions.

[0166] S312, Based on the risk salience ranking and the key propagation paths identified in the graph attention mechanism, a structured diagnostic report is generated. The diagnostic report includes high-risk components, potential causes and failure impact paths, and the propagation weight calculation logic in the graph attention mechanism is adaptively adjusted based on the feedback information of the operation and maintenance personnel on the diagnostic report.

[0167] Furthermore, components are arranged from high to low according to their local risk scores to form a risk component list;

[0168] During the attention propagation process of the extracted graph, the connection edges whose normalized attention weights are consistently higher than the set level constitute the main fault propagation path;

[0169] The list of risky components and the main failure propagation paths are integrated into a structured diagnostic report. The report includes the name of the high-risk component, the associated abnormal indicators, the possible starting point of the failure and the scope of its impact.

[0170] Receive confirmation or correction feedback from operations and maintenance personnel on diagnostic reports, and use the feedback information to incrementally update the parameters in the graph attention mechanism;

[0171] Among them, the incremental update process adopts the elastic weight solidification method to prevent the model from forgetting historical knowledge. The updated attention parameters are used to adjust the propagation weight calculation logic in the next round of prediction, so that the model can more accurately identify the real fault propagation mode.

[0172] It should be noted that the incremental update mechanism based on operation and maintenance feedback adjusts model parameters while retaining historical knowledge through the elastic weight solidification method, preventing model performance degradation due to insufficient or biased new feedback data, realizing incremental optimization of diagnostic logic, and enabling the system to have adaptive evolution capabilities under long-term operation.

[0173] This embodiment also provides a server fault prediction system integrating AI diagnostic functions, such as... Figure 4 As shown, it includes: topology modeling module, data alignment module, feature fusion module, graph propagation module, risk assessment module, and diagnostic update module;

[0174] The topology modeling module is used to obtain the hardware topology of the target server and build a dependency graph model containing multiple hardware components and their connection relationships based on the hardware topology. The dependency graph model is used to represent the physical dependencies and signal transmission paths between the components.

[0175] The data alignment module is used to collect multi-source sensor data generated by the target server during operation, perform time synchronization and sampling frequency unification processing on the sensor data, and generate a time-series state sequence corresponding to each component in the dependency graph model.

[0176] The feature fusion module is used to extract second-level fluctuation features and long-term trend features based on the time-series state sequence through a dual-channel time-series analysis architecture. The second-level fluctuation features and long-term trend features are then fused into a comprehensive health characterization to reflect the changes in the operating status of components at different time scales.

[0177] The graph propagation module is used to map the comprehensive health representation to the corresponding component nodes in the dependency graph model. It combines the connection relationship of the dependency graph model and uses the graph attention mechanism to perform multiple rounds of state propagation between nodes, dynamically adjusts the influence weight of each connection path in fault propagation, and identifies high-probability fault propagation paths.

[0178] The risk assessment module is used to determine the risk significance of each component based on the node status after propagation is completed, generate server-level fault risk assessment results based on risk significance, and output a warning signal when the fault risk assessment results meet the warning trigger conditions.

[0179] The diagnostic update module generates a structured diagnostic report based on the risk salience ranking and the key propagation paths identified in the graph attention mechanism. The diagnostic report includes high-risk components, potential causes, and failure impact paths. It also adaptively adjusts the propagation weight calculation logic in the graph attention mechanism based on feedback from operations and maintenance personnel.

[0180] Based on the hardware dependency graph model, multi-scale temporal analysis, and graph attention propagation technology framework constructed in this embodiment, it can be further extended to a wider range of intelligent operation and maintenance and system health management scenarios, breaking through the boundaries of a single server device and forming a cross-level and cross-system predictive maintenance system. For example, the dependency graph model of a single server can be expanded into a data center-level multi-layer coupled graph network of "device-rack-network-power supply-cooling", realizing the transmission path tracing from component failure to data center-level hot zones and power bottlenecks, and improving the overall infrastructure resilience management capabilities.

[0181] In terms of temporal modeling, in addition to the currently used TCN and sequence modeling units, an event-driven sparsity processing mechanism can be introduced. This mechanism activates model computation only when there are sudden state changes or when key indicators exceed limits, significantly reducing the computational power consumption of edge deployments and making it suitable for lightweight deployments of large-scale server clusters. Simultaneously, it is possible to explore encoding non-numerical semantic information such as operation and maintenance logs, firmware versions, and configuration changes into auxiliary node or edge attributes in the graph, integrating knowledge graph reasoning capabilities to enhance the ability to identify complex faults such as software configuration errors and driver incompatibility.

[0182] Regarding the diagnostic feedback mechanism, the current incremental updates based on manual feedback can be upgraded to a closed-loop self-evolutionary architecture: by introducing a counterfactual reasoning module, the failure evolution path under different intervention measures can be simulated, the effectiveness of diagnostic suggestions can be evaluated, and correction rules can be automatically extracted to reduce the dependence on manual annotation; or by combining a federated learning framework, model updates can be shared among multiple data centers without leaking the original data, thereby achieving cross-domain knowledge transfer and collaborative optimization.

[0183] Furthermore, this technological paradigm can be transferred to other complex electromechanical systems, such as industrial control equipment, electric drive systems for new energy vehicles, and rail transit signaling devices. As long as there are clear physical dependencies and multi-source sensor data, a corresponding "functional unit dependency graph" can be constructed, enabling cross-domain generalized intelligent diagnosis. Ultimately, this technology is not limited to fault prediction but can also be extended to full lifecycle management functions such as performance bottleneck location, resource scheduling optimization, and life assessment, becoming the core engine for autonomous health management of intelligent systems.

[0184] In this embodiment, hardware configuration information is read by the server baseboard management controller, key components are identified as graph nodes, and electrical dependencies are determined based on circuit connection logic to establish directed connections. An initial propagation weight is configured for each edge, based on the strength of physical influence between components, to characterize the likelihood of fault signals propagating from the source component to related components in subsequent graph calculations. This step addresses the problem of traditional methods ignoring hardware topology and is fundamental to achieving accurate fault propagation path identification. A temporal convolutional network is used to extract transient anomaly patterns and output transient anomaly feature vectors. A historical state offset accumulation mechanism is combined to calculate component aging indicators and output slow degradation feature vectors. A learnable gating fusion mechanism is introduced to weightedly integrate the two, generating a comprehensive health characterization reflecting sudden disturbances and gradual degradation. This mechanism dynamically adjusts the contribution ratio of different features, enhancing the model's ability to express complex failure modes. It maps the comprehensive health representation to corresponding nodes in the dependency graph model, serving as the initial state vector. In each propagation round, it calculates the attention score between adjacent nodes, reflecting the impact of source node state changes on the target node. Based on the attention score, it weights and aggregates the state information of neighboring nodes and updates the hidden state of the current node. Through multiple iterations, it identifies high-probability failure propagation paths. This process adaptively adjusts the information transmission intensity, enhancing the accuracy of causal reasoning about failures.

[0185] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0186] Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0187] This embodiment also provides a server fault prediction device, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as described herein. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0188] Figure 5This is a structural block diagram of a server fault prediction device according to an embodiment of this application; as shown... Figure 5 As shown, it includes:

[0189] The building unit 502 is used to build a component dependency graph based on the hardware topology of the target server. The component dependency graph includes multiple interconnected nodes. The nodes are used to indicate a server component in the target server, and the connection relationship between the nodes is used to indicate the communication connection relationship between multiple server components.

[0190] The first determining unit 504 is used to determine the state feature vector of at least one server component according to the respective running state sequence of at least one server component in the target server, wherein the running state sequence is used to indicate the change of the running state of the server component within the target period.

[0191] The second determining unit 506 is used to determine the connection weights between multiple nodes in the component dependency graph based on at least one state feature vector to obtain a target component dependency graph, wherein the connection weights are used to indicate the path of state propagation of fault states between multiple server components.

[0192] The third determining unit 508 is used to determine the fault prediction result based on the target component dependency graph, wherein the fault prediction result is used to indicate at least one abnormal component among multiple server components.

[0193] As an optional solution, the third determining unit 508 includes: an acquisition module, used to acquire the fault probability coefficient corresponding to each node in the target component dependency graph, wherein the fault probability coefficient is used to indicate the probability of the server component corresponding to the node failing; and a traversal module, used to traverse each node in the target component dependency graph, and if the fault probability coefficient corresponding to the current node satisfies the probability condition, determine the current node as the target node, and determine the server component corresponding to the current node as the abnormal component.

[0194] As an optional approach, the acquisition module is also used to acquire the state vector corresponding to the first node in the target component dependency graph, wherein the state vector is used to indicate the operating state of the first node; and to perform a nonlinear transformation on the state vector to obtain the fault probability coefficient corresponding to the first node.

[0195] As an optional approach, the acquisition module is further configured to determine the state feature vector corresponding to the first node as the state vector of the first node; perform at least one state propagation on the target component dependency graph, and in each state propagation, perform the following operations: calculate the first attention score corresponding to the first node and at least one second node, wherein the second node and the first node have a dependency relationship, and the first attention score is used to indicate the degree of influence of the first node on the second node; based on at least one first attention score, perform a weighted summation of the state vectors of at least one second node to obtain the neighbor state vector; and update the state vector of the first node based on the neighbor state vector.

[0196] As an optional solution, the third determining unit 508 further includes: a calculation module, used to calculate the second attention score corresponding to the third node and at least one fourth node in the target component dependency graph, wherein there is a communication connection between the server component corresponding to the third node and the server component corresponding to the fourth node, and the second attention score is used to indicate the degree of influence of the third node on the fourth node; and a determining module, used to determine the edge between the fourth node corresponding to the second attention score that satisfies the attention condition and the third node as the fault propagation path, wherein the fault prediction result includes the fault propagation path.

[0197] As an optional solution, the device further includes: an acquisition unit for acquiring multi-source sensor data generated by the target server during operation through at least one sensor; an alignment unit for aligning the multi-source sensor data with timestamps according to a network time protocol; an operation unit for performing resampling operations on data with different sampling frequencies in the multi-source sensor data to obtain a multi-dimensional state vector sequence; and an allocation unit for allocating the multi-dimensional state vector corresponding to each time step in the multi-dimensional state vector sequence to nodes in the component dependency graph to obtain a running state sequence.

[0198] As an optional solution, the first determining unit 504 includes: a first input module, used to input the running state sequence into a temporal convolutional network to obtain a short-time feature vector output by the temporal convolutional network, wherein the short-time feature vector is used to predict the possible failure of at least one server component, and the state feature vector includes the short-time feature vector; a second input module, used to input the running state sequence into a sequence modeling network to obtain a long-time feature vector output by the sequence modeling network, wherein the long-time feature vector is used to predict the remaining lifetime of at least one server component, and the state feature vector includes the long-time feature vector; a second acquisition module, used to acquire a first weight value corresponding to the short-time feature vector, wherein the first weight value is used to indicate the importance of the short-time feature vector; a third acquisition module, used to acquire a second weight value corresponding to the long-time feature vector, wherein the second weight value is used to indicate the importance of the long-time feature vector; and a fusion module, used to perform weighted fusion of the short-time feature vector and the long-time feature vector based on the first weight value and the second weight value to obtain the state feature vector.

[0199] As an optional solution, the device further includes: a matching unit for matching the fault prediction result with historical fault repair data to obtain a target repair operation, wherein the historical fault repair data is used to indicate the repair operations performed on the target server, and the target repair operation is used to prevent possible anomalies that may occur during the operation of the target server.

[0200] As an optional solution, the matching unit includes: a first matching module, used to match the fault prediction result with historical fault repair data to obtain a first repair operation, wherein the first repair operation is used to refresh the memory cache of the target server, and the target repair operation includes the first repair operation; a second matching module, used to match the fault prediction result with historical fault repair data to obtain a second repair operation, wherein the second repair operation is used to reallocate the computing resources of the target server, and the target repair operation includes the second repair operation; and a third matching module, used to match the fault prediction result with historical fault repair data to obtain a third repair operation, wherein the third repair operation is used to modify the firewall rules of the target server, and the target repair operation includes the third repair operation.

[0201] For a description of the features in the embodiment corresponding to the server fault prediction device, please refer to the relevant description of the embodiment corresponding to the server fault prediction method, which will not be repeated here.

[0202] Embodiments of this application also provide an electronic device. Figure 6 This is a schematic diagram of an electronic device according to an embodiment of this application, such as... Figure 6As shown, the electronic device includes a memory and a processor, the memory storing a computer program, and the processor being configured to run the computer program to perform the steps in any of the above-described embodiments of the server fault prediction method.

[0203] In one exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0204] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0205] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described embodiments of the server fault prediction method when running.

[0206] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0207] Embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the methods in various embodiments of this application; the computer program product further includes a non-volatile computer-readable storage medium storing the computer program, which, when executed by a processor, implements the steps of the server fault prediction method in various embodiments of this application.

[0208] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0209] The above provides a detailed description of a server fault prediction method provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A failure prediction method of a server, characterized by, The method comprises the following steps: constructing a component dependency graph based on the hardware topology of a target server, wherein the component dependency graph comprises a plurality of interconnected nodes, the nodes are used to indicate a server component in the target server, and the connection relationship between the nodes is used to indicate the dependency relationship between a plurality of server components; inputting a time convolution network of a high-frequency sampled performance indicator in a running state sequence of each of at least one server component in the target server, and extracting an instantaneous abnormal feature vector of each of the at least one server component through a plurality of layers of causal dilated convolution, wherein the running state sequence is used to indicate the change of the running state of the server component in a target period; modeling a health parameter input sequence sampled at a low frequency in the running state sequence of each of at least one server component in the target server, and calculating a component aging degree indicator by combining a historical state offset accumulation mechanism to output a slow degradation feature vector of each of the at least one server component; fusing the instantaneous abnormal feature vector and the slow degradation feature vector of a first component to obtain a comprehensive health vector of the first component, wherein the comprehensive health vector is used to indicate the change of the running state of the first component at different time scales, and the at least one server component comprises the first component; determining a connection weight between a plurality of nodes in the component dependency graph according to at least one comprehensive health vector to obtain a target component dependency graph, wherein the connection weight is used to indicate the path of the state propagation of the fault state between a plurality of server components; determining a fault prediction result based on the target component dependency graph, wherein the fault prediction result is used to indicate at least one abnormal component in a plurality of server components.

2. The method of claim 1, wherein, Determining a fault prediction result based on the target component dependency graph comprises: obtaining a fault probability coefficient corresponding to each node in the target component dependency graph, wherein the fault probability coefficient is used to indicate the probability of the fault of the server component corresponding to the node; traversing each node in the target component dependency graph, and determining a current node as a target node and a server component corresponding to the current node as the abnormal component when the fault probability coefficient corresponding to the current node meets a probability condition.

3. The method of claim 2, wherein, Obtaining a fault probability coefficient corresponding to each node in the target component dependency graph comprises: obtaining a state vector corresponding to a first node in the target component dependency graph, wherein the state vector is used to indicate the running state of the first node; performing nonlinear transformation on the state vector to obtain a fault probability coefficient corresponding to the first node.

4. The method of claim 3, wherein, Obtaining a state vector corresponding to a first node in the target component dependency graph comprises: determining a comprehensive health vector corresponding to the first node as the state vector of the first node; performing at least one state propagation on the target component dependency graph, and in each state propagation, performing the following operations: compute a first attention score corresponding to each of the first node and at least one second node, wherein the second node has a dependency relationship with the first node, and wherein the first attention score indicates a degree of influence of the first node on the second node; perform a weighted summation on state vectors of the at least one second node based on the at least one first attention score to obtain a neighbor state vector; update a state vector of the first node based on the neighbor state vector.

5. The method of claim 1, wherein, determine a fault prediction result based on the target component dependency graph, and the method further comprises: compute a second attention score corresponding to each of a third node and at least one fourth node in the target component dependency graph, wherein the third node corresponds to a server component and the fourth node corresponds to a server component, and wherein the third node and the fourth node have a dependency relationship, and wherein the second attention score indicates a degree of influence of the third node on the fourth node; determine an edge between the fourth node corresponding to the second attention score satisfying an attention condition and the third node as a fault propagation path, wherein the fault prediction result comprises the fault propagation path.

6. The method of claim 1, wherein, Before determining a comprehensive health vector of each of at least one server component in the target server based on a running state sequence of each of the at least one server component, the method further comprises: acquire multi-source sensor data generated by the target server during running through at least one sensor; perform time stamp alignment on the multi-source sensor data according to a network time protocol; perform a resampling operation on data with different sampling frequencies in the multi-source sensor data to obtain a multi-dimensional state vector sequence; assign a multi-dimensional state vector corresponding to each time step in the multi-dimensional state vector sequence to a node in the component dependency graph to obtain the running state sequence.

7. The method of claim 1, wherein, determine a comprehensive health vector of each of at least one server component in the target server based on a running state sequence of each of the at least one server component, and the method further comprises: obtain a first weight value corresponding to the instantaneous abnormal feature vector, wherein the first weight value indicates importance of the instantaneous abnormal feature vector; obtain a second weight value corresponding to the slow degradation feature vector, wherein the second weight value indicates importance of the slow degradation feature vector; perform a weighted fusion on the instantaneous abnormal feature vector and the slow degradation feature vector based on the first weight value and the second weight value to obtain the comprehensive health vector.

8. The method according to any one of claims 1 to 7, characterized in that, After determining a fault prediction result based on the target component dependency graph, the method further comprises: match the fault prediction result with historical fault repair data to obtain a target repair operation, wherein the historical fault repair data indicates a repair operation performed on the target server, and wherein the target repair operation prevents an abnormality that may occur during running of the target server.

9. The method of claim 8, wherein, matching the fault prediction result with historical fault repair data to obtain a target repair operation, including at least one of: matching the fault prediction result with historical fault repair data to obtain a first repair operation, wherein the first repair operation is used to refresh a memory cache of the target server, and the target repair operation includes the first repair operation; matching the fault prediction result with historical fault repair data to obtain a second repair operation, wherein the second repair operation is used to reallocate a computing resource of the target server, and the target repair operation includes the second repair operation; matching the fault prediction result with historical fault repair data to obtain a third repair operation, wherein the third repair operation is used to modify a firewall rule of the target server, and the target repair operation includes the third repair operation.

10. An electronic device, comprising: comprising: a memory for storing a computer program; a processor for implementing the steps of the fault prediction method of the server according to any one of claims 1 to 9 when executing the computer program.

Images