An intelligent device anomaly identification method for data center operation and maintenance

Abstract

The application discloses a kind of equipment abnormality intelligent identification methods for data center operation and maintenance, comprising: collecting multidimensional operating parameters and computing resource occupancy and delimiting preliminary screening abnormal window;Through cross-node situation linkage perception model, each node parameter is mapped as situation characteristic tensor and the global situation deviation degree is calculated;Topology-dependent fault propagation algorithm is executed, and root cause node is traced back level by level along equipment topology-dependent relationship and quantifies each level fault injection amount;Call computing resource distortion multiscale detection function, measure computing resource occupancy distortion degree under multiple time scales and output distortion flag;Situation deviation degree, fault injection amount and distortion flag are input into computing resource cluster fault intelligent diagnosis platform, and candidate fault type is determined by matching fault tree;Reverse check the degree of parameter statistics outlying in preliminary screening abnormal window, determine whether abnormal event is established.The application effectively improves the accuracy and timeliness of data center computing resource cluster anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment anomaly identification technology, and in particular to an intelligent equipment anomaly identification method for data center operation and maintenance. Background Technology

[0002] As data centers continue to expand, the number of devices within computing clusters is surging, and the topology is becoming increasingly complex, resulting in highly dynamic and coupled operational status. Anomalies in a single device or parameter often propagate along dependencies within the cluster, causing localized failures to rapidly escalate into regional service degradation or even complete computing power paralysis. Traditional maintenance methods, primarily based on manual inspections and single-threshold alarms, are no longer sufficient to handle massive concurrent fluctuations in metrics and implicit cross-node correlations of anomalies. There is an urgent need to develop an intelligent anomaly identification solution that integrates multi-dimensional operational parameters, topological dependencies, and the dynamic characteristics of computing resources.

[0003] Currently, mainstream solutions in the industry generally adopt time-series database-based monitoring systems. These systems collect metrics such as central processing unit utilization, memory usage, and network throughput of each device at fixed intervals and store them in a time-series database. Several independently running anomaly detection algorithms are then deployed on top of this database. These algorithms are mostly based on machine learning models such as statistical process control or isolated forests, calculating the degree of deviation for each device or parameter individually. When the deviation exceeds a pre-set static or dynamic threshold, an alarm is triggered. Some more advanced solutions introduce simple causal inference or association rule mining, attempting to locate suspected root cause devices after an alarm outbreak using a manually labeled fault propagation rule base, and then push the diagnostic results to maintenance personnel for manual verification.

[0004] Existing technologies have two main drawbacks. First, the processing of multi-dimensional parameters for each device in existing solutions is fragmented, lacking a holistic measurement of cross-node status quo changes. This makes it difficult to effectively identify anomalies when there is a coordinated shift in the overall computing power distribution of the cluster, easily leading to missed fault precursors involving the combined effects of multiple nodes. Second, fault root cause localization relies on a pre-set static propagation rule base. When the cluster topology is dynamically adjusted or the fault propagation path changes due to load variations, the rule base cannot be updated adaptively, causing diagnostic results to lag significantly behind the actual fault propagation process and greatly prolonging anomaly recovery time. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a method for intelligent identification of equipment anomalies in data center operations and maintenance.

[0006] The technical solution adopted in this invention is an intelligent identification method for equipment anomalies in data center operation and maintenance, comprising the following steps: S1, collecting multi-dimensional operating parameters and computing resource occupancy rates of data center operation and maintenance equipment, and defining an initial screening anomaly window based on the temporal fluctuation amplitude of the parameters; S2, based on a cross-node situational linkage perception model, mapping the parameter sequences of each node into situational feature tensors, and calculating the global situational deviation using the parameter cooperation between nodes; S3, executing a topology-dependent fault propagation algorithm, constructing directed edges for fault propagation based on the equipment topology connection relationship, tracing the root cause node along the directed edges level by level, and quantifying the fault injection volume at each level; S4, calling a multi-scale detection function for computing resource distortion, performing distortion degree measurement on the computing resource occupancy sequence at multiple time scales, and outputting distortion flags at each scale; S5, inputting the global situational deviation, the fault injection volume at each level, and the distortion flags into a computing cluster fault intelligent diagnosis platform, and determining candidate fault types through the platform's built-in fault tree matching rules; S6, using the candidate fault types as indexes, reverse-checking the statistical outlier degree of each parameter within the initial screening anomaly window, and determining whether the anomaly event is valid.

[0007] Furthermore, the cross-node situational awareness model is represented as follows: ,in At the current assessment point, The length of the historical time window. Let be the parameter state matrix of all nodes at time t. The spatial gradient of the state matrix. It is the Frobenius norm. This contributes to the historical accumulation of activity levels. ,in The total number of nodes. For node indexing, For nodes The response decay coefficient, For nodes Historical cumulative trend activity weight, For nodes; the partial derivative of situational activity with respect to time. To achieve overall situational coordination and balance.

[0008] Furthermore, the topology-dependent fault propagation algorithm is expressed as: ,in This is the number of the currently evaluated node. For a set of upstream nodes that provide direct dependent inputs to a node, Number the upstream node. Let be the propagation attenuation coefficient from node k to node j. For node k in delay The residual fault strength afterward, The cumulative fault injection amount received by node j at time t; ,in Let be the real-time residual fault intensity of node j at time t. Let be the spontaneous fault baseline value of node j at time t. For integration time variable, Let be the fault cumulative gain coefficient for node j. Let be the fault dissipation rate coefficient of node j.

[0009] Furthermore, the computing power resource distortion multi-scale detection function is expressed as: ,in For scale-level indexing, This represents the number of difference segments at the current scale. The difference segment number, Let m be the interval step size at the m-th scale. For the computing resource occupancy sequence at time 1 The value at that location, For the m-th scale Local standard deviation centered on Here is the regularization constant. denoted as the distortion detection quantity at the m-th scale.

[0010] Furthermore, the multi-dimensional operating parameters collected in S1 include central processing unit utilization, memory read / write throughput, disk input / output latency, network interface packet loss and retransmission rate, temperature sensor readings, and fan speed feedback. The parameters are acquired at fixed sampling intervals, with the sampling interval ranging from 1 to 5 seconds. When defining the initial screening abnormal window in S1, the deviation multiple of the current sampled value relative to the mean of the previous sliding window is calculated for each parameter. The sliding window length is 60 sampling points. When the deviation multiple of any parameter exceeds a preset dynamic threshold, the sliding window to which the current sampling time belongs is marked as the initial screening abnormal window. The preset dynamic threshold is determined by multiplying the difference between the upper quartile and the median of the deviation multiple distribution of the parameter in the past 24 hours by an adjustable coefficient, with the adjustable coefficient ranging from 3 to 5.

[0011] Further, S2 includes: S21, for each node, arranging the multi-dimensional operating parameters collected in step S1 into local feature row vectors, and incorporating the computing power resource occupancy rate as an additional dimension into the tail of the row vectors; S22, stacking the local feature row vectors of all nodes vertically according to the node number to form an original feature matrix, and performing spatial gradient operation on the original feature matrix to obtain a spatial gradient matrix; S23, taking the sum of the squares of each element in the spatial gradient matrix as the spatial gradient modulus at that moment, and performing time integration on the spatial gradient modulus within the historical time window to obtain the historical cumulative situational activity; S24, taking the partial derivative of the historical cumulative situational activity components of each node with respect to time, and then performing exponential weighting and multiplication to obtain the global situational linkage balance.

[0012] Further, S3 includes: S31, establishing a topological directed graph based on the physical connections and logical dependencies between data center devices, where directed edges in the topological directed graph point from upstream dependent nodes to downstream dependent nodes; S32, assigning a propagation attenuation coefficient and a propagation delay parameter to each directed edge, with the propagation attenuation coefficient dynamically adjusted based on the physical distance and bandwidth occupancy between the two ends; S33, for the current evaluated node, accumulating the product of the residual fault intensity of all upstream dependent nodes after delay and the corresponding propagation attenuation coefficient, as the cumulative fault injection amount of the current node; S34, amplifying the spontaneous fault base value and the cumulative fault injection amount of the current node through a fault accumulation gain coefficient, then subtracting the product of the fault dissipation rate coefficient and the real-time residual fault intensity, integrating this difference in the time domain to obtain the updated real-time residual fault intensity.

[0013] Further, S4 includes: S41, setting multiple time scale levels, each scale level corresponding to a different interval step size, the interval step size increasing as the scale level increases; S42, for the current scale level, dividing the computing power resource occupancy sequence into multiple difference intervals according to the interval step size, each difference interval including two sampling points separated by the interval step size; S43, calculating the absolute difference between the sampling points at both ends of each difference interval, and dividing the absolute difference by the sum of the local standard deviation centered at the current time and the regularization constant under the current scale level, to obtain the normalized difference value between each difference interval; S44, performing a multiplication operation on the normalized difference values ​​between all difference intervals under the current scale level to obtain the distortion detection quantity under the current scale level.

[0014] Further, S5 includes: S51, combining the global situational deviation, the amount of fault injection at each level, and the distortion flag into a multi-dimensional evidence vector, and inputting the multi-dimensional evidence vector into the front-end interface of the intelligent fault diagnosis platform for the computing power cluster; S52, the platform matches the multi-dimensional evidence vector with the evidence intervals corresponding to each leaf node in the built-in fault tree, and calculates the membership degree of the multi-dimensional evidence vector falling into the evidence interval of each leaf node; S53, for leaf nodes whose membership degree exceeds the preset acceptance threshold, tracing upwards along the fault tree to the intermediate node, and summarizing the membership degree of each leaf node according to the logic gate type of the intermediate node; S54, outputting the fault type corresponding to the intermediate node with the highest summarized membership degree as a candidate fault type.

[0015] A method for intelligent identification of equipment anomalies in data center operations and maintenance is disclosed. This method is implemented through an intelligent equipment anomaly identification system for data center operations and maintenance, comprising: a heterogeneous parameter synchronous acquisition unit, used to synchronously acquire multi-dimensional operating parameters and computing resource occupancy rates of each device using a unified clock reference and temporarily store them in a circular cache; a spatiotemporal situation feature tensor construction unit, with its input connected to the output of the heterogeneous parameter synchronous acquisition unit, used to map the parameter sequences of each node into a situation feature tensor and calculate the global situation deviation; and a topology dependency recursive tracing unit, with its input connected to the outputs of the heterogeneous parameter synchronous acquisition unit and the spatiotemporal situation feature tensor construction unit respectively, used to recursively calculate the topology dependency of each node. The system includes: a cumulative fault injection quantity and root cause node tracing; a multi-scale distortion parallel detection unit, whose input is connected to the output of the heterogeneous parameter synchronous acquisition unit, used to calculate the distortion detection quantity of the computing resource occupancy sequence in parallel at multiple time scales and output distortion flags at each scale; a fault tree matching adjudication unit, whose input is connected to the output of the spatiotemporal situation feature tensor construction unit, the topology dependency recursive tracing unit, and the multi-scale distortion parallel detection unit, used to adjudicate candidate fault types according to evidence matching rules; and a reverse verification confirmation unit, whose input is connected to the output of the fault tree matching adjudication unit, used to reverse verify the statistical outlier degree within the initial screening anomaly window based on the candidate fault type and output the final anomaly identification result.

[0016] Beneficial Effects: This invention proposes an intelligent identification method for device anomalies in data center operations and maintenance. By using a cross-node situational awareness model to perform spatial gradient calculations and temporal integration on the parameter state matrices of all devices within the cluster, a deviation index representing the global collaborative situational shift is generated. This allows for the overall capture of latent anomalies caused by the combined action of multiple nodes, overcoming the shortcomings of existing technologies that miss cluster collaborative shift anomalies due to fragmented parameter processing, thus improving the early detection capability of anomaly precursors. Simultaneously, this invention employs a topology-dependent fault propagation algorithm to dynamically maintain a directed graph based on the physical connections and logical dependencies between devices. Each propagation edge is assigned an attenuation coefficient and delay parameter that adaptively adjusts with load and distance. By accumulating the upstream residual fault intensity and performing time-domain integration, the fault injection quantity at each level is updated in real time. This allows the fault propagation path to dynamically evolve with changes in cluster topology and load, no longer relying on a static rule base, thus improving the timeliness and accuracy of root cause tracing. This invention introduces a multi-scale detection function for computing power resource distortion. It performs piecewise difference and local standard deviation normalization multiplication operations on the resource occupancy sequence at different time scales and outputs distortion flags at each scale, which further enriches the diversity of anomaly criteria. Combined with the multi-evidence fusion matching and reverse verification mechanism of the intelligent fault diagnosis platform for computing power clusters, it reduces the false alarm rate and shortens the anomaly recovery cycle. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S2 of the present invention; Figure 3 This is a flowchart of method step S3 of the present invention; Figure 4 This is a flowchart of method step S4 of the present invention; Figure 5 This is a flowchart of step S5 of the method of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1As shown, an intelligent identification method for equipment anomalies in data center operations and maintenance includes the following steps: S1, collecting multi-dimensional operating parameters and computing resource occupancy rates of data center operations and maintenance equipment, and defining an initial screening anomaly window based on the temporal fluctuation amplitude of the parameters; S2, based on a cross-node situational awareness model, mapping the parameter sequences of each node into situational feature tensors, and calculating the global situational deviation using the parameter cooperation between nodes; S3, executing a topology-dependent fault propagation algorithm, constructing directed edges for fault propagation based on the equipment topology connection relationship, tracing the root cause node along the directed edges level by level, and quantifying the fault injection volume at each level; S4, calling a multi-scale detection function for computing resource distortion, performing distortion degree measurement on the computing resource occupancy sequence at multiple time scales, and outputting distortion flags at each scale; S5, inputting the global situational deviation, the fault injection volume at each level, and the distortion flags into a computing cluster fault intelligent diagnosis platform, and determining candidate fault types through the platform's built-in fault tree matching rules; S6, using the candidate fault types as indexes, reverse-checking the statistical outlier degree of each parameter within the initial screening anomaly window, and determining whether the anomaly event is valid.

[0020] During step S1, multi-dimensional operating parameters and computing resource utilization of the data center's maintenance equipment are collected. These multi-dimensional operating parameters include CPU utilization, memory read / write throughput, disk I / O latency, network interface packet loss and retransmission rate, temperature sensor readings, and fan speed feedback. Each parameter is acquired synchronously at a fixed sampling interval of 1 to 5 seconds. Computing resource utilization includes GPU utilization and FPGA (Field-Programmable Gate Array) resource utilization. Based on the temporal fluctuations of each parameter, the deviation factor of the current sampled value relative to the mean value within the preceding sliding window is calculated for each parameter. The length of the preceding sliding window is fixed at 60 sampling points. When the deviation factor of any parameter exceeds a preset dynamic threshold, the sliding window to which the current sampling time belongs is marked as an initial screening abnormal window. The preset dynamic threshold is determined by extracting the distribution of all deviation multiples of the parameter in the previous 24 hours, calculating the difference between the upper quartile and the median of the distribution, and multiplying the difference by an adjustable coefficient to obtain the preset dynamic threshold. The adjustable coefficient ranges from 3 to 5, and the actual value is selected as 3.5 or 4.2 according to the sensitivity requirements of the device.

[0021] In step S2, a situational analysis is performed on all nodes within the data center based on the cross-node situational awareness model. Each node organizes the multi-dimensional operating parameters collected in step S1 into a local feature row vector, and appends the node's computing power resource utilization rate as an additional dimension to the end of this row vector, thus obtaining the extended feature row vector of each node. All nodes are stacked vertically in ascending order of node number to form a unified original feature matrix. The difference between adjacent elements in each row and column of this original feature matrix is ​​calculated to obtain the row direction difference matrix and the column direction difference matrix. The elements at corresponding positions in the row direction difference matrix and the column direction difference matrix are squared, summed, and then the square root is taken to obtain the spatial gradient magnitude at that moment. A historical time window is traced back from the current evaluation moment on the time axis, with the length of the historical time window set to 300 seconds. The spatial gradient magnitude at each moment within this historical time window is squared, and all squared results are integrated over time to obtain the historical cumulative situational activity. The historical cumulative situation activity is decomposed into components of each node. The partial derivative of each node's component with respect to time is calculated. The partial derivative is multiplied by the response decay coefficient of that node and then exponentially calculated. Finally, the exponential results of all nodes are multiplied together to obtain the global situation linkage equilibrium.

[0022] In step S3, the topology dependency fault propagation algorithm is executed. A directed topology graph is constructed based on the physical connection relationships and logical service dependencies of the data center devices. Each node in the directed topology graph represents a physical device or a logical service unit, and directed edges point from upstream dependent nodes to downstream dependent nodes. Each directed edge is assigned a propagation attenuation coefficient and a propagation delay parameter. The initial value of the propagation attenuation coefficient is set to 0.7, and it is dynamically adjusted according to the physical distance between the two ends and the current bandwidth utilization. For every 1 meter increase in physical distance, the propagation attenuation coefficient decreases by 0.01 from its original value; for every 10 percent increase in bandwidth utilization, the propagation attenuation coefficient decreases by 0.02 from its original value. The propagation delay parameter ranges from 1 second to 10 seconds and is determined based on the actual network round-trip delay measurements between the two ends. For the currently evaluated node, search for all upstream nodes pointing to the node in the topology directed graph. For each upstream node, obtain the residual fault strength value of the upstream node after the propagation delay parameter before the current time. Multiply the residual fault strength value by the propagation attenuation coefficient from the upstream node to the currently evaluated node. Add up the above product results of all upstream nodes to obtain the cumulative fault injection amount of the currently evaluated node at the current time.

[0023] In step S4, the multi-scale detection function for computing resource distortion is invoked, with at least three time scale levels pre-defined, corresponding to short-scale, medium-scale, and long-scale operations. The interval step size for short-scale is set to 5 seconds, for medium-scale to 30 seconds, and for long-scale to 120 seconds. For the current moment being evaluated, at each scale level, the computing resource occupancy sequence prior to the current moment is extracted. The length of this sequence is the product of the interval step size for that scale level and the number of segments, which is set to 8. Starting from the current moment and moving backward, 8 difference intervals are sequentially divided according to the interval step size for that scale level. Each difference interval includes two sampling time points, and the time interval between the two sampling time points is exactly equal to the interval step size for that scale level. The absolute difference in computing resource occupancy values ​​between the two sampling time points within each difference interval is calculated. Simultaneously, calculate the standard deviation of all computing resource occupancy values ​​within a local window centered on the current time and extending 10 sampling points before and after it at this scale level. Add this standard deviation to a regularization constant with a value of 0.001. Then, divide the absolute difference between each difference interval by the sum of the above sums to obtain the normalized difference value corresponding to each difference interval. Multiply all the normalized difference values ​​between the eight difference intervals together to obtain the distortion detection quantity at this scale level.

[0024] In step S5, the global situational deviation calculated in step S2, the fault injection amounts at each level calculated in step S3, and the distortion flags at each scale calculated in step S4 are combined into a multidimensional evidence vector. The number of dimensions of the multidimensional evidence vector is equal to the sum of the total number of nodes and the total number of scale levels plus 1. This multidimensional evidence vector is input into the intelligent fault diagnosis platform for the computing power cluster. The intelligent fault diagnosis platform for the computing power cluster has a pre-set fault tree. Each leaf node of the fault tree corresponds to a fault type, and each leaf node is associated with an evidence interval. The upper and lower limits of the evidence interval are determined based on the statistical distribution in the historical fault case database. The platform calculates the membership degree of each component value of the multidimensional evidence vector falling into the evidence interval corresponding to each leaf node. The membership degree is calculated using linear interpolation, and the membership degree value ranges from 0 to 1. For all leaf nodes whose membership degree is greater than a preset acceptance threshold of 0.6, backtrack up the fault tree to the direct parent node of these leaf nodes. Summarize the membership degree of the child nodes according to the logic gate type of the parent node. When summarizing, if the logic gate type is AND gate, take the minimum membership degree of each child node; if the logic gate type is OR gate, take the maximum membership degree of each child node. Finally, output the fault type corresponding to the parent node with the highest membership degree after summarization as the candidate fault type.

[0025] When implementing step S6, using the candidate fault types output in step S5 as the retrieval index, the system queries the fault type and parameter anomaly pattern lookup table for at least two key parameter names associated with the candidate fault type and the preset outlier threshold for each key parameter. The key parameter names include at least two of the following: CPU utilization, memory read / write throughput, and disk I / O wait time. The preset outlier threshold is the boundary value determined by adding or subtracting three standard deviations from the mean of the parameter under normal operating conditions. For all sampling time points marked as the initial anomaly screening window in step S1, all values ​​of parameters with the same key parameter names as described above are extracted from these sampling time points and compared one by one with the preset outlier threshold. The number of values ​​exceeding the preset outlier threshold is counted. If, within the initial anomaly screening window, the cumulative number of times the value of at least one key parameter associated with the candidate fault type exceeds the preset outlier threshold reaches three or more times, the current anomaly event is determined to be valid, the candidate fault type is confirmed as the final anomaly identification result, and the final anomaly identification result is output to the operation and maintenance alarm system.

[0026] Preferably, the cross-node situational awareness model is represented as follows: ,in At the current assessment point, The length of the historical time window. Let be the parameter state matrix of all nodes at time t. The spatial gradient of the state matrix. It is the Frobenius norm. This contributes to the historical accumulation of activity levels. ,in The total number of nodes. For node indexing, For nodes The response decay coefficient, For nodes Historical cumulative trend activity weight, For nodes; the partial derivative of situational activity with respect to time. To achieve overall situational coordination and balance.

[0027] Specifically, the cross-node situational awareness model consists of two formulas with a progressive derivation relationship. The first formula represents the historical cumulative situational activity. Abnormal disturbances of any node in the data center cluster will inevitably show changes in the spatial gradient. Therefore, the spatial gradient is solved for the state matrix composed of all node parameters at each moment. The spatial gradient is obtained by taking the square root of the sum of the squares of the differences between adjacent node parameters in the row direction and the differences between adjacent node parameters in the column direction of the state matrix. The larger the magnitude of the gradient, the more severe the difference in parameters between adjacent nodes. After squaring the magnitude of the gradient, it is integrated with time over a historical time window of 300 seconds to obtain the historical cumulative situational activity within the window. This integration operation can eliminate the interference of instantaneous random fluctuations on situational judgment. The second formula establishes a global situational linkage equilibrium based on the output of the first formula. Its principle is that when the cluster is operating normally and coordinating consistently, the rate of change of the historical cumulative situational activity component of each node over time should be similar. If the rate of change of a certain node deviates abnormally from the mean, it indicates that the node is out of sync with the overall situation of the cluster. Therefore, the partial derivative of the historical cumulative situational activity component with respect to time is calculated for each node. The partial derivative is multiplied by the response attenuation coefficient of the node and the exponent is taken. The response attenuation coefficient is determined according to the location of the cabinet where the node is located. The value is 0.02 to 0.08 for nodes in high vibration or high temperature areas and 0.10 for other nodes. Finally, the exponents of all nodes are multiplied together. The multiplication operation can amplify the difference in the rate of change between nodes. When the product deviates significantly from 1, it indicates that the cross-node situational linkage is unbalanced. The two formulas form a recursive logical chain from local gradient accumulation to global coordination judgment.

[0028] Preferably, the topology-dependent fault propagation algorithm is expressed as: ,in This is the number of the currently evaluated node. For a set of upstream nodes that provide direct dependent inputs to a node, Number the upstream node. Let be the propagation attenuation coefficient from node k to node j. For node k in delay The residual fault strength afterward, The cumulative fault injection amount received by node j at time t; ,in Let be the real-time residual fault intensity of node j at time t. Let be the spontaneous fault baseline value of node j at time t. For integration time variable, Let be the fault cumulative gain coefficient for node j. Let be the fault dissipation rate coefficient of node j.

[0029] Specifically, the first formula of the topology-dependent fault propagation algorithm calculates the cumulative fault injection amount received by the current node. In the directed graph of the device topology, the fault input of any downstream node must be the sum of the residual fault strengths of all upstream nodes after propagation attenuation and transmission delay. Therefore, for the current evaluation node, all upstream nodes pointing to the node in the directed graph are traversed. The contribution of each upstream node is the residual fault strength of the node at one propagation delay moment in advance multiplied by the propagation attenuation coefficient corresponding to the path between the two nodes. The propagation attenuation coefficient is initially set to 0.65 to 0.75. According to the physical distance between the two nodes, the attenuation coefficient decreases by 0.01 for every 1 meter increase and the bandwidth utilization rate increases by 10 percent, the attenuation coefficient decreases by 0.02. The propagation delay is set between 1 second and 10 seconds based on the network round-trip time measurement value. The cumulative fault injection amount of the current node is obtained by summing the contributions of all upstream nodes. The second formula constructs a dynamic update mechanism for the real-time residual fault intensity. Its principle is that the residual fault intensity of the current node is determined by the spontaneously generated fault base value plus the cumulative injection part minus the dissipation part. The spontaneous fault base value is set to 0.05 to 0.15 through the equipment's factory failure rate parameter. The cumulative fault injection amount is subtracted from the product of the fault dissipation rate coefficient and the current residual fault intensity. The fault dissipation rate coefficient is set to the range of 0.01 to 0.03 based on the equipment's cooling efficiency. The difference is then multiplied by the fault accumulation gain coefficient, which ranges from 0.08 to 0.12. This product is integrated in the time domain from 0 to the current time. The residual fault intensity at the initial time of integration is 0. The integration operation can accumulate the net injection amount throughout the entire historical process, thereby obtaining the updated real-time residual fault intensity at the current time. The two formulas constitute a complete fault propagation recursive chain from input injection to state update.

[0030] Preferably, the computing power resource distortion multi-scale detection function is expressed as: ,in For scale-level indexing, This represents the number of difference segments at the current scale. The difference segment number, Let m be the interval step size at the m-th scale. For the computing resource occupancy sequence at time 1 The value at that location, For the m-th scale Local standard deviation centered on Here is the regularization constant. denoted as the distortion detection quantity at the m-th scale.

[0031] Specifically, the multi-scale detection function for computing resource distortion uses a single formula to measure the degree of distortion across multiple scales. This formula is based on the premise that the difference between adjacent intervals in the computing resource occupancy sequence under normal conditions should be within the range of local fluctuations. If significant deviations from the local standard deviation occur continuously at multiple intervals, it indicates that resource occupancy has become distorted. The formula is calculated for three pre-defined time scales: a short-scale interval of 5 seconds to capture transient spike distortion, a medium-scale interval of 30 seconds to capture continuous oscillating distortion, and a long-scale interval of 120 seconds to capture slow drift distortion. At each scale level, the current time is divided into 8 difference intervals according to the interval step size. The absolute difference in computing power resource usage at both ends of each difference interval reflects the change range within that interval. At the same time, the local standard deviation is calculated by extending 10 sampling points before and after the current time as the center. The local standard deviation represents the normal fluctuation benchmark near that time. The regularization constant is set to 0.001 to avoid overflow of division operation due to excessively small local standard deviation. The absolute difference between each difference interval is divided by the sum of the local standard deviation and the regularization constant to obtain the normalized difference value of each interval. This division makes the change range dimensionless relative to the local fluctuation benchmark, eliminating the dimensional influence of the fluctuation level at different times. Finally, the 8 normalized difference values ​​are multiplied together. The multiplication can accumulate the abnormal deviation degree within multiple intervals. A large normalized difference value in any interval will significantly amplify the final distortion detection amount. Each scale is calculated independently and compared with the preset threshold of the corresponding scale level to determine whether the distortion is valid at each scale.

[0032] Preferably, the multi-dimensional operating parameters collected in S1 include central processing unit utilization, memory read / write throughput, disk input / output waiting time, network interface packet loss and retransmission rate, temperature sensor readings, and fan speed feedback values. The parameters are acquired at a fixed sampling interval, with the sampling interval ranging from 1 to 5 seconds. When defining the initial screening abnormal window in S1, the deviation multiple of the current sampled value relative to the mean of the previous sliding window is calculated for each parameter. The sliding window length is 60 sampling points. When the deviation multiple of any parameter exceeds a preset dynamic threshold, the sliding window to which the current sampling time belongs is marked as the initial screening abnormal window. The preset dynamic threshold is determined by multiplying the difference between the upper quartile and the median of the deviation multiple distribution of the parameter in the past 24 hours by an adjustable coefficient, with the adjustable coefficient ranging from 3 to 5.

[0033] Preferred, such as Figure 2As shown, S2 includes: S21, for each node, arranging the multi-dimensional operating parameters collected in step S1 into local feature row vectors, and incorporating the computing power resource occupancy rate as an additional dimension into the tail of the row vectors; S22, stacking the local feature row vectors of all nodes vertically according to the node number to form an original feature matrix, and performing spatial gradient operation on the original feature matrix to obtain a spatial gradient matrix; S23, taking the sum of the squares of each element in the spatial gradient matrix as the spatial gradient modulus at that moment, and performing time integration on the spatial gradient modulus within the historical time window to obtain the historical cumulative situational activity; S24, taking the partial derivative of the historical cumulative situational activity components of each node with respect to time, and then performing exponential weighting and multiplication to obtain the global situational linkage balance.

[0034] Specifically, in step S21, for each node in the data center, the six parameters collected in step S1—CPU utilization, memory read / write throughput, disk I / O latency, network interface packet loss and retransmission rate, temperature sensor readings, and fan speed feedback—are arranged in a fixed order to form a local feature row vector with six elements. At the same time, the current GPU utilization and FPGA resource utilization of the node are added as additional dimensions to the end of the local feature row vector, so that the extended feature row vector of each node includes eight elements. In step S22, the extended feature row vectors of all N nodes are stacked vertically in ascending order of node number to form an original feature matrix with dimensions of N rows and 8 columns. The row direction difference is obtained by subtracting adjacent column elements within each row of this original feature matrix, and the column direction difference is obtained by subtracting adjacent row elements within each column. The squares of the corresponding values ​​of the row and column direction differences are then added together and the square root is taken to obtain the spatial gradient magnitude at that moment. In step S23, the spatial gradient magnitude calculated at each moment is first squared, and then 300 seconds are taken back from the current evaluation moment as a historical time window. The squared result at each moment within this historical time window is integrated over time using the trapezoidal integration method, with a time step equal to the sampling interval of 1 second, to obtain the historical cumulative situational activity. In step S24, the historical cumulative situational activity is decomposed into independent components for each node. The partial derivatives of each node component are calculated with respect to the evaluation time. The values ​​of the partial derivatives are calculated using the central difference method with a step size of 2 seconds. The partial derivatives are multiplied by the response attenuation coefficient of the node and the natural exponent is obtained. The response attenuation coefficient is determined by looking up a table based on the vibration level and temperature range of the cabinet where the node is located, and the value ranges from 0.02 to 0.10. The exponent results of all nodes are multiplied together to obtain the global situational linkage equilibrium.

[0035] Preferred, such as Figure 3As shown, S3 includes: S31, establishing a topological directed graph based on the physical connections and logical dependencies between data center devices, with directed edges in the topological directed graph pointing from upstream dependent nodes to downstream dependent nodes; S32, assigning a propagation attenuation coefficient and a propagation delay parameter to each directed edge, with the propagation attenuation coefficient dynamically adjusted based on the physical distance and bandwidth occupancy between the two ends; S33, for the current evaluated node, accumulating the product of the residual fault intensity of all upstream dependent nodes after delay and the corresponding propagation attenuation coefficient, as the cumulative fault injection amount of the current node; S34, amplifying the spontaneous fault base value and the cumulative fault injection amount of the current node through a fault accumulation gain coefficient, then subtracting the product of the fault dissipation rate coefficient and the real-time residual fault intensity, integrating this difference in the time domain to obtain the updated real-time residual fault intensity.

[0036] Specifically, in step S31, based on the physical connection relationship of the high-speed serial expansion bus between devices in the data center and the logical service call dependency relationship recorded by the upper-layer container orchestration system, each physical device or logical service unit is treated as a vertex in the directed topology graph. The direction from the upstream dependent device or service to the downstream dependent device or service is taken as the direction of the directed edge, thus constructing a complete directed topology graph and storing it as an adjacency list structure. In step S32, each directed edge in the directed topology graph is traversed, and each directed edge is assigned an initial propagation attenuation coefficient of 0.70. This coefficient is dynamically adjusted based on the physical distance between the two nodes of the directed edge and the current bandwidth occupancy. The attenuation coefficient decreases by 0.01 for every 1 meter increase in physical distance (in meters), and decreases by 0.02 for every 10 percent increase in bandwidth occupancy. Simultaneously, a propagation delay parameter is set based on the network round-trip delay measurement between the two nodes. This delay parameter continuously takes values ​​between 1 and 10 seconds. In step S33, for the node to be evaluated, the set of all upstream nodes pointing to the node is searched in the adjacency list of the directed topology graph. For each upstream node in the set, the residual fault strength value of the upstream node at the current evaluation time is obtained by subtracting the propagation delay parameter corresponding to the upstream node from the current node. This residual fault strength value is multiplied by the corresponding propagation attenuation coefficient. The above multiplication of all upstream nodes in the set is summed to obtain the cumulative fault injection amount of the current node. In step S34, the spontaneous fault base value of the current node is set to 0.10. The spontaneous fault base value is calculated based on the mean time between failures (MTBF) of the equipment. The cumulative fault injection amount is subtracted from the product of the fault dissipation rate coefficient and the real-time residual fault strength before the current time. The fault dissipation rate coefficient is set to 0.02 based on the speed level of the equipment cooling fan. This difference is multiplied by the fault accumulation gain coefficient of 0.10. Time integration is performed from the initial time 0 to the current time, with an integration step size equal to the sampling interval of 1 second. The forward Euler method is used for iterative calculation to obtain the updated real-time residual fault strength.

[0037] Preferred, such as Figure 4 As shown, S4 includes: S41, setting multiple time scale levels, each scale level corresponding to a different interval step size, the interval step size increasing as the scale level increases; S42, for the current scale level, dividing the computing power resource occupancy sequence into multiple difference intervals according to the interval step size, each difference interval including two sampling points separated by the interval step size; S43, calculating the absolute difference between the sampling points at both ends of each difference interval, and dividing the absolute difference by the sum of the local standard deviation centered at the current time and the regularization constant under the current scale level, to obtain the normalized difference value between each difference interval; S44, performing a multiplication operation on the normalized difference values ​​between all difference intervals under the current scale level to obtain the distortion detection quantity under the current scale level.

[0038] Specifically, in step S41, three time scale levels are defined and labeled as short-scale, medium-scale, and long-scale levels. The interval step size for the short-scale level is fixed at 5 seconds to capture transient spikes in computing resource usage; the interval step size for the medium-scale level is fixed at 30 seconds to capture continuous oscillating resource fluctuations; and the interval step size for the long-scale level is fixed at 120 seconds to capture slowly drifting resource trend changes. The interval step size for each scale level increases strictly with the scale level. In step S42, for the current moment being evaluated, at each scale level, starting from the current moment, eight difference intervals are sequentially divided backward according to the interval step size of that scale level. The first difference interval consists of two sampling points separated by one interval step from the current moment; the second difference interval consists of two sampling points separated by another interval step, and so on until the eighth difference interval. The time difference between the two sampling points within each difference interval is exactly equal to the interval step size of that scale level. In step S43, for each difference interval, the computing power resource occupancy values ​​at the two sampling time points of the interval are taken and the absolute value of the difference between them is calculated. At the same time, a local window is formed by extending 10 sampling points forward and backward with the current time as the center position. The total length of the local window is 21 sampling points. The standard deviation of all computing power resource occupancy values ​​in the local window is calculated. The standard deviation is added to the regularization constant 0.001 to obtain the denominator. Then, the absolute difference between the difference intervals is divided by the denominator to obtain the normalized difference value between the difference intervals. In step S44, the normalized difference values ​​corresponding to all 8 difference intervals under the current scale level are multiplied sequentially. The initial value of the multiplication is set to 1.0. The first to eighth normalized difference values ​​are multiplied one by one, and the final product is the distortion detection quantity under the current scale level. Steps S42 to S44 are executed independently for each of the three scale levels, and the distortion detection quantity of each of the three scale levels is output. Each distortion detection quantity is compared with the preset threshold of the corresponding scale level. If the distortion detection quantity exceeds the threshold, the distortion flag position of the scale is set to logical true.

[0039] Preferred, such as Figure 5 As shown, S5 includes: S51, combining the global situational deviation, the amount of fault injection at each level, and the distortion flag into a multi-dimensional evidence vector, and inputting the multi-dimensional evidence vector into the front-end interface of the intelligent fault diagnosis platform for the computing power cluster; S52, the platform matches the multi-dimensional evidence vector with the evidence intervals corresponding to each leaf node in the built-in fault tree, and calculates the membership degree of the multi-dimensional evidence vector falling into the evidence interval of each leaf node; S53, for leaf nodes whose membership degree exceeds the preset acceptance threshold, tracing upwards along the fault tree to the intermediate node, and summarizing the membership degree of each leaf node according to the logic gate type of the intermediate node; S54, outputting the fault type corresponding to the intermediate node with the highest summarized membership degree as a candidate fault type.

[0040] Specifically, in step S51, the global situational deviation output in step S2 is used as a scalar evidence component. The real-time residual fault intensity of all N nodes output in step S3 is arranged sequentially to form an N-dimensional evidence sub-vector. The three scale-level distortion flags output in step S4 are arranged in a fixed order of short scale, medium scale, and long scale to form a 3-dimensional evidence sub-vector. The above scalar evidence component, N-dimensional evidence sub-vector, and 3-dimensional evidence sub-vector are concatenated in the order of scalar first, N-dimensional vector in the middle, and 3-dimensional vector last to form a multi-dimensional evidence vector with dimension N plus 4. This multi-dimensional evidence vector is then transmitted to the front-end data receiving interface of the intelligent fault diagnosis platform for the computing power cluster. In step S52, the intelligent fault diagnosis platform for the computing power cluster has a pre-set fault tree, which includes 30 leaf nodes and 15 intermediate nodes. Each leaf node is associated with a fault type and stores the upper and lower limits of the evidence interval corresponding to that fault type. The upper and lower limits of the evidence interval are determined based on the statistical distribution of each evidence component in the historical fault case library. The platform compares each component of the multidimensional evidence vector with the evidence interval of each leaf node one by one, and uses a linear interpolation method to calculate the membership degree of each component falling into the interval. The membership degree value range is 0 to 1. In step S53, all leaf nodes with membership degree values ​​greater than the preset acceptance threshold of 0.6 are extracted. The direct parent nodes of these leaf nodes are found from the fault tree and classified according to the parent node number. For all child leaf nodes under the same parent node, the logic gate type of the parent node is checked. If the logic gate type is AND gate, the minimum value of the membership degree of each child leaf node is taken as the total membership degree of the parent node. If the logic gate type is OR gate, the maximum value of the membership degree of each child leaf node is taken as the total membership degree of the parent node. In step S54, the membership values ​​of each intermediate node after aggregation are compared among all 15 intermediate nodes. The intermediate node with the highest aggregation membership is selected, and the predefined fault type name of the intermediate node is read. The fault type name is then output as a candidate fault type to the result cache area of ​​the diagnostic platform for use in step S6.

[0041] A method for intelligent identification of equipment anomalies in data center operations and maintenance is disclosed. This method is implemented through an intelligent equipment anomaly identification system for data center operations and maintenance, comprising: a heterogeneous parameter synchronous acquisition unit, used to synchronously acquire multi-dimensional operating parameters and computing resource occupancy rates of each device using a unified clock reference and temporarily store them in a circular cache; a spatiotemporal situation feature tensor construction unit, with its input connected to the output of the heterogeneous parameter synchronous acquisition unit, used to map the parameter sequences of each node into a situation feature tensor and calculate the global situation deviation; and a topology dependency recursive tracing unit, with its input connected to the outputs of the heterogeneous parameter synchronous acquisition unit and the spatiotemporal situation feature tensor construction unit respectively, used to recursively calculate the topology dependency of each node. The system includes: a cumulative fault injection quantity and root cause node tracing; a multi-scale distortion parallel detection unit, whose input is connected to the output of the heterogeneous parameter synchronous acquisition unit, used to calculate the distortion detection quantity of the computing resource occupancy sequence in parallel at multiple time scales and output distortion flags at each scale; a fault tree matching adjudication unit, whose input is connected to the output of the spatiotemporal situation feature tensor construction unit, the topology dependency recursive tracing unit, and the multi-scale distortion parallel detection unit, used to adjudicate candidate fault types according to evidence matching rules; and a reverse verification confirmation unit, whose input is connected to the output of the fault tree matching adjudication unit, used to reverse verify the statistical outlier degree within the initial screening anomaly window based on the candidate fault type and output the final anomaly identification result.

[0042] A novel intelligent identification method for equipment anomalies in data center operations and maintenance is proposed. This method utilizes a cross-node situational awareness model to synchronously perform spatial gradient calculations and temporal integration on the state matrices of all equipment parameters, generating a deviation index characterizing the global collaborative situational shift. This allows for the comprehensive capture of latent anomalies arising from the combined effects of multiple nodes, overcoming the shortcomings of existing technologies that miss cluster collaborative shift anomalies due to fragmented parameter processing. This significantly improves the early detection capability of anomaly precursors. Furthermore, a topology-dependent fault propagation algorithm dynamically maintains a directed graph based on the physical connections and logical dependencies between devices. Each propagation edge is assigned an attenuation coefficient and delay parameter that adaptively adjusts with load and distance. By accumulating the upstream residual fault intensity and performing temporal integration, the fault injection quantity at each level is updated in real time. This allows the fault propagation path to dynamically evolve with changes in cluster topology and load, eliminating reliance on a static rule base. This significantly improves the timeliness and accuracy of root cause tracing, effectively solving the problem of existing technologies lagging behind the actual fault propagation process.

[0043] This invention introduces a multi-scale detection function for computing resource distortion. It performs piecewise difference and local standard deviation normalization multiplication operations on the resource occupancy sequence at different time scales, outputting distortion flags at each scale. This enriches the diversity of anomaly criteria and avoids the omission of short-term transient distortions or long-term slow drifts by single-scale detection. Finally, the intelligent fault diagnosis platform for computing clusters integrates global situational deviation, fault injection amounts at each level, and distortion flags into a multi-dimensional evidence vector. Through dynamic matching and membership aggregation using a built-in fault tree, it outputs candidate fault types. A reverse verification mechanism then confirms the anomaly's validity, reducing the false alarm rate, improving the reliability of diagnostic results, shortening the anomaly recovery cycle, and ensuring the continuous and stable operation of the data center computing cluster.

[0044] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of equipment anomalies in data center operations and maintenance, characterized in that, Includes the following steps: S1 collects multi-dimensional operating parameters and computing resource utilization of data center operation and maintenance equipment, and defines the initial screening abnormal window based on the parameter time-series fluctuation amplitude. S2, based on the cross-node situational awareness model, maps the parameter sequences of each node into situational feature tensors and uses the parameter cooperation between nodes to calculate the global situational deviation. S3 executes the topology-dependent fault propagation algorithm, constructs directed edges for fault propagation based on the device topology connection relationship, traces the root cause node along the directed edges level by level and quantifies the fault injection amount at each level; S4 calls the multi-scale detection function for computing resource distortion, performs distortion measurement on the computing resource occupancy sequence at multiple time scales, and outputs distortion flags at each scale; S5 inputs the global situational deviation, the amount of fault injection at each level, and the distortion flag bit into the intelligent fault diagnosis platform of the computing power cluster, and determines the candidate fault type through the fault tree matching rules built into the platform. S6 uses the candidate fault type as an index to reverse-check the statistical outlier degree of each parameter in the initial screening anomaly window and determine whether the anomaly event is valid.

2. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, The cross-node situational awareness model is represented as follows: ,in At the current assessment point, The length of the historical time window. Let be the parameter state matrix of all nodes at time t. The spatial gradient of the state matrix. It is the Frobenius norm. This contributes to the historical accumulation of activity levels. ,in The total number of nodes. For node indexing, For nodes The response decay coefficient, For nodes Historical cumulative trend activity weight, For nodes; the partial derivative of situational activity with respect to time. To achieve overall situational coordination and balance.

3. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, The topology-dependent fault propagation algorithm is expressed as follows: ,in This is the number of the currently evaluated node. For a set of upstream nodes that provide direct dependent inputs to a node, Number the upstream node. Let be the propagation attenuation coefficient from node k to node j. For node k in delay The residual fault strength afterward, The cumulative fault injection amount received by node j at time t; ,in Let be the real-time residual fault intensity of node j at time t. Let be the spontaneous fault baseline value of node j at time t. For integration time variable, Let be the fault cumulative gain coefficient for node j. Let be the fault dissipation rate coefficient of node j.

4. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, The multi-scale detection function for computing power resource distortion is expressed as follows: ,in For scale-level indexing, This represents the number of difference segments at the current scale. The difference segment number, Let m be the interval step size at the m-th scale. For the computing resource occupancy sequence at time 1 The value at that location, For the m-th scale Local standard deviation centered on Here is the regularization constant. denoted as the distortion detection quantity at the m-th scale.

5. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, The multi-dimensional operating parameters collected in S1 include central processing unit utilization, memory read / write throughput, disk input / output latency, network interface packet loss and retransmission rate, temperature sensor readings, and fan speed feedback. The parameters are acquired at fixed sampling intervals, with the sampling interval ranging from 1 to 5 seconds. When defining the initial screening abnormal window in S1, the deviation multiple of the current sampled value relative to the mean of the previous sliding window is calculated for each parameter. The sliding window length is 60 sampling points. When the deviation multiple of any parameter exceeds a preset dynamic threshold, the sliding window to which the current sampling time belongs is marked as the initial screening abnormal window. The preset dynamic threshold is determined by multiplying the difference between the upper quartile and the median of the deviation multiple distribution of the parameter in the past 24 hours by an adjustable coefficient, with the adjustable coefficient ranging from 3 to 5.

6. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, S2 includes: For each node, the multidimensional running parameters collected in step S1 are arranged into a local feature row vector, and the computing power resource utilization rate is added as an additional dimension to the end of the row vector; The local feature row vectors of all nodes are stacked vertically according to the node number to form the original feature matrix. Spatial gradient operation is performed on the original feature matrix to obtain the spatial gradient matrix. The sum of the squares of each element in the spatial gradient matrix is ​​taken as the spatial gradient modulus at that moment. Time integration is performed on the spatial gradient modulus within the historical time window to obtain the historical cumulative situational activity. The partial derivative of the historical cumulative situational activity components of each node with respect to time is obtained by exponential weighting and multiplication to obtain the global situational linkage balance.

7. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, S3 includes: A topological directed graph is established based on the physical connections and logical dependencies between data center devices. In the topological directed graph, directed edges point from upstream dependent nodes to downstream dependent nodes. Each directed edge is assigned a propagation attenuation coefficient and a propagation delay parameter. The propagation attenuation coefficient is dynamically adjusted based on the physical distance between the two nodes and the bandwidth occupancy rate. For the currently evaluated node, the product of the residual fault intensity of all upstream dependent nodes after delay and the corresponding propagation attenuation coefficient is accumulated as the cumulative fault injection amount of the current node. The spontaneous fault base value and the cumulative fault injection amount of the current node are amplified by the fault accumulation gain coefficient, and then the product of the fault dissipation rate coefficient and the real-time residual fault intensity is subtracted. The difference is then integrated in the time domain to obtain the updated real-time residual fault intensity.

8. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, S4 includes: Multiple time scale levels are set, each scale level corresponds to a different interval step size, and the interval step size increases as the scale level increases; for the current scale level, the computing power resource occupancy sequence is divided into multiple differential intervals according to the interval step size, and each differential interval includes two sampling points separated by the interval step size. Calculate the absolute difference between the sampling points at both ends of each difference interval, and divide the absolute difference by the sum of the local standard deviation centered at the current time and the regularization constant at the current scale level to obtain the normalized difference value between each difference interval; perform a multiplication operation on the normalized difference values ​​between all difference intervals at the current scale level to obtain the distortion detection quantity at that scale level.

9. The intelligent identification method for equipment anomalies in data center operation and maintenance according to claim 1, characterized in that, S5 includes: The global situational deviation, the amount of fault injection at each level, and the distortion flag are combined into a multidimensional evidence vector, which is then input into the front-end interface of the intelligent fault diagnosis platform for the computing power cluster. The platform matches the multidimensional evidence vector with the evidence intervals corresponding to each leaf node in the built-in fault tree and calculates the membership degree of the multidimensional evidence vector falling into the evidence interval of each leaf node. For leaf nodes whose membership degree exceeds the preset acceptance threshold, the fault tree is traced upwards to the intermediate node, and the membership degree of each leaf node is summarized according to the logic gate type of the intermediate node. The fault type corresponding to the intermediate node with the highest membership degree after aggregation is output as the candidate fault type.

10. A method for intelligent identification of equipment anomalies in data center operation and maintenance according to any one of claims 1-9, characterized in that, This method is implemented through an intelligent equipment anomaly identification system for data center operations and maintenance, including: The heterogeneous parameter synchronous acquisition unit is used to synchronously acquire the multi-dimensional operating parameters and computing resource utilization of each device with a unified clock reference and temporarily store them in the circular cache. The spatiotemporal situation feature tensor construction unit, whose input end is connected to the output end of the heterogeneous parameter synchronous acquisition unit, is used to map the parameter sequence of each node into a situation feature tensor and calculate the global situation deviation. The topology dependency recursive tracing unit has its input end connected to the output end of the heterogeneous parameter synchronous acquisition unit and the output end of the spatiotemporal situation feature tensor construction unit, respectively. It is used to recursively calculate the cumulative fault injection amount of each node along the device topology dependency relationship and trace the root cause node. The multi-scale distortion parallel detection unit, with its input end connected to the output end of the heterogeneous parameter synchronous acquisition unit, is used to calculate the distortion detection quantity of the computing power resource occupancy sequence in parallel at multiple time scales and output the distortion flag bits at each scale. The fault tree matching adjudication unit has its input connected to the output of the spatiotemporal situation feature tensor construction unit, the topological dependency recursive tracing unit, and the multi-scale distortion parallel detection unit, and is used to adjudicate candidate fault types according to evidence matching rules. The reverse verification confirmation unit, whose input is connected to the output of the fault tree matching decision unit, is used to reverse verify the statistical outlier degree within the initial screening anomaly window based on the candidate fault type and output the final anomaly identification result.

Images