| CPC G06F 11/079 (2013.01) [G06F 11/0712 (2013.01)] | 20 Claims |
|
1. A computer-implemented method, comprising:
training, by a computer system, a graph neural network (GNN) to generate a trained GNN from training data that includes historical time series data;
retrieving output from the trained GNN that has performed inference on a plurality of time series, wherein the output includes an indication, for individual ones of the plurality of time series, as to whether the time series contains an anomaly, dependencies between different ones of the plurality of time series, and values indicating a strength of the dependencies;
generating, by the computer system, a dependency graph based on the output from the GNN, wherein the dependency graph is a data structure comprising a plurality of nodes and a plurality of directed edges, each of the plurality of nodes representing a time series and the plurality of directed edges representing dependencies between the time series;
removing from the data structure, by the computer system, one or more nodes representing time series in which no anomalies are found, as well as directed edges associated with the one or more nodes, to create one or more sub-data structures; and
for a first sub-data structure comprising multiple nodes:
applying, by the computer system, a root cause analysis (RCA) algorithm to the first sub-data structure to determine a root cause sub-data structure;
identifying, by the computer system, a root cause anomaly within a time series corresponding to a root node of the root cause sub-data structure; and
resolving the root cause anomaly by automatically repairing a system associated with the root node.
|