Anomaly Detection of Nodes in Cloud Based on Transfer Learning

The cloud platform is made up of a series of computing nodes and storage nodes. The timely detection and remediation of node anomalies in the cloud is of critical importance. Traditional anomaly detection methods struggle to effectively manage discontinuous changes in node status. To address this issue, we propose a method for anomaly detection of multidimensional time series data based on pre-trained models and transferring the problem to the image domain, named Sequence to Image Anomaly Detection (StI-AD). This method implements a multidimensional data anomaly detection method based on visual domain pre-training models. By transforming time series data into polar coordinates, the temporal correlation of the coordinates can be reflected. A series of experiments show that this method has a 5–7 percent improvement in accuracy compared to traditional anomaly detection methods and significantly shortens the model training cycle.