KPI Anomaly Detection Method for Multi-Scale Feature Fusion in Time-Frequency Domain

KPI anomaly detection is a key task in the field of intelligent operation and maintenance. Existing methods are based on time series anomaly detection, which have the problem of weak feature extraction ability in operation and maintenance scenarios, especially the insufficient utilization of frequency domain information. This paper proposes an unsupervised KPI anomaly detection method based on multi-scale feature fusion in timefrequency domain for intelligent operation and maintenance of data center. In this method, the one-dimensional convolutional neural network with different scales convolution kernels is configured to extract the multi-scale features in time domain and frequency domain respectively, and the cross-domain attention fusion module is used to realize the dynamic weighted interaction of time-frequency features, so as to enhance the feature expression ability. The variational autoencoder is used to reconstruct the time series data, and anomalies are determined by the reconstruction probability. Experimental results on three real server KPI datasets show that the proposed method achieves the best recall rate and F1 value compared with other baseline methods, and the F1 value is $0.95,0.97$ and 0.94, respectively. It is worth noting that the introduction of frequency domain features significantly improves the recall rate, which is $0.98,0.96$ and 0.96, respectively, which verifies the key role of frequency domain information in KPI anomaly detection, and provides an effective solution to the problem of weak feature extraction ability of existing methods.