Continual Learning Algorithm for Online Cloud Anomaly Detection

Abstract

Addressing the challenges of continuously arriving monitoring data, evolving anomalies, and the inadequacy of traditional static models to adapt to long-term online scenarios in cloud computing environments, this paper proposes a continuous learning method for online cloud anomaly detection. This method takes multi-dimensional monitoring time series as input, constructs continuous observation samples through a sliding window, utilizes 1DCNN to extract local mutations, short-term fluctuations, and fine-grained anomaly features, and combines LSTM to model the temporal context and state evolution relationship, thereby improving the ability to identify complex time-series anomalies. Furthermore, a continuous learning constraint is introduced to mitigate the knowledge forgetting problem during online updates, ensuring the model maintains a stable representation of existing anomaly knowledge while absorbing new anomaly patterns. This paper focuses on model structure design, data representation methods, and online detection requirements, forming a unified framework that balances local feature learning, long-term dependency modeling, and continuous adaptation capabilities, providing effective methodological support for online cloud anomaly detection tasks.