Anomaly detection by using multimodal deep learning
Guangxin Jiang, Zhenyu Wan, Kunhan Wang, Jinsi Han, Shengli Wu, Min 、Tong · 2023
In the field of power systems, the detection of multidimensional fault data is essential. Introducing the concept of IT Operations Artificial Intelligence (AIOps), combined with big data and machine learning techniques, aims to take place extensive IT operational tasks, including service availability and performance monitoring. Multiple models are utilized to process monitoring data patterns, such as text attributes and real-value response times extracted from logs and traces, enabling the detection of faults and potential anomalies in cloud services. To detect anomalies in the execution of system components, a dual-mode distributed tracing data from large-scale cloud infrastructure is employed, and a novel method for anomaly detection is proposed. The application of LSTM (Long Short-Term Memory) multimodal neural networks is demonstrated to learn the sequential characteristics of the two modes of data in traces. The capability to detect dependencies and concurrent events is showcased through a method that reconstructs execution paths using the proposed models. In experimental evaluations using large-scale production cloud data, the new approach outperforms traditional architectures and other deep learning methods.