Incremental Learning Through Fusion of Discrete Anomaly Models from Odometry Signals in Autonomous Agent Navigation

This paper presents a dynamic data-driven approach for efficient anomaly detection, extraction, and fusion of multiple heterogeneous anomaly models in a generative fashion. First, we propose an adaptive Bayesian filtering technique based on a combination of Null force hypothesis and Particle filtering to accurately track the trajectories of normal and abnormal cases. We then analyze the generalized vectors and clusters generated from adaptive filtering and sequential clustering procedures to effectively detect areas with high abnormalities. To achieve this, we use probabilistic distance measurements. Finally, to increase the agent's vocabulary, we fuse different anomaly distributions to generate coupled anomaly models that allow the agent to have incremental learning capabilities. Our approach is completely data-driven and does not require any previous knowledge of the data or the environment. We show that our proposed method can effectively detect anomalies using low-dimensional odometry data and can eventually improve itself over time through iterative generation of fused anomaly models.