Adaptive State-Space Mamba for Real-Time Sensor Data Anomaly Detection

Modern sensor networks generate large volumes of sequential data in real time. Detecting unusual patterns or anomalies within these streams is crucial for numerous applications, ranging from industrial process monitoring to environmental surveillance and predictive maintenance. Traditional anomaly detection approaches often rely on static feature engineering or rigid statistical assumptions, limiting their applicability in dynamic environments. Recently, deep learning models have emerged as powerful alternatives for sequence modeling, with architectures such as recurrent neural networks (RNNs) and transformers [4] achieving impressive performance. Nevertheless, these methods can be resource-intensive for streaming data scenarios, where real-time or near-real-time processing is essential. State-space models (SSMs) offer a principled approach to describing the evolution of a hidden state as a function of inputs and outputs [1]. While these models have been actively studied for signal processing and time-series forecasting, they have gained traction in broader AI tasks as well. For instance, Wang and Liu [5] introduced a novel usage of a state-space model (called "Mamba") for efficient text-driven image style transfer. Inspired by the notion of modeling sequential dependencies via hidden states, we adapt the core Mamba idea to the entirely different application of real-time anomaly detection in sensor data.