Efficient sensor anomaly detection using Markov-LSTM architecture for methane sensing

Abstract

The integration of the Internet of Things (IoT) into industrial activities has unlocked myriad possibilities, particularly in applications like environmental monitoring, which facilitates effective landfill management. Nevertheless, IoT environments present challenges, including resource constraints, heterogeneity and potential hardware/software failures. These issues often lead to sensor anomalies, triggering false alarms and stalling data-driven systems. Existing models for edge devices frequently overlook the sensor life cycle, leading to extensive training times and significant computational demands. In this paper, a collaborative approach is proposed wherein a Markovian architecture gauges the operational state of a sensor, assisting the Long Short-Term Memory (LSTM) model in outlier detection within real-world data. Commercially available MQ-4 sensor alongside a microwave RADAR-based Methane (CH4) sensor in a tandem setup is employed to evaluate our methodology. The Bathtub curve and the Pearson Correlation Coefficient (PCC) function as the switching mechanisms for the Markov chain. Real-time data validation yielded an impressive 92.57% accuracy and 94.86% efficiency in anomaly detection. When benchmarked against the Autoregressive Integrated Moving Average (ARIMA) and the Prophet algorithm, our method demonstrated superior anomaly rejection rates of 9.63% and 3.01%, respectively. Implementing the Markov-LSTM model in methane sensing significantly enhances the accuracy of recorded sensor values compared to standard methane sensors.