Enhanced Adaptive Autoencoder Framework for Continuous Monitoring and Fault Management in Industrial Sensors

Ensuring the reliability and safety of industrial processes requires robust monitoring systems capable of detecting and diagnosing issues in complex, dynamic, and high-dimensional data environments, even when some sensors are faulty, under maintenance, or being recalibrated. Autoencoder (AE)-based methods face a major limitation: they require retraining whenever a sensor fails or is temporarily unavailable, which limits their practicality in large-scale systems. This need for reconfiguration can be impractical for large-scale industrial systems, leading to increased downtime and computational overhead. To address this challenge, we propose an AE-based framework with virtual switches that dynamically isolate faulty sensors without requiring retraining. This enhances adaptability, ensuring continuous monitoring and improving operational efficiency. The motivation for this approach lies in its ability to maintain fault detection accuracy while reducing the complexity of managing sensor failures in real-world industrial scenarios. This study introduces a novel framework that integrates virtual switches, allowing the AE model to function seamlessly despite missing sensors. The model is developed and trained using the TensorFlow framework, ensuring continuous monitoring and adaptability to changing process conditions without the need for retraining. The proposed framework’s effectiveness is validated using the Tennessee Eastman Process (TEP) benchmark, demonstrating its operational efficiency and fault detection accuracy. This approach addresses critical challenges in large-scale industrial monitoring, providing a scalable and efficient solution for sensor fault management and system reliability.