OD-ViTFScL: Asynchronous Few-Shot Continual Learning for Intelligent Process Monitoring and Fault Diagnosis in Petrochemical Plants

Abstract

Petrochemical plants are complex facilities composed of interconnected equipment to produce essential products for daily human activities. In view of the adoption of the Industrial Internet of Things (IIoT), these facilities use various sensors, including those for level, flow, temperature, and pressure, to monitor and control operations. Several process faults can be detected and classified using data from these sensors. Detecting and classifying these faults helps mitigate production losses, improve product quality, decrease environmental concerns, and improve human safety. Conventional static learning, which learns from historical data, and continual learning, which learns incremental tasks with known preset boundaries, fall short on asynchronous online unknown streaming boundary tasks such as the OpenWorld Problem, which is the true representation of real-world dynamic plant scenarios. To tackle these challenges, we propose a novel online asynchronous framework termed OD-ViTFScL. Our framework utilizes a dual-attention backbone, which concatenates orthogonal information from attention on the vertical time-series data and horizontal inter-sensor relationships. The fused MLP from the temporal and sensor streams learns independent weights for each orthogonal axis. Our framework also uses ODIN, an out-of-distribution technique, to trigger the arrival of a new fault class in the streaming data. Extensive experiments on the Tennessee Eastman Process (TEP) and Case Western Reserve University (CWRU) bearing datasets validate that our proposed OD-ViTFScL outperforms other state-of-the-art fault detection and diagnosis (FDD) approaches.