Industrial Anomaly Detection Based on Factorized Temporal-Channel Fusion

In today’s rapidly advancing technological landscape, intelligent manufacturing factories play a crucial role in enhancing productivity and reducing costs. A key challenge in this field is minimizing machine downtime to improve economic efficiency, enhance system reliability, and reduce maintenance costs, which can be effectively addressed by detecting anomalies in machine-related sensor data. However, the data generated by industrial sensors are characterized by highly nonlinear time dependencies and complex interactions between multidimensional variables, making anomaly detection particularly challenging. This article introduces a novel time-series anomaly detection model—temporal-channel fusion (TCF) framework—designed to address these challenges. By leveraging the Transformer architecture, the TCF framework models the complex multivariate correlations present in sensor data. The core of our approach is the factorized TCF (FTCF) block, which integrates temporal and channel information through a unique factorization strategy. This method thoroughly mines the dynamic characteristics of the data, efficiently extracting critical industrial time-series information. To enhance the model’s performance, we implement a dual-stage enhancement strategy that amplifies the differences between reconstructed and original data, thereby improving the model’s generalization and stability. This capability enables the TCF framework to more effectively differentiate anomalous data, facilitating predictive fault classification in intelligent manufacturing systems and paving the way for reliable predictive maintenance. Our extensive evaluation on five real-world industrial datasets demonstrates the TCF framework’s significant superiority over state-of-the-art time-series anomaly detection methods, highlighting its practical value in real-world manufacturing environments, such as equipment fault warning systems in manufacturing and anomaly detection in energy pipeline monitoring.