A survey of data-driven fault-diagnosis methods for large-scale industrial production processes

Fault diagnosis for large-scale industrial production systems has attracted considerable research interest in response to the complex, multisource, and precision requirements of these processes. Fault diagnosis is crucial for the safe, reliable, and real-time maintenance of industrial production processes. This work presents a comprehensive survey of fault-diagnosis methods and emphasizes two cornerstone strategies: data-driven paradigms and distributed methods. Traditional fault-diagnosis methods based on mechanism have limited applications because precise modeling of the systems considered is required. Data-driven approaches avoid the dependence on precise modeling; thus, the research focus on fault diagnosis for industrial production processes has gradually shifted from mechanism-based to data-driven methods that integrate historical data, real-time data, and multisource information to enhance the accuracy and efficiency of the fault detection and identification approaches. A comprehensive overview of data-driven fault-diagnosis methods is given in the first part of this work. Specifically, smooth data from industrial processes is collected and used for fault diagnosis. The internal state variables may continuously change, and accompanied with time correlation among the process measurements. Integrating data-driven and dynamic analyses, necessitated by the dynamic nature of the process variables, offers a more accurate representation of system behavior. This can be achieved through the introduction of time-series modeling in basic multivariate statistics and the capture of dynamic properties by subspace identification. Furthermore, deep learning and kernel methods address nonlinearity, and non-Gaussian traits are tackled by independent component analysis (ICA) and other methods. Second, distributed fault-diagnosis methods for large-scale industrial production processes are reviewed. The usual fault-diagnosis methods for large-scale systems rely on centralized sensor network monitoring. Centralization necessitates consolidated data processing, which can create immense computational stress. Applying distributed fault-diagnosis methods spreads the monitoring capacities among all the subsystems, enabling each subsystem to independently assess its safety and performance based on its own data and interactions with neighboring subsystems. The latest advances in system decomposition and data fusion, correlation analysis, and consensus mechanisms are elaborated on subsystems based on the distributed structure and sensor networks in large-scale systems. System decomposition focuses on the local features of each subsystem although adopting a purely subsystem-centric approach to data processing often leans toward decentralization, in contrast to the essence of distribution. Thus, the effective integration of data is critical for achieving comprehensive information about large-scale systems. Correlation analysis examines the intricate relationships among subsystems or nodes. It elucidates mutual interactions and uncovers dependencies and influences. Meanwhile, consensus analysis focuses on the communication topology of the nodes, ensuring that all nodes converge to a unified data state at any instant. Finally, the practical applications for evaluating the performance of distributed data-driven fault-diagnosis methods are summarized. The potential trends are also highlighted, including qualitative and quantitative methods integration, improved diagnosis robustness, and data security assurance.