Unified Stationary and Nonstationary Data Representation for Process Monitoring in IIoT

Abstract

The Industrial Internet of Things (IIoT), which integrates industrial systems with advanced computing, communication, and control technologies, has become the mainstream of industrial manufacturing. Due to the large scale and complexity of the modern industry, industrial processes are characterized by multimode and mixed stationary and nonstationary variables. At the same time, faulty data in industrial processes, especially the small ones, are easily concealed by the normal variation trend of nonstationary data, which brings challenges to the process monitoring task. To facilitate the process monitoring within the framework of IIoT, a stationary and nonstationary data representation method for process monitoring is proposed, which combines the cointegration analysis and the representation learning synergistically. In detail, a cointegration model is established to extract the long-term equilibrium relationship between nonstationary variables to eliminate their negative effects. The equilibrium relationship, namely, stationary residuals, is fused with stationary variables and then reconstructed by a joint dictionary learning method. Hereafter, using the kernel density estimation method, the control limit can be calculated by the reconstruction error. Consequently, when online data samples arrive, we use the cointegration model and dictionary to reconstruct the data. Process monitoring can be realized timely by the reconstruction error. Extensive experiments, including a numerical simulation, a benchmark penicillin fermentation process, and an industrial roasting process, are used to verify the superiority and effectiveness of the proposed method for process monitoring based on IIoT. Our experimental results also demonstrate that the proposed method can detect small faults of the multimode process with mixed stationary and nonstationary variables.