Fault detection using Fourier neural operator

In order to generate higher-quality products and increase process efficiency, there has been a strong push in the processing and manufacturing sectors. This has called for the creation of methods to identify and fix faults to ensure optimal performance. As a result, it is essential to develop monitoring systems that can effectively detect and identify these faults so that operators can quickly resolve them. This article proposes a novel fault detection method that adopts a deep learning approach using a Fourier neural operator (FNO) in a probabilistic way, an operator learning model that aims to learn the distribution of multivariate process data and apply them for fault detection. Herein, the historical data under normal process conditions were first utilized to construct a multivariate statistical model; after that, the model was used to monitor the process and detect faults online. The proposed FNO combines the integral kernel with Fourier transformation in a probabilistic way. As the Fourier transform helps in the time-frequency localization of time series, FNO takes advantage of them to discover the complex time-frequency characteristics underlying multivariate datasets. On the benchmark Tennessee Eastman process (TEP), a real-world chemical manufacturing dataset, the performance of the proposed method was demonstrated and compared to that of the widely used fault detection methods.