Deep Learning with Tensor Factorization Layers for Sequential Fault Diagnosis and Industrial Process Monitoring

Abstract

Fault diagnosis technology is crucial to ensure the long-term reliability of the industrial process control system. With the increase of industrial data availability, conventional monitoring approaches may not function well under the assumption that the training and the application data come from the same distribution. Following the intuition that industrial data exhibit time dependency and inherent complex characteristics, this paper proposes an adaptive sequential fault diagnosis method based on a tensor factorization layer merged with deep neural network model (TF-DNN). Tensor representation is firstly applied to preserve the number of the raw data entries and their sequential dependence between observations. Multilinear mapping with tensor-to-tensor projection is then to transform the input and hidden tensor to the low-dimensional tensors, which makes the learned representations sparser with the grid-like structures. With the tensor factorization layers, the proposed deep network shares efficient knowledge across the spatiotemporal features of fault data. The outstanding performance of this method is demonstrated and compared to the existing models in the benchmark Tennessee Eastman process and a real industrial methanol plant.