A novel one‐dimensional convolutional neural network architecture for chemical process fault diagnosis

Abstract

In recent years, industrial production has become increasingly automated, with the widespread application of informational and digital technology. Fault detection and diagnosis (FDD) technology is also playing an increasingly important role in the chemical process industry. However, owing to the weak generalization ability of prior models, or prior methods not being suitable for industrial sensor signal data, the fault detection rate is not satisfactory, which is a significant limitation of many fault diagnosis methods in practical applications. In response to this problem, the one-dimensional convolutional neural network (1D-CNN) model can directly process signal samples without changing the one-dimensional characteristics of the data, which may be more suitable for processing such signal data. Therefore, a new 1D-CNN architecture is proposed for FDD. The network architectures, including convolutional layers, pooling layers, fully connected layers, and various parameters, are optimized in the proposed method. The Tennessee Eastman process (TE process) is employed to assess prominent performance of the method. To comprehensively reflect the performance of the model, three evaluation indexes are selected in this study: accuracy, F1-score, and fault detection rate. The experimental results indicate that compared with other diagnostic methods, the 1D-CNN model has excellent feature extraction ability, which can remarkably improve diagnostic capability in the TE process.

ACNNapproach based on correlation metrics to chemical process fault classifications with limited labelled data

“…Fault diagnosis approaches for chemical processes are widely circulated in the literature. [1][2][3] The main investigation directions can be classified into knowledge-based, [4,5] data-based, [6,7] and hybrid-based [8] fault diagnosis methods. In addition, it is seen that intelligent and data-driven signal processing has achieved fruitful results in the fields of artificial intelligence and machine learning in recent years.…”

A comparative study of different data representations under CNN and a novel integrated FDD architecture

“…Its network is also in numerical form with size . [ 12 ] Wu and Zhao proposed a 2D‐CNN architecture for FDD in which the input has the size of ; here, T represents the length of the time window or time scale. [ 13 ] Gao et al also addressed a 2D CNN‐based FDD model that can extract dynamic feature characteristics of process data at different time scales whose input size is also .…”