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

In recent years, deep‐learning‐based fault detection and diagnosis (FDD) methods have received extensive attention. As we all know, different input forms have a great impact on the final performance. In this paper, three categories and seven representation methods are discussed: numeric representations, image mapping representations (radar chart mapping and Gramian angular summation field (GASF) mapping), and signal transforming representations (fast Fourier transform (FFT) and wavelet). The tests on the Tennessee Eastman process (TEP) dataset prove that the FFT method has achieved the best performance on average. Based on this, a general FDD integration framework is proposed to integrate multiple base learners together to make decisions by weighted voting or maximum voting. Finally, the comparison between our proposed method and other five typical models (FFT, a GASF and a multi‐scale neural network (GASF–MSNN), convolutional neural network (CNN), Long Short‐Term Memory (LSTM), and Support Vector Machine (SVM)) illustrates the effectiveness of our method for FDD on the TEP. The proposed integrated method provides an effective platform for deep‐learning‐based FDD. [ABSTRACT FROM AUTHOR]