A Deep Convolutional Neural Network Model for Nonlinear Drift Fault Diagnosis in Sensors using Transfer Learning and Multiresolution Analysis

Reduction in manufacturing cost due to the implementation of advanced technology resulted in the large-scale deployment of wireless sensors. As the sensors are subject to harsh operating conditions they develop faults. The faulty sensor readings and the faulty sensors need to be isolated in real-time and remedial measures to be taken for maintaining system stability. Drift faults are slow-changing constant faults that get induced into the sensor measurements and are difficult to isolate. Drift faults can be linear or may be nonlinear. Though studies have been conducted for linear drift fault isolation, studies related to nonlinear drifts are limited. Studies have achieved a very high level of accuracy while deploying Deep Convolutional Neural Networks (DCNNs). Continuous Wavelet Transform (CWT) has also shown promising results as a tool for noise reduction and feature extraction. Keeping these things in view, in the present study the ability of pre-trained DCNN through transfer learning is exploited, and scalograms as the feature extractor from the sensor signals. Three nonlinear drift fault modes are simulated to test the sensitivity of the DCNN models under different fault scenarios. The performance of the proposed model for the classification of nonlinear drift faults is compared with other models in terms of accuracy and loss. The proposed Inception V3-based DCNN model achieved the highest accuracy of 99.05% in classifying nonlinear drift faults. An algorithm for simulating nonlinear drift faults is also proposed.