DOI: 10.14704/nq.2022.20.8.NQ44564

Deep Convolutional Neural Network and Wavelet Transform based Model for Intermittent Fault Diagnosis in Sensors

Aruna Kumar Mishra, Dr.Subrat Kumar Mohanty

Abstract

The use of wireless sensors for patient monitoring and extensive use of telemedicine has thrown a challenge for medical professionals as they have to rely on sensor signals for disease classification. Timely detection and classification of faulty sensor signals from healthy ones are vital for the accurate diagnosis of the disease. Intermittent fault detection and classification had always been a challenge because of their unpredictable nature. Recent advancements in deep convolutional neural networks have opened a new dimension to the sensor fault diagnosis approach. Researchers are also using multiresolution analysis for denoising and feature extraction from the sensor signals. Datasets for testing fault diagnosis algorithms are scarce because of the difficulty in collecting pre and post-fault occurrence data. As suitably labeled benchmark datasets for testing intermittent fault diagnosis algorithms are not available, we generated two intermittent fault modes using the temperature sensor signals of the Intel Berkeley Lab dataset. The time-frequency representation of faulty and non-faulty sensor signals was generated using the continuous wavelet transform. The two-dimensional scalograms were used as input to the headless pretrained deep convolutional neural network models for generating feature vectors which were subsequently used as input to the Dense layer for the classification of intermittent faults. Using the proposed model validation accuracy of 100% was achieved in both the intermittent fault modes. The performance of the proposed model was also compared in terms of validation accuracy and loss with other DCNN-based models.

Keywords

DCNN; Intermittent fault; Scalogram; Transfer learning; Wavelet transform.

Full Text

References