Convolutional Neural Network based Automatic Detection of Visible Faults in a Photovoltaic Module

ABSTRACT

Background/Objective: The primary objective of the present study is to distinguish several visual faults which hinder the performance, reliability and lifetime of photovoltaic (PV) modules. Research question: Conventional fault detection techniques require specific operating conditions which also consumed a lot of time, manpower and expenditure. Innovative techniques and technological advancements in the highly paced world expect instant results. Advanced and automatic fault diagnosis is such a process that delivers instant results and guarantees an extended lifetime for numerous critical photovoltaic module (PVM) components. Hypothesis: This study performs an automatic detection of faults in PVM with convolutional neural networks (CNN) that accurately classifies various faults based on the images captured from unmanned aerial vehicles (UAVs). Methodology: Dataset creation is one of the primary constraints when it comes to working with CNN. To overcome this drawback, a data augmentation method is adopted to enlarge the dataset from the limited number of available aerial images of PVM. These augmented images are fed into an automatic fault detection CNN model for deep feature extraction and classification. Results and Conclusion: The presented method exhibits an increase in the accuracy and performance of PVM health monitoring when compared with other conventional solutions. The performances of uniform and non-uniform datasets are also presented. Various pre-trained models like VGG16 and ResNet50 are compared with the proposed solution for performance evaluation. The results demonstrate that the overall classification accuracy of the proposed model for uniform and non-uniform datasets was found to be 95.07% and 94.14% respectively with lesser training time and number of epochs.

KEYWORDS:

• Fault diagnosis
• visible faults
• photovoltaic modules
• convolutional neural network
• deep learning
• unmanned aerial vehicle
• condition monitoring
• multiple faults

Abbreviations

PV	=	photovoltaic
PVM	=	photovoltaic module
UAV	=	unmanned aerial vehicle
UAVs	=	unmanned aerial vehicles
CNN	=	convolutional neural network
CASAE	=	cascading auto encoder
DCNN	=	deep convolutional neural network
EL	=	electroluminescence
RF	=	random forest
SVM	=	support vector machines
GAN	=	generative adversarial network
IEA	=	international energy agency
ReLU	=	rectified linear units
RMSProp	=	root mean square propagation
IoT	=	internet of things
$y_{ij}$	=	image label
${\hat{y}}_{ij}$	=	predicted value of the model
$N$	=	PVM condition
$M$	=	number of aerial images
$\theta$	=	weight parameter
$\eta$	=	learning rate
$t$	=	time step
$g_t$	=	moving average of squared gradient
$\gamma$	=	decay term
$\epsilon$	=	a constant

Additional information

Notes on contributors

Naveen Venkatesh Sridharan

Naveen Venkatesh Sridharan is a PhD student in the School of Mechanical Engineering at Vellore Institute of Technology-Chennai Campus. His research interests are fault diagnosis, solar energy and renewable energy systems.

V Sugumaran

V Sugumaran is a Professor in the School of Mechanical Engineering at Vellore Institute of Technology-Chennai Campus. His research fields are fault diagnosis, applications of machine learning, predictive analytics, machine condition monitoring, renewable energy technologies and so on. He has published more than 200 papers including conference proceedings and articles published in SCI-indexed journals.