Image processing involved specifically with image recognition and image classification has taken a huge stride with the advent of high-performance GPU's and the increase in the processing speeds of images by the computer systems. Also, the availability of huge datasets on different classes of images has helped in solving issues which can be addressed by huge number of annotated datasets in teaching a supervised Convolutional Neural Network (CNN) model to classify the images. The present method of Structural Health Monitoring (SHM) of Wind Turbine Blade (WTB) inspection manually is involved with a great amount of risk and also takes a lot of time for inspection of each turbine and the operation time of each turbine comes down as the turbine should be at halt while inspecting. These problems are overcome by the inspection through drones and the basic idea of this research is to use deep learning with keras frame work written in python working on top of tensor flow to use the drone captured images to train a neural network model and classify into faulty and not faulty images of blades, which when deployed can be used in classifying new images. This method in turn reduces the maintenance time and inspection time and less risk for the inspection of WTB and in turn the SHM. In this research CNN's are used to find whether the given image of WTB is having damage or not. Since the image data of wind turbine blades are not available easily with annotated images as like ImageNet and AlexNet, this project was developed from scratch right from acquiring the image data. Along the way, the focus was on the influence of certain hyper-parameters and on seeking theoretically founded ways to adapt them, all with the objective of progressing to satisfactory results as fast as possible. In the end, also a promising attempt in classifying WTB damages into a few different classes is presented and deploying the saved model using Flask such that the model can be used anywhere when required. Accuracies of around 94.94% for binary fault classification and 91% accuracy for multiple class fault classification has been achieved. The significance of the proposed method is that it is the first of its kind for WTB damage classification and detection for different classes of damage without using transfer learning but by training the model from the image dataset prepared by image augmentation methods and manual editing. © 2019 Elsevier Ltd