Analysis of the scalability of siamese neural network for performing quality inspection of the welding nuts

Abstract

Fabrication is a common process in the manufacturing industry. It is important to inspect the nuts before welding in order to maintain product quality and avoid the need for rework. Nowadays, artificial intelligence based visual validation have demonstrated that machines can efficiently perform quality inspection. This thesis provides a siamese neural network based solution to identify incorrectly placed nuts before welding process. The network classifies incorrect nuts, missing nuts and flipped nuts as invalid cases. The network is designed in such a way that the model trained to perform quality analysis of one task can be scaled to a new task by using six or ten training images and one minute of training duration. The decision to use siamese network is made because of its ability to learn from semantic similarity. To validate the adaptability of the solution, we have focused on three use cases in two different environments. First case consists of test nuts which are different in size, shape and position. The second use case validates the model on different orientation of test nuts. In third case, the test nuts are captured in different environment conditions. We have proposed two different approaches, one with custom convolutional neural network and the other with EffecientNet-B0 as feature extractor. The model with custom convolutional feature extractor has better performance and scales to all the three use cases. The model with EfficientNet-B0 as feature extractor has exceptional performance in third use case which consists of data captured in factory environment. We evaluate the proposed solutions by recording the accuracy and confusion matrix of different use cases and architectures. However this approach is limited in terms of production efficiency and needs more validation in factory lighting conditions.