Högskolan i Skövde

This thesis explores the application of artificial intelligence, specifically using autoencoder (AE) and variational autoencoder (VAE), for anomaly detection, when dealing with data imbalance. The primary aim is to develop and evaluate models that can predict human errors during the assembly process of engine valves and springs, while addressing research questions on performance enhancement when pretrained models are incorporated into the architecture of convolutional autoencoders, VAEs for anomaly detection, focusing on the utility its latent space for identifying anomalies and identifying the limitations of existing convolutional autoencoder approaches and explore emerging techniques to address these challenges.The literature review identifies several limitations of existing convolutional autoencoder approaches. These limitations include the tendency to focus on local patterns and struggle with capturing long-range dependencies, the reconstruction error metrics, often based on pixel-wise differences, may not align well with human perception, necessitating the exploration of perceptual loss metrics. Furthermore, traditional AEs may overfit the training data, reducing their generalisability and effectiveness in real-world applications. The literature review further explores non-conventional approaches for anomaly detection, including the use of Generative Adversarial Network (GAN) and other advanced techniques, which provide insights into potential improvements and guide future research efforts. These emerging techniques aim to enhance robustness of anomaly detection models, addressing the limitations identified in traditional convolutional autoencoder approaches. The research leverages pre-trained models for the encoder part of the network, specifically MobileNetV2 and InceptionResNetV2, while keeping the decoder architecture consistent and investigates the effectiveness of these models in identifying anomalies and compares their performance with traditional approaches. The results indicate that integrating pretrained models like InceptionResnetV2 and MobileNetV2 into the encoder the architecture improves anomaly detection performance. MobileNetV2 demonstrated the fastest training times, making it a potential practical choice for scenarios requiring quick model updates while VAE with InceptionResnetv2 shows better accuracy. VAEs showed superior performance, with the use of Mahalanobis Distance and Gaussian Mixture Models (GMM) as thresholding techniques, yielding high accuracy in identifying anomalies that weren’t distinguished by traditional autoencoders with MSE threshold.