Monitoring the production of plastic consumables for laboratories: an artificial intelligence approach to balance efficiency and sustainability

This thesis addresses the use of Artificial Intelligence algorithms to improve the accuracy and efficiency of a production process. The application is based on a real-world scenario, driven by the need of a company in the consumable plastics manufacturing sector. The company aims to automate one of their quality control system, which is currently performed by human operators. To automate the control process, an electronic device equipped with AI algorithms has been developed. This device will be integrated into various stages of the production line in order to make the process more efficient and sustainable. The focus of the investigation is on plastic laboratory vials containing a transparent anticoag- ulant substance. The goal of the automated control process is to verify the actual presence of the substance inside the vials. These vials can vary in shape, and the anticoagulant may be present either as a droplet or in a nebulized form. In light of this, two different approaches to control the manufacturing process will be investigated: detecting the presence or the absence of the antico- agulant regardless of the size of the test tube (referred as 2-output-labels case); and determining whether the anticoagulant is present or absent while also identifying the size of the test tube (either large or small, termed as 4-output-labels case). To address these challenges, we opted for Com- puter Vision (CV) and Artificial Intelligence (AI) techniques. Several studies have demonstrated the positive impact of CV and AI-based monitoring systems on Industrial sector. Computer Vision systems offer a faster, more objective solution and can continuously inspect thousands of vials per minute. When fine-tuned, these systems provide more accurate results than human operators. By implementing CV systems, manufacturers can reduce errors, enhance production efficiency, and en- sure regulatory compliance, ultimately lowering costs and increasing profitability. Moreover, these techniques enable real-time monitoring and proactive detection of anomalies or defects, allowing for timely interventions and minimizing costly errors or delays. As far as we know, the problem addressed in this thesis has not been explored in the literature, and two different strategies will be presented to investigate it. The first involves creating a Convo- lutional Neural Network (CNN) from scratch. The second is the Transfer Learning technique, which uses a pre-trained network and re-trains only the last layer with the data from the application of interest. The pre-trained models considered include both CNNs and the current state-of-the-art, Transformers (specifically Vision Transformers). The results obtained are extremely promising. However, since we needed to optimize the monitoring process, we also conducted an analysis of the resources consumed by the models. A comparison of the various AI frameworks will be presented, followed by the introduction of the optimal solution that balances accuracy and sustainability in the monitoring process. Finally, as attention to AI model explainability continues to grow, and given its importance when discussing the certification of the quality of the monitoring process, an explainability analysis will be presented to illustrate the reasoning behind the models’ decisions (both in the simple two-output case and in the more complex scenario). Keywords: Artificial Intelligence, Convolutional Neural Networks, Vision Transformers, Industry 4.0, Monitoring Systems, Sustainability, Innovation, Green AI, Explainable AI.