Optimizing maintenance processes with AI and modern techniques

One of today’s most prominent issues is the environmental impact of industrial activities, with the manufacturing sector being a big contributor to greenhouse gas emissions, pollution, and waste. As industries face increasing pressure to adopt more sustainable practices, predictive maintenance has been presented as a part of the solution. This study, conducted at the biopharmaceutical company AstraZeneca, develops and evaluates a one-dimensional convolutional neural network (1D-CNN) and a combined one-dimensional convolutional neural network with bidirectional long short-term memory (1D-CNN-BiLSTM), for fault classification in complex industrial packaging machines. As part of predictive maintenance, fault classification can significantly reduce machine downtime by enabling early failure detection, improving productivity, and minimizing costs. Additionally, accurate fault diagnosis helps avoid unnecessary component replacements, reducing waste and supporting sustainable manufacturing. Fault classification in packaging machines is an overlooked area of research. While existing studies mainly focus on fault classification in larger industrial systems using vibration sensors on isolated motors and in simplified test rigs, little research has been conducted on fault classification within small packaging machinery. Packaging machines present unique challenges due to their intricate mechanical structures, multiple moving parts, and changing operational conditions, such as varying loads and speeds, all of which contribute to complex vibration patterns. This study evaluates model performance across different machine speeds and assesses its ability to generalize fault classification under varying operational conditions. It also examines whether integrating data from multiple vibration sensors and incorporating long-term temporal context improves classification accuracy compared to models processing single-sensor data without long-term dependencies.