28 n AUTOMATION ETHICS June 2025 www.drivesncontrols.com Navigating the ethical and regulatory maze As AI and automation become increasingly prevalent in manufacturing, the potential for algorithmic bias to influence quality control processes is a growing concern. For example, AI-powered visual inspection systems, designed to identify and flag defects in products, may inadvertently perpetuate existing biases embedded in their training data. The collection of training data for these systems often relies on existing manufacturing processes, which may be skewed towards certain production lines. If the dataset consists primarily of products manufactured on a specific production line, for instance, the AI algorithm may learn to associate certain defect patterns with that particular line – such as overlooking defects, misinterpreting product features or unfair targeting of certain production lines – leading to inaccurate assessments of products manufactured on other production lines. Diversifying training datasets is crucial for ensuring that AI systems developed for manufacturing applications are fair, equitable and effective. By incorporating data from a wide range of production lines, or using data augmentation techniques, manufacturers can mitigate the risk of algorithmic bias. There are two main approaches to diversifying training datasets. The first is active data collection. Manufacturers can actively seek out production lines that are currently under-represented in their training datasets. This may involve reaching out to suppliers, distributors, or partner factories to collect data from their operations. Manufacturers can also collect data from external sources, such as academic databases, public repositories, or open-source datasets. This can help to ensure that the training data is representative of a wider range of manufacturing environments and practices. Manufacturers can use data augmentation techniques to expand their existing training datasets artificially. This can involve techniques such as image translation, rotation, and mirroring to create new variations of existing data points. Generative adversarial networks (GANs) are a type of neural network that can generate new data that is indistinguishable from real data. Manufacturers can use GANs to create synthetic data for their training datasets, which can help to improve the generalisation ability of their AI systems. Transparent AI systems Furthermore, the intricate workings of AI algorithms can be opaque, making it challenging for manufacturers to understand how decisions are made and to identify The integration of AI with automation technologies could revolutionise how we work, manufacture and interact with our environment. Yet, as the boundaries of innovation expand, so too do the ethical and regulatory implications. Stephen Hayes, managing director of Beckhoff UK, explores the ethical dilemmas and regulatory challenges raised by AI and automation.
RkJQdWJsaXNoZXIy MjQ0NzM=