Visual Inspection with Artificial Intelligence

AI is redefining automated visual inspection, making defect detection more precise and reliable than ever. By leveraging neural networks, manufacturers can significantly reduce false rejects and improve efficiency. But how can AI be seamlessly integrated into inspection machines? And what challenges arise when implementing AI in regulated environments?

Artificial Intelligence (AI) is an umbrella term encompassing various technologies that enable machines to perform tasks typically requiring human intelligence. Within AI, Machine Learning (ML) is a specialized field where algorithms analyze data to identify patterns and make decisions with minimal human intervention. Unlike traditional rule-based systems, ML models are trained on large datasets to optimize performance for specific tasks. However, in most industrial applications, ML models do not continuously adapt in real-time but instead follow a structured process: data is collected, models are trained offline, and only updated when necessary after careful evaluation. ML is particularly useful for tasks requiring complex decision-making and pattern recognition.

A deeper subset of ML is Neural Networks, which are inspired by the structure and function of the human brain. Neural networks consist of layers of interconnected nodes (neurons) that process information in a way similar to biological neurons. These networks excel at recognizing intricate patterns within large datasets, making them ideal for tasks such as image analysis, defect detection, and quality control in manufacturing environments.

In practical applications, neural networks enhance automated visual inspection systems by improving accuracy and adapting more efficiently to variations in defect types. Unlike traditional machine learning approaches, neural networks excel at recognizing patterns across a broader spectrum, reducing the need for manually defining distinguishing features. While neural networks do not inherently offer greater consistency than rule-based systems for identical inputs, they can be more resilient to slight environmental variations, such as changes in lighting or setup. By leveraging large, labelled datasets and real-world production environments, neural networks enhance inspection efficiency, while minimizing operator dependency and false rejects.

We at WILCO leverage neural networks to enhance defect detection accuracy and efficiency. By summarizing intricate structures into easily distinguishable features, we improve detection capabilities without the extensive manual effort required by classical methods. This adaptability enables our machines to identify even subtle variations in product quality that might be overlooked by conventional inspection systems. 

To fully harness the potential of AI in defect detection, different learning approaches come into play, each with a specific function. 

Supervised Learning: Teaching AI to Recognize Defects

In this approach, humans manually label thousands of defect images, helping the AI model learn what to detect. Over time, the model becomes proficient in identifying similar defects with high precision. This method is effective for well-defined defects but requires a significant amount of labelled data.

Anomaly Detection: Identifying Deviations from the Norm

Unlike supervised learning, anomaly detection focuses on distinguishing good samples from abnormal ones. The AI is trained using defect-free products and learns to flag anything that deviates from this standard. While this method requires less labelled data, it poses challenges in regulated environments, where a detailed classification is sometimes asked for.

Applications for AI in an Automated Visual Inspection machine

AI can be successfully deployed to: 

• Achieve superior defect detection performance compared to rule-based vision tools, effectively addressing challenges associated with specific defect types.
• Maintain a high detection rate in being less sensitive to the natural variabilities of the process (False Reject Rate due to primary packaging and manufacturing variabilities).
• Improve defect classification capabilities compared to rule-based algorithms vision tools capabilities, if supervised training is used. 

Challenges of AI Implementation in Inspection Machines: Revalidation

Once AI is integrated into an inspection machine, the system must undergo revalidation to ensure compliance with regulatory standards. However, if AI is used only to replace specific functions rather than the entire inspection process, only a partial requalification is necessary. This reduces the validation burden and improves implementation efficiency.

Upgrading certain machines with AI is possible from a hardware and software perspective, but a revalidation process is necessary. 

How WILCO Supports AI Integration for Automated Visual Inspection

AI integration in automated visual inspection follows a structured four-step process. First, Data Acquisition involves capturing high-quality images of products during inspection. Next, Data Labeling ensures that these images are properly categorized, allowing the AI model to understand different types of defects. Then, Model Training and Evaluation take place, where AI learns from labeled data and is tested for accuracy and reliability. Finally, the trained model is Deployed on the Machine, enabling real-time defect detection.

As a machine manufacturer, our expertise lies in building high-performance inspection machines. Therefore, we focus on two critical steps in this process:

1. Data Acquisition – Capturing high-quality images of products during inspection.
2. Model Deployment – Integrating the trained AI model into our machines for real-time defect detection.

The intermediary steps—data labeling, model training and evaluation—are typically handled by specialized AI developers. We strive to offer a flexible deployment framework such that our customers can perform the intermediary steps - data labeling, model training and evaluation - with the tools of their choice. Our framework allows you to experiment with new models and training regimes to find the best fit for your products while our in-house AI experts are– always just a mail away to support in the development process. Once a reliable AI model is trained, we ensure its seamless integration into our machines to enhance inspection capabilities.

The EVO CAX is designed to seamlessly integrate AI models, ensuring pharmaceutical manufacturers can harness the benefits of AI-driven visual inspection with minimal effort. Built with human-inspired robotic movements, the EVO mimics the precision and adaptability of manual inspection while offering the consistency and efficiency of automation.

Several advantages can be gained by integrating AI into EVO CAX. Examples are:

• Consistent Inspection Quality – Unlike manual inspection, AI ensures uniform defect assessment without human variability.
• Seamless AI Deployment – EVO CAX is engineered to support AI models, making integration straightforward without extensive hardware modifications.

By combining the latest in automated inspection with AI readiness, EVO CAX future-proofs quality control processes, providing a scalable and intelligent solution for pharmaceutical manufacturers.