DETECTING DEFECTS AND ANOMALIES THROUGH ARTIFICIAL INTELLIGENCE

The need for visual inspection of products and inventory is crucial for any modern manufacturing and logistics company, with big players investing more and more on the development of efficient and better suited equipment to minimize the chances of imperfections and contamination during production or storage. Anomalies can create all sorts of issues, not only impacting the quality of the final product in an assembly line, but also slowing down production, creating possible environmental or safety hazards and increasing costs.

Specifically, dealing with quality control inside an assembly line is an expensive and time-consuming task, that requires several hours of contiguous checking by human inspectors, which may be rendered redundant by just a single missed anomaly. Further, the growing complexity of production pipelines sometimes makes it impossible for a manual inspection to be sufficiently accurate, and herein lies the problem, because accuracy not only leads to a more cost-effective production line, but also prevents loss of expensive equipment and material, rework and, in the worst cases, even injuries or fatalities.

This explains the market’s increasing request for better ways to perform tasks that, in the past, were exclusively executed by humans, particularly the ones that require our unique ability to quickly identify details within patterns, recognizing in a split second what is right and what is wrong in a given scenario. Technology has come a long way to try and solve these issues, with deep learning-based solutions being the preferred choice by the big manufacturers, because of their strong adaptability and reliability.

ARTIFICIAL INTELLIGENCE, MACHINE LEARNING AND DEEP LEARNING

To discuss artificial intelligence is a little like discussing algebra or geometry: it is an extremely diversified field, that through the years has developed several disciplines. As defined by the Oxford English Dictionary: “AI is the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages”. We can think of AI as a Russian doll, with deep learning sitting inside machine learning and machine learning sitting inside AI.

The first big step for AI in terms of performing human-like tasks, was, in fact, the advancement of machine learning. This discipline, at its core, utilizes algorithms and statistical models to analyse data, making predictions of its own, without the need of an instruction or training set. This gave birth to the first instances of computer vision, which at this stage was still in need of hand-coded classifiers to recognize certain patterns or objects. It was clear that another step was needed in order for machines to behave more like a human brain and learn from the provided data more independently. This is when deep learning came into play.

Deep learning is a different way to implement machine learning that takes inspiration from the human brain, where complex neural networks made by a large number of layers are used to identify patterns. The key here is that deep learning works by itself and learns from experience, tweaking the algorithm each time to improve the outcome. Imagine teaching a child the difference between the word cat and the word cap; at the beginning, the child will point at what is on his dad’s head saying “cat,” and the parent would correct him/her right away saying, “No, that’s a cap.” This instant feedback creates a model inside the brain, one that can be used in the same scenario the next time to know what is wrong and what is right; and this is exactly how neural networks work in deep learning, with different layers providing more and more information derived from previous mistakes, that are later used to provide the most precise output. And with more training, the better the result achieved. Such a powerful tool can turn out to be priceless while searching for a quick and automatic way to detect faulty or contaminated products, especially in an environment where many different things can go wrong at any time.

If you want to know the thoughts of an expert regarding this topic, you can check our interview to MIT’s Department of Brain and Cognitive Sciences and the director of the NSF Center for Brains, Minds and Machines at MIT, Tomaso Poggio.

SAFETY AND EFFICIENCY AS PILLARS FOR INNOVATION

A deep learning-based monitoring system can significantly improve the production process, both in terms of safety and efficiency. These two dimensions are arguably the pillars upon which any manufacturing process is structured and are significantly interconnected, with studies showing that productivity is increased when the conditions of the work environment are safe and controlled, as a consequence of the increasing levels of confidence and engagement that the workers experience. In light of this, we strongly believe that machine vision and deep learning can drastically change the way people live their working lives, and that they are a total game changer for production pipelines and warehouses.

In fact, AI and efficiency are two realities that go hand-in-hand: not only can an AI-equipped line help to save human resources, and thus lowering production costs, but it can also reduces the chances of human error leading to a more consistent and higher quality product. In addition, relevant data can be easily gathered from these tools providing real-time results that can be monitored and benchmarked against other sites in order to access quality standards across multiple facilities. These insights enable a more thorough analysis of the various issues and defects that can occur during the production process.

KONICA MINOLTA’S FIRST IN CLASS AI TECHNOLOGY FOR QUALITY INSPECTION

With the launch of our latest solution, Box Defect Detection, we have combined our long-standing efforts in the deep learning field with class-leading Mobotix IoT cameras to achieve the best results not only in terms of smart anomaly detection, but also from a build quality, ease of use and picture quality perspective.

What we have achieved is a solution that:

• Enables to distinctively identify and enumerate anomalies such as defects, foreign objects, and contaminations
• Rates the severity of the anomaly on a scale from 0% to 100%
• Can check for any barcodes or labels, recognizing different products
• Controls the separator to automatically send the defective box out of the production line
• Can be easily monitored thanks to a real-time dashboard that provides long-term statistics
• Is completely customizable for any task required

The goal is to offer a product that can reliably and safely detect any issue within a product line, and this is guaranteed by the high accuracy of our field-tested algorithm. This, along with the raw speed at which the AI-model can learn a new environment and the attention to detail we put into the UI, which makes it operable even with no coding or machine learning knowledge, gives our solution all the most up-to-date features to optimize a product’s life cycle.

To find out how these solutions could apply to your industry, get in touch with our researchers today.