Artificial Intelligence (AI) has now become a global buzzword that is being regularly used in the media. In Japan, more and more products with built-in AI, including voice-controlled apps in smartphones, robot cleaners, self-driving cars and AI speakers are available, and our daily life is increasingly becoming connected to these devices. This trend is set to continue with AI expected to make our lives even more convenient with the introduction of many new services such as a control system for home appliances, a taxi dispatch service, an unmanned delivery service and unmanned convenience stores among others.
AI is also being used in areas that are not directly related to our daily lives. More and more companies are using AI to expand their businesses and/or services. One of the advantages of using AI is to analyze significant volumes of data, which helps companies to understand their customers’ needs in greater detail. One example of this is the use of targeted advertisements that are based on a customer’s Internet browsing history. The same information can be used to identify potential customers and develop a tailored marketing strategy. Another advantage of using AI is efficiency. With AI, many tasks that used to depend on human’s eyes and/or hands can now be automated and can be done even more quickly and accurately. For example, AI with the capacity to detect specific types of images can easily identify all of the cracks in a pavement and road.
“Robots” may be one of the first things that come to mind when you think about improving the efficiency of the manufacturing sector. Indeed, the introduction of an industrial robot and the improvement of its capacity, and more recently the introduction of cobots are helping to make both factories and production processes more efficient. From around 2010, companies in Europe and the U.S. also started making efforts to realize a “smart factory” where all devices and equipment are connected. As the movement of humans and goods is becoming more and more streamlined, it is bringing significant transformations to factories.
What’s to come next is an integration with AI, as many companies have already started to use AI technology to improve their manufacturing. In Japan, Renesas Electronics Corporation (hereafter referred to as Renesas) announced that it will introduce AI-driven innovation in three key factories by 2020. Renesas has also announced its plans to develop AI with the capacity of statistical inference like experienced engineers do. FANUC and Hitachi announced that they are in a joint venture with Preferred Networks, a start-up specializing in AI research and development. These three companies will work together with the goal of creating a factory in which machines in the production line learn themselves.
In many production sites in Japan, product quality inspections and a machine/equipment’s error detection or anomaly movements are often carried out by humans alone. The introduction of AI will not only reduce the necessary manpower, but will also increase the productivity. For example, in Renesas’ factories, engineers “find” 50 errors per month, which in reality have not been errors. After AI was introduced to perform the same role, the occurrence of false error detection has become less than one-tenth. In only a half year, Renesas has saved JPY 500 million thanks to AI.
In the application of AI for factory inspections, MVTec, a German software company, provides an image analysis system to detect defective products. The system is specifically designed for use in factories and uses a deep learning network that is tailored to the analysis of the industrial product image. The algorithm learns what is good and what is defective. Thanks to this mechanism, the system can identify the condition of the product and the types of defects without complicated featured value analysis or an extraction algorithm.
In factories, shipping defective products should be avoided at all costs. Therefore, in the conventional inspection process with an image analysis system, everything that has the potential to be defective is considered as defective. Whenever the system finds 50 defective products and 50 potentially defective products out of 10,000 products, then the system has treated these 100 products as defective. Humans have then checked whether these 100 “defective” products are indeed defective or not. This process is both time consuming and undermines productivity. MVTec’s system is the answer to this dilemma. In the example above, the system has been able to reduce the number of defective products from 100 to 50 thanks to a better image analytic capability. This means that humans now only need to check half the number of products, resulting in cost and labor savings as well as productivity improvement.
Having said this, however, some differences can be very subtle and human experiences and expertise can matter a lot. For example, “This smear is OK, but that smear is not OK”, or “This small problem is OK, but that problem is not OK” – this kind of decision depends on expert judgment. Whenever the image analysis system is used in such a process, clear criteria such as “the area is greater than X” or “the color is Y” is provided, but subtle differences have always been a major challenge for such a system. Thanks to deep learning, today’s image analysis system can make more human-like decisions. Take an example of colors and think of a product with a red and black gradation. The pattern of gradation may be slightly different, so previously, the image analysis system was incapable of detecting subtle differences in gradation and thus human judgement was needed. MVTec’s system can do that because the capacity to handle such subtle features and attributes can be incorporated to the system. Even though the features and attributes may differ in each factory and may also change over time, the system can be re-trained by providing additional data that enables the system to learn what is/isn’t defective.
MVTec has rich experience in industrial image analysis, and thus images that help the system to distinguish normal products and defective products have been built-in to the system. The users can add more images that help the system to understand their specific products’ features and attributes to be used in the analysis and to create a tailored network. The system can also be launched quickly because only 100 to 200 images are needed to fine-tune the system to the users’ environment.
An electric parts manufacturer, which uses MVTec’s system reports that the system has a 99.9% accuracy rate to detect defective products out of potentially defective products. Another advantage is that it is easy to introduce. If you have enough images to allow the system to learn the key attributes to identify any defects, you can start using it. Therefore, the system is not only getting attention from electric parts manufacturers, but also from automobile, metal, steel and tablet makers too.
An increasing number of companies are now using AI as part of their IIoT businesses and/or services, and the trend is expected to accelerate in the future.