How to solve problems with sparse modeling (Part 3)

How To Solve Problems With Sparse Modeling (Part 3)

Hello everyone, My name is Kenshin Fujiwara and I am the CEO and founder of HACARUS Inc. 

Through this series of blogs, I will discuss a wide range of topics of AI, from the history of AI to practical tips for the successful application of AI projects. I hope my blog posts will help you better understand AI and solve your business issues.

In the last blog, we studied the case of Mitsubishi Tanabe Pharma, in particular, its challenges in the implementation of AI and how they solved it with sparse modeling. Today, we will look at another case study of an electronic equipment manufacturer (referred to as Company A) that inspected printed circuit boards.

Case Study 2: Challenges in Reducing Inspection Times for High-Volume, High-Variety Electronic Equipment Manufacturing

Printed circuit boards are used in various products including automobiles and smartphones. In the beginning, Company A was using a conventional visual inspection system. This system operated using a pattern matching mechanism using images, which was common in the manufacturing industry. 

Inspection using pattern matching is performed using a multi-step process. First, “normal” products are registered into the system. Next, an image of the target component is compared to the normal product image. The system judges the differences and identifies sections as “abnormal” if they pass a certain threshold which is converted into a numerical value. 

In the manufacturing industry, it is unacceptable to ship out potentially defective components. For this reason, the threshold for abnormal readings is generally set strictly to ensure that even the slightest difference is not overlooked. 

However, the setting of strict thresholds also caused another issue for Company A. Since the manufacturing process often creates minor differences in components, they were often identified as being “abnormal.“ In many cases, these components were taken out of production even though they fell within the acceptable range.

While identifying defective components is still of the utmost importance, discarding normal products deemed as defective reduces the overall productivity of the company. Company A chose to solve this problem by incorporating a secondary check performed by human operators. These technicians would manually check the products that were determined to be defective by the inspection system. 

Challenges with AI Implementation

There are several drawbacks to working with a manual visual inspection. The first issue is that inspection skills are often limited to a single person. Another issue is that there is currently a shortage of highly skilled inspectors. 

At this point, Company A was considering using AI to solve these two issues in their factories. However, there was a major issue with the scope of the process. Company A manufactures a wide variety of products, and each product uses a different printed circuit board. 

Originally, we looked at applying deep learning methods to solve the problem. However, it was necessary to collect a large number of images for a wide variety of products to make it work. In addition, new data would also need to be collected every time a new product was introduced. 

This issue is common in the manufacturing industry and isn’t unique to the example of Company A. For low-volume, high-mix production sites, this data barrier is encountered when trying to use AI to handle a diverse product line. 

To address these issues, we focused on inspecting specific parts of the board during the actual inspection. By concentrating on the workmanship of certain components mounted on the board, different board types could be inspected interchangeably. Even though the boards themselves varied, the inspection targets could be grouped into a common pattern. 

Advantages of AI Implementation in Inspection System

Therefore, by preparing an AI for each of those groups, we were able to apply it to inspect multiple types of boards. With this method, the number of man hours spent on visual inspection could be cut in half. 

In addition, there were two more advantages that this method provides. First, there was no need to prepare product data from every product to create the AI model. This also enabled efficient AI implementation and reduced the amount of man-hours needed to collect data by over 50%. 

The second benefit involved the future-proofing of the AI model and the existing inspection systems. In this case, even if new parts are introduced in the future, there is no need for Company A to prepare large amounts of data to retrain the system. We have essentially built an AI inspection system that is not bound by product type. 

Another key point in this case study was that Company A did not aim for complete AI implementation out of the blue. When introducing a new AI, it is important to understand that it won’t instantly solve the problem in its entirety. 

Even with AI, misjudgments could still occur, and we proceeded with the project with the goal of simply reducing the visual inspection workload. With Company A, we began AI implementation while retaining the manual double-checking workflow. 

In the end, we were able to meet our expectations of reducing the man-hours required for visual inspection by half. This was a good example of an AI project with speedy implementation. 

In the next blog, I will introduce another case of AI implementation for AI diagnosis support at Kobe University and Kyoto University. For your updated knowledge and insight about AI technology, subscribe to our newsletter or visit HACARUS website


Subscribe to our newsletter

Click here to sign up