Hitachi: Using visual inspection AI for quality control and achieving 100% accuracy

About Hitachi

Founded in February 1, 1920, Hitachi is a electronics manufacturer that develops advanced products and services in various fields, centering on business segments, such as IT, energy, industry, mobility, and life. The Omika Works, which started operations in 1969, provides information control systems for social infrastructure and industrial fields, such as power generation, transmission, distribution systems, railway operation management systems, operation and maintenance management systems for water supply and sewerage facilities, and production systems for factories and steelworks. It seeks to address various social issues and create new innovations through IoT and data analysis.

Industries: Manufacturing
Location: Japan

Tell us your challenge. We're here to help.

Contact us

Omika Works, a line of factories by Hitachi, has developed a visual inspection system that connects the Hitachi industrial edge computer CE50-10 and Visual Inspection AI by Google Cloud. The system was first introduced to ensure quality control over the crimping of the electric wires' connectors used to assemble control panels, and the proof-of-concept created with Google Cloud achieved a 100% defect detection rate.

Google Cloud results

  • Create a machine learning model in a day from approximately 100 sample images, with Visual Inspection AI
  • Achieve perfect accuracy in visual inspection with AI with a defect detection rate of 100%
  • Verify the configuration of a visual inspection system successfully such that it can be implemented for on-site operations

Achieves 100% defect identification rate for its visual inspection system

Omika Works by Hitachi, Ltd. is a line of factories that manufactures, operates, and maintains information control systems for the social infrastructure and industrial sector. This includes power generation, transmission, distribution systems, railway operation management systems, water supply and sewerage monitoring systems, and production systems for steelworks. As Omika Works is central to system development at Hitachi, it facilitates Hitachi in creating new solutions and service offerings through Internet of Things (IoT) and data analysis, while implementing innovative initiatives to tackle social issues.

One such initiative is the Omika Green Network, which sees Hitachi pursuing carbon neutrality through a network that connects regions and supply chains, with Omika Works being a central hub for these activities. To this end, Omika Works now serves as a platform for testing, with dry runs on reducing environmental impact and improving production being conducted, and the results shared with suppliers and the region.

"The control panels we manufacture are designed to be used for at least a decade, so we exercise great care in ensuring the connectors' quality control to prevent problems such as poor contact. But visual inspection is time-consuming. Plus, no matter how careful we are in our inspection, we cannot completely prevent mistakes."

Masayuki Takagaki, MONOZUKURI Management Department, Production Management Division, Hitachi

As such, improving productivity at Omika Works has become a priority for Hitachi. The aim is to improve productivity of Omika Works such that its factories are more energy efficient, while reducing carbon dioxide emissions, making Omika Works the model for an environmentally friendly factory. To this end, Hitachi developed a proof of concept (PoC) for an visual inspection system that uses artificial intelligence (AI) to automatically detect defects in the manufacturing process, while measuring the effectiveness of this system at the factory.

Part of the verification was a visual inspection of electric wires' crimp terminal connectors, which are used to assemble the control panel. At that time, employees have to verify this component manually, but this process is time-consuming and prone to human errors.

Examples of crimped terminals and verification details
Examples of crimped terminals and verification details

"The control panels we manufacture are designed to be used for at least a decade, so we exercise great care in ensuring the connectors' quality control to prevent problems such as poor contact. But visual inspection is time-consuming. Plus, no matter how careful we are in our inspection, we cannot completely prevent mistakes," says Masayuki Takagaki, MONOZUKURI Management Department, Production Management Division at Hitachi. "Secondly, we needed to review the current visual inspection system, since it only grants access to existing users, which means that only specific people can use it. This would not be feasible when we hire more employees in the future or hand over the management of the system to new staff."

This is a conundrum that AI seems poised to address, according to Takeshi Saito, Senior Engineer, Industrial IoT and Robotics Engineering Department, Control and Service Platform System Division at Hitachi. "Inspections aren't always carried out the same way, and it takes an experienced eye to accurately discern these things. For example, what and how we check changes depending on various elements, such as the shape and size of a terminal, or thickness of a cable. I had high hopes that AI would produce results similar to a trained eye."

Through the Hitachi industrial edge computer CE50-10 that is able to implement AI and the visual inspection application built on AI, Hitachi delivers solutions that tap on edge computers to perform image recognition with AI for quality control. But when they are introduced to manufacturing sites, building individual machine learning models for every customer becomes necessary, which complicates their deployment across factories. "When we wanted to enhance their versatility, we thought of turning to a cloud service," says Naohiko Irie, Ph.D, Senior Strategist, Control System Platform Development Division at Hitachi.

Image of connection between Hitachi industrial edge computer CE50-10 and Google Cloud at Omika Works
Connection between Hitachi industrial edge computer CE50-10 and Google Cloud at Omika Works

The PoC was developed through the Lumada Alliance Program by Hitachi. An initiative for driving closer collaboration with partners to create innovative solutions. For Omika Green Network, the initiative is also leveraged to address and resolve various social and regional issues that's beyond the capacity of a single company. Working with Google Cloud, Hitachi decided to adopt Visual Inspection AI after examining the tools that can be integrated with its solutions.

When it comes to configuring the visual inspection system, Irie recognizes the importance of leveraging AI at manufacturing sites. "For general AI services, creating a machine learning model requires preparing a large number of images containing various patterns, and then spending time tuning them to improve the accuracy of detection for each type of inspection. However, it is not realistic to have an AI professional stationed at the factories to tune the machine learning model at all times. Therefore, we needed an AI service that on-site operators could easily use," says Irie. "I heard that Visual Inspection AI by Google Cloud can easily build a machine learning model for visual inspection from a small number of sample images, so combining it with the Hitachi industrial edge computer CE50-10 was ideal."

Omika Works control device production line
Omika Works control device production line

Creating a machine learning model with 100 sample images

As part of the visual inspection system, the Cosmetic Inspection feature in Visual Inspection AI was used to detect scratches on parts, and the Assembly Inspection tool to inspect whether parts are properly bonded. The on-site edge computer and Google Cloud are then connected via a virtual private network (VPN) to upload sample images. These are the basis of the machine learning models created by Visual Inspection AI.

Image of system configuration
System configuration

Approximately 100 images were used as samples for verification. Excluding the time taken to collect the images, Takagaki says that the process of creating the machine learning model took about a day. "We made about 500 to 600 images in total, and used about 100 of them. We first had the impression that using AI would be somewhat difficult, but creating the machine learning model was easy, since it's just a matter of a few simple operations on the AI ​​console. We felt that on-site staff would be able to easily use this."

"We have achieved a 100 percent defect identification rate, which means that there are no false negatives. In the operational model that we built, there is a manual visual inspection in the later phases of the process, so the false reporting rate is permissible to a certain level. We believe the model has already reached a usable level for implementing the system on manufacturing sites."

Takeshi Saito, Senior Engineer, Industrial IoT and Robotics Engineering Department, Control and Service Platform System Division, Hitachi

Saito explains that a crucial purpose of the PoC was to work out what more was needed to incorporate the system into existing on-site operations. "Manufacturing lines are based on meticulous calculations, and the time required for specific processes needs to be strictly adhered to. We need to build the operation on that condition as to where to put the machine in the process, and the timing for taking pictures and conducting inspections. Not only is it necessary to ensure that the technology works, but so is formulating a hypothesis about the operations that will create the desired production line and the expected outcomes. Being able to carry out all these will be a huge accomplishment."

Through Visual Inspection AI, Hitachi saw its inspections being performed with high-level accuracy, as it can obtain results at a level that would not disrupt the actual operation. "We have achieved a 100 percent defect identification rate, which means that there are no false negatives. However, the false reporting rate was a little high at 13%, but this was partly due to us intentionally including data that would easily prompt false reporting, in order to verify the accuracy of the system. We believe the rate will be much lower in real life. In the operational model that we built, there is a manual visual inspection in the later phases of the process, so the false reporting rate is permissible to a certain level. We believe the model has already reached a usable level for implementing the system on manufacturing sites," says Saito.

Overview of electric wiring process
Overview of electric wiring process

But some issues still need to be addressed before introducing the system to the site, such as the identification speed. "During verification, it took the system 4.5 seconds to identify each item from inference execution to displaying a result. This was at an acceptable level, but there was a request from the field to shorten it to two seconds so we need to further tune the display result and more," says Shunichi Kagaya, Senior Engineer, New Business Development & Promotion Center, Control and Service Platform System Division at Hitachi. "In addition, we need to connect the system with the existing infrastructure, such as the electric wire processing system. As there is additional equipment involved, the production flow line also needs to be inspected."

"Solving these problems means thinking about the necessary next step, and then repeating further verification steps," adds Kagaya. "Fortunately, Visual Inspection AI can quickly create a machine learning model. We can say that this speed can only be achieved with cloud services, and it's extremely vital to our operations."

"Solving these problems means thinking about the necessary next step, and then repeating further verification steps. Fortunately, Visual Inspection AI can quickly create a machine learning model. We can say that this speed can only be achieved with cloud services, and it's extremely vital to our operations."

Shunichi Kagaya, Senior Engineer, New Business Development & Promotion Center, Control and Service Platform System Division at Hitachi

Upgrading wastewater treatment facilities with Visual Inspection AI

As part of the Omika Green Network initiative, Hitachi is also conducting a dry run to upgrade its wastewater treatment facilities at its Omika Works factories and other facilities to reduce downtime, ensure that the facilities are meeting regulatory compliance and promote sustainability. This is an initiative using sensing technology and edge AI to predict water leakages in pipes, detect signs of equipment failure, visualize wastewater treatment status, and deliver treatment status updates automatically. With advanced management over the wastewater quality standards, these facilities can further reduce the factories' environmental impact.

This testing also taps on Visual Inspection AI to create machine learning models for its edge AI. This is such that the image recognition using edge AI automatically inspects whether the process to treat the wastewater with chemicals is properly handled, so that it can be returned to rivers. "The purpose of upgrading our wastewater treatment facility was to improve holistically through this environmentally-friendly design, and our edge computer and Google AI play an important role in this. Through this solution, we plan to use the cloud for features such as sensor information collection and data processing, which will then allow us to make full use of Google Cloud services in addition to AI," says Kagaya.

"This PoC has helped us create new end-to-end value by fusing the strength of our edge technology and the cloud. We would like to continue developing a system like this that's so seamless that users won't even realize the boundary between the edge and cloud," says Irie. "To develop mission-critical IoT, we will also be integrating our solutions with the advanced technology of Google Cloud."

Hitachi spokespersons
Hitachi spokespersons

Tell us your challenge. We're here to help.

Contact us

About Hitachi

Founded in February 1, 1920, Hitachi is a electronics manufacturer that develops advanced products and services in various fields, centering on business segments, such as IT, energy, industry, mobility, and life. The Omika Works, which started operations in 1969, provides information control systems for social infrastructure and industrial fields, such as power generation, transmission, distribution systems, railway operation management systems, operation and maintenance management systems for water supply and sewerage facilities, and production systems for factories and steelworks. It seeks to address various social issues and create new innovations through IoT and data analysis.

Industries: Manufacturing
Location: Japan