Recently, Zhejiang Daily published a front-page article titled “AI Hits the Factory Floor to Secure the Final Gate of Quality Control,” focusing on how auto parts manufacturers in Yinzhou District, Ningbo, are using algorithms to tackle challenges on the production line. The original article is translated below.

Yinzhou District in Ningbo is a major manufacturing hub, home to more than 900 companies involved in the automotive parts industry. It is also one of Zhejiang Province’s first pilot areas for empowering manufacturing with artificial intelligence (AI), with a particular focus on the automotive parts sector.

So, what exactly is AI changing on the production line? Recently, reporters visited several representative companies and found some answers on the factory floor.

At the smart manufacturing workshop for new energy products at Sanlong Intelligent Technology, reporter followed Ge Qijie, Director of Digitalization and Logistics, into the production area. On a screen, a section highlighted in a red box could be seen. It was a defect identified during product inspection: a slight twist in a rubber sealing ring that was almost impossible to detect with the naked eye. Yet the AI system immediately classified it as a defective product.

“A sealing ring that is not properly installed can pose a major safety risk to the entire vehicle. In the past, these defects had to be identified by experienced workers and could not be effectively detected through conventional machine vision inspection. Now, machines are much more accurate than people,” Ge said.

Ge recalled an incident that had a major impact on the company. The company once had an export order that was at risk of being returned in full after the customer discovered a potential quality problem involving twisted sealing rings. With the products and return shipping costs potentially amounting to hundreds of thousands of yuan, the company faced significant losses.

In response, the company urgently asked its overseas staff to photograph the products one by one and send the images back. Its self-developed AI model was then used to screen the images, successfully identifying defective products in time for them to be recalled, while qualified products were delivered to the customer as scheduled.

According to Ge, the incident convinced the company to turn AI inspection from an emergency tool into a standard part of the production line. In the first half of this year, AI inspection was officially integrated into the company’s fully automated, 100% inspection process, allowing defects to be intercepted at an earlier stage.

Ge also showed the reporter a machine that was being trained to identify defects. Under different lighting conditions, viewing angles and operating environments, the characteristics of the same type of crack can sometimes disappear completely.

“We are doing everything we can to collect defect samples under extreme conditions, so that AI can learn to make accurate judgments regardless of the circumstances,” Ge said.

The company has now labeled tens of thousands of defect images and established its own defect database.

Defect characteristics vary significantly from one product to another, meaning that data from a single factory is far from sufficient. According to Ge, the company is now working on a more fundamental upgrade: moving AI from simply “replacing the human eye in taking pictures” toward multidimensional data fusion and comprehensive decision-making.

In the past, visual inspection, air-tightness testing and tolerance measurement were conducted independently. The company is now integrating visual data, pressure and air-tightness data, as well as assembly-condition data, to make comprehensive judgments.

“For example, if a sealing ring is installed slightly less than required, the internal pressure may be higher than normal, but the product may still meet the standard after subsequent pressing and assembly. Looking at just one indicator could lead to a misjudgment. With multidimensional data fusion, we can identify genuine defects more accurately,” Ge explained.

Making AI Work on the Production Line

How can AI truly “work” on the factory floor?

Sanlong’s strategy is pragmatic: the company develops core industrial AI capabilities in-house while relying on external partners for other areas of research and development.

The company has worked with university research teams led by PhD researchers to develop AI monitoring systems in laboratories. With support from the local government, it has also connected with leading companies such as Geely Research Institute to draw on their experience and expertise.

Ge said the company’s goal for this year is particularly practical: “We want the company to see that AI really works, and we want our employees to feel that AI is actually helping them in their work.”

During the interviews, some entrepreneurs noted that industrial AI is still in a stage of exploration and further development. There is still a considerable gap between current capabilities and the ideal state of AI being able to “learn autonomously, verify autonomously and optimize autonomously.”

From a technological perspective, robots operating continuously with high precision still face challenges in areas such as control accuracy, stability and heat dissipation.

From an economic perspective, companies also have to bear relatively high costs related to computing power, equipment training and real-world application testing.

From a Single Workshop to an Entire Industry

AI-powered breakthroughs in individual manufacturing applications are already underway. But how can these advances move from a single workshop to an entire industrial sector?

The problems faced by leading companies—particularly insufficient data and the high cost of trial and error—are common across the industry.

Yinzhou District has adopted a model featuring “state-owned enterprises providing the platform and specialized division of labor.”

District-owned enterprises work together with local service providers and leverage the real production-line resources of leading companies such as Sanlong. Together, they have initially established one general-purpose dataset and six specialized datasets, accumulating more than 800,000 sets of core data related to 3D vision and quality inspection.

As a result, small and medium-sized enterprises no longer have to start from scratch. They can share data resources and algorithm capabilities, while the costly process of trial and error can be spread across the broader industrial ecosystem.

One remark from a company executive during the interviews was particularly striking:

“AI is not here for show. It is here to solve real problems.”

In the workshops of Yinzhou, we are seeing those words become reality.

Disclaimer: Portions of the content and/or images in this article are sourced from enterprises or publicly available online resources. If any material infringes upon your copyright or other legal rights, please contact us for prompt removal.