Staff in full protective gear work in the BICV Technology smart cockpit display workshop. (Photo/Liu Li)
Chongqing – Chongqing produced 4.285 million new energy vehicles (NEVs) in the first four months of 2026, nearly 45% of total output, according to according to the China Association of Automobile Manufacturers (CAAM). As NEV production and smart vehicle tech advance, traditional inspection tools face limits, boosting demand for AI-powered visual quality checks.
BICV Technology, headquartered in Chongqing, is among the companies addressing this challenge. The company specializes in the research, development, and production of intelligent connected automotive products.
Inside its production workshop, a palm-sized circuit board contains hundreds of electronic components. Workshop engineer Xing Donglin said some of the smallest capacitors are even smaller than sesame seeds, making them difficult to identify with the naked eye.
Traditional Automated Optical Inspection (AOI) equipment struggles to inspect products with such precision. Under conventional inspection methods, engineers must manually mark, position, and adjust parameters for every new component on a circuit board. Programming a single board typically takes one to two hours. During production, even slight changes in lighting conditions or component color can trigger equipment alarms.
Industry data show that traditional AOI systems often record false detection rates of up to 5% when inspecting complex Printed Circuit Board Assembly products.
“This not only reduces the overall first-pass yield rate and consumes production capacity through manual reinspection but also increases the risk of operators overlooking real defects due to fatigue, creating potential quality risks,” said a project manager at BICV Technology. “Introducing more intelligent AI visual inspection technology has become an urgent need for the company.”
However, applying AI to visual inspection requires more than simply installing off-the-shelf software.
“The challenge lies in the fact that standardized AI products available on the market cannot truly understand our circuit boards,” the project manager said.
Differences between component batches, combined with subtle changes in workshop lighting conditions, create distinct characteristics that cameras and standardized AI algorithms often struggle to interpret.
To enable AI to accurately identify defects measuring only fractions of a millimeter in environments affected by reflections, shadows, and color variations—and achieve near-perfect accuracy—the company needed to continuously train the system using real production scenarios.
To address this challenge, BICV Technology and its partner launched a joint development program at the end of 2025. Technical teams from both sides worked directly in the production workshop, training the AI model with hundreds of thousands of images of qualified and defective products.
After training on this large dataset of annotated images, the AI model gradually developed the ability to identify complex hidden defects and entered operation several months later.
Today, when BICV introduces a new product, the AI system automatically recognizes components on the circuit board.
“Our programming time for component recognition has been reduced by 50%, and overall production line efficiency has improved significantly,” Xing said.
After learning from massive datasets, the AI system no longer reacts to surface reflections or minor color differences. Its ability to identify common defects has improved substantially. The workshop’s first-pass yield rate has reached 99.9%, while the false detection rate has fallen from 5% to below 0.1%.
As AI technology becomes more deeply integrated into production, BICV Technology’s manufacturing model and workforce responsibilities have also evolved.
AI systems now work alongside employees. The AI handles image scanning, feature extraction, and defect identification, while workers focus on AI training and optimization. Their responsibilities have shifted from parameter adjustment and rule setting to data optimization and process improvement.
Xing said that when the AI encounters a new type of defect that it cannot accurately assess, workers only need to label the issue and provide feedback through the system.
“Once the AI receives feedback, it becomes smarter the next time,” he said.