The process of buying new home appliances should be exciting, but for one customer in Guangzhou, it quickly turned into a frustrating experience. When the delivery team unpacked her newly purchased dishwasher, she spotted a fingerprint on the door and a half-dried water stain inside, and a closer look at her new air conditioner revealed a fine scratch on the panel. This cascade of defects led to a cascade of complaints and a request for a replacement, a scenario that plays out in homes across the country every single day.
For years, the home appliance industry has relied on human eyes to spot cosmetic flaws during quality checks. Even the most experienced inspectors can miss things, as the human eye can only reliably detect marks larger than 10 microns, and many subtle defects are only visible when the product is angled just right. In recent years, camera-based visual inspection has become more common, but the industry standard involves scattered single-point cameras. Each camera can only capture a small portion of the product, leaving curved surfaces, edges, and corners as blind spots where subtle imperfections frequently go undetected, making missed or false detections a constant problem.
Fortunately, AI has arrived to change the game.
In Wuxi, Jiangsu, a pilot project is underway at Midea's dual high-end factory for washing machines, specifically on the Toshiba production line. This project aims to develop the highest-standard exterior visual inspection system ever used in the home appliance industry, and it must be implemented flawlessly from the start. The term "flawless from the start" means the system must pass two rigorous test rounds before it goes live, a stark contrast to the industry norm where systems are deployed and then iteratively improved by collecting defect data during real-world operation. The first test round involves marking known defect points on a product, and the system must identify all of them. The second round presents a more difficult challenge: defects are placed randomly with no hints, requiring the system to autonomously detect them all.
The challenge was handed to the Intelligent Equipment Research Institute at Midea's Intelligent Manufacturing Research Institute (IMRI). At the initial project meeting, no one was eager to volunteer, given the extraordinarily high bar. The director, Yin Bo, began looking for someone with the right skillset. He zeroed in on Peng Bo, a former internet industry algorithm developer who joined Midea in 2022. Peng's background was in AI voice and search algorithms, and Yin Bo believed this experience could be the key to solving the project's central problem. The internet's AI modeling approach, which relies on massive datasets, could potentially compensate for the chronic lack of defect samples in industrial settings, making him the perfect person to lead the project.
The initial idea was promising, but the fundamental logic of internet AI and industrial vision are worlds apart. While internet companies can easily access hundreds of millions of user data points to train models, a new production line struggles to gather even a handful of real defect images. Peng Bo admits the biggest challenge was the severe shortage of defect samples. To make matters worse, the washing machine bodies vary in height from 850mm to 1780mm, with mixed-model production requiring constant changes, effectively doubling the difficulty of the setup. The production line's single-unit cycle time is only 22 seconds, mandating a 100% inspection rate with no sampling allowed. There were no off-the-shelf solutions to reference; everything from the algorithm to the imaging architecture had to be built from scratch. Peng Bo and his team immersed themselves on the production line, meticulously breaking down risks and compiling a comprehensive list of model sizes, defects, and processes.
Meanwhile, Yin Bo was not idle. He delved into academic papers, consulted university experts specializing in industrial vision, and translated their core algorithms into practical applications. The team even developed proprietary AIGC technology to simulate defects. They were racing against the clock because a traditional approach could take two to three months, time the eagerly awaited Toshiba line simply didn't have. The AIGC integration was a game-changer, drastically accelerating development. With just five to ten real defect images, they could generate thousands of realistic simulated samples, trimming the model training cycle by two-thirds. This proprietary method quickly pushed model recognition accuracy to 90%, just crossing the threshold required for Toshiba's launch.
The visual inspection system developed by IMRI features three key innovations. The first is 360-degree, full-surface AI visual automated inspection. This involves four 8K line-scan cameras and two area-scan cameras mounted on a synchronized gantry system, enabling a single pass to capture all five sides (front, back, left, right, and top) with no blind spots. The team also refined a fly-by shooting technique that captures high-definition images without stopping the line, achieving a minimum cycle time of just 2.78 seconds.
The second innovation is a flexible, multi-model 3D gap measurement system. It uses 2D cameras to first measure the overall position of the unit, then a collaborative robotic arm automatically switches to a dedicated scanning trajectory. The gap measurement precision reaches 0.1mm, with a repeatability error of only 0.04mm. A single workstation completes the measurement in under 15 seconds, making it more than 100 times faster and more accurate than manual inspection.
The third innovation is full life-cycle digital traceability. Each appliance is assigned a unique body barcode upon assembly, and all images and inspection records are stored and archived accordingly, allowing for traceability at any time. If a new, previously unseen defect appears, samples are collected on-site and the AI model is updated online in real time.
Behind these three innovations lies a long and difficult journey marked by countless setbacks and compromises. Take, for example, the development of the 360-degree system. In the industry, area-scan cameras are the standard choice for appliance inspection due to their familiarity and simplicity. Sticking with the proven method seemed safe, but it quickly proved inadequate. The cameras struggled with products featuring large curved surfaces, as the lighting would create uneven brightness that hid subtle scratches and dirt in shadows. Attempting to solve this by adding more area-scan cameras would have required at least a dozen units, and even then, detection on curved surfaces was not guaranteed. Peng Bo concluded that more cameras didn't equate to better detection capability. Through rigorous lab testing, his team proved that 8K line-scan cameras, which work like high-speed scanners by capturing and stitching images during motion, offered over three times the resolution of a 2.5K area-scan camera and produced significantly better image quality for curved surfaces. The trade-off was the need to build a completely new piece of equipment: a gantry lifter. This was unprecedented in the industry and sparked considerable internal concern about potential failures that could halt production. After months of debate, the team meticulously addressed all risks, bringing in factory and supplier experts for three-party reviews, and implementing multiple safety features like safety edges, light curtains, and laser ranging to eliminate collision risks. They also thoroughly mapped out the entire product range to account for all sizes and materials. After nearly half a year of relentless laboratory testing and line debugging, this truly industry-first technology was born.
The industry widely believes that IMRI's self-developed AI inspection system has established a unified, high-precision, and replicable quality control benchmark for outgoing products. The defect detection rate on current projects is stably above 99%, and in some cases surpasses 99.5%, with a long-term target of 99.99%. Peng Bo has a clear roadmap for continuous improvement: the production lines will keep collecting samples of new exterior defects, and the AI model will be retrained periodically, ensuring the system benefits from lifetime algorithm optimization. By the spring of 2026, two production lines at the Wuxi dual high-end factory are scheduled for full acceptance and operation. The COLMO line will have eight inspection points, and the Toshiba line will have six, completely replacing manual quality inspection with AI. In the six months following launch, the system is projected to have reduced losses from defects by over 12 million yuan.
While the factory's cost reduction and efficiency gains are significant, the core value is ultimately passed on to the consumer. The AI's micron-level, full-dimensional screening blocks flawed machines from ever leaving the factory, ensuring customers open their boxes to find pristine, perfectly assembled appliances with no scratches or uneven gaps. This eliminates the hassle and time cost associated with returns and exchanges. Looking further ahead, the deep integration of AI will continue to lower quality control costs, and these savings are expected to trickle down to consumers, making high-quality home appliances more affordable in the future.