Real-Time Conveyor Belt Defect Detection via Edge TPU-Accelerated YOLOv11 on Raspberry Pi 5
This paper presents and validates an end-to-end, low-power edge AI system deployed on a Raspberry Pi 5 with a Google Coral Edge TPU that utilizes a quantized YOLOv11n model to achieve real-time, high-accuracy detection of conveyor belt defects in harsh industrial environments, enabling cost-effective predictive maintenance without reliance on power-hungry GPUs.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the heavy industries that build our modern world, from mining vast quantities of ore to manufacturing cement, massive rubber belts serve as the arteries of production. These conveyor systems move thousands of tons of abrasive materials day after day, often running hundreds of meters through dusty, vibrating, and harsh environments. Over time, the constant stress of heavy loads, falling rocks, and friction causes these belts to develop hidden weaknesses. A small tear or a puncture can quickly grow into a catastrophic rupture, snapping the belt entirely. When this happens, the entire production line stops, costing the facility thousands of dollars in lost time and expensive repairs, while also creating dangerous situations for the workers nearby. For decades, the only way to find these flaws was for human inspectors to walk along the moving machinery, a method that is slow, risky, and prone to missing small but critical damage.
To solve this, researchers have turned to artificial intelligence, teaching computers to "see" defects the way a human would. However, most of these smart vision systems are too heavy and power-hungry to run directly on the factory floor. They usually require massive, expensive computers or a connection to the cloud, which is impractical for remote industrial sites with limited electricity and network bandwidth. A team of engineers in Indonesia has now bridged this gap by creating a compact, low-power system that can watch a conveyor belt in real time and spot damage instantly. They built a device that fits inside a standard metal box, uses a tiny camera to watch the belt, and runs a sophisticated visual detection program on a small, specialized chip designed for speed and efficiency.
The researchers focused on a specific type of artificial intelligence model known as YOLO, which stands for "You Only Look Once." This type of model is famous for its ability to identify objects in a video stream almost instantly. While previous versions of this technology worked well on powerful desktop computers, the team wanted to see if the newest, most efficient version could run on a tiny single-board computer, similar to those used by hobbyists, but paired with a specialized accelerator chip. This accelerator, called an Edge TPU, is a small piece of hardware designed specifically to perform the math required for artificial intelligence using very little power. The team took a camera feed from a real cement plant in West Java, where the air is thick with dust and the machinery vibrates constantly, and fed it into their system.
To teach the computer what to look for, the team gathered a massive collection of images. They did not rely solely on photos taken in a clean laboratory. Instead, they combined pictures taken directly from the cement plant with images from public datasets and new photos captured by an automated system that recorded the belt moving over time. This mix ensured the computer learned to recognize tears, holes, and patched repairs even when they were covered in dust or illuminated by harsh, uneven factory lights. In total, they used over 3,000 images to train the model, carefully marking every defect so the computer could learn the difference between a damaged belt and a normal one. They then simplified the computer program, converting it into a format that the small accelerator chip could understand and execute without needing a powerful main processor.
When the system was deployed in the actual cement plant, the results were immediate and stable. The device, housed in a sealed metal box to protect it from the dust and vibration, successfully watched the conveyor belt and identified defects as they passed by. It processed the video at a rate of 30 to 40 frames per second, which is fast enough to keep up with the speed of the moving belt. The entire process, from the camera capturing an image to the system identifying a flaw, took between 70 and 90 milliseconds. To put this speed in perspective, if the belt is moving at a typical industrial speed, the system identifies a problem before the damaged section has moved more than a few inches, giving the plant enough time to react before the damage spreads.
The system proved to be remarkably reliable under difficult conditions. Over long periods of testing, the small computer and the accelerator chip remained cool enough to operate safely without needing fans or air conditioning, maintaining steady temperatures even in the hot, dusty environment. The system did not lose any video frames or crash, demonstrating that it could run continuously without human intervention. It successfully detected four types of damage: long tears, small punctures, holes, and uneven patch repairs. While it was slightly less effective at spotting the tiniest punctures when they were partially hidden by dust, it performed with high accuracy on the larger, more dangerous defects that pose the greatest risk to the machinery.
This work demonstrates that advanced artificial intelligence does not need to be locked away in a data center to be useful. By optimizing the software to run on small, affordable hardware, the researchers have shown that factories can monitor their own equipment in real time, anywhere in the world, without relying on expensive infrastructure. The system acts as a tireless, unblinking eye that never gets tired, never misses a shift, and can spot the early signs of failure before they become disasters. This approach offers a practical path toward safer, more efficient industrial operations, proving that the future of predictive maintenance can be small, fast, and built right on the factory floor.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.