Intelligent defect detection system in manufacturing

Intelligent video analytics can be used to automate quality control in various types of manufacturing. Using cameras, AI captures the blanks or products passing along the conveyor and detects defects. In most cases, visual inspection of products is conducted by quality department employees, which carries the risk of human error, and the inspection itself takes a considerable amount of time. Moreover, most defects are not detected in advance leading to financial losses and negatively impacting the production process Let's consider an example from a specific industry – metallurgy. To produce high-quality rolled products it's crucial to detect defects in steel billets promptly, as this affects the quality of the final product. Such control is often carried out visually by the technical personnel of the workshop and the quality management staff. Since the steel billets are stored closely together, during inspections, employees do not have the opportunity to examine them from all sides. Additionally, when loading onto the line before entering the furnace, the movement speed of the billets is too high, reaching 2 meters per second, which does not provide enough time for visual inspection. Billets with detected defects are also marked manually with chalk. Under poor lighting conditions and at high speeds, operators must continually monitor the roller conveyor to reject the billets at the furnace loading stage and redirect them not to rolling but to the processing of the defective area. The process is automated by a defect detection and recognition system based on convolutional neural networks by VisionLabs. The installation of cameras and use of video analytics enables the analysis of every centimeter of the steel billets. The process is structured as follows: the billet moves along the mill's loading roller conveyor. Before weighing, the billet passes through a surface defect inspection point, where cameras are installed to capture the condition of the billet's surface on each of its four sides. When the billet enters the cameras' field of view, the neural networks detect it and check for defects. When a defect is detected in real time, the operator receives an audible alert and an image along with the serial number of the defective billet is displayed on the screen. This allows for timely rejection and re-routing for defect area processing instead of rolling. Additionally, the system counts the number of accepted billets and classifies detected defects for further analysis, including distinguishing between critical production defects and acceptable ones that don't affect the final product quality. The collected statistics can be used to modify the technological processes in previous workshops.