Piston Ring Inspection
Cognex Deep Learning defect detection tool simplifies the automated detection and characterization of defects on textured metal surfaces
A piston’s compression rings serve several functions in a reciprocating engine, sealing the combustion chamber and regulating oil consumption. Defects on a compression ring are difficult to detect because of the piston’s reflective metallic surface. The cylindrical shape of the piston sometimes renders as blurry and un-focused in images. Normal variations in the metal’s surface texture are to be expected as part of the manufacturing process, and some—including rust spots, white areas, and even surface cracks and fissures--are permitted to pass inspection. Long scratches that affect the piston’s performance and threaten compression levels in the cylinder, however, are indications of true defects. An inspection system must be able to tolerate normal variations and insignificant anomalies on compression rings’ surfaces while identifying any long scratches.
Programming an inspection of this complexity into a rule-based algorithm would require complex defect libraries. Human inspection, though more flexible, would be too slow. Cognex Deep Learning offers an effective inspection solution, combining a human’s ability to appreciate minor variations with the reliability, consistency, and speed of an automated system. Using the Cognex Deep Learning Defect Detection Tool in supervised mode, an engineer trains the deep learning-based software on a representative set of known “good” and “bad” compression ring images. A technician annotates known “bad” images where long scratches occur and “good” images with normal variations and tolerable defects, including rust spots and small cracks. Based on these images, Cognex Deep Learning learns a piston’s natural form and surface texture, as well as the normal appearance of scratches. Additional images can be added to the training set during validation testing to reflect additional examples and help optimize the system. Parameters can continually be adjusted during the training phase and validation period until the trained model correctly detects and segments all images with long scratches. During run-time, the software characterizes images with long scratches as defective, having learned to recognize and ignore irrelevant variations.