Abstract:A real-time, non-contact circuit boards fault diagnosis algorithm based on deep learning is presented to solve the problem that existing contact circuit board fault detection methods is difficult be applied to large scale integrated circuit fault detection. Establish an image data set of PCB board defect detection and component recognition, and adopt data enhancement technology to enhance the data volume of training to improve the accuracy and robustness of model detection.Component detection model is got by training based on Darknet framework and YOLO4 algorithm, and reasonable Anchors is designed by K-means clustering algorithm to make the model have multi-scale defect detection function. Image registration algorithms are used to register and fuse infrared and visible images. According to the functional area divided by PCB board design, the average temperature of five areas is collected continuously by thermometry thermal imager, and the short circuit or short circuit status is judged by judging the relationship between the five average temperatures. After testing, using pre-set faulty circuit board as the experimental object, by collecting infrared and visible image data during the operation of the experimental object, based on the designed fault detection model, not only the real-time and effective identification of component location, but also the intuitive identification of components with short-circuit and short-circuit faults.After practical application, it can satisfy the engineering application of real-time circuit board fault detection when the equipment is running.
