Detecting faults on images of solderings is a challenging task that requires a high level of precision and accuracy. The complexity of the soldering process, combined with the variety of potential defects that can occur, make it difficult to detect faults with traditional methods. Moreover, small changes in lighting, orientation, or surface reflection can greatly impact the quality of the image, making it harder for human inspectors to identify faults.

Throughout the project, we identified three primary types of faults. To determine the most effective approach for identifying which type the individual solderings belong to, we tested various AI algorithms, including CNN-based classification, anomaly detection, semantic segmentation, and unsupervised clustering. Notably, unsupervised solutions proved useful in identifying previously unnoticed fault groups in addition to the previous ones.

Fault detection accuracy of the Artillence AI detector outperforms the current computer vision (CV) and operator detection. Our detector correctly rejects faulty solderings that were accepted by other solutions, and correctly accepts good solderings that were incorrectly rejected by other solutions. The first image array shows that the Artillence detector blocks many truly bad solder joints which were let through by the CV and operator screening. The last array shows that many good solder joints are screened out by the current system, which should be let through.

For instances where no current camera solution is available, we are able to design and implement custom solutions. For demo purposes, we developed a custom built optical inspection cell that uses a Keyence 64MP camera, and Artillence's advanced AI processor for detecting soldering faults during manufacturing.