Quality inspection of specific electronic boards by deep neural networks.

Reliability and lifetime of specific electronics boards depends on the quality of manufacturing process. Especially soldering splashes in some areas of PCB (printed circuit board) can cause change of selected electrical parameters. Nowadays, the manual inspection is massively replaced by specialized visual systems checking the presence of different defects. The research carried out in this paper can be considered as industrial (industry requested) application of machine learning in automated object detection. Object of interest-solder splash-is characterized by its small area and similar properties (texture, color) as its surroundings. The aim of our research was to apply state-of-the art algorithms based on deep neural networks for detection such objects in relatively complex electronic board. The research compared seven different object detection models based on you-look-only-once (YOLO) and faster region based convolutional neural network architectures. Results show that our custom trained YOLOv8n detection model with 1.9 million parameters can detect solder splashes with low detection speed 90 ms and 96.6% mean average precision. Based on these results, the use of deep neural networks can be useful for early detection of solder splashes and potentially lead to higher productivity and cost savings.

Novel computer algorithm for cough monitoring based on octonions.

The objective assessment of cough frequency is essential for evaluation of cough and antitussive therapies. Nonetheless, available algorithms for automatic detection of cough sound have limited sensitivity and the analysis of cough sound often requires input from human observers. Therefore, an algorithm for the cough sound detection with high sensitivity would be very useful for development of automatic cough monitors. Here we present a novel algorithm for cough sounds classification based on 8-dimensional numbers octonions and compare it with the algorithm based on standard neural network. The performance was evaluated on a dataset of 5200 cough sounds and 90000 of non-cough sounds generated from the sound recordings in 18 patients with frequent cough caused by various respiratory diseases. Standard classification algorithm had sensitivity 82.2% and specificity 96.4%. In contrast, octonionic classification algorithm had significantly higher sensitivity 96.8% and specificity 98.4%. The use of octonions for classification of cough sounds improved sensitivity and specificity of cough sound detection.