Ultrasound based glass fragments detection in glass containers filled with beverages using neural networks and short time Fourier transform

Glass fragments may accidentally make their ways to glass juice bottles, which has been a serious concern of beverage manufacturers. Using ultrasound to detect these fragments in the bottle poses a challenging task for signal processing and classification. When glass fragments settle down on the bottom or are sticked to the wall of a bottle, ultrasound signals returned from the fragments will superimpose themselves onto the echoes from the inner surface of the bottle. This superposition makes it impossible to distinguish the signals of bottles with fragment(s) from those without fragment(s) by commonly used methods such as time gating and spectrum analysis. A method is developed by combining radial basis function neural networks (RBF-NN) with short time Fourier transform (STFT) for ultrasound signal classification to detect glass fragments behind glass walls. The STFT algorithm was used to extract signal features while the RBF-NN was trained using the features to distinguish signals of bottles in presence from those in absence of glass fragments. Successful classification rate of 95% was achieved by this method.