Kocaeli University, Department of Environmental Engineering, 41380 Kocaeli, Turkey.
Kocaeli University, Department of Mechanical Engineering, 41380 Kocaeli, Turkey.
Waste Manag. 2016 Oct;56:46-52. doi: 10.1016/j.wasman.2016.06.030. Epub 2016 Jul 1.
In this study, we investigate the usability of sound recognition for source separation of packaging wastes in reverse vending machines (RVMs). For this purpose, an experimental setup equipped with a sound recording mechanism was prepared. Packaging waste sounds generated by three physical impacts such as free falling, pneumatic hitting and hydraulic crushing were separately recorded using two different microphones. To classify the waste types and sizes based on sound features of the wastes, a support vector machine (SVM) and a hidden Markov model (HMM) based sound classification systems were developed. In the basic experimental setup in which only free falling impact type was considered, SVM and HMM systems provided 100% classification accuracy for both microphones. In the expanded experimental setup which includes all three impact types, material type classification accuracies were 96.5% for dynamic microphone and 97.7% for condenser microphone. When both the material type and the size of the wastes were classified, the accuracy was 88.6% for the microphones. The modeling studies indicated that hydraulic crushing impact type recordings were very noisy for an effective sound recognition application. In the detailed analysis of the recognition errors, it was observed that most of the errors occurred in the hitting impact type. According to the experimental results, it can be said that the proposed novel approach for the separation of packaging wastes could provide a high classification performance for RVMs.
在这项研究中，我们研究了声音识别在回收机（RVM）中用于包装废物源分离的可用性。为此，准备了一个配备声音记录机制的实验装置。使用两个不同的麦克风分别记录了三种物理冲击（自由落体、气动冲击和液压破碎）产生的包装废物声音。为了根据废物的声音特征对废物类型和尺寸进行分类，开发了基于支持向量机（SVM）和隐马尔可夫模型（HMM）的声音分类系统。在仅考虑自由落体冲击类型的基本实验设置中，SVM 和 HMM 系统为两个麦克风提供了 100%的分类准确性。在包括所有三种冲击类型的扩展实验设置中，动态麦克风的材料类型分类准确率为 96.5%，电容麦克风的准确率为 97.7%。当同时对废物的材料类型和尺寸进行分类时，两个麦克风的准确率为 88.6%。建模研究表明，对于有效的声音识别应用，液压破碎冲击类型的记录非常嘈杂。在对识别错误的详细分析中，观察到大多数错误发生在冲击冲击类型。根据实验结果，可以说用于包装废物分离的这种新方法可以为 RVM 提供高分类性能。
