Intelligent Automation of Acoustic Classification Using Independent Component Analysis and Competitive Neural Nets

在本研究中，我们的最终目标是利用独立分量分析（ICA）和竞争神经网络实现具有紧凑尺寸的智能声学分类系统。利用ICA方法，开发了ICA噪声分离系统，获得了良好的信号与噪声分离效果，并将其应用于在强噪声条件下观察的紧凑型剃须机的实际声学数据。然后利用傅里叶变换将分离出来的声信号变换到频域。从频谱上确定了基于竞争神经网络的特征向量，将紧凑型剃须刀分类为好或不好。所采用的神经网络是Kohonen的自组织特征映射网络（SOM）。利用SOM，我们选择最有效的声学数据频谱特征向量。然后我们应用了另一种竞争神经网络学习向量量化（LVQ）。最后，我们还测试了其他产品，如按摩机、照明灯灯罩、轴承轴等。

In this research, we have set the final goal to achieve the intelligent acoustic classification system with compact size by using independent component analysis (ICA) and competitive neural networks. Using the ICA, we have developed the ICA noise separation system to obtain excellent separation of signals and noises and applied it to real acoustic data of compact shaving machine, which were observed under very noisy condition.Then the separated acoustic signals were transformed into frequency domain by using Fourier transform. From the spectrum we have decided the feature vector to classify the compact shavers into good or not based on the competitive neural networks. The adopted neural network is the Kohonen's network of self-organizing feature map (SOM). Using the SOM, we select the most effective feature vector of the spectrum of acoustic data. Then we apply another competitive neural network of learning vector quantization (LVQ). Finally, we have tested the other products such as massage machine, caps of illumination lamps, bearing shaft, etc.