【摘要】 为了解决红外光谱定量分析中的特征提取和校正规模问题,提出了一种输入层自构造神经网络。这种网络能够利用训练数据的某些先验知识,自动选择输入层神经元的个数。在学习过程中,输入神经元个数从最小值1开始,根据网络误差的变化逐步增加,最终确定最佳神经元数量。这种网络模型将特征提取和参数学习过程融为一体,有利于提高建模效率。利用仿真红外光谱的定量分析实验表明,这种网络模型不仅能够对光谱数据实现高效率的波长选择,并具有抑制随机噪声和非线性干扰的能力。更多还原

【Abstract】 In order to solve the problems of feature extraction and calibration modeling in the area of quantitatively infrared spectral analysis, a structural adaptive neural network is proposed. In this network, some prior knowledge can be utilized and the network can be constructed automatically . During the learning process, the number of the input neurons is increased from 1 according to the change of the model error and thus the best number of the input neurons is determine at the end of the process. In this algorithm, the feature extraction and the parameter learning are handled at the same time. This increases the efficiedcy of modeling. The experiment of quantitative analysis using simulated infrared spectra shown that this network algorithm not only realized a high performance of wavelength selection, but also reduced the disturbances of random noise and non-linearity the spectral data.更多还原