人工神经网络用于近红外光谱测定柴油闪点

采用主成分－人工神经网络对不同留程柴油的近红外光谱进行校正，预测其闪点。采用监控集控制网络训练过程，以避免过训练。探讨了人工神经网络（ＡＮＮ）、直接线性连接人工神经网络（ＬＡＮＮ）的校正效果，并与局部权重回归（ＬＷＲ）、主成分回归（ＰＣＲ）及偏最小二乘（ＰＬＳ）等校正方法进行了比较，认为人工神经及直接线性关联的较好手段。

Application of Artificial Neural Network for the Determination of Diesel Fuel Flash Point by Near-infrared Spectroscopy

Back-propagation artificial neural network (ANN) and ANN with direct linear connection (LANN) were used as the mathematical models, which correlate the near-infrared spectra with the flash point of diesel fuel with different distillation range. Principal components of spectrum were used as input variate of the network to reduce noise and the number of input nodes. Monitoring set was applied to avoid over-training. The models established by LANN, ANN and other multivariate calibration methods such as locally weighted regression (LWR), partial least squares (PLS) and principal component regression (PCR), were compared for predicting the flash points of a set of Samples with wide variation of physical properties. Results were also compared, when the spectra were obtained with different light ball (source). The artificial neural network methods have better accuracy and more ruggedness. Result obtained by locally weighted regression method was accurate, but it was not rugged. When broad sample range was encountered, which usually occur in refining, the artificial neural network can ha used to construct a single model with good accuracy and ruggedness.