提出了特征矩阵联合对角化(JADE)结合超限学习机(ELM)的稳健建模方法, 并应用于砂梨成熟度的鉴别。砂梨近红外光谱是多种独立化合物光谱信号的随机线性混合, 首先采用多元散射校正和小波变换去除原始光谱噪声, 再利用 JADE提取独立光谱, 得到包含独立化合物浓度信息的混合矩阵; 随后使用 ELM算法, 通过调节隐层节点个数建立稳健性强的成熟度鉴别模型。JADE利用高阶累积量全面提取原始光谱的幅值、相位信息, 降低不同化合物之间的光谱干扰, 而 ELM隐层节点参数随机生成, 两者的有机结合可使所建模型稳健性强, 有利于模型的传递与共享。该方法应用于砂梨 4种不同成熟度的鉴别, 所建模型预测准确率为 96.67%。

The paper proposes an method of application of Joint Approximative Diagonalization of Eigenmatrix(JADE) algorithm and Extreme Learning Machine(ELM) for modeling steadily to discriminate maturity of different Chinese pears. The near infrared spectra of Chinese pears were linear combination of the different chemical components. To eliminate the noise, MSC and wavelet transform were used. Then the source signals were extracted from initial data set by JADE and a linear representation of non-Gaussian data was founded. By changing the number of neurons, ELM was used to build a discrimination maturity of different pears stability model. JADE was able to find more complete information of samples, and reduced the spectral interference. ELM has a high measurement precision and sets a few parameters. This method is the foundation of model transfer and sharing. Parameters of ELM algorithm were random, which made the model more stable. With Chinese pears as experimental samples, the model prediction accuracy was 96.67%.