霉变玉米电子鼻检测中信号降噪及特征提取方法研究

发布时间：2018-06-22 13:23

本文选题：电子鼻 + 霉变玉米　；参考：《河南科技大学》2017年硕士论文

【摘要】：为了提高电子鼻检测霉变玉米的正确率,探究了电子鼻信号不同多特征组合表征模式方法对鉴别结果的影响,并给出了一种基于Wilks统计量的特征参量鉴别能力评价方法。同时,考虑到不同气敏传感器选择特性的差异,在多特征表征模式下给出了一种基于Wilks统计量主元消去变换的传感器信号表征特征的筛选方法。为了提高电子鼻预测霉变玉米真菌毒素含量,在五特征表征模式下发展了一种基于核Fisher判别分析融合BP神经网络的预测模型构建方法,以期提高霉变玉米真菌毒素含量的电子鼻检测能力。具体研究工作说明如下:首先对电子鼻中每个气敏传感器的霉变玉米气敏信号分别提取积分值、小波能量值、方差、相对稳态平均值、平均微分值作为特征值。然后,利用WilksΛ统计量计算单一特征表征模式和多特征组合表征模式下所构造的特征向量的鉴别能力。计算结果表明,多特征表征模式的鉴别效果优于单特征表征模式的鉴别效果,并且随着表征特征数的增多,所对应的特征向量的鉴别能力也进一步提高。研究结果也指出了多特征表征模式下如何进行特征组合是不具有规律性的,但可通过对比不同特征组合的WilksΛ值,来获得较好的表征特征组合。同时,在多特征组合表征模式下,借助于所给出的特征筛选方法,优化筛选了不同传感器气敏信号的特征组合。结果显示:在多特征表征模式下不同传感器特征表征是不同的,说明了特征筛选的必要性。为了揭示上述特征鉴别能力评定方法的有效性,运用Fisher判别分析(FDA)直观考察了不同特征表征模式下的鉴别结果。FDA结果显示,无论是单一特征表征模式还是多特征表征模式,它们的鉴别效果与基于WilksΛ统计量的鉴别能力评价结果相吻合,而且随着表征特征数的增多,鉴别正确率也逐渐提高,5特征组合下的FDA鉴别正确率升至98%。FDA分析结果表明,所给出的特征鉴别能力评价方法是有效的。最后,分别借助于BP神经网络、核变换FDA(KFDA)融合BP神经网络方法,研究了黄曲霉毒素B1含量、呕吐毒素含量、玉米赤霉烯酮含量的预测模型构建方法。预测结果显示:单纯的BP神经网络预测玉米赤霉烯酮、呕吐毒素含量、黄曲霉毒素B1,预测误差在5%以内的正确样本数所占比例最高为85%;而基于KFDA的BP神经网络的预测误差在0.6%以内的正确样本数所占比例为100%。且两种模型的预测值和实测值的拟合决定系数由0.95提高到1.00。研究结果表明,运用基于KFDA的BP神经网络方法预测霉变玉米真菌毒素含量是有效的,提高了电子鼻的检测精度。论文研究结果可获得4个方面的结论:1)用多特征融合表征模式可以更有效地反映霉变玉米样品的电子鼻响应信息。2)多特征表征向量的鉴别能力可用WilksΛ统计量进行有效评价。3)多特征表征模式下,每个气敏传感器的表征特征可用Wilks统计量主元消去变换的方法进行筛选。4)基于KFDA融合BP神经网络方法预测霉变玉米真菌毒素含量是有效的。
[Abstract]:In order to improve the accuracy of detection of mouldy corn by electronic nose, the influence of different multi feature combination characterization methods on the identification results of electronic nose signals is explored, and an evaluation method for distinguishing feature parameters based on Wilks statistics is given. At the same time, considering the difference of the selection characteristics of different gas sensitive sensors, the multi feature characterization model is taken into account. In order to improve the prediction of mycotoxin content in mouldy corn by Wilks, a method of building a prediction model based on nuclear Fisher discriminant analysis fusion BP neural network is developed to improve the mildew in order to improve the mildew. The electronic nose detection ability of the content of mycotoxin in corn. The specific research work shows as follows: first, the integral value, the wavelet energy value, the variance, the relative steady state value and the average differential value are taken as the eigenvalues respectively, and the single characteristic table is calculated by the Wilks statistics. The results show that the discriminant effect of the multi feature representation model is better than that of the single feature representation model, and the identification ability of the corresponding eigenvector is further improved with the increase of the characteristic feature number. The feature combination under multi feature representation is not regular, but the better characterization combination can be obtained by comparing the Wilks values of different features. At the same time, the feature combination of different sensor gas sensing signals is optimized by using the feature selection method given in the multi feature combination representation mode. The results show that the feature representation of different sensors is different under the multi feature representation model, indicating the necessity of feature selection. In order to reveal the effectiveness of the evaluation method for the characteristics of the above features, the Fisher discriminant analysis (FDA) is used to visualized the.FDA results of the identification results under the different characterization patterns, regardless of the single feature. Characterization mode or multi feature representation model, their identification results are consistent with the evaluation results based on Wilks based statistics, and with the increase of characterization number, the correct rate of identification is gradually improved. The FDA identification accuracy under the 5 feature combination is raised to the 98%.FDA analysis results. The method is effective. Finally, with the help of BP neural network and nuclear transformation FDA (KFDA) fusion BP neural network, the method of predicting the content of aflatoxin B1, the content of vomit toxin and the prediction model for the content of Zea zearalenone is studied. The prediction results show that the simple BP neural network predicts the content of zearalenone, the content of vomit toxin and Aspergillus flavus Toxin B1, the maximum number of correct samples within 5% of the predicted error is 85%, while the proportion of the correct samples within 0.6% of the KFDA based BP neural network is 100%., and the fitting decision coefficients of the predicted and measured values of the two models are increased from 0.95 to the 1.00. research results, and the KFDA based BP nerve is used. The network method is effective to predict the content of mycotoxin in mouldy corn, which improves the detection precision of the electronic nose. The results of this paper can be obtained from 4 aspects: 1) the multi feature fusion characterization model can more effectively reflect the electronic nose response information.2 of mouldy corn samples). The identification ability of the multi characteristic vector can be calculated by the Wilks statistics. Under the multi characteristic.3) model, the characterization features of each gas sensor can be screened by the method of Wilks statistic principal element elimination transformation. It is effective to predict the content of mycotoxin in mouldy corn based on the KFDA fusion BP neural network method.
【学位授予单位】：河南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TS207.3;TP212

【参考文献】