Research On Feature Extraction Of Tobacco Leaf Oil Content Based On Touch

Oil grading judgment of flue-cured tobacco has important production significance for flue-cured tobacco grading.At present,the grading method of tobacco leaf purchasing stations is still mainly manual,but this grading method is not efficient and highly depends on the grading technical level of graders.Although in the scope of scientific research there are also using machine vision,hyperspectral technology to determine the tobacco oil,but these methods are expensive equipment,high requirements for detection environment,is still in the stage of laboratory research.Therefore,in this paper,according to different oil grade flue-cured tobacco classification problem,from the surface of the flue-cured tobacco of tactile information collection platform,flue-cured tobacco and two surface tactile information features extraction,the training sample data set selection,the features of training samples and the perspective of classification,explore the use of tactile sensor for flue-cured tobacco new method to determine oil grade,the main work is as follows:(1)Using three axis acceleration sensor and pressure sensor as a tactile information flue-cured tobacco surface receptors,and through the ANSYS analysis software sensor installation piece for the statics analysis and random vibration analysis,identified by external force,the stress of the sensor installation piece of sensitive point,namely the area of large strain and vibration,to integrate the three axial acceleration sensor and pressure sensor installed in the sensor installation piece of stress sensitive point,created a bionic tactile sensor.(2)Spectrum analysis of tactile information on flue-cured tobacco surface.The surface vibration acceleration information of flue-cured tobacco was synthesized in three dimensions,and the spectral analysis of the synthesized acceleration signals was carried out.When the initial positive pressure or sliding velocity increases,the spectral amplitude of the frictional vibration acceleration increases gradually for the same grade of flue-cured tobacco.When the sliding velocity increases,the spectral amplitude of the frictional vibration acceleration increases with the decrease of the grade of flue-cured tobacco oil in the grades of "more","there" and "slightly",and decreases with the grade of "less".The spectrum of friction vibration acceleration on the positive and negative sides of flue-cured tobacco shows that the spectrum amplitude of the negative side of flue-cured tobacco is greater than that of the positive side of flue-cured tobacco.A total of 25 features in time domain,frequency domain and time-frequency domain were extracted from the collected friction vibration acceleration signals.Through correlation analysis,it can be seen that there is a certain correlation between the features.(3)80% of the tactile information sample set of flue-cured tobacco surface was used as the training data set,and the remaining 20% was used as the prediction data set,which was input into the SVM classifier model for training.When the initial positive pressure was300 m N and the sliding speed was 20mm/s,the highest classification accuracy was93.75%.Based on the tactile information on the back side of flue-cured tobacco,the classification accuracy was 93.75% when the initial positive pressure was 200 m N and the sliding speed was 20mm/s and the initial positive pressure was 300 m N and the sliding speed was 5mm/s.The classification results show that the tactile information acquisition device of flue-cured tobacco surface made in this paper can be used to distinguish flue-cured tobacco with different oil grades,and has a good classification accuracy.(4)Combined with the SVM classifier,the extracted 25 features were selected.First,the RF-score value of each feature was calculated by the RF algorithm,and the RF-score value was used to sort the contribution degree of the extracted surface vibration acceleration features.The features were input into the SVM classifier according to their contribution degree,and it was found that when the classifier had the highest classification accuracy,the dimension of the feature subset was less than 25,indicating that the redundant features had no effect on the classification performance of the classifier.So this paper puts forward an improved RF feature selection strategy,through the RF feature selection algorithm to calculate the various characteristics of RF-Score,will feature in the RF-Score sort,in turn,add to the size of the feature subset,if classifier classification accuracy is keep the characteristic,if classifier classification accuracy does not increase or decrease to remove the feature.The improved RF feature selection strategy combined with SVM classifier was used to screen feature subsets.The features that improved the classification accuracy of the classifier were selected as wavelet scale entropy E2,average value,gravity center frequency,wavelet scale entropy E6,degree of scabbard,wavelet scale entropy E3,and the classification accuracy was 93.75%.The results show that the optimal feature subset selected by the improved RF feature selection strategy can reduce the data dimension of the feature subset while maintaining the original classification accuracy.