Recognition Of Surface Properties Of Fruits And Vegetables Based On Tactile Information

Agricultural robots are more and more widely applied to fruit and vegetable picking, processing and other operating environments, the end effector usually need to grip all kinds of objects, but in an unstructured environment, the object’s stiffness, shape, texture, temperature, etc. information are often unknown, so we can not adapt appropriate strategies to implement crawl, it will make the object slip or damage to fruits and vegetables. Therefore when the robots crawl something in an unknown environment, the robot needs to be able to get the texture information of the object in order to determine the stable strategy. To this end, this paper picking robot crawling system as a research platform, developed from tactile sensors, tactile signal feature extraction, such as training and classification of samples conducted in-depth research, the main work is as follows:1)In this paper, the PVDF piezoelectric film and the resistance strain gauge were choosen as the sensor element which were suitable for detecting tactile information, four PVDF piezoelectric film sensor element and four strain gauges were used to make the tactile sensor, which were arranged by the way of random distribution,the humanoid tactile sensors was made by mimicing the characteristics of human skin and used to the research on surface properties of fruits and vegetables. To make sure the effectiveness of the output signal of the sensor, Sensor model is introduced into ANSYS to make the mechanical analysis.According to the results, the sensor element arrangement position was determined, the small output of sensor element should be arranged in the middle region of the sensor.2) Tactile information detection platform was built, in hardware, DH5853 charge amplifier and DH3841 amplifier were choose for amplifying the piezoelectric film strain and resistance strain gauge signal. Force sensitive resistive material was used as the sensing means and used to produce a force sensor, the force sensor output signal was used as feedback for controlling the clamping force between the sample and the fingers; In software, multi-channel data acquisition program was written in LAB VIEW, while for multi-channel data acquisition and storage. The actual experiments showed that the choice of the amplifier can achieve a good signal amplification,the force sensor can be used to control the clamping force between the gripper and the sample well, the data acquisition program can be used to implement synchronous multi-channel data acquisition and storage.3) Three kinds of subjects:apples, cantaloupe, cucumber were choosen for the experiment.The tactile information detect platform was used to collecte data for each 200 samples, or a total of 420 sets of data. For each sample contained four data signals and four strain gauge signals,16 features were extracted for each of sample signal, respectively, respectively, was the signal maximum Max, signal minimum Min; the difference between the maximum and minimum signal dk,dk=|Max-Min| of four piezoelectric film signal, the difference between the maximum and minimum signal dk, dk=|Max-Min| of four resistance strain gauge signals. Three kinds of classifiers were selected for classification, namely: support vector machines, BP neural networks, decision trees, classification results were: 89.99%,93.49% and 92.29%,the results showed that the classification accuracy of the three classification methods had up to about 90%, proved that the tactile sensor can be used to distinguish objects with different surface characteristics.4) For the extracted 16 features were characteristic evaluation and selection, an improved evaluation methods NG-score feature values were used to evaluate the characteristics, each characteristic value of the size of NG-score were calculated, The results show that the value of 1,2,3,4,5,6’s NG-score value was large,it means that these features had a bigger contribution rate for classification. Filter models and wrapper modles were adopted, a subset of features were constituted descending in accordance with NG-score values,if it improved the classification performance then it was selected, if lower then not elected. After feature selection algorithm to select the (3,6,2,5) for the support vector machine optimal feature subset 89.394% accuracy rate, (3,6,5,4,12) is characterized by sub-optimal BP neural network collection, accuracy was 93.737%, (3,6,2,5,4) as the optimal feature subset tree, accuracy was 93.978%.The result proved that in the subset of features as the optimal feature subset can maintain the classification performance of the original system, but also reduce the dimensionality of data sets and simplify the calculations.