Abstract:In this study, to investigate the performance of Fiber Bragg Grating (FBG) sensors in texture classification tasks, a method utilizing two FBG tactile sensors for collecting texture signals is proposed. The sensors are encapsulated in flexible polydimethylsiloxane (PDMS) blocks, and optimal encapsulation position is explored using COMSOL Multiphysics finite element simulation software, locating at 2mm from the contact surface of PDMS blocks to enhance the capability of texture information acquisition. Two flexible sensors are attached respectively to the index and middle fingers of a glove, and the glove is worn to directly touch texture samples for acquiring feature waveforms. A sliding window method is employed to traverse signal waveforms to construct feature vector sets. Principal Component Analysis (PCA) is applied to reduce redundancy and correlation of feature vector sets, followed by employing Squeeze-and-Excitation Convolutional Neural Network (SE-CNN) for data processing. In classification experiments on eight types of tile surface textures, the dual FBG perception effectively improves classification accuracy, with increases of 11.6% and 8.7% com-pared to single FBG usage, achieving a maximum accuracy of 95.2%. The average classification accuracy of 10-fold cross-validation experiments is 93.7%, which is 7.9% and 9.4% higher than traditional CNN and LSTM models, respectively. The experimental results provide valuable insights for texture recognition in robot fingertip tactile perception.