Research On Pressure Recognition Based On Bionic Tactile Sensor

Tactile sensation is an important medium for humans to perceive the world.Bionic haptics is the reconstruction of the human tactile perception system,which is widely used in the fields of biomedicine,wearable devices,intelligent machinery and human-computer interaction.Currently,research on bionic haptic is still at the early stage,and it is difficult to achieve highly accurate perception.Limited by the low detection sensitivity and low integration density of sensing elements,the sensing quality of the raw signal cannot meet the requirement.And the subsequent information processing methods cannot be fully adapted to the information of bionic haptics.And the existing information processing methods are not designed for haptic information.Therefore,aiming to improve the accuracy of bionic haptic perception,the design and manufacturing of sensors,high-accuracy perceptual recognition methods,and the lightweight of the recognition models have been studied in depth.In order to enhance the quality of sensing signals and achieve accurate force-electric sensing,a highly sensitive pressure-strain sensors based on hollow carbon spheres（HCS）modified with nitroxyl radical（NO·）was designed.Firstly,HCS-g-NO·was successfully synthesized by in situ grafting NO·onto HCS via esterification reaction,and its structure,morphology and chemical components were systematically characterized and analyzed.Secondly,the HCS-g-NO·was filled in a flexible matrix of polydimethylsiloxane（PDMS）to prepare HCS-g-NO·@PDMS force-sensitive composite.The electrical properties and mechanical properties were measured by a self-constructed calibration platform.Finally,the prepared HCS-g-NO·@PDMS was used to fabricate a mechanical sensor,and 3×3,8×8 sensor arrays for mapping the sensitivity and force spatial distribution.The results showed that NO·significantly enhanced the conductivity of HCS,with HCS-g-NO·@PDMS resistivity being an order of magnitude lower than HCS@PDMS at the same concentration.The pressure sensitivity of the sensor reached-0.55 k Pa^-1 with a strain coefficient of 211,and the sensing array was able to map the spatial distribution information of the force.Then,the spatial distribution information of force was characterized as tactile images,and the characteristics and non-ideal effects of tactile images were analyzed such as low resolution,small size,and blurred edges.To address the non-ideal effects,the full-sharing bilinear feature fusion convolutional neural network（FBF-CNN）was proposed to improve the recognition accuracy of tactile images.Fully shared bilinear features were generated using the outer product of single-stream unimodal features in different channels at the same location.So the second-order information was introduced to enhance feature sensitivity to texture and edges,with a 16.5%improvement in accuracy in ablation experiments.Feature fusion improved the efficiency of feature utilization,exploiting the complementary of higher-level semantic features with lower-level local features to improve accuracy by 4.5%.The proposed FBF-CNN achieved 98.64%accuracy in the 26 letter-shape tactile letter dataset.Finally,to improve the real-time response of bionic tactile recognition,a symmetric bilinear hybrid order convolutional neural network（SBH-CNN）was proposed for light-weight of the model.Firstly,the symmetric bilinear features were generated by the outer product of the feature map within a single channel,and the feature extraction of edges and textures were improved with an accuracy enhancement of 16.9%in ablation experiments.Since the symmetric bilinear feature is structurally symmetric,the triangular matrix of the feature maps can be used in forward transfer and storage,which significantly reduces the time complexity and space complexity of the network.Secondly,the improvements of splicing fusion and element addition on network fitting were studied,with an accuracy improvement of 5.7%and 4.8%respectively.The SBH-CNN achieved highest recognition accuracy of98.77%in the 26 letter-shape tactile dataset.Finally,the fused gradient-weighted class activation mapping（F-Grad-CAM）method was proposed to locate the region of interest of the sample with a high-resolution heat map for visual interpretation of the results of SBH-CNN.