A Customized Design Method for Bionic Spider Crack Strain Sensors Based on Active Learning and Data Augmentation

To enable intelligent and efficient demand-oriented customization of flexible strain sensors, this article proposes a new machine learning paradigm integrating active learning and data augmentation for the inverse design of bionic spider crack strain sensors. First, the sensing mechanism is validated to establish a physical foundation for data augmentation. Subsequently, a random forest-based prediction model is constructed, and a navigation function is introduced to guide the active learning loop for efficient exploration of the design space. Multiple data augmentation techniques are systematically compared, among which the generative adversarial network (GAN) is effectively employed to expand the dataset, addressing the challenge of small-sample modeling and ultimately establishing a reliable performance prediction model. Based on this model, optimal structural parameters can be reversely recommended according to target strain ranges and gauge factor (GF) requirements. Experimental results demonstrate that the sensor fabricated based on predefined requirements exhibits excellent performance: a representative GF of 20 575, response/recovery times of 45 and 49 ms, respectively, detectable strain as low as 0.1 %, and stable operation over 1,000 cycles. Applications in human physiological signal monitoring, joint motion detection, and robotic hand gesture detection fully validate the effectiveness and application potential of this customized design approach. This study provides a novel technical pathway for the rapid and on-demand development of high-performance flexible sensors, which is expected to advance flexible electronics and intelligent sensing systems toward greater efficiency, intelligence, and personalization, offering significant theoretical and application prospects.