InGaSnO Synaptic Device Array for In-Sensor Convolution and Nondestructive Defect Detection

Abstract

In-sensor neuromorphic computing integrates sensing, storage, and computation at the device level, effectively overcoming the energy consumption and latency bottlenecks imposed by the memory-processing separation in the traditional von Neumann architecture. Here, we propose a hardware convolutional sensing architecture based on an InGaSnO synaptic device array. Benefiting from the pulse number dependent plasticity under UV illumination, the array enables direct spatiotemporal feature weighting and accumulation at the sensing layer, thereby significantly reducing data transfer and external processing burdens. In natural gas pipeline defect detection, the system achieves real time local feature extraction and anomaly identification by leveraging the parallelism of multi-node arrays, exhibiting low latency and low power consumption. This work establishes a lightweight and highly integrated brain inspired computing approach, and provides an effective hardware foundation for future smart monitoring systems and Internet of Things nodes with in sensor computing capabilities.