Fault-Tolerant Character Recognition in Neuromorphic Systems Using RRAM Crossbar Arrays

Resistive Random-Access Memory (RRAM) crossbar arrays provide a high-density, low-power platform for neuromorphic computing. In this work, we implement an RRAM-based architecture for alphabet recognition using the EMNIST dataset, where all 26 English letters are represented as <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$28 \times 28$</tex> binary images. Beyond ideal conditions, we study the impact of hardware imperfections, including stuck-at faults, random bit flips, and process variations, on recognition performance. To improve resilience, we evaluate two fault tolerance strategies: Triple Modular Redundancy (TMR) and Algorithm-Based Fault Tolerance (ABFT). TMR delivers strong reliability by masking faults through replication, while ABFT efficiently detects and corrects at a lower storage overhead, but at a higher computational cost. Our results demonstrate that RRAM crossbars combined with lightweight fault tolerance provide accurate, energy-efficient, and resilient neuromorphic computing, highlighting their promise for robust and efficient edge AI deployment.