Real-Time and Rapid Dynamic Missile Identification Utilizing a TiOx Memristor Array

Abstract

Real-time missile identification using artificial intelligence (AI) is becoming a crucial element in modern warfare that can significantly affect the national air defense. In this study, a real-time missile target identification (MTI) AI model is developed using step-weighted long-short-term memory networks based on a bit quantization scheme of the fabricated 1 kbit TiOx memristor array to classify five missile types: nonthreat (Non), field gun (FG), mortar (Mt), rocket (Rk), and rocket-assisted projectile (RAP). To enhance accuracy and address dataset imbalance during training, data augmentation techniques are employed, including random trajectory rotation and Gaussian noise into the radar cross-section, as well as introducing a custom loss function and dynamic learning rate (LR) to enhance early-stage prediction and accelerate learning. Employing these strategies, the proposed MTI AI model achieves a 94.4% accuracy at 3.2 s in identifying Non class, while average accuracy for five classes is 94.4% at 12.8 s. The model exhibits approximate to 43.6% greater accuracy at 3.2 s than that of the conventional model, and the estimated false-negative rate can be kept less than 2.5%. This MTI AI model can reduce the uncertainty of premature alerts for unidentified targets and exhibit superior detection capabilities for identifying and targeting missiles.