Q-DELTA: A Novel GA-Autoencoder Anomaly Detection for Logistic Tracking Using Quantization

Abstrak

Delivery delays are becoming more prevalent in logistic tracking, and irregularities can negatively impact both the packet being sent and the recipient. Detecting delivery anomalies early can prevent delays. However, these detections are computationally heavy and inefficient. Complex deep learning models that can learn anomalies, such as autoencoders, are difficult to run on resource-limited Internet of Things (IoT) devices because they require high-spec hardware for deployment. Therefore, this paper proposes a quantized delay detection in logistic tracking autoencoder (Q-DELTA), a computationally efficient, lightweight deep learning model deployed on IoT systems that leverages quantization in autoencoders to detect anomalies in logistic tracking services. Quantized autoencoders are fast and compact while minimizing performance loss. To achieve an optimized, efficient autoencoder-based anomaly detection, a dual-step approach is used: the genetic algorithm (GA) and Youden’s J statistic (J-stat). With benchmarked methods, the quantized autoencoder is the most accurate, fastest, and smallest anomaly detection model. Experimental results demonstrate that the novel GA-autoencoder, combined with J-stat and an edge-based architecture for logistics tracking, enables anomaly detection using quantization.

Identifier

• OpenAlex: https://openalex.org/W7140948363
• DOI: https://doi.org/10.1109/access.2026.3678089