Time-generative AI-enabled temporal fusion transformer model for efficient air pollution sensor calibration in IIoT edge environments

The deployment of real-time sensor calibration models for air pollution monitoring on resource-constrained Industrial Internet of Things (IIoT) edge devices presents significant challenges due to the computational complexity and memory requirements of deep learning models. This paper addressed these challenges by proposing a time-series-generative approach that integrated model quantization, generative artificial intelligence (AI), and temporal deep learning architectures to ensure efficient deployment. Specifically, we introduced a TimeGAN-augmented temporal fusion transformer (TFT) model optimized for edge devices. By leveraging model quantization, the approach reduces the memory footprint and computational demands of the model without compromising calibration accuracy. Furthermore, the integration of generative adversarial networks (GANs) enhances the robustness of the model by generating high-quality synthetic time-series data, compensating for sparse or noisy sensor readings. This ability to generate synthetic data mirrors the real sensor trends, ensuring reliable model performance even in data-limited environments. A comprehensive evaluation of the proposed model, comparing its performance against both float and quantized versions, demonstrates the effectiveness of the TimeGAN-augmented quantized TFT. This model achieves a significant 88% reduction in size (from 800.04 KB to 97.34 KB) while maintaining excellent predictive performance, evidenced by a mean squared error (MSE) of 0.3212 and a mean absolute error (MAE) of 0.4375. Additionally, the TimeGAN-augmented Float TFT model emerges as a strong contender for real-time applications, offering an optimal balance between inference speed and accuracy, with a rapid inference time of 23.4 ms, making it ideal for real-time pollution monitoring.