In recent years, security landscape of Industrial Internet of Things (IIoT) has emerged as a pivotal area of concern within the realms of cyberwarfare and engineering research, particularly after the rise of advanced and sophisticated malware (e.g., Stuxnet) deliberately crafted to target such systems. Although artificial intelligence has achieved a great success in the field of intrusion detection, the limited computing capabilities and storage of IIoT devices hinder the actual deployment of high-complexity models. In this paper, a lightweight intrusion detection system (LIDS) empowered by machine learning technologies is proposed to detect intrusions in an effective and efficient manner to enhance network security of IIoT. Firstly, we apply a wrapper-based approach via LightGBM model to identify the most important features. Features with the highest importance scores are effectively selected and constructed as the optimal feature subset. Then, we develop a advanced model of extreme gradient boosting (XGBoost) with Bayesian optimization (BO) for intrusion detection, where the important parameters of XGBoost are optimized and fine-tuned using BO algorithm to improve the detection accuracy, computational efficiency, and stability of the model. Experimental results on NSL-KDD, a benchmark dataset to demonstrate that the proposed method achieves an accuracy of 99.79% in binary classification and 99.84% in multi-classification, showcasing its potential as a reliable security solution for IIoT.