The Industrial Internet of Things (IIoT) will transform industries as we know them today by processing data in real-time, and enabling intelligent automation across the entire ecosystem. Nonetheless, the combination of IIoT systems and the ever-growing connectivity via IIoT systems presents serious cybersecurity threats, making the case for real-time intrusion detection systems (IDS) to protect critical infrastructures. This paper aims to present and analyze a machine-learning based Intrusion detection system (IDS) in real-time to prevent cyber threats. We evaluate performance across four differing approaches (ensemble learning (XGBoost and Random Forest), deep learning (BiLSTM) and classical machine learning (Logistic Regression). These models utilize NSL-KDD dataset to evalu-ate by optimizing detection accuracy along with minimizing latency for real-time applications. The performance metrics are accuracy, precision, recall, F1-score, ROC-AUC, training time, prediction time, real-time detection. Our findings re-veal that ensemble learning models, especially XGBoost, outperform others with the best accuracy (99.0289%) and prediction time (0.0422 s) indicating their suit-ability for real time detection. In comparison, BiLSTM has proven to be more suited for capturing complex temporal patterns and as such achieves high recall (0.9883) on anomaly detection (but is more computationally intensive). Logistic Regression less accurate (93.1546%) and fastest prediction times (0.0016 s), serves as a baseline under real-time applications. Finally, the study found that the dream of deploying adaptive and robust cybersecurity solutions in IIoT envi-ronments lies in the experimental results such as complexity/accuracy and com-plexity/accuracy.