With smart connectivity and integration into IT networks, Industrial Control Systems (ICS) are becoming increasingly vulnerable to cyberattacks. Traditional Intrusion Detection Systems work on the premise of having a library of known attacks and are therefore ineffective against unknown attacks. This paper introduces an adaptive cyber defence system to protect ICS using the Machine Learning (ML) based Network Intrusion Detection Systems (NIDS). The proposed framework tests six machine learning algorithms namely, Random Forest, Decision Tree, Gradient Boosting, K-Nearest Neighbors, Logistic Regression, and Multi-Layer Perceptron algorithms on the dataset CICIDS-2017. For feature-level interpretation of predictions, it features Explainable AI (XAI) using the SHAP technique to increase transparency and operator trust. The framework also features an adaptive defence engine of automated mitigation recommendations, real-time threat intelligence integration via IP geolocation and automated SOC reports generation. The experimental results show that the best accuracy and F1-score were obtained when using Random Forest with the accuracy of about 99.1% and F1 of 0.99. The proposed system offers a highly resilient detection-to-response mechanism to protect modern industrial environment.

Network Intrusion Detection, Industrial Control Systems, Machine Learning, SHAP, Adaptive Defence, CICIDS-2017