<p>The Industry 4.0 production needs predictive defective detection to enhance product quality, performance, and cost. Conventional reactive quality control is time-losing, wastsome, and has unforeseen downtimes. Gradient boosting machine learning, such as LightGBM, has provided manufacturers with the ability to generate high-quality, timely defects detection in real-time in Industry 4.0 automated and data-driven decision making. The paper aims to use optimised LightGBM-based predictive defect detection models on large scale production data and early defect detection. The proposed method makes use of an optimized LightGBM model, which is optimized by hyperparameter optimization methods to improve its performance in defect detection tasks. Preprocessing of data involved the treatment of missing data and feature scaling as well as balancing to make sure that the patterns are reflected in the model. Different LightGBM hyperparameters such as learning rate, number of leaves and feature fraction were optimized by the Crow search optimization. The fine-tuned LightGBM model, optimized with the Crow Search Optimization (CSO), had a high accuracy of 97.7 with a F1-score of 0.97 on the majority class and a F1-score of 0.82 on the minority class compared to other baseline models such as XGBoost, CatBoost, and Random Forest using real-world manufacturing defect data. The evaluation of the model was performed through cross validation on a set obtained through the use of the IoT sensors and quality inspection systems.</p> Graphical abstract <p></p>