Meta-Learning for Rapid Model Adaptation in Dynamic Environments

Abstract

The significance of Industrial fault detection has become crucial in the current smart manufacturing environment considering that the systems operate under constantly changing circumstances and labeled data is limited. Conventional methods involving Machine Learning (ML) techniques and the Deep Learning (DL) models face problems like the poor adaptability, generalization, and are slow convergence when detecting complex patterns of industrial faults. To solve these challenges, this research presents an approach to meta-learning which involves Scalable Learning Memory Network (SL-MN) for effective Industrial fault prediction in dynamic environments characterized by limited datasets. SL-MN incorporates MANN and Scalable Learning Optimization (SLO) for improved fault adaptation capability, feature representation and stability in learning process. A fault detection dataset sourced from the industrial IoT environment, involving about 6,000 observations obtained from Kaggle is used for experimental analysis at 80:20 for the purpose of training and testing. Data preparation processes involve normalization through Min–max Scaling and missing value estimation using Cubic Spline Interpolation (CSI) techniques. Feature selection and reduction is done using PCA. Results from the experiments reveals that the Introduced SL-MN model delievers the outstanding performance with the precision of 0.9375, recall of 0.9346, F1-score of 0.9317, and accuracy of 0.9370, outperforming conventional machine learning models. The model is implemented using Python version 3.8, with the model that is being used including PyTorch, NumPy, Pandas, SciPy, and the Scikit-learn. Overall, the proposed model provides an effective solution for reliable industrial fault prediction in dynamic environments.