Autonomous Industrial Intelligence via Multi-Modal Learning in CloudConnected Manufacturing Ecosystems

Abstract

The rapid evolution of Industrial Internet of Things (IoT) has transformed modern manufacturing by enabling continuous monitoring of machines through sensor-generated data. Traditionally, industrial systems relied on manual inspection and rule-based monitoring, where operators analyzed parameters such as temperature, vibration, and pressure to make decisions. However, these approaches often lacked real-time analytical capability and were unable to handle large-scale data effectively. As industrial environments became more complex, the need for intelligent data-driven systems emerged. Existing traditional systems suffer from limitations such as delayed fault detection, reactive maintenance strategies, poor scalability, and inability to capture hidden patterns in multidimensional data. To address these challenges, this study presents an advanced industrial analytics framework based on machine learning techniques integrated within a web-based environment. The proposed system utilizes a unified Two Classification and Two Regression Tasks (2CA2RT) framework to perform maintenance flag and production status prediction (classification), along with downtime and efficiency score prediction (regression). Multiple models including Passive Aggressive (PA), Extreme Gradient Boosting (XGB), Adaptive Boosting (AB), and a hybrid Neural Architecture Search–Greedy Rule Forest (NAS-GRF) are implemented to analyses industrial sensor data. The system incorporates data preprocessing, Exploratory Data Analysis (EDA), model training, evaluation, and prediction modules to ensure endto-end functionality. Experimental results demonstrate that the proposed NAS-GRF model achieves superior performance across all four target variables. For classification tasks, it achieves 100% accuracy (1.0000) in maintenance flag prediction and 100% accuracy (1.0000) in production status prediction.