Multi-Source Heterogeneous Data Fusion in Smart Industrial Systems: Intelligent Monitoring, Predictive Quality Assessment, and Fault Diagnosis for Safer Operations

Abstract

The rapid growth of connected devices, intelligent sensors, and digital record-keeping systems has made modern industrial facilities simultaneously data-rich yet insight-poor. At the heart of solving this paradox is Multi-Source Heterogeneous Data Fusion (MSHDF) the systematic integration of streams of data, which are dissimilar in format, sampling rate, dimensionality, and physical meaning, into one, coherent, and practical layer of knowledge. The current review paper is an in-depth analysis of MSHDF in three high-impact application domains: (1) Intelligent Process Monitoring, which combines real-time structured sensor data with unstructured operator data records to detect operational anomalies earlier and more reliably than single-source methods; (2) Predictive Quality Assessment, which integrates spectroscopic measurement data with operational condition data record to predict final product or output quality without costly offline testing; and (3) Fault Diagnosis and Safety, which leverages uncertainty quantification and explainability approaches in data fusion pipelines to build trustworthy and human-interpretable safety monitoring systems. The review compares the data-level, feature-level and decision-level of fusion architectures and surveys core algorithms such as Bayesian networks, deep multimodal networks, graph neural networks and natural language processing models. Among the major issues, the heterogeneity of data, the interpretability of models, sensor drift, and limited fault-labelled data are discussed. Future opportunities such as digital twins, physics-informed neural networks, federated learning, and more are discussed. In this paper, it was concluded that MSHDF is a radical paradigm shift in operational intelligence that can be applied in manufacturing, energy, transportation, and infrastructure fields.