Research Progress in Machine-Learning-Based Fault Diagnosis for Mechanical Equipment

Fault diagnosis of mechanical equipment is essential for ensuring safe and reliable operation and reducing losses caused by unplanned downtime. With the large-scale deployment of sensors and the Industrial Internet of Things, condition-monitoring data have grown rapidly in both volume and dimensionality. Conventional diagnostic approaches that rely on human expertise and signal processing are increasingly inadequate, while machine learning, especially deep learning, has become the dominant research paradigm because of its automatic feature extraction and powerful nonlinear modeling capabilities. Following a data-feature-model-application framework, this paper systematically reviews advances in machine-learning applications for mechanical equipment fault diagnosis. It first outlines the basic diagnostic workflow, commonly used monitoring signals, and representative public datasets. It then reviews diagnostic methods based on conventional machine learning, including feature engineering, support vector machines, and ensemble learning, as well as deep-learning methods based on convolutional neural networks, recurrent neural networks, autoencoders, attention mechanisms, and Transformers, comparing their principles, applicable conditions, strengths, and limitations. For real industrial environments, the paper focuses on emerging directions such as few-shot learning, transfer learning, multisensor fusion and graph neural networks, federated learning, and generative-adversarial-network-based data augmentation, and summarizes their applications to representative systems including rotating machinery, motors, computer numerical control machine tools, and pumps. Finally, current challenges are analyzed from the perspectives of data quality, model generalization, interpretability, and real-time edge deployment, and future trends involving physics-data integration, foundation models, and intelligent operation and maintenance are discussed. The analysis shows that the field is evolving from discriminative diagnosis dependent on large amounts of labeled data toward few-shot, cross-condition, interpretable, and deployable solutions. Data governance and engineering implementation will therefore become the principal bottlenecks constraining practical adoption.