[Paper Review] Deep learning models for predictive maintenance: a survey, comparison, challenges and prospect

Oscar Serradilla, Ekhi Zugasti+1 more|arXiv (Cornell University)|Oct 7, 2020

Industrial Vision Systems and Defect DetectionEngineering195 references17 citations

TL;DR

This survey reviews state-of-the-art deep learning models for predictive maintenance, categorizing architectures by technique, data type, and application to anomaly detection, root cause analysis, and remaining useful life estimation. It identifies key challenges like interpretability, real-time inference, and novelty detection in industrial settings, and proposes integrating domain knowledge and explainable AI to bridge the gap between lab performance and real-world deployment.

ABSTRACT

Given the growing amount of industrial data spaces worldwide, deep learning solutions have become popular for predictive maintenance, which monitor assets to optimise maintenance tasks. Choosing the most suitable architecture for each use-case is complex given the number of examples found in literature. This work aims at facilitating this task by reviewing state-of-the-art deep learning architectures, and how they integrate with predictive maintenance stages to meet industrial companies' requirements (i.e. anomaly detection, root cause analysis, remaining useful life estimation). They are categorised and compared in industrial applications, explaining how to fill their gaps. Finally, open challenges and future research paths are presented.

Motivation & Objective

• To provide a comprehensive review of deep learning architectures applicable to predictive maintenance in industrial settings.
• To compare the performance and suitability of different deep learning models across key PdM tasks such as anomaly detection and remaining useful life estimation.
• To identify critical gaps in current research, particularly in interpretability, real-time processing, and handling of unsupervised or one-class data.
• To propose a hybrid approach combining domain expertise with data-driven deep learning to improve model trustworthiness and industrial applicability.
• To highlight the need for explainable AI (XAI) techniques to bridge the gap between high-accuracy black-box models and industrial requirements for transparency and validation.

Proposed method

• Systematically reviewed 200+ publications from Scopus, Engineering Village, and Google Scholar to identify state-of-the-art deep learning models in predictive maintenance.
• Classified deep learning models by underlying architecture (e.g., autoencoders, CNNs, RNNs/LSTMs, DBNs), data type (time-series, sensor data), and application (anomaly detection, RUL estimation).
• Mapped the predictive maintenance pipeline into distinct stages—data collection, preprocessing, model training, inference, and decision-making—assessing how models integrate at each stage.
• Conducted a comparative analysis of SotA models on benchmark datasets, particularly the turbofan engine degradation dataset, to evaluate performance across architectures.
• Evaluated the suitability of models for real industrial deployment, focusing on streaming data processing, low-latency inference, and handling of non-failure data.
• Proposed a hybrid modeling framework integrating domain knowledge with deep learning, emphasizing the role of explainable AI (XAI) to enhance model interpretability and trust.

Experimental results

Research questions

• RQ1 Which deep learning architectures are most effective for predictive maintenance tasks such as anomaly detection and remaining useful life estimation?
• RQ2 How do data-driven deep learning models compare to traditional statistical and machine learning methods in terms of accuracy and industrial applicability?
• RQ3 What are the key limitations of current deep learning models in real-world industrial settings, particularly regarding interpretability and real-time deployment?
• RQ4 How can domain knowledge and explainable AI (XAI) techniques be integrated to improve model transparency and trust in industrial maintenance systems?
• RQ5 Why do existing unsupervised and one-class classification models fail to link detected failures to physical root causes, and how can this gap be addressed?

Key findings

• Deep learning models achieve state-of-the-art performance in predictive maintenance tasks when trained on sufficient data, particularly in time-series modeling using LSTMs, CNNs, and autoencoders.
• Despite high accuracy, most models fail to address critical industrial requirements such as real-time inference, uncertainty quantification, and interpretability.
• Unsupervised and one-class learning models, such as autoencoders and deep belief networks, are widely used due to the scarcity of failure data in industrial environments.
• A significant gap exists between lab-based SotA models and industrial deployment, primarily due to lack of integration with domain expertise and insufficient interpretability.
• Explainable AI (XAI) techniques are underexplored in predictive maintenance but represent a promising path to bridge the gap between high-accuracy black-box models and industrial trust.
• The integration of domain knowledge with data-driven models—especially through hybrid frameworks—can enhance model reliability, explainability, and practical adoption in real-world maintenance systems.