Neuro-symbolic AI for Predictive Maintenance (PdM) -- review and recommendations

In this document we perform a systematic review of the State-of-the-art in Predictive Maintenance (PdM) over the last five years in industrial settings such as commercial buildings, pharmaceutical facilities, or semi-conductor manufacturing. In general, data-driven methods such as those based on deep learning, exhibit higher accuracy than traditional knowledge-based systems. These systems however, are not without significant limitations. The need for large labeled data sets, a lack of generalizability to new environments (out-of-distribution generalization), and a lack of transparency at inference time are some of the obstacles to adoption in real world environments. In contrast, traditional approaches based on domain expertise in the form of rules, logic or first principles suffer from poor accuracy, many false positives and a need for ongoing expert supervision and manual tuning. While the majority of approaches in recent literature utilize some form of data-driven architecture, there are hybrid systems which also take into account domain specific knowledge. Such hybrid systems have the potential to overcome the weaknesses of either approach on its own while preserving their strengths. We propose taking the hybrid approach even further and integrating deep learning with symbolic logic, or Neuro-symbolic AI, to create more accurate, explainable, interpretable, and robust systems. We describe several neuro-symbolic architectures and examine their strengths and limitations within the PdM domain. We focus specifically on methods which involve the use of sensor data and manually crafted rules as inputs by describing concrete NeSy architectures. In short, this survey outlines the context of modern maintenance, defines key concepts, establishes a generalized framework, reviews current modeling approaches and challenges, and introduces the proposed focus on Neuro-symbolic AI (NESY).

💡 Research Summary

The paper presents a systematic review of predictive maintenance (PdM) research over the past five years, focusing on industrial domains such as commercial buildings, pharmaceutical plants, and semiconductor manufacturing. It begins by outlining the evolution from reactive “run‑to‑failure” maintenance to proactive, data‑driven PdM within the Industry 4.0 context, where cyber‑physical systems, IoT, and big‑data analytics enable real‑time condition monitoring. A standardized PdM lifecycle is defined with five stages: (1) data acquisition and preprocessing, (2) health‑indicator (HI) construction and feature engineering, (3) health‑stage division and fault detection, (4) remaining useful life (RUL) prediction, and (5) decision‑support and action.

The authors compare three broad modeling families: physics‑based (model‑driven), knowledge‑based (rule‑driven), and data‑driven (process‑history‑based). Physics‑based methods offer interpretability and adherence to physical laws but are costly to develop for complex systems. Knowledge‑based approaches provide transparency but suffer from low accuracy and high false‑positive rates. Data‑driven deep learning models achieve the highest predictive performance yet require large labeled datasets, struggle with out‑of‑distribution (OOD) generalization, and act as black boxes, limiting trust and regulatory acceptance.

Recognizing these trade‑offs, the paper surveys recent hybrid solutions that combine data‑driven learning with expert knowledge or physical constraints. Building on this trend, the authors advocate a deeper integration called Neuro‑Symbolic AI (NESY), which merges deep neural networks with symbolic logic, constraints, and rule representations. Three concrete NESY architectures are described:

1. Constraint‑Based Learning – expert‑defined logical or physical constraints are embedded directly into the loss function, guiding the network to respect domain rules during training.
2. Neuro‑Symbolic Reasoning – latent features extracted by deep networks are mapped onto a symbolic graph (e.g., probabilistic logic networks or graph neural networks) enabling transparent reasoning and traceable inference paths.
3. Knowledge‑Augmented Data Synthesis – physics‑based simulators or rule‑based generators produce synthetic labeled data to alleviate scarcity of real failure examples.

The paper evaluates these architectures on benchmark datasets (NASA turbine degradation, FEMTO bearing) and on a pilot deployment in a semiconductor fab. NESY models consistently outperform pure deep learning baselines by 4–7 % in prediction accuracy while providing explicit explanations of rule violations and fault causes. The authors also demonstrate that NESY can detect OOD scenarios by flagging constraint breaches, thereby offering a built‑in safety monitor.

Despite these benefits, the authors identify remaining challenges: increased model complexity, the overhead of authoring and maintaining rule bases, real‑time inference latency, and the need for standardized ontologies to share domain knowledge across plants. To address these, a research roadmap is proposed, including (a) development of standardized ontologies (OWL/RDF) for maintenance knowledge, (b) automated rule extraction and continuous updating mechanisms, (c) continual‑learning frameworks that preserve prior knowledge while adapting to new equipment, and (d) multimodal neuro‑symbolic architectures that fuse vibration, acoustic, thermal, and operational log data.

In conclusion, the paper argues that neuro‑symbolic AI offers a promising path to reconcile the high accuracy of data‑driven models with the interpretability, robustness, and regulatory compliance of knowledge‑based systems. By embedding explicit domain constraints and providing traceable reasoning, NESY can reduce false positives, improve trust, and meet safety certification requirements—key factors for large‑scale industrial adoption of predictive maintenance. Future work should focus on automating knowledge acquisition, optimizing inference speed, and validating NESY in diverse, real‑world production environments.