Technical Summary: Neuro-symbolic AI for Predictive Maintenance (PdM)

Problem Statement

The paper addresses the limitations of current State-of-the-art (SOTA) approaches in Predictive Maintenance (PdM) within industrial settings (e.g., commercial buildings, pharmaceutical facilities, semiconductor manufacturing). The authors identify a dichotomy in existing modeling strategies:

1. Data-Driven Methods (Deep Learning): While exhibiting high accuracy, these systems suffer from a "black-box" nature (lack of transparency), a heavy reliance on large labeled datasets (which are scarce in industrial "run-to-failure" scenarios), and poor generalization to out-of-distribution environments.
2. Knowledge-Based/Physics-Based Methods: These approaches offer high explainability and incorporate domain expertise but often struggle with accuracy, generate false positives, and require continuous expert supervision and manual tuning.

While hybrid approaches (combining data-driven and knowledge-based models) are emerging, the authors argue they do not go far enough. The core problem is the lack of systems that deeply integrate deep learning with symbolic logic to create architectures that are simultaneously accurate, explainable, interpretable, and robust to data scarcity.

Methodology

The paper employs a Systematic Literature Review (SLR) following the guidelines of Kitchenham and Charters (2007).

• Search Strategy: A query was executed on the Scopus database in August 2025, covering English-language engineering and computer science publications from 2020 to 2025.
• Filtering: Initial results (9,005 records) were filtered by excluding review papers and non-relevant domains (e.g., sports, audio).
• Automated Extraction: A Large Language Model (LLM, OpenAI gpt-4.1-nano) was used to extract metadata from titles and abstracts, including domain, input/output types, modeling categories, techniques, and explainability status.
• Manual Refinement: Domain labels were manually cleaned and clustered using OpenRefine into 46 categories, reducing the final dataset to 3,495 relevant studies.
• Analysis: The authors synthesized descriptive statistics and analyzed the typology of modeling approaches, specifically focusing on the evolution toward hybrid and neuro-symbolic systems.

Key Contributions

The paper makes several distinct contributions to the PdM literature:

1. Generalized PdM Framework: The authors define a standardized five-stage lifecycle for PdM:

   • Data Acquisition and Processing.
   • Health Indicator (HI) Construction and Feature Engineering.
   • Health Stage Division and Fault Detection.
   • Prognostics and Remaining Useful Life (RUL) Prediction.
   • Decision Support and Action.

2. Typology of Modeling Approaches: The review categorizes current approaches into four distinct types, detailing their strengths and weaknesses:

   • Physics-Based: Relies on first principles (e.g., Kalman Filters, Bond Graphs). High interpretability but difficult to generalize.
   • Knowledge-Based: Uses rules, ontologies, Bayesian Networks, and Fault Trees. Transparent but brittle and hard to maintain.
   • Data-Driven: Utilizes statistical models, Machine Learning (ML), and Deep Learning (DL). High accuracy but data-hungry and opaque.
   • Hybrid: Combines the above. The authors note that while hybrid is a trend, true integration is often superficial.

3. Proposal of Neuro-symbolic AI (NESY) for PdM: The central contribution is the proposal to move beyond standard hybrid models to Neuro-symbolic AI, defined as the tight integration of deep learning with symbolic logic. The authors categorize NESY architectures into three types relevant to PdM:

   • Nested: A symbolic system containing neural modules (or vice versa), such as using a neural network for perception within a symbolic planner.
   • Cooperative: Multiple neural and symbolic modules learning iteratively from each other (e.g., Reinforcement Learning agents).
   • Compiled: Symbolic knowledge is fully integrated into the neural network structure or loss function. This includes:
     • Compiled into Nodes (NESY-CN): Logic Neural Networks (LNN), Equation Learners (EQL), and Hamiltonian Neural Networks (HNN) where rules or physics equations are embedded in the architecture.
     • Compiled into Loss (NESY-CL): Physics-Informed Neural Networks (PINNs) and Signal Temporal Logic (STL) constraints added as penalty terms in the loss function to enforce physical laws or temporal rules.
     • Compiled into Tensors (NESY-CT): Logic Tensor Networks (LTN) that encode symbolic relations into vector spaces.

4. Agentic AI Vision: The paper outlines a future direction where NESY enables "Agentic AI" in maintenance. These agents would not only predict failures but also perform root cause analysis, plan interventions, and coordinate with other systems while adhering to explicit safety and regulatory constraints encoded in symbolic logic.

Results and Findings

Based on the analysis of 3,495 studies, the authors report:

• Dominance of Data-Driven Approaches: The vast majority of recent literature focuses on data-driven methods, particularly Deep Learning (CNNs, LSTMs, Transformers).
• Rise of Hybrid Systems: There is a clear trend toward hybrid models, though many do not explicitly label themselves as such.
• Scarcity of NESY: Only a small fraction of the reviewed literature explicitly addresses Neuro-symbolic approaches for time-series PdM data.
• Persistent Challenges: The review confirms that data scarcity (lack of run-to-failure data), the "black-box" nature of DL models, and the difficulty of fusing heterogeneous data remain significant barriers to widespread adoption.
• NESY Potential: The authors argue that compiled NESY architectures specifically address these pain points by:
  • Reducing false positives through logical constraints (e.g., STL).
  • Improving generalization by embedding physical laws (e.g., PINNs, HNNs).
  • Providing inherent interpretability through rule-weighted reasoning and decision trees.

Significance and Claims

The paper claims that Neuro-symbolic AI is the central enabling technology for the next frontier in industrial maintenance: fully autonomous Agentic systems.

The authors assert that while current hybrid solutions combine models, NESY offers a deeper integration that allows systems to:

1. Bridge the Gap: Combine the predictive power of neural networks with the explicit reasoning and constraint handling of symbolic logic.
2. Ensure Trust and Compliance: Provide traceable rationales for decisions, satisfying regulatory, safety, and contractual requirements by showing which rules and sensor patterns led to a specific maintenance action.
3. Enable Autonomy: Move beyond simple prediction to systems that can diagnose root causes, generate auditable maintenance plans, and coordinate complex actions within operational workflows with minimal human oversight.

The paper concludes that adopting NESY is a necessary step to overcome the limitations of current SOTA approaches, specifically addressing data scarcity, the need for robust generalization, and the critical requirement for interpretability in high-stakes industrial environments.