💻 computer science

A SHAP-Guided Retrieval-Augmented Generation Framework for Predictive Maintenance Decision Support

This paper proposes a proof-of-concept framework that integrates SHAP-based model explanations with Retrieval-Augmented Generation (RAG) to transform predictive maintenance failure predictions into actionable, interpretable, and case-specific maintenance recommendations by retrieving relevant knowledge based on dominant failure factors.

Original authors: AHMED AIT HADJ IAICH, Smail ZAKI, Mohamed ABOUSSALEH

Published 2026-10-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: AHMED AIT HADJ IAICH, Smail ZAKI, Mohamed ABOUSSALEH

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the humming heart of a modern factory, machines run on a delicate balance of temperature, speed, and pressure. When these systems work, they keep the world moving, but when they fail, the cost is measured in broken production lines and expensive downtime. For decades, engineers have tried to predict these failures before they happen, using computers to scan sensor data for early warning signs. While these computer models have become very good at spotting trouble, they often act like black boxes: they can tell a human operator that a machine is about to break, but they cannot easily explain why or what to do about it. This lack of clarity creates a barrier; a factory manager cannot fix a machine based on a warning they do not understand. To bridge this gap, researchers are now exploring ways to combine the raw predictive power of machine learning with the ability to search through vast libraries of technical manuals and expert knowledge, turning a simple alarm into a clear, actionable instruction.

A team of researchers from Morocco has built a new system designed to solve this exact problem. They created a framework that takes a prediction from a computer model, asks the model to explain its reasoning, and then uses that explanation to search a digital library of maintenance guides. The result is a system that does not just say "failure is coming," but instead says "failure is coming because the tool is worn and the torque is too high, so you should check the bearings and reduce the load." The researchers tested their idea using a simulated dataset of ten thousand machine operations, a collection of data that mimics the real-world mix of normal running and rare breakdowns. Their goal was not just to build a better predictor, but to create a complete chain of logic that connects a raw number from a sensor to a specific, human-readable recommendation.

The process begins with a standard computer model trained to watch the machine's vital signs. In this study, the researchers used a method called a Random Forest, which works by building many decision trees to look for patterns in the data. They found that this approach was particularly good at catching rare failures, identifying them correctly nearly ninety percent of the time. However, the real innovation happens after the model makes its prediction. Instead of stopping there, the system asks the model to break down its decision, identifying which specific factors—such as high temperature or excessive tool wear—pushed the machine toward failure. This step is crucial because it translates a complex mathematical calculation into a list of physical causes.

Once the system knows the cause, it does not just display a chart for an engineer to interpret. Instead, it converts those causes into a natural language question. If the model identifies that high torque and worn tools are the main drivers of risk, the system formulates a query like "high torque combined with advanced tool wear." This query is then sent to a search engine that looks through a specialized database of maintenance documents. These documents are not generic web pages; they are structured guides based on international engineering standards, describing specific failure modes and the exact steps needed to fix them. The system uses a technique to find the documents that best match the meaning of the query, rather than just matching keywords.

The final step brings everything together. The system takes the original prediction, the list of causes, and the retrieved maintenance documents, and feeds them into a large language model. This model acts as a translator, synthesizing the information into a clear, step-by-step recommendation. For a case of tool wear, the output might instruct the operator to inspect the cutting edge and replace the tool if it exceeds a certain limit. For a case of mechanical overstrain, it might suggest reducing the machine's load and checking the bearings. The researchers tested this entire pipeline and found that it worked as intended. When they asked human experts to review the results, the experts agreed that the retrieved documents were relevant to the predicted failures in eighty-one percent of the cases. Furthermore, the system was smart enough to skip the search entirely when the machine was running normally, avoiding unnecessary noise.

The study confirms that it is possible to link the mathematical logic of a predictive model directly to the practical knowledge of a maintenance manual. The researchers emphasize that this is a proof of concept, a demonstration that the architecture works in a controlled environment using simulated data. They acknowledge that real factories are messier than their test dataset, with more missing data and more complex machine behaviors. However, the core finding is robust: by using the model's own explanation to drive the search for solutions, the system creates a direct line from data to decision. This approach suggests a future where industrial AI does not just act as a silent alarm, but as a knowledgeable partner that helps human workers understand the "why" behind a failure and guides them toward the right fix. The work remains a blueprint for a more transparent and useful type of industrial intelligence, one that respects the complexity of the machines it monitors and the people who keep them running.

Technical Summary: A SHAP-Guided Retrieval-Augmented Generation Framework for Predictive Maintenance Decision Support

Problem Statement
While predictive maintenance (PdM) models can achieve high detection performance, their industrial deployment is hindered by a lack of interpretability and actionable guidance. Existing approaches often fail to establish a direct link between failure predictions, model explanations, and maintenance-oriented recommendations. In high-stakes industrial environments, "black-box" outputs are insufficient for decision-making; engineers require technically justifiable predictions that connect to known degradation mechanisms. Furthermore, current Explainable AI (XAI) methods, such as SHAP and LIME, typically focus on feature attribution and visualization but do not systematically translate these attributions into maintenance-relevant interpretations or connect them to domain knowledge bases. This gap limits the operationalization of PdM models, as predictions remain difficult to trust and translate into specific maintenance actions.

Methodology
The authors propose a proof-of-concept framework that integrates predictive modeling, SHAP-based explainability, and Retrieval-Augmented Generation (RAG) to bridge the gap between numerical predictions and maintenance knowledge. The framework operates through a sequential pipeline:

Predictive Modeling: A supervised machine learning model (specifically Random Forest and Gradient Boosting) is trained on tabular sensor data (the AI4I 2020 dataset) to predict machine failures. The model outputs a binary decision and a failure probability.
SHAP-Based Explainability: SHAP (SHapley Additive exPlanations) values are computed at the instance level to decompose predictions into feature-wise contributions. Instead of using these solely for visualization, the top- $k$ features (ranked by absolute SHAP values) are selected to identify dominant drivers of the prediction.
SHAP-Guided Query Construction: This is the core innovation. The selected features and their contribution directions (increasing or reducing failure risk) are transformed into structured natural language queries. A semantic mapping function converts numerical features (e.g., torque, tool wear) into physical degradation descriptors (e.g., "high torque increasing failure risk"). A rule-based inference step assigns a high-level failure hypothesis (e.g., "Tool Wear Failure" or "Overstrain Failure") to the query.
Retrieval-Augmented Generation (RAG):
- Retrieval: The structured query is embedded using a sentence transformer (all-MiniLM-L6-v2) and used to search a vector database containing a maintenance knowledge corpus. The corpus consists of documents describing failure mechanisms and recommended actions, derived from international standards (ISO, DIN).
- Generation: A Large Language Model (GPT-4o) consumes the predictive output, the SHAP-derived signals, and the retrieved documents to generate case-specific maintenance recommendations. The generation is conditioned on the retrieved content to ensure alignment with domain knowledge and limit unsupported outputs.
- Post-Processing: A lightweight rule-based layer formats the output into a standardized structure including failure risk assessment, probable mechanism, and prioritized actions.

Key Contributions
The paper outlines four primary contributions:

Formalization of SHAP-Guided Query Construction: A method to convert tabular feature attributions into structured, semantic retrieval queries that align with maintenance reasoning rather than raw numerical values.
Retrieval Mechanism for PdM: The introduction of a retrieval-based mechanism that links predictive model outputs and SHAP explanations directly to relevant maintenance documentation.
End-to-End RAG Pipeline: The development of a complete decision-support pipeline connecting predictive outputs, model explanations, retrieval, and LLM-based generation specifically adapted for tabular PdM data.
Evaluation Framework: An evaluation of the framework using both predictive performance metrics and retrieval relevance assessments, including independent expert annotation.

Results
The framework was evaluated on the AI4I 2020 benchmark dataset, which exhibits significant class imbalance (3.4% failure rate).

Predictive Performance: The Random Forest model achieved a recall of 0.898 and an F1-score of 0.85, significantly outperforming the baseline Bagged Trees model (recall 0.708) in detecting rare failure events. Gradient Boosting showed higher precision (0.89) but lower recall. Random Forest was selected for the pipeline due to its superior failure detection capability.
Retrieval Performance: The SHAP-guided retrieval module demonstrated strong performance. Internal consistency metrics showed an Average Similarity@3 of 0.71 and a Hit Rate@3 of 0.87.
Expert Validation: An independent evaluation by six domain experts on a stratified sample of query-document pairs yielded a "fully relevant" retrieval rate of 81.0% and a mean inter-annotator agreement (Cohen's $\kappa$ ) of 0.709. All experts correctly rejected distractor pairs (Distractor Rejection Rate = 1.0), confirming the system's ability to distinguish relevant maintenance knowledge from irrelevant content.

Significance and Claims
The authors position this work as a proof-of-concept rather than a fully validated industrial decision-support system. The primary significance lies in demonstrating the feasibility of integrating predictive modeling, SHAP-based explanation, retrieval, and LLM-based generation for maintenance-oriented interpretation.

The paper claims that the main barrier to deployable explainable PdM is not merely predictive accuracy, but the translation of model outputs into operational guidance. By converting numerical SHAP explanations into retrieval queries, the framework provides an architectural blueprint for connecting sensor-based predictions to domain knowledge. The results suggest that SHAP-guided retrieval can produce semantically coherent and practically relevant outputs under controlled experimental conditions. However, the authors explicitly note that the framework relies on a synthetic dataset and a limited maintenance corpus; future work requires validation on real industrial data with authentic variability and expanded expert-annotated corpora.

Technical Summary: A SHAP-Guided Retrieval-Augmented Generation Framework for Predictive Maintenance Decision Support

More like this