A Risk-Aware Explainable Multi-Agent Decision Intelligence Framework for Human-Centric Predictive Maintenance in Smart Manufacturing

In the high-stakes world of commercial aviation, the safety of a flight depends on the silent, steady endurance of its engines. For decades, engineers have relied on predictive maintenance, a strategy that uses data to guess when a machine part might fail before it actually does. This approach moves beyond simply fixing things after they break or replacing them on a rigid schedule. Instead, it listens to the hum of the machinery, looking for subtle signs of wear that suggest a component is nearing the end of its life. The goal is to intervene at the perfect moment: not too early, which wastes resources, and not too late, which risks catastrophe. However, a significant gap has long existed between the raw numbers generated by these computer models and the actual decisions made by human maintenance crews. A computer might predict that an engine has a certain amount of life left, but it often cannot explain why it made that guess, how much it trusts that guess, or what specific risks are involved if the crew decides to wait. Without this context, the most accurate prediction remains just a number, difficult to act upon with confidence.

A team of researchers at MS Ramaiah University of Applied Sciences in India has developed a new system designed to bridge this gap, transforming raw data into a clear, trustworthy story for human decision-makers. They call their creation the Risk-Aware Explainable Agentic Prognostic Intelligence Framework. Rather than simply outputting a single estimate of how long an engine will last, this system acts like a team of specialized digital assistants, each with a specific job. One part of the system watches the engine's sensors to predict its health and remaining life. Another part acts as a skeptic, constantly checking how confident the first part should be in its own predictions. A third part evaluates the danger of the situation, while a fourth explains exactly which sensor readings led to the conclusion. Finally, a decision-making agent brings all these perspectives together to suggest a concrete action, such as continuing to fly, monitoring the engine closely, or scheduling an immediate repair. The researchers tested this framework using a well-known dataset of simulated turbofan engine failures, a standard benchmark in the field, to see if it could turn complex data into reliable, human-readable advice.

The core of this new framework is a deep learning model that learns from the history of engine degradation. It looks at a stream of data from seventeen different sensors, tracking how the engine behaves over time. Instead of just guessing a number, the model learns to recognize patterns of wear and tear, predicting both a health score and the remaining useful life of the engine. But the innovation lies in what happens after that prediction is made. The system does not stop there. It immediately runs the prediction through a reliability check, using a method that compares multiple independent models to see if they agree. If the models disagree or if the data is noisy, the system flags the prediction as uncertain. This is crucial because in aviation, a wrong guess can be dangerous. The system calculates a confidence score, and if that score drops too low, it automatically triggers a human review, ensuring that no unsafe automated decision is ever made without a human expert looking over the shoulder.

Once the system is confident in its numbers, it moves to the next layer: risk and explanation. A specialized agent uses fuzzy logic, a way of reasoning that handles shades of gray rather than just black and white, to assess the operational risk. It considers the engine's health, its remaining life, the confidence in the prediction, and how stable the engine's behavior has been. This agent categorizes the situation into levels like low, medium, high, or critical risk. Simultaneously, an explainability agent works to make the invisible visible. It uses advanced techniques to highlight which specific moments in the engine's history and which specific sensor readings were most important in reaching the conclusion. For instance, it might point out that a specific sensor reading from the last few cycles was the primary reason for a high-risk warning. This transforms the "black box" of artificial intelligence into a transparent process where a human engineer can see the evidence and understand the reasoning behind the recommendation.

The final step in this intelligent workflow is the decision-making agent. It takes the health scores, the risk levels, the confidence metrics, and the explanations, and weighs them all together to suggest a course of action. The system does not just say "the engine is sick"; it says "the engine is sick, we are 82 percent sure of this, the risk is high, and here is the specific sensor that is worrying us, so you should schedule maintenance." In their tests, the framework demonstrated high accuracy, with very small errors in its predictions of both engine health and remaining life. More importantly, it successfully categorized the risk of engine failure and generated maintenance recommendations that aligned with the severity of the situation. In one specific case study involving an engine that was near failure, the system correctly identified the critical risk level and, noticing that its own confidence was lower than usual, wisely recommended a human review rather than an automated fix. This ability to recognize its own uncertainty and defer to human judgment is a key feature of the design.

The researchers found that their approach offers a significant step forward from traditional methods that focus solely on prediction accuracy. By integrating reliability checks, risk assessment, and clear explanations into a single system, they created a tool that is not just a calculator but a partner in decision-making. The system proved capable of handling the complexity of engine degradation data and translating it into actionable intelligence. While the study was conducted using simulated data from a standard benchmark, the results suggest that such a framework could be highly valuable in real-world aerospace maintenance. It addresses the human need for trust and understanding, ensuring that when a machine says an engine needs attention, the people responsible for the aircraft know exactly why and how much they can rely on that warning. The work highlights a shift in how artificial intelligence is applied to safety-critical industries, moving from simple prediction to a more holistic, human-centric approach where the machine helps the human make the right choice.