10^th Intelligent Maintenance Conference

Artificial intelligence and data science are changing the maintenance of railway infrastructure. Instead of manual track inspections and isolated data sources, AI powered methods are supporting step by step automatic deviation and failure detection, assessment of condition data, and consistent quality assurance across the entire value chain. Modern analytical techniques turn large volumes of data from sensors, diagnostic vehicles, and digital models into actionable information.

The open data format RCM DX also creates a shared digital foundation that makes data consistently usable for the first time—regardless of systems, applications, or vendors. Complementary AI technologies such as digital structure gauge analysis, object detection, and sensor fusion open new perspectives on the condition of the tracks and support a holistic understanding of railway infrastructure.

This brings us closer to the shift from reactive to predictive maintenance: deviations are detected earlier, resources are used more purposefully – safety and efficiency are increased. AI and data science are thus key enablers for an efficient, scalable, and future oriented railway maintenance.

Maintenance has transformed from a "necessary evil" to a critical industry priority, driven by economic awareness and technological advances. In our view,this evolution spans three distinct periods: before 2000; from 2000 to 2017, and from 2017 to now. Historical Perspective

Before 2000: foundation years. After “ corrective-only” maintenance and then scheduled maintenance based on “equipment potential”, Reliability-Centred maintenance has emerged, as a rational way of selecting maintenance strategy on the basis of functional analysis and requirements specifications. In parallel, a number of condition-monitoring techniques have enabled the “ condition-based maintenance” concept.

From 2000 to 2017: the PHM Revolution. “Prognostics & Health Management” and predictive maintenance emerged first in the aerospace industry, then in some other sectors .This period is also when giant steps took place in machine learning and its epoch-making applications to image processing (Image Net) and natural language processing . In parallel, a number of physics-based degradation models started to be applied. The first PHM standard ( IEEE-std -1856) was published in 2017.

From 2017 to now: towards more intelligence .A 2020 paper in Engineering Applications of AI (“Potential, Challenges and Future Directions for Deep Learning in PHM Applications”, O.Fink et al.) described the successes of deep learning in the worlds of images and words, and highlighted the perspectives and challenges that exist in the world of “things” , such as maintenance .Quite a few of those challenges have now been successfully addressed already, at least in part. Enthusiasm for “big data” and purely data-driven concepts started to give way to hybrid approaches : physics-informed machine learning, and also reliability-informed deep learning. With Chat-GPT ( November 2022) and its competitors, the concept of Large Language Models (LLM) has entered our lives - and in particular the maintenance field. Key strides have been made in addressing explainability, causal inference, trustworthy AI for PHM, uncertainty quantification, and have been presented in successive editions of this Intelligent Maintenance Conference. Last but not least, the vision of the role of humans in the maintenance process has evolved ,as evidenced by the transition from Industry 4.0 to Industry 5.0, and now 6.0. Cooperation between humans and intelligent machines seems to be the trend.

Industry adoption of PHM has progressed at very unequal speeds in different fields, and, in many cases, much remains to be done before it becomes 'mainstream'.

Future Outlook

Key remaining challenges include PHM for safety-critical systems, designing PHM for complex systems and systems of systems, merging PHM and RAMS, dealing with autonomous systems, and health-aware control. Undoubtedly, those challenges will drive future IMC discussions and industry innovations.

Industrial systems across land, air, and sea rely on safety-critical assemblies, such as engines, fuel tanks, and pipelines, that require periodic inspection to detect issues before they necessitate costly, unscheduled repairs or service removals. However, performing in-situ maintenance is exceptionally challenging because these components are often crammed into confined, high-complexity environments with limited access. A continuum robot is an ultra-flexible robotic structure that uses internal actuation (like cables or concentric tubes) to achieve continuous, snake-like bending, allowing it to navigate complex, cramped spaces that traditional rigid-joint robots cannot reach. Recently, several academic breakthroughs have successfully transitioned into commercial products. This presentation explores the latest advancements in snake/continuum robotics for an entirely new application: the "healthcare" of industrial systems. I will provide a brief overview of the evolution of continuum robots, followed by the specific research challenges and solutions involved in deploying these systems across the aerospace and nuclear sectors.

As a cornerstone of the Plants of Tomorrow initiative, Holcim is transitioning from localized AI pilots to a continental-scale predictive maintenance ecosystem. While a Proof-of-Concept (PoC) proves an algorithm, scaling to 1,000+ critical assets—including high-torque gearboxes, vertical roller mills (VRMs), and kilns—requires solving the "Industrial AI Paradox": maintaining high precision while managing massive environmental and operational variance across 100+ global sites. This technical presentation breaks down the architecture and deployment of C3 AI Reliability on a global scale. We move beyond theoretical data science to explore the systems engineering required to achieve operational resilience:

Hybrid Edge-to-Cloud Topology: A deep dive into deploying AI on resource-constrained, semi-air-gapped edge devices. We discuss the optimization of Kubernetes clusters at the plant level to ensure real-time inferencing and high availability, even during network latency or outages.

Heterogeneous Asset Modeling: Strategies for managing "Model Drift" across diverse manufacturing environments. We analyze how we leverage "Golden Models" while incorporating unit-specific baselines to account for varying loads, speeds, and ambient clinker temperatures.

Automated Feature Engineering & Signal Processing: How the pipeline automates the transformation of raw high-frequency vibration and thermal telemetry into actionable "Probability of Failure" (PoF) metrics across thousands of concurrent streams.

Closing the Loop with Ground Truth: The technical integration between AI alerts and the Technical Information System (TIS), ensuring that maintenance feedback from the floor directly refines the supervised learning sets for improved precision.

Attendees will leave with a technical blueprint of how Holcim is moving toward a self-healing plant, shifting maintenance from a reactive cost center to a competitive, data-driven weapon in the global building solutions market.

As reducing maintenance costs and enhancing operational efficiency become increasingly critical, remote monitoring with automated analytics has emerged as a key strategy for managing turbomachinery installations. While the use of complex physics-based algorithms like thermodynamic models can be beneficial for fleets with machines of similar types and instrumentation, this becomes increasingly difficult for fleets comprising multiple machine types, such as compressors and turbines, operating under diverse conditions, configurations, ages, and service histories. To address this, an end-to-end data processing chain is presented that transforms raw sensor signals into actionable diagnostic insights. This spans from sensor measurements through analytics to decision support. As specialized approaches only have a limited impact on the overall fleet, the first step is to reliably detect anomalies relative to expected behavior. For that, a combination of semi-supervised, ML-based anomaly detection and physicsinformed regression algorithms is employed. The regression algorithms predict important KPIs like efficiency, depending on the operating condition and comparing them to target values. The output of these algorithms is then fed into a fuzzy logic framework that is based on formalized knowledge mostly collected via expert workshops. This enables the integration of insights from root-cause analyses into an automated data flow, creating actionable advice for detected anomalies. The presentation also emphasizes how well-suited different analytics solutions are for diverse machine fleets and how large language models can be leveraged to accelerate the knowledge-collection process by giving suggestions based on documentation and service reports.

Root cause analysis (RCA) in industrial systems aims to identify the underlying physical or operational cause of abnormal behavior. In complex systems with inter-variable dependencies and a control policy, RCA is often slow, costly, and heavily dependent on expert knowledge, particularly under varying or unseen operating conditions. In practice, traditional RCA relies on alarm logs, historical reports, and physical system inspection, which limits scalability and generalization to new operating conditions or unseen faults. To address those issues, we propose an agentic framework that combines system-specific diagnostic tools with large language models (LLMs) for interpretable root cause analysis. Our framework decomposes RCA into two components: (i) relationship-based anomaly surrogates trained on normal process data to model global variable dependencies and detect deviations, and (ii) an LLM-based reasoning agent that iteratively queries these tools to gather evidence, construct causal explanations, and rank fault hypotheses. We evaluate the approach on both synthetic and real-world industrial process datasets using steady-state process measurements. This work outlines a scalable and interpretable pathway toward hybrid AI systems for industrial diagnosis.