9^th Intelligent Maintenance Conference

In the dynamic world of aircraft and engine maintenance, predictive analytics and artificial intelligence (AI) are transforming how we manage and optimize our MRO services. This presentation will explore the innovative applications of these technologies at KLM, with specific focus on the KLM Engine Service. We will cover predictive health monitoring, prescriptive analytics, and AI in automation. Predictive health monitoring benefits KLM and its clients by using PROGNOS, which leverages big data and advanced analytics to monitor aircraft systems and engine components. This provides early warnings of potential issues to prevent AOG situations, flight delays, and cancellations, thus improving airworthiness and aircraft availability. In prescriptive analytics, we are designing tools for maintenance planning and resource optimization to improve visit schedules and manage aircraft and engine lease contracts. Phaser, an optimizer recently developed at KLM, generates efficient shop visits and phase-out plans for engines, thereby maximizing utilization and reducing costs. It is effectively overseeing the phase-out of KLM's 737 aircraft, ensuring a seamless and cost-effective transition. We will also cover how GenAI is automating engine income inspections in our shop. Using speech-to-text and a smart voice bot, we streamline administrative tasks, improve data quality, increase mechanics' hands-on time, and reduce non-performance costs.

In industrial environments, efficient and reliable diagnosis of bearing failures is critical for predictive maintenance strategies. This talk presents a computer vision deep learning tool designed for the automatic recognition of bearing failure modes, building upon our previously developed framework. The system has been expanded to incorporate state-of-the-art architectures, including ResNet and Vision Transformer (ViT) models, significantly enhancing classification accuracy and robustness on an industrial sourced bearing damage dataset. Beyond model development, the talk will explore the practical challenges of image data management in industrial contexts—such as data labelling, quality control, and integration with maintenance systems. Emphasis will be placed on how scalable data workflows, combined with powerful deep learning models, can transform condition monitoring and failure analysis processes. The session aims to provide both a technical deep dive and practical insights for applying AI solutions effectively in real-world industrial environments.

Modern infrastructure systems such as district heating networks, water systems, and structural assets are becoming increasingly complex, distributed, and sensor-enabled. However, dense physical sensing remains costly and often infeasible, motivating the need for computational sensing approaches that reduce cost while maintaining system observability. These methods aim to infer unmeasured quantities from limited sensor data by exploiting the underlying structure and correlations in the system. This talk explores how graph-based methods offer powerful tools to model, infer, and optimize sensing in such spatially distributed systems. We first introduce the concept of soft sensing in district heating networks, where temperature or pressure distributions across the network must be estimated with limited measurements. We show how infrastructure systems can be naturally represented as graphs, where measurements evolve as signals over nodes and edges. We also demonstrate how physics and domain knowledge can be incorporated into graph neural network-based soft sensing models to improve accuracy and interpretability. In addition, we address the challenge of optimal sensor placement—deciding where sensors should be located to maximize inference accuracy—by combining spectral graph theory with learning-based strategies.

The aviation industry is increasingly challenged to improve operational efficiency and reduce unscheduled downtime while ensuring safety and regulatory compliance. This presentation introduces a hybrid maintenance framework that integrates reliability knowledge and flight data, into operational decision-making. Bridging two research efforts — one focused on uncertainty-aware and interpretable Remaining Useful Life (RUL) prediction, and the other on integrated optimization of tail assignment, maintenance scheduling, and uncertainty management — we demonstrate how a combined strategy can support the transition from reactive and preventive maintenance to predictive and condition-based approaches. We present methods for embedding reliability-informed degradation priors into deep learning models for prognostics, while enhancing interpretability through health indicators. These prognostic insights are integrated into a stochastic optimization framework for multi-day planning, enabling maintenance decisions that reflect both operational constraints and system health. The result is an adaptable system capable of supporting day-to-day operational planning and leveraging predictive insights for long-term asset management. Use cases from airline operations highlight the practical impact of this integrated approach, showcasing how data fusion and uncertainty quantification can support a more robust, scalable maintenance decision framework.

The predictive maintenance framework implemented at dsm-firmenich ingredients operations seeks to enhance asset reliability by transitioning from reactive interventions to proactive, data-driven decision-making. Utilizing real-time operational data and advanced machine learning algorithms, the framework accurately forecasts the health status of critical equipment, initially targeting vacuum pump systems. These predictive insights facilitate early anomaly detection, promptly alerting engineering/maintenance teams through automated notifications. Additionally, integrated condition monitoring indicators dynamically adjust process parameters to significantly reduce equipment overload risks. This comprehensive approach optimizes maintenance scheduling, minimizes unplanned downtime, extends asset lifespan, and ensures seamless operational continuity.

Environmental sciences rely heavily on accurate, timely and complete data sets which are often collected manually at significant risks and costs. Robotics and mobile sensor networks can collect data more effectively and with higher spatial-temporal resolution compared to manual methods while benefiting from expanded operational envelopes and added data collection capabilities. In future, robotics and AI will be an indispensable tool for data collection in complex environments, enabling the digitalisation of forests, lakes, off-shore energy systems, cities and the polar environment. However, such future robot solutions will need to operate more flexibly, robustly and efficiently than they do today. This talk will present how animal-inspired robot design methods can integrate adaptive morphologies, functional materials and energy-efficient locomotion principles to enable this new class of environmental robotics. The talk will also include application examples, such as flying robots that can place sensors in forests, aerial-aquatic drones for autonomous water sampling, drones for aerial construction and repair, and impact-resilient drones for safe operations in underground and tunnel systems.

The Norwegian Continental shelf has been a substantial oil and gas production province since the 1970's. Recently, the area is undergoing a major transition, with the electrification of the existing assets that used to be powered by gas turbine generated energy. The operating model is changing with electrification of the offshore assets, from traditionally, 'island mode' of energy production and maintenance of systems. Each, asset has been responsible for providing their own power supply and plan/execute the maintenance of the local systems on their installation. As of today, there are several area electrification solutions (e.g.: Utsira High) emerging where the assets are bundled together via a shared onshore grid power supply or with offshore wind farms/parks. The oil and gas assets are more dependent on a 'collaboration mode' of energy supply and demand and they need to coordinate and integrate their maintenance schedules. One oil and gas operator asset is typically defined as the TSO or system responsible for the offshore grid and responsible for coordinating the utilization of the grid energy across the different assets and also create an integrated plan for maintenance of the systems that comprise the offshore grid. The system responsible need to align the maintenance plans across the bundled assets and their local plans. There are emerging tools for developing optimized maintenance schedules, knowing the technical condition and down-time of electric equipment are critical to maintain a good availability of the offshore power grid in an area where several assets share the same onshore power supply. The presentation will introduce the integrated planning of maintenance in such an offshore grid and provide some insights on the software tools that are used to optimize the utilization of the energy supply and for integrated planning of maintenance of the offshore grid.

As predictive maintenance shifts from reactive strategies to proactive intelligence, the convergence of novel sensing technologies, including ultra-low-power, neuromorphic, and energy-harvesting sensors, with Edge AI is transforming the field. In this talk, I present an AI-based approach to predictive maintenance, driven by fast, local decision-making and energy-efficient sensing architectures tailored for battery-operated devices as well as robotic platforms. Drawing from our research at ETH Zurich, I will showcase embedded and efficient intelligence across multiple systems, from self-sustaining sensor nodes to autonomous robotic drones and dogs equipped with multimodal perception capabilities (radar, depth, and inertial sensing). These platforms enable robust condition monitoring, anomaly detection and predictive maintenance in dynamic environments, leveraging real-time edge processing to reduce latency, bandwidth, and energy consumption. Finally, I will explore how sensor fusion, neural model optimization, and custom hardware-software co-design unlock scalable, sustainable solutions for the next generation of maintenance in industrial, agricultural, and infrastructure monitoring domains.

Within railway track asset management, data analytics meanwhile is the one approach to plan maintenance - predictive maintenance is implemented. It is possible to predict both the right point in time and the precise location of the net necessary maintenance work. If it is both technically and economically wise to maintain only the spots identified is at least questionable. The planning process for the maintenance works that should be executed based on the evaluations needs more focus in future. Additionally, the still increasing possibilities to analyse track condition and its behaviour over time leads to questions concerning the type of maintenance foreseen. Looking into the technical consequences of - very general - poor quality opens room for changing from one type of maintenance towards another. Preventive maintenance based on failure root cause analyses seems to provide much more sustainability than nowadays still failure triggered maintenance. This will change maintenance procedures in future and maybe also will ask for maintenance machinery and processes not available today.

The integration of computer vision and machine learning into industrial environments presents both significant opportunities and unique challenges. This presentation shares key lessons learned from the development and deployment of AI-driven software solutions designed to support operators and managers in waste management facilities. Drawing from a range of real-world projects, the talk will explore critical aspects such as user-centric design, system validation, and operational integration. Emphasis will be placed on translating research and development into robust, scalable products, with concrete examples illustrating both successes and setbacks, highlighting the important of interdisciplinary collaboration.

Building energy modeling plays a vital role in optimizing the operation of building energy systems by providing accurate predictions of the building’s real-world conditions. In this context, various techniques have been explored, ranging from traditional physics-based models to data-driven models. An emerging trend is to combine physics-based and data-driven models into hybrid approaches with the aim of leveraging the advantage of each. This talk will provide an overview of such hybrid approaches, explore their mechanisms and showcase their application on a real-world building model. Specific aspects of evaluation include predictive performance, data dependency and interpretability.

The evolution from predictive maintenance to comprehensive life-cycle management of track infrastructure is a significant advancement in railway operations. Automated track inspection technologies have revolutionized maintenance strategies for railway infrastructure through precise measurements and data collection. By leveraging cutting-edge sensors and data analytics, potential failures can be predicted, and maintenance schedules optimized, enhancing the reliability and safety of railway operations. Building on these predictive maintenance techniques, the concept of life-cycle management is introduced. This approach not only focuses on immediate maintenance needs but also considers the long-term performance and sustainability of the track infrastructure. Integrating life-cycle management principles aims to extend the lifespan of assets, reduce overall costs, and improve the efficiency of resource allocation. The presentation will also provide a forward-looking perspective on how life-cycle management can be further enhanced with emerging technologies and innovative practices. The transition from predictive maintenance to life-cycle management can transform the way track infrastructure is managed and maintained, ensuring a more sustainable and efficient future for railway systems.

Federated learning is a promising approach for data preserving machine intelligence for predictive maintenance in the vehicle industry. However, industrializing this technology comes with significant technical challenges. I will share insights in our ongoing efforts to implement federated learning and discuss practical and technical challenges involved in building a scalable FL infrastructure, deploying edge-based ML machine learning models for multivariate anomaly detection, and implementing online learning for sensor-time series data. If times permits, I will touch on approaches for uncertainty-aware remaining useful life (RUL) prediction and causal discovery—both crucial for developing robust and interpretable predictive systems.

In the face of increasing demands and a congested railway network, the Swiss railway system is confronted with significant challenges, including a growing backlog of maintenance and infrastructure work. This talk explores how Agent Based Modeling (ABM) and reinforcement learning can be leveraged to enhance scenario creation and evaluation, providing valuable insights into the complexities of network management. By simulating various operational scenarios, we can identify optimal strategies to address current pressures and future demands. Our approach aims to ensure a stable and robust railway system, fulfilling our commitment to high mobility performance well into the second half of the century. The talk will discuss the integration of advanced modeling techniques into our planning processes, enabling us to navigate the complexities of modern railway management effectively.

190 years after the start of railway in Germany, the ageing and highly stressed rail infrastructure is facing enormous challenges: short time windows for inspections, complex maintenance and staff shortages require innovative solutions. This presentation examines how predictive maintenance technologies could replace manual and resource-intensive processes. Our focus is on how to meet the real-world challenges of integrating newly developed predictive maintenance technologies into the safety-related rules and process landscapes of the railway industry that have evolved over decades. We will have a look at the first approaches to solutions at DB InfraGO and how we harvest our data to provide effective solutions. The complexity of such an implementation can only be addressed together with industry partners and the scientific community. Together, we want to shape the future of railway infrastructure maintenance and secure an economical and reliable infrastructure. Join us on this exciting journey of innovation, lessons learned, and let us define the way to the future of rail maintenance!