Artificial Intelligence
Our goal is to transform and develop the latest technologies into advanced products and services using artificial intelligence (AI) to protect people and the environment. We see enormous potential in merging human intelligence with different areas of artificial intelligence to use this in turn to enable decisions that support this goal.
Harnessing AI for smarter asset management
In today's world of big data, each inspection or technical service generates vast amounts of measurement and operational data. This raw data undergoes cleansing, interpretation, and analysis to convert it into valuable information, which in turn leads to informed decision-making. Harnessing the power of artificial intelligence has become indispensable in this process. Similar to how AI-driven software tailors purchase suggestions and advertising strategies to individual consumers, we leverage AI techniques to furnish experts with asset-related data, streamlining real-time decisions for asset safety. Moreover, machine learning and AI models not only enhance decision support but also enable proactive prediction and mitigation of potential threats, allowing us to stay ahead of challenges and minimize their impact effectively.
Explore our AI research areas
MFL technologies are the backbone of the pipeline inspection industry for comprehensive corrosion defect assessment. ROSEN has always been a market leader in providing tools in both magnetizing directions – axial (MFL-A) and circumferential (MFL-C). Conventional data evaluation provides for a separate evaluation of MFL data and generates two inspection reports. However, with the development of data fusion machine learning models, ROSEN can now take advantage of the complementary information of two MFL inspections and generate high-resolution 3D metal loss depth maps. The high-quality 3D depth maps dramatically increase the accuracy of safe remaining life and pressure assessments, ultimately reducing mitigation cost for our clients.
By putting the pipeline asset at the center of our data management system, we can offer unique, world-class data services. With an asset-centric view, we connect all business domains to ensure cross-functional data analytics and deliver the most valuable insights to our customers. Autonomous teams can work to provide data management solutions that are unified and standardized through a flexible integration layer. This also helps us to deliver reliable artificial intelligence models.
Security is always an important aspect of our data management system. Therefore, we have implemented a multi-layered data governance framework to ensure that our data only gets into authorized hands and meets the highest quality standards. With our data management system, we provide the best foundation for our services to ultimately offer our customers added value.
ROSEN supports asset operators in making data-driven decisions in critical business processes such as pipeline integrity management. Artificial intelligence plays an important role here, as many of these decisions are based on the results of AI models. It must be ensured that AI models are trained on high-quality data, are frequently updated and are reliable and robust. In addition, constant monitoring of model metrics and infrastructure must ensure consistent model performance.
At ROSEN, we use Machine Learning Operations (MLOps) to achieve this goal. MLOps describes techniques, processes and the use of tools to bring machine learning models into production, monitor and maintain them. In practice, we have implemented tools such as Azure DevOps, Artifactory and Kubeflow as well as standardized development processes as part of our AI landscape to meet industry best practices and compliance standards in the area of MLOps.
The laws of physics and thus most measurement technology are governed by Partial Differential Equations (PDEs). The main idea of physics informed machine learning is to incorporate domain knowledge in form of PDEs into the learning process of neural networks. As it is easy to calculate derivatives of neural networks not only with respect to their parameters but also with respect to their input variables, we can reduce the question "How well does the output of the neural network follow a specific physical or mathematical law?" to an easily computable loss function. As research in this area is still on a rather foundational level, there is a close cooperation with the University of Osnabrück and a joint industrial PhD student is conduction research on this topic.