Research Projects
VIPFLUID - Vorausschauende Instandhaltung für Pumpensysteme auf Basis von Federated Learning und Synthese von multiplen Sensordaten
Summary
Climate change with extreme weather events, including in Germany, such as heat or heavy rainfall, require countermeasures to deal with them acutely and steps to prevent further CO2 emissions. The project objective of predictive, adaptive maintenance of wastewater pumps and pumping stations is intended to save resources such as energy, people (fewer inspection trips) and materials (longer service life) and to create a resilient infrastructure by preventing unexpected technical failures. In addition, the development steps, including vibration simulation and the digital solution itself, are designed to be sustainable using open source software or efficient, distributed algorithms. Adaptive sensor technology and an intelligent sensor hub record the pump data and process it locally (edge) in order to send a compressed data stream to local computing resources (fog). Synthetic data (GANs) enable local and resource-saving ML models for prediction. In compliance with data privacy, the model is transferred to a cloud system for global training. Using a federated/transfer learning approach, the shared, pre-emptive ML models are iteratively improved by other users, mirrored back into the fogs and all participants benefit in terms of process reliability and failure prediction.
Relation to political objectives
The present project addresses the 17 formulated lines of the UN Agenda 2030, particularly in the areas of ecology (SDG 13) and the economy (SDG 9), and creates the opportunity to sustainably operate existing infrastructure in the area of decentralized wastewater disposal. The digital structures/models to be developed are intended to counteract excessive consumption of resources (energy, materials, personnel) in the area of maintenance. This means greater efficiency for maintenance processes and a longer service life for the equipment. The aim of the project is to design the digital solution in such a way that software and hardware resources are used optimally through minimized data transfer and the applied AI models are used efficiently through federated learning (FL), data synthesis and energy-saving algorithms. This leads to significant resource savings and consequently to significant CO2 reductions (SDG 13), which supports the achievement of climate and environmental protection goals.
Economic and / or socio-political relevance of the project
In times of resource scarcity, ensuring distributive justice is of great socio-political relevance. This applies to water as a resource as well as to the use of energy, materials and personnel to maintain and optimize water supply and wastewater disposal (WVE). The savings primarily contribute to the above-mentioned aspect of ecological sustainability. Freed-up capacities ensure a more balanced distribution of resources, have a stabilizing effect and prevent bottlenecks. This affects energy, raw materials and the availability of skilled workers. As decentralized systems, pumping stations for wastewater require a high level of manual maintenance by skilled workers. In the long term, the technical solution can also be extended to other areas of application (spill-over). In the area of wastewater drainage, use for dynamic control of wastewater flows is conceivable. Existing infrastructure could, for example, be dynamically controlled in the event of climate-related heavy rainfall events and damage could be prevented. The development goals of this project prevent the failure of essential technology in unexpected stress phases and make the existing infrastructure resilient (SDG 9). In addition, the use of ML processes in the plant sector harbors great strategic potential that currently remains untapped due to the high use of resources in conventional learning processes.
Promise of benefits and concrete, targeted results of the planned project
The aim of the project is to supplement the status data of wastewater pumps with suitable sensor technology and use this for machine learning (ML) training. This enables proactive maintenance and minimizes reactive action. In particular, this prevents cost-intensive and system-critical failures and aims to significantly increase process reliability. In addition to sensor technology, the core of this approach is the digital implementation of data generation and processing, as well as the reuse of existing ML models. The software solutions to be developed based on Generative Neural Networks and FL are intended to establish ML technology in the field of pre-emptive maintenance in an economical and ecological way, i.e. to enable the new technologies with minimal use of resources for data acquisition and to consolidate learning successes of partners into joint models in order to provide rapid implementation and reliable prediction results in specific cases. In this way, maintenance resources are to be significantly reduced.
Openness of the project with regard to the use / generation of open standards and / or use / generation of open source solutions
The implementation and solution approach of the project are primarily based on open source software. In addition, the FL and the associated training on several devices implicitly create a common model that can be used by everyone. The analogue technical Sesnlr solution is also implemented in a manufacturer-open manner. Furthermore, the primarily targeted area of application of wastewater treatment of mostly municipal companies (organized in special-purpose associations) promotes the open exchange of knowledge, whereby additional effective synergistic effects in data generation and a strengthening of the model can be expected.
- Funded by: BMWK, Referat VIB3, Entwicklung digitaler Technologien
- Contact person: Karsten Wendt
- Project Website: link
- Funding period: 05/2023 - 04/2026