Affordable and durable energy with improved energy production, energy distribution, and energy consumption.
We work on demonstrators of the AI Research for following applications:
Industrial machines are continuously evolving towards more complex systems, often featuring multiple interacting subsystems and components. These assets are often operating in complex industrial environments, influenced by a variety of different factors in interaction with a human end user (e.g. the operator). Since failures or downtime can have serious consequences (e.g., with respect to safety or economic impact), it is important to adequately monitor the health condition of these machines. Not in the least, this holds for wind turbines, for which the operational management not only is expensive, but often also crucial for a stable power supply. Therefore, modern wind turbines are equipped with sensors. The data from these sensors is analyzed to predict abnormal behavior and in this way enable to plan maintenance ahead of time (predictive maintenance).
Prognostics and Health Management (PHM) is the research field that links studies of failure mechanisms of industrial assets to system lifecycle management. The goal of Prognostics and Health Management (PHM) is to provide methods and tools to design optimal maintenance policies for a specific asset under its distinct operating and degradation conditions, achieving a high availability at minimal costs.
The research within the Flanders AI Research Program
The increasing complexity of industrial assets and their operating environments increasingly requires a more real-time and dynamic health assessment. Traditional PHM approaches, which are primarily knowledge-based, no longer suffice. Therefore, the research groups in this use case study how data-driven and knowledge-based models can reinforce one another by combining them in a hybrid methodology to perform more accurate prognostics and management of ageing industrial machinery. This includes research towards improved techniques for
Challenges and researched methods
Analyzing (sensor) data from such complex industrial machinery comes along with several specific challenges that form the subject of this use case. These include amongst others the lack of labelled data on failures and machine degradation, due to which un- and semisupervised techniques are being studied in combination with transfer learning to exploit the limited available labelled data in the best possible way. In some cases, it might be important to collect feedback on particular aspects from the user, for example to learn about the subtle difference between normal and anomalous behavior, for which the use of active learning techniques will be researched. Also the translation of the results of the algorithms into interpretable insights for decision support is crucial. The biggest challenge and main focus of the use case is however combining data-driven and knowledge-based models, such that they can reinforce one another in a hybrid learning strategy.
The first phase of this research was focused on data-driven techniques for diagnostics and prognostics. Amongst others, a methodology was developed to assess the condition of a machine based on data-driven health indicators. These indicators can also serve as the basis to detect the degradation in performance of an asset, in which the different operating modes of a machine are accounted for. To tackle the aforementioned challenge of a lack of labelled data, also a number of transfer learning approaches have been researched, applied to the problem of detecting icing on the blades of a wind turbine. Furthermore, also the first steps were made towards a data-driven digital twin for hybrid PHM.
A number of these topics are also being researched in more depth in the AI ICON projects CONSCIOUS and TRACY, that were recently started in collaboration with research partners from the Flanders AI Research program and a consortium of Flemish companies. A short description of these projects can be found below.
More info and related projects
Producing green energy is an expensive process. In the eight offshore parks off the Belgian coast, there are about 400 wind turbines that supply energy to 2.2 million families. Keeping this infrastructure operational is not only extremely expensive, but also crucial for a stable power supply. The modern wind turbines are therefore equipped with about 800 sensors that transmit data every second. The data is interpreted in real-time, on the one hand to monitor the condition of each machine and to be able to plan the necessary maintenance work, and on the other hand also to improve energy production.
The research within the Flanders AI Research Program
Thanks to AI, wind turbines can inform the database and each other to adapt to real weather conditions. In this way, AI can ensure that the wind turbines are not overloaded and that production is more efficient.
More info and related projects
The Low Voltage (LV) distribution grid (the part of the electric grid to which the end consumers and prosumers are connected) is a critical infrastructure for our society and an important enabler for the energy transition in Flanders. Hence, we have to make sure that the LV grid does not jeopardize the transition to a sustainable energy future, due to limitations when dealing with a significant increase in electricity demand. The use of our distribution networks is changing rapidly due to, among other things, the growth of renewable energy production (solar panels) and the electrification of transport and heat sources (charging of electric and hybrid cars and heat pumps). In Flanders, the charging of electric cars and the use of household appliances for additional frequency reserves are a particular concern given that they lead to higher loads on the grid, an increased synchronicity, which in turn leads to higher peak loads.
The LV grid as we know it today, was built over the course of the last century using a “fit and forget” strategy: install (over dimensioned) cables with sufficient capacity to cover all demand peaks. However, with the massive rollout of technology linked to the energy transition, the required capacity when using present safety margins would lead to a very high investment cost. A business-as-usual approach can hence lead to the LV distribution grid becoming a financial and practical obstacle to the transition to a sustainable energy system. The alternative is the development of technology to use the installed capacity more optimally by operating our grids closer to their limit, and to technically support measures to mitigate the impact of the energy transition. Although the workings of electric grids are well understood, the specific challenge with the LV distribution grid is the limited, incomplete, or incorrect available data: the layout and electric properties of the grids are only partially known, and measurements are very limited and rarely real time. Also, the scale of the problem is vast: Flanders counts with 40k LV transformers, 240k LV feeders, and 3.5M connection points. By combining newly available grid data, e.g., from digital meters, with AI techniques, we aim to calculate the grid load in detail both in time and space to assist distribution grid operators in their infrastructural and operational decisions to make more optimal use of the existing grid capacity and to avoid massive investments in new cables.
Members of our team obtained the third position in the IEEE-CIS Competition titled “Technical Challenge on energy prediction from smart meter data” with a ratio-based approach. The proposed method consisted of different steps applied sequentially: pre-processing, data augmentation and normalization, clustering, prediction based on ratios and finally ensemble learning and post processing. In another work, we introduced a new two-layer hybrid Neural Network architecture called “Decomposition-Residuals Deep Neural Network (DR-DNN)” that was successfully used for day-ahead electricity demand forecasting. This architecture consists of one decomposition layer in which the raw time series are decomposed into trend, seasonal and residual signals, and one Deep Neural Network (DNN) layer for modelling unknown nonlinear patterns.
We have developed and implemented an interactive ILP (Inductive Logic Programming) method for learning relational concepts, such as house-cable connections, from examples. We are currently working towards the application of this method to the GIS (Geographic Information System) dataset provided by Fluvius (the Flemish distribution grid operator), to better determine the house-cable connections of a city in Flanders.
We have clustered consumption user profiles using an active semi-supervised method, COBRAS. The method incorporated a new dissimilarity measure proposed exclusively for this use case, as well as a new time series representation for flexible supervision expert feedback. Finally, by applying K-medoids clustering on voltage time series from smart metering, we were able to validate grid topology data and derived which household is connected to what electrical phase