Electromechanical Engineering - Decision, Sensing & Control


Our competences relate to sensing, control and condition monitoring of cyberphysical systems. We combine in-depth physical insights into electromechanics with insights obtained from data that becomes available in today industry 4.0 setting applying state-of-the-art and innovative artificial intelligence (AI) routines (Hybrid AI).

As such the group contributes to building

  • explainable and transferable digital twins, and 
  • explainable and trustworthy AI for the manufacturing industry. 

Sensing & monitoring of (a fleet of) drivetrains

  • development of (multiplicative) hybrid data-physical models for explainable condition monitoring
  • virtual sensing of electrical machines based on electrical signal analysis for condition monitoring 
  • stochastic and uncertainty modelling to evaluate and improve the effectiveness of sensing and monitoring strategies
  • camera based sensing and monitoring of oil film in and degradation of bearings and gearbox  

Control & decision making of (a fleet of) drivetrains

  • development of probabilistic decision making routines that are adaptive, explainable and transferable 
  • development of (multiplicative) hybrid data-physical models for explainable system identification
  • controller structures that leverage on the well-known, safe and robust traditional control structures enhanced with AI to include data driven learnings or (high fidelity) models to improve performance 
  • techniques to do automatic (e.g. context aware) PID tuning
  • development of reinforcement learning control enhanced with physics suitable for the manufacturing industry
  • distributed optimal control techniques 

(fleet of) production & assembly lines

  • development of low-probability event estimations and predictions
  • use of AI techniques to correlate product-process feature 


Hybrid AI modeling for predicting abnormal behavior of complex mechatronic applications

Hybrid AI for design and controlAlthough data-driven AI models (e.g. recurrent Neural Networks) are typically known to learn intricate patterns of complex dynamic systems directly from raw data, enhanced prediction performances are observed when providing physics-inspired features to the model. The effect is especially more pronounced when learning from a small amount of data or from data that does not reflect all operating conditions or system variations. We have shown that on an experimental cam-follower mechanism, predicting the undesired detachment of the vertical follower from the rotating cam is more accurate when introducing physics-inspired features into a purely data-driven AI model (LSTM). Additionally, using the DeepSHAP framework as an explanation model we could objectively quantify the influence of these features on the model predictions. This approach reveals that the model prioritizes the inclusion of physics-inspired expert knowledge, explaining the observed enhanced generalization performances. The results can serve as an incentive to augment other data-driven prediction models with expert knowledge to enable improved monitoring of various other complex phenomena occurring in mechatronic applications.

Research on robotic applications

Haptic Reinforcement LearningWe conduct research at the intersection of system identification, control and machine learning for mechatronic and industrial robotic systems. Our hardware infrastructure is indispensable in our research goals to endow physical dynamic systems with improved functionalities and capabilities when interacting with uncertain environments, other systems and humans. Foundational research activities are being carried out on advancing upon system-level modeling of dynamic systems based on data and physical knowledge, as well as control methodologies based on machine learning and nonlinear optimal control techniques. Strategic basic research activites are on-going to improve the efficiency, performance, robustness, autonomy of safety-crticial physical dynamic systems. For example, we are developing a learning framework where guided and haptically demonstrated dexterous interaction with objects are learned and autonomously reproduced (force control) by the robotic setup.

Assessing similarity of process runs of a complex machine based on not-directly measurable process KPIs in view of warmstarting a new process run

Use of tSNE for KPIFor complex mechatronic machines with many machine settings it is often difficult to relate which settings contribute most to a certain desired process KPI. This KPI can often not be measured directly, and additionally the relation between the KPI and sensor measurements could be unknown. We have developed a method and routine to assess similarity of process runs based on a process KPI with an a priori unknown relation to sensor measurements. The method iteratively learns a 2D/3D embedding of the process runs and assesses the similarity based on the run’s embeddings. This toolbox has the capability to 1) Uncover which sensor signals are most relevant to be used to cluster the historical process runs according to the process KPI (using t-SNE manifold learning) 2) Identify which reference process runs had a similar process KPI to the new run (similarity assessment using k-Nearest Neighbors) 3) Determine adequate process settings (warmstart) for this process KPI (correlating feature similarity with settings similarity). This allows to warmstart the settings for a new run with the settings of a reference run having a similar behavior and a high efficiency, in view of optimizing the process KPI.

Hybrid modelling for digital twins in view of system identification, monitoring and control

modaDigital twins are a very promising concept, key is there is a unique digital twin for each physical asset such that the behaviour of the digital twin matches that of the physical asset. The data connectivity in industry 4.0 of these physical assets enables this and machine learning techniques and data driven models typically ensure the desired match in behaviour. However, they are hard to interpret and cannot be extrapolated. By intimately linking data driven components with physics based models, as such building real hybrid models, we introduce these characteristics.

We have developed a toolbox for hybrid modelling featuring: building a hybrid model, visualisation extracting analytical models from data. Our toolbox has been validated on a lab setup including a slider-crank with spring mechanism and in an industrial context (confidential).  

Hybrid AI: creating trustworthy AI for manufacturing 

hybridaiWithin our hybrid AI track we develop control systems and decision support systems which combine AI techniques with domain specific knowledge. In traditional process control this could be MPC control schemes, in motion products these are traditional PID controllers. Such formal control schemes are intimately combined with AI techniques such as reinforcement learning or Bayesian networks. As such learning and exploratory systems typical to AI systems are created that perceive the system through data capturing while being constrained and maintaining at least partially the advantages of traditional controllers (e.g. deterministic).   

Basic techniques have been applied on an dual drive system as well as on a learning controller for wind turbine drive trains.  

Data enhanced physical models (convex mapping) for optimisation: use case Optimal Power Split

codesign.pngA model-based strategy to design the drive train of all electric vehicles using a nested optimization approach wherein parameter exploration is attained using an evolutionary algorithm and the optimal power flows are determined by abstracting the high-fidelity behavioral models into appropriate convex loss mappings. This allows not only for an accelerated design procedure based on convex optimization without compromising accuracy; it also allows during operation to adapt key parameters of the convex maps to obtain optimal behaviour even when parameters change over time (due to temperature, degradation, ... ).   

As an example we sized an electric drivetrain for maximal range extension. A tractable convex formulation is obtained and optimization time is reduced by 99.3% compared to the traditional approach. Optimal control of the incorporated power split increases the operational range by 0.7% compared to the isolated operation of a single motor. The proposed methodology thus paves the way for extensive designs of drive trains and complex mechatronic systems in a general context

Data enhanced physical models (lumped parameters) for virtual sensing and more accurate control: use case induction machine

moformIn electrical machines, knowledge of rotor temperature is crucial for accurate torque estimation and control. However, rotor temperature measurements are often not available. Our tool chain allows to determine and consequently compensate for the actual virtually sensed rotor temperature. A first routine allows for the inverse identification of a thermal (induction) motor model. The developed routine is a data-driven inverse identification method: it identifies the thermal parameters of a second-order Lumped-Parameter Thermal Network (LPTN) of an electric motor. Once such a model is available for one such motor, our second routine enables rotor temperature virtual sensing using Kalman Filtering techniques and a measurement of the stator windings' temperature on a similar motor. In a final routine, we compensate for the effect of temperature on torque estimation such that it can be incorporated in the torque control loop.

This way, we created a virtual sensor to obtain a dynamic estimation of the considered motor's rotor temperature and compensating the torque control. We have applied our tool chain on induction motor and validated for both fixed and intermittent motor operation (up to nominal operation point). 

Adaptive Approximate Dynamic Optimization​: high fidelity models in the loop

aadoThe Adaptive Approximate Dynamic Optimization (AADO) algorithm is tailored to dynamic optimization problems where the system model contains high-fidelity (HF) model components of subsystem(s) that slow down the system model's evaluation time. Such high-fidelity models could e.g. include a finite-element model. We developed routines to either (i) replace the high-fidelity subsystem models with an approximate subsystem model set; (ii) replace the system model with a local (i.e. only valid in subregions of the state space) approximate dynamic system model based on sampled model evaluations, whereby the sampling is adaptive and localized to the optimal trajectory in correspondence with the trajectory optimizer. 

We have applied our techniques on a trajectory optimization scenario applied to a laboratory setup being a limit cycle slider-crank including a high-fidelity nonlinear load model. We found that the AADO algorithm requires only 0.1% system model evaluations compared with conventional DO and only 30% compared with non-adaptive AADO. 

Coping with system uncertainty/variability in design and measurements

System uncertainty

In all systems the output is partly defined by variability:  unmeasurable inputs or system variability. This brings several challenges:

  • During the design of motion products, it is key to understand and quantify the impact of variability on key system outputs to ensure optimal robust designs [=forward calculation]
  • During test/validation of motion products one whishes to be able to link variability in measured outputs to the modelled behaviour either the variability [=inverse estimation]

While state-of-the-art techniques use Monte Carlo simulations, we adopted the Polynomial Chaos Expansion (PCE) technique. This ensures not only a fast forward calcultion, combined with the max log likelihood methods it also allows to perform the inverse calculation effectively. Specifically for fluid mechanics, we developed efficient methodologies for the robust optimization under uncertainty. 

Our tools have been applied to real industrial applications. E.g. within the Flanders Make project EVIT they have been used to test in a more efficient manner new controllers for weaving looms. E.g. within the SBO project EUFORIA we have used our tools to design heat exchangers and unmanned Aerial Vehicles (UAV).   

Virtual sensors for electric machines (SRM, PMSMP, BLDC)

virtual sensorFor SRM’s we have experience with resonance-based position estimation where feedback of an equivalent circuit with injected test pulses is used to determine position. The test circuit is an 1D-diffusion model of lamination coupled with a Preisach model. For PMSM’s and BLDC’s we use the existing current sensor and information embedded in the response to the PWM signals. By slightly adapting the PWM signals while not disturbing the we can obtain an accurate position estimation. This methodology avoids the need of expensive position sensors

Series Distance Matrix for Anomaly detection

Series Distance MatrixWe developed an unsupervised learning toolbox (Generalized Matrix Profile) that processes a sequences of discrete-time data. Toolbox has a variety of Python implemented techniques such as Z-normalized Euclidean Distance, Matrix Profile (MP), Multidimensional MP (MMP), Contextual MP (CMP), Radius Profile, Series Distance Matrix, etc. suited to address specific challenges of the discrete-time data. Due to the variety of implemented techniques, the toolbox has a potential to become near universal data mining solution. One specific tool includes series distance matrices (SDM), a black-box type framework (data-driven) capable to efficiently process anomalies and trends in the time series. In its core, the method works by searching for pairwise ‘mathematical’ distances (=correspondence) between selected subsequences within the time series. Next, step involves processing of the resulted distance matrix, we offer a framework that includes a fragmented processing of distance matrix making it much more efficient and suitable for (i) processing long time series as well as (ii) online processing. Our toolbox is currently deployed for anomaly detection in NewYorkDataset, VentilationDataset. This work is done in collaboration with imec.

Link to Github

Surrogate models - AI for small data sets

Surrogate ModelsIn a manufacturing context, often models are unavailable (e.g. for very complex processes) or are too challenging to run in the loop (e.g. high fidelity models). In such case, surrogate models provide fast-running approximations of complex time-consuming computer simulations. Surrogate models are also known as response surface models (RSM), metamodels, proxy models or emulators. They mimic the complex behavior of the underlying simulation model. They bridge the gap between the numerical or experimental, and the analytical. Surrogate models are used for parametric studies, optimization, design-space exploration, visualization, prototyping, uncertainty quantification and sensitivity analysis.

Further, to build such models, there is a need for data. However, obtaining such data can be expensive (costly experiments, numerically challenging simulations, ... ). One solution to cope with this is to apply adaptive sampling as a method within a design of experiments strategy. We have developed a whole suite of tools to develop surrogate models. This work is done in collaboration with imec.

More information

Sensor Fusion & Dynamic Dashboards

DynamicDashboard.jpgBy designing multi-modal and multi-sensor architectures we provide collaboration among sensors (classical, video-based, time-series, mobile sensing and virtual/data mining sensors) in order to feed back the available information and intelligence of all sensors to optimize their functionality and enhance the detection and interpretation of advanced events. A wide range of applications can benefit from these multi-modal and multi-sensor architectures that fuse amongst other visual, audio, thermal, vibration and/or data mining information. Examples are industrial process control or condition monitoring. 

The heterogeneity and vast amount of sensors, as well as the difficulty of creating interesting sensor data combinations, also hinder the deployment of fixed structure dashboards as they are unable to cope with the accordingly vast amount of required mappings. Therefore, we additionally develop dynamic dashboards that precisely visualize the interesting data for the end-user produced by sensors in multi-sensor environments by dynamically generating meaningful service compositions, allowing the detection of complex events that used to remain undetected. This work is done in collaboration with imec.

More information

More in this video

Research on robotic applications

under construction

Research & test infrastructure

We strongly believe in validating and developing research that works on real hardware, in real mechatronic setups. Therefore we have several setups available for testing and validating our research. As part of this we have developed hardware implementations on Field Programmable Gate Arrays (FPGAs) of complex algorithms for dynamic applications that require computationally time consuming resources. Examples are Kalman filtering, approximate optimal control with online optimization, machine learning based control, … This work resulted in a tool chain that transforms these algorithms from high level descriptions into a physical implementation. The existing DSPACE-FPGA-platform allows you to create a proof-of-concept and to evaluate the value of such an FPGA implementation for your application.  

See here our list of infrastructure