Projects

Information:

• End-of studies internship (Master's) 

• Workplace: GIPSA-lab, French National Center for Scientific Research (CNRS) and Université Grenoble Alpes

• Timeline: May – October 2021

• Supervisors: Prof. Olivier Sename, Dr. Thanh Phong Pham

• Collaborators: Marcelo Menezes Morato, Juan C. Tudon-Martinez

Abstract:

The semi-active suspension system plays a crucial role in the automotive industry as it serves both driving comfort guaranteeing and road holding objectives. However, there are some scientific gaps in the sense of state and fault estimation, as well as the development of control laws under faulty conditions for this particular system. In this research work, these issues are examined. This work is two-fold, contributing to both theory and experiment aspects.

In the theory part, we develop multiple state and fault estimation methods with applications to the semi-active suspension system. These include polytopic/grid-based LPV/qLPV observers using criteria from H∞ to H2 to a mixed combination of both (multi-objective). Then, we extend into grid-based, existing unified (algebraic) observers, including the Lipschitz NLPV observer, whose matrices are all parameter-dependent. Thanks to the parameterization of the observer matrices, the estimation error is unaffected by input disturbances, while LMI solving attenuates the effects of the control input and measurement noise on the error. Additive fault estimation algorithms for suspension are also developed, viz., the LPV filter for unknown input estimation (designed in the same formulation as an output-feedback controller) and LPV PI observer (assuming that the fault is slow-varying). We validate the proposed approaches using analyses in the frequency (using Bode diagrams) and time domain (using simulations).

In the experiment part, we conduct fault-tolerant control (FTC) experiments on our vehicle test platform. To obtain the PWM input (the control input of our testbed) from the corresponding damper force (the FTC output), we propose and validate various inverse models using lookup tables and deflection velocity-dependent polynomials obtained using real data. We also propose signal processing methods for denoising measurements while working with the testbed. Lastly, but most importantly, we generate the fault by subtracting the PWM signal corresponding to the fault from the actual PWM signal. The first results of our experiments illustrate the graceful degradation achieved with FTC by showing that the damper force, in this case, lies between that of the faultless case and the case of a nominal controller under fault.

Keywords: LPV observers, unified observers, fault estimation filter, PI observer, fault-tolerant control, semi-active suspension system

Information:

• Student research project (second  year of Master's)

• Workplace: GIPSA-lab, French National Center for Scientific Research (CNRS) and Université Grenoble Alpes

• Timeline: September 2020 – January 2021

• Supervisor: Prof. Olivier Sename 

• Collaborators: Thanh Phong Pham, Marcelo Menezes Morato

• Publications: here

Abstract:

This work is about the design of different unified observers and the application of linear parameter-varying (LPV) model predictive control (MPC) to an autonomous vehicle's semi-active suspension system, improving driving comfort while being more energy economical than active suspension. In semi-active suspensions, it is necessary to estimate the vehicle vertical dynamics states and the damper force exerted using state observers. We design a multi-objective unified qLPV observer by first establishing its general form, then parameterizing this form to decouple the effects of bounded unknown input disturbance, and then using the H∞ condition to treat the control input (which is a disturbance) and the H2 one to treat the noise. Doing this, we provide different methods tailored to the characteristics of each type. The Pareto plot is used to examine further the H∞/H2 trade-off. When solving the multi-variable LMIs, we apply the projection lemma to break them into smaller LMIs that are solved successively to reduce the high level of conservatism, thus preventing the observer from becoming a proportional one. Next, we apply and compare two LPV MPC strategies, using the quarter-car model that models the vertical dynamics at one of the vehicle's four corners. The first method is frozen guess LPV MPC, where the varying parameters are assumed constant throughout the prediction horizon. The second one is sequential QP-based LPV MPC, where additional iterations are performed inside each traditional MPC iteration to estimate the varying parameter values used in the optimization problem. The MPC approaches provide a suitable trade-off between cost and passengers' driving comfort performances of the semi-active suspension system in a sampling period as limited as 5 ms, thus allowing for possible real-time application.

Keywords: Observers, model predictive control (MPC), linear parameter-varying (LPV), semi-active suspension system

Information:

• Research internship (first year of Master's)

• Workplace: K. Nakano Lab, The University of Tokyo

• Timeline: June – September 2020

• Supervisor: Prof. Kimihiko Nakano

• Collaborators: Zheng Wang, Koge Yusuke, Bo Yang

• Publications: here

• GitHub: here

Abstract:

Advanced driver-assistance systems have proven their capability of assisting drivers in certain aspects, such as driving safety or comfort. In an attempt to help the (partially) disabled to drive, relieve long-distance driving fatigue, as well as reduce risks of muscle injuries, we seek to replace the conventional automotive control interface with a modern human-machine interface through the use of surface electromyography (sEMG) sensors attached to the driver's muscles. In this study, we propose several aspects of using sEMG-controlled interfaces for driving assistance, particularly for braking control of vehicles. We build a database of sEMG signals measured from the activities of five leg muscles corresponding to the driver braking and preparing for braking. Power spectral analysis is first performed to identify the components of these signals, which include noises from the electrical network (periodic, 50 Hz), from the driver's heartbeat, and unknown noises. Different signal processing methods are proposed and tested, namely the traditional and optimal linear (Wiener) filtering algorithms, and independent component analysis, which allow us to separate the independent components contained in a set of signals. We then develop and test, by simulation, different machine learning-based classification algorithms, including the k-nearest neighbors, deep reinforcement learning, and deep learning using a long short-term memory network to classify the preprocessed signal into two groups: braking and not braking the pedal. This output is then used for vehicle braking control.

Keywords: Human-machine interface (HMI), surface electromyography (sEMG), machine learning, independent component analysis (ICA), long short-term memory (LSTM), advanced driver-assistance systems (ADAS)

Information:

• Student research project (first year of Master's)

• Workplace: GIPSA-lab, French National Center for Scientific Research (CNRS) and Université Grenoble Alpes

• Timeline: January – June 2020

• Supervisor: Prof. Olivier Sename 

• Collaborators: Thanh Phong Pham, Eduarda Costa

• Publications: here

• GitHub: here

Abstract:

In this project, we develop a self-scheduled longitudinal cruise control of an autonomous vehicle so as to guarantee driving comfort faced with speed and mass variations. The longitudinal car model is first written as a linear parameter-varying (LPV) system, where the parameters are the speed and mass of the vehicles. Then, a grid-based H2 optimal control method is developed to ensure the required performances. The problem, formulated as a set of linear matrix inequalities (LMIs), is solved through a grid-based approach to minimize conservatism. In order to improve the tracking performances with realistic road profiles, we propose a feedback + feed-forward controller assuming that the car’s mass is measured on-line using multiple built-in sensors that detect the additional load and that the road slope can be known/estimated in real-time. In the second part, a comfort objective is introduced in order to adapt the reference speed to the passenger comfort part of the whole body vibration. Indeed, it is known that the level of discomfort perceived by users is a result of road surface roughness and varies according to the velocity of the car. Such a criterion can be evaluated using the international roughness index (IRI) and/or the frequency weighted vertical acceleration (defined in the ISO 2631 standard). A study is conducted to establish a correlation between the self-driving car’s speed and the passengers’ comfort, from which the optimal speed profiles are identified. We here propose a new way to formulate the ride quality as a function of the car velocity parameter. Performances and robustness assessment are carried out through time-domain simulations assuming poor knowledge of the scheduling parameters (vehicle mass and speed), some uncertainty on the road condition parameters (road slope, rolling resistance coefficient, wind speed), and some neglected dynamics (actuator dynamics, communication delays). This work is later extended into integrated cruise + semi-active suspension control, both using the H2 approach for LPV systems. Note that the vehicle speed links between the longitudinal and vertical dynamics. 

Keywords: Vehicle dynamics, semi-active suspension control, cruise control, LPV, passenger comfort 

Information:

• Student research project

• Workplace: Control and Automation Lab, HCM City University of Technology

• Timeline: April – August 2019

• Supervisor: Prof. Cong Bang Pham

• Collaborator: Quoc Bao Pham

Abstract:

In this project, we discuss the mechanical design and modeling of a self-balancing motorcycle’s reaction wheel. First, equations of motions are formed and analyzed, and then we test the methods for finding an optimal mechanical design of the reaction wheel (pendulum length and wheel geometry) and selecting the suitable material. Then, we identify the criteria for motor selection (stall torque and maximum angular velocity).

Keywords: Mechanical design, inertia wheel pendulum, material selection, actuator selection