Search

Individuals using upper extremity prostheses often struggle with activities of daily living. The most challenging tasks to execute with a prosthesis are those requiring cooperation between the prosthesis and the intact hand, e.g., cutting bread or tying shoelaces. In our ongoing research, we apply a user-centred design approach to develop and implement a machine learning algorithm (MLA) that can recognize and classify movement intentions based on EMG signals of both the stump, as well as the intact upper extremity. The goal is that the MLA recognizes task intentions and that the prosthesis can naturally execute the corresponding movements at high speed and precision. To get a deeper understanding of the difficulties, needs and experiences of myoelectric prosthesis users, we are conducting a pilot study with three semi-structured interviews and a qualitative content analysis. In a first analysis, we identified different aspects with potential for improvement regarding task performance. Interviewees indicated that the lack of speed of prosthesis movement limited its usefulness to the extend where users preferred doing a task single-handed rather than with the prosthesis. Further, the selection of different grips through the human-prosthesis interface was perceived as non-natural and cumbersome. In addition, several specific tasks were identified, where the biggest potential for improvement is expected, such as typing on a cell phone or turning pages in a book. Interviewees also mentioned tasks that are often conducted at the same time and hence need cooperation between the hands, such as using a blow dryer and hairbrush. Results from these pilot interviews will, in combination with pilot EMG analysis, inform the planning of our larger study. We plan to design an MLA that is compatible with commercially available myoelectrical arm prostheses. We further see future potential to extend the algorithm to other use cases, such as exoskeletons.

Introduction: Individuals missing a part of their upper extremity often use prostheses to increase their ability to conduct activities of daily living. However, the function of the upper limb is challenging to replace, as it includes gross and fine motor activities. The most challenging tasks to execute with a prosthesis are cooperative hand movements. Prostheses can be controlled by recorded EMG signals of the remaining muscles, potentially also by utilising EMG signals of the contralateral side. However, it is not trivial to adequately record and decode these signals in order to turn them into appropriate movement commands for the prosthesis. Consequently, the goal of this proof-of-concept study was to utilize a machine learning algorithm to identify tasks based on bilateral EMG-recordings in able-bodied individuals. Methods: Bilateral surface electromyography of two arm muscles (forearm extensors and flexors) was measured in two participants while performing two tasks of daily living (cut bread and cutting/eating with fork and knife). Each task was repeated 15 times on both sides. The raw signal of both muscles was processed (removal of DC offset, rectification, 3rd order low pass filter) before dividing the data into 80% used for training a sequential model with 4 layers. 20% of the data were used to test the model. Results and Discussion: The machine learning algorithm resulted in a correct rate of task classification of 83.33%. This result is slightly lower than previous studies which showed an accuracy of 92.60% (1) or 98.78% respectively (2) using machine learning to classify different hand movements through surface electromyography. Conclusion: The machine learning algorithm used in this pilot study showed promising results of being able to identify tasks of daily living based on surface electromyography. Future research will extend the algorithm to additional tasks and aim at improving the algorithms by incorporating additional muscle groups.

Surface Electro-MyoGraphic (sEMG) signals acquired on the forearm can provide information on the hand movement, which can help control a prosthetic implant for disabled people. To do so, the sEMG signals must be accurately classified despite the signals' non-stationnarity, noise from sensors, involved muscles, and patient's peculiarities. This study deals with the classification of hand movement using sEMG signals, and our main goal is to compare several feature extraction methods. We focus especially on the use of time-frequency domain for feature extraction and on several linear and non-linear methods for the dimension reduction. Methods as the Discrete Orthonormal Stockwell Transform (DOST) and Multidimensional Scaling (MDS) are applied for the first time on sEMG signals, and an extensive comparison study is performed on the combinations of the proposed methods. Classical classifiers are then used on a public dataset in order to evaluate the applied methods. Both short-time Fourier transform and Stockwell transform performed well, with respectively 90% and 91% accuracy, while promising results are obtained with DOST and MDS with classification rate 87% and significant improvement in feature extraction computation time.

International audience; This paper proposes an adaptive voltage-oriented control (VOC) with onlineac filter parameters identification for three-phase voltage-source rectifier (VSR). A new methodbased on adaptive linear neuron (ADALINE) is first designed to identify the ac filter parameters.For accurate identification, the VSR nonlinearity is included in the ADALINE structure.Thereafter, the developed ADALINE is inserted in the VOC to realise an adaptive VOC. Thus,the decoupled terms and the proportional-integral current controller gains are updated online.Finally, the ADALINE ability to track properly the ac filter parameters is investigated byexperimental analysis. It shows that the VSR nonlinearity consideration has significant influenceon the resistance identification. Compared to the VOC, the enhancement of the proposedadaptive VOC is experimentally proved. The originality of this paper is the building of a VSRmodel including VSR nonlinearity that is suitable for implementation with ADALINE. This leadsto ease of implementation and accurate identification.

Publication in the conference proceedings of EUSIPCO, Lisbon, Portugal, 2014

International audience; The objective of this study is to minimize the classification cost using Support Vector Machines (SVMs)Classifier with a double hinge loss. Such binary classifiers have the option to reject observations whenthe cost of rejection is lower than that of misclassification. To train this classifier, the standard SVMoptimization problem was modified by minimizing a double hinge loss function considered as a surrogateconvex loss function. The impact of such classifier is illustrated on several discussed results obtained withartificial data and medical data.

This paper presents an original method, based on artificial neural networks, to reduce the torque ripple in a permanent-magnet non-sinusoidal synchronous motor. Solutions for calculating optimal currents are deduced from geometrical considerations and without a calculation step which is generally based on the Lagrange optimization. These optimal currents are obtained from two hyperplanes. The study takes into account the presence of harmonics in the back-EMF and the cogging torque. New control schemes are thus proposed to derive the optimal stator currents giving exactly the desired electromagnetic torque (or speed) and minimizing the ohmic losses. Either the torque or the speed control scheme, both integrate two neural blocks, one dedicated for optimal currents calculation and the other to ensure the generation of these currents via a voltage source inverter. Simulation and experimental results from a laboratory prototype are shown to confirm the validity of the proposed neural approach. CPER Région Alsace 2007-2013

Le travail présenté dans ce mémoire concerne l'élaboration d'une stratégie complète d'identification et de commande neuronale d'un filtre actif parallèle (FAP). L'objectif visé est l'amélioration des performances par rapport aux systèmes classiques de dépollution des installations électriques basse tension. Basée sur l'utilisation des techniques neuromimétiques, notre approche de compensation des harmoniques se fait en trois étapes. Les deux premières étapes identifient respectivement les composantes de la tension et les courants harmoniques à l'aide de réseaux de neurones du type Adaline. La troisième étape injecte les courants harmoniques dans le réseau électrique par un module de commande à base de réseaux de neurones multicouches. Plusieurs architectures neuronales ont été développées et comparées pour chacune des étapes. La structure proposée s'adapte automatiquement aux variations de la charge du réseau et donc aux fluctuations du contenu harmonique des perturbations. Elle permet également la compensation sélective des harmoniques et la correction du facteur de charge. Finalement, ces stratégies ont été validées sur un banc expérimental et leur aptitude à l'intégration matérielle a été testée en simulation.