The objective of this thesis work was to build a deep neural network capable of distinguishing the two main gait phases based on electromyographic recordings. The developed network makes it possible to perform a step analysis using only the surface electromyography signals of the legs, without resorting to photogrammetric systems or other devices for the detection and reconstruction of the patient’s movement. The data was collected by 30 walking healthy subjects aided by an exoskeleton. EMG signals were detected on the soleus, femoral biceps, anterior tibial and broad lateral muscles of the dominant leg. To distinguish the gait phases, the angle of rotation of the hip was reconstructed using a system consisting of two wearable IMU. The exercise is done by walking back and forth along a pair of parallels and on a treadmill. In the first case, in addition to the contact and flight phases of the step, the net was also trained to distinguish the phase of upright posture corresponding to the reversal of the direction of travel during which the subject is passively rotated by the experimenter, with accuracy greater than 80%. Using the data collected while walking on a treadmill (constant and non-stop exercise with only two classes: swing and stance), the network accuracy exceeds 90%. In this work the network has learned from a few noisy data; in the future new recordings, made even without exoskeleton, will increase the accuracy of the network, allowing also, potentially, to predict the angles of rotation of the joints of the legs.
L'obiettivo di questo lavoro di tesi è stato il costruire una rete neurale profonda in grado di distinguere le due principali fasi del passo durante la camminata sulla base di registrazioni elettromiografiche. La rete sviluppata consente di effettuare un’analisi del passo utilizzando solo i segnali di elettromiografia di superficie delle gambe, senza ricorrere a sistemi fotogrammetrici o altri dispositivi per il rilevamento e ricostruzione del movimento del paziente. I dati sono stati raccolti facendo camminare 30 soggetti sani aiutati da un esoscheletro. I segnali EMG sono stati rilevati sui muscoli soleo, bicipite femorale, tibiale anteriore e vasto laterale della gamba dominante. Per distinguere le fasi del passo, è stato ricostruito l’angolo di rotazione dell’anca utilizzando un sistema costituito da due IMU indossabili. L’esercizio viene svolto camminando avanti e indietro lungo una coppia di parallele e su tapis roulant. Nel primo caso, oltre alle fasi di contatto e di volo del passo, la rete è stata allenata anche a distinguere la fase di postura eretta corrispondente alla inversione della direzione di marcia durante la quale il soggetto viene rotato passivamente dallo sperimentatore, con accuratezza superiore all’80%. Utilizzando i dati raccolti durante la camminata su tapis roulant (esercizio costante e senza pause caratterizzato da sole due classi: volo e appoggio), l’accuratezza della rete supera il 90%. In questo lavoro la rete ha appreso da pochi dati rumorosi; in futuro nuove registrazioni, realizzate anche senza esoscheletro, potranno aumentare l’accuratezza della rete, consentendo anche, potenzialmente, di predire gli angoli di rotazione delle articolazioni delle gambe.
Classificazione automatica delle fasi del passo basata su segnali sEMG
IMBALZANO, ALESSANDRO
2021/2022
Abstract
The objective of this thesis work was to build a deep neural network capable of distinguishing the two main gait phases based on electromyographic recordings. The developed network makes it possible to perform a step analysis using only the surface electromyography signals of the legs, without resorting to photogrammetric systems or other devices for the detection and reconstruction of the patient’s movement. The data was collected by 30 walking healthy subjects aided by an exoskeleton. EMG signals were detected on the soleus, femoral biceps, anterior tibial and broad lateral muscles of the dominant leg. To distinguish the gait phases, the angle of rotation of the hip was reconstructed using a system consisting of two wearable IMU. The exercise is done by walking back and forth along a pair of parallels and on a treadmill. In the first case, in addition to the contact and flight phases of the step, the net was also trained to distinguish the phase of upright posture corresponding to the reversal of the direction of travel during which the subject is passively rotated by the experimenter, with accuracy greater than 80%. Using the data collected while walking on a treadmill (constant and non-stop exercise with only two classes: swing and stance), the network accuracy exceeds 90%. In this work the network has learned from a few noisy data; in the future new recordings, made even without exoskeleton, will increase the accuracy of the network, allowing also, potentially, to predict the angles of rotation of the joints of the legs.È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/15343