Abstract

Regular physical activity can help people improve both physical and mental health. As people pay more attention on their health, fitness has become a popular activity. For individuals who have specific training goals, such as losing fat, gaining weight, and preparing to participate in competitions, it is important to avoid injury during exercise and improve the efficiency of training. Physiological monitoring during exercise, such as heart rate, blood lactate, oxygen uptake, and tissue oxygenation, is helpful for improve the effectiveness and safety of training. Many researchers are devoting their efforts to propose methodology and invent instruments of measuring these metrics. It is well known that heart rate is a commonly used measurement indicator, which can be measured on many fitness equipment and wearable devices. However, heart rate is a global parameter of the trainer's body and cannot represent the training intensity of a specific muscle. Because the muscles are directly affected by the exercise, it is necessary to measure the metrics of muscles to determine whether a specific muscle can tolerate the exercise load or not. If the muscles are overworked, there is a high probability of injury, which must be avoided. Hence, it is imperative to measure local muscles.

Muscle oxygen saturation (SmO2) is an indicator of the altering between oxygen delivery and consumption in the muscles. The more intense the exercise, the more oxygen is consumed by the muscles. So SmO2 is a good indicator to assess how fatigued a specific muscle is. In sports science, it is usually measured non-invasively by near-infrared spectroscopy (NIRS). Many instruments were developed by researchers previously based on different NIRS techniques and algorithms. In this thesis, a methodology of measuring absolue value of SmO2 was proposed for a wearable measurement device with one source and two detectors using 5-wavelength NIRS. The algorithm of fitting the light attenuation to the Taylor expansion model by bound-constrained non-linear least squares fitting was evaluated with simulated tissues. For in vivo measurement, an orthogonalization technique was introduced to reduce the effect of the absorption and scattering of overlying tissues. With comparison and analysis, the measuring SmO2 values of two designed running procedures were reasonable.

During exercise, the trainer may not always wear the device correctly and the device may move or fall off. And some of the individuals may not exercise as the designed training procedures. So the measured data will be unreliable in these cases. In order to remind users to wear the device properly and to train as the designed procedures, a fault diagnostic method was proposed by machine learning approach in this thesis. With labelling data by its reliability and splitting data into different training status, a support vector machines (SVM) model with Gaussian radial basis function kernel was trained. According to two evaluation curves, ROC curve and cross-validation learning curve, the SVM classifiers in both training states can achieve an accuracy of over 97%. These trained models can be applied as a fault diagnostic for the measurement device.

There is no screen on the device, so the results need to be displayed on a computer or mobile phone. In this thesis, an application that integrated the SmO2 calculation and fault diagnostic was developed in Matlab App Designer. With this application, after three clicks by users, the SmO2 curve during training and the absolute values could be displayed in the interface. Since the device didn't have the feature of real-time wireless transmission, a simulation of real-time mode was done to show the possibility of real-time measurement in the future.