Localized muscle fatigue is of major kinematic and medical significance in the field of sports and rehabilitations. However, only few works reported fatigue recognition for field training, and via analysis of only the noise-susceptible electromyography. In this study, along with portable EMG sensors, a wearable smart arm band powered by fabric strain sensors was implemented to detect the thickness variation of biceps. Dumbbell curl training protocol was designed, and 32 subjects were enrolled, contributing to a database of 580 curl samples. Fatigue-related characteristics of muscle thickness were proposed according to muscle physiology and combined with that of sEMG signals, forming feature vectors of the samples. Clustering results showed that the fatigued cycles can be successfully recognized and labeled. For potential applications, fatigue prediction models based on supervised learning methods were further proposed and compared, of which the SVM-based model was observed with satisfactory overall accuracy of 83.3%, yet the fatigue samples can be predicted at much higher rate, with effective recall 90%, F1-SCORE 95% and AUC 98.7%. This work not only exploits feasibility of using latent physiological indicators such as muscle thickness in muscle fatigue monitoring, the result and methodologies will also inspire wider horizon of human-centered applications using novel flexible sensors.