<p>Localized muscle fatigue is an exercise-induced decline in the force-generating capacity of muscles. The patients with severe motor disabilities, such as tetraplegics, can use their facial muscles to control assistive devices. However, repeated use of these muscles often leads to fatigue which in turn alters the characteristics of facial EMG. The facial EMG signals are stochastic, nonstationary, and multicomponent, and their characteristic changes under fatiguing contractions are not yet well established. In this work, facial EMG signals are recorded from the left and right frontalis muscles of fifty healthy subjects under a standard experimental protocol. The first and last six-second segments of the signals correspond to nonfatigue and fatigue conditions, respectively. These signals are preprocessed and decomposed using a three-level maximum overlap discrete wavelet packet transform (MODWPT). The first two different wavelets are employed for the decomposition, namely Daubechies-4 (db4) and discrete Meyer (dmey). In order to quantify the time-scale representations, wavelet energy is extracted from all scales. Finally, these features are used to develop a multilayer perceptron (MLP) network for detecting the fatigue state. MODWPT is capable of representing the nonstationary and multicomponent variations of facial EMG under both fatigue and nonfatigue conditions. Wavelet energy is higher in the nonfatigue state across all scales and both the wavelets (p &lt; 0.05). The MLP based on db4 achieves a maximum accuracy of 94.35% with a sensitivity of 94.9% and a specificity of 93.8%. The performance of dmey is slightly lower than that of db4. This study shows that the proposed MODWPT features combined with an MLP network can accurately detect fatiguing contractions of facial muscles and could be readily implemented in the human–machine interface. It also appears that the proposed approach could also be extended to detect the fatigue in musculoskeletal disorders and sports applications.</p>