In recent years, an increasing number of people suffer from lifestyle diseases, such as metabolic syndrome. Eating habits monitoring is an important parameter for life-style diseases prevention. Since, we have been developing a system to monitor meal time activities. Our system consists of two bone conduction microphones connected to a portable IC recorder that collects internal body sound data. In the meal time activities differentiation process, we adopted a wavelet function for the feature extraction. We extracted 70 feature vectors from coefficients of discrete wavelet transformation, then we selected the optimal feature vectors set using minimal-redundancy-maximal-relevance criterion (mRMR), and finally we used probabilistic neural network (PNN) to classify meal-related activities. Experiments were carried on sound data from six persons. Our model proved to achieve better classification accuracy, and selected features to be independent from individual differences.