Comparative Analysis of Wavelet and MFCC Features, and Machine Learning Techniques for the Robust Classification of Seismic Signals

The seismic signal classification has numerous real-time applications for objects. Human and animal footsteps identification is crucial in order to avoid false alarms in security systems which can be costly and disruptive. This study focuses on the exploitation of frequency domain features and employs various machine learning methods to achieve robust detection of seismic signals. The dataset is acquired in a rural area farm and contains the ground vibration of five animals and six humans during multiple events, e.g., running and walking. Non-imaging sensors are used in data acquisition yielding a low-cost and reliable solution. A comparative analysis of features extracted by wavelet packet decomposition (WPD) and mel-frequency cepstral component (MFCC) has been performed. The mean and standard deviation of WPD and MFCC are calculated to avoid the curse of dimensionality. Individual signatures of humans and animals are differentiated by applying support vector machines, random forests, and Naïve Bayes. Random forest achieved a 91.6% F1-score outperforming other algorithms. The study can be utilized in wildlife monitoring, animal intrusion, predator animals’ recognition, and other applications beyond security systems. The robustness of these algorithms is limited to the size of the dataset, variability of reverberant environmental, and geological conditions.