Low Resource Algorithms for Abnormal Instances Detection in the Internet of Things Framework

Electrocardiography (ECG) signals are widely used to appraise the health of the human heart. The resulting time series signal is visually analyzed by a cardiologist to detect any unusual beat, which could be referred to a series of abnormal behavior going from a motion artifact to arrhythmia that the patient may have suffered. In this dissertation, a low-complexity and low-resources method for anomaly detection based on principal component analysis is presented. The weightlessness of the proposed algorithm well- fits implementation at low resources devices, with a view to edge computing. In particular, energy of the signal is observed in a properly chosen subspace, and normal processes are discriminated from abnormal ones, based on their accordance with certain criteria. The evaluation of the method is firstly conducted on synthetic ECG, and performance are compared with Literature's related works. The efficacy of the algorithm is finally proved through a qualitative evaluation on ECG signals coming from real patients. Addressing the particular case of motion artifact detection, this dissertation proves how, despite the extremely low computations costs, the method would help in reducing diagnosis time, by reducing to the minimum false alarms.