A least square approach for bidimensional source separation using higher order statistics criteria

The anomaly detection based on processing of distributed temperature sensors data is a new research problem. The acquired data is highly influenced by the response of the ground in which the sensors are buried. It therefore becomes essential to remove the influence of this response. This response, being the most coherent factor in the acquired signal, appears as the most energetic source vector. However, its classical estimation by SVD runs the risk of taking into account energetic phenomena like precipitations. We propose to characterize such phenomena using higher order statistics thus giving a criteria of selecting only the data not influenced by such phenomena. An overlapping window approach then allows estimation of characteristic ground response source. Moreover, the corresponding ground response subspace is constructed by least squares based unmixing approach on the characteristic source. This avoids also the physically unjustifiable orthogonality condition of temporal variations of the estimated sources imposed by SVD.