On-line Dynamic Sparse Recovery of Streaming Signals from Correlated Multiple Measurements

Blocking effects are inevitably introduced when compressed sensing is applied in the multi-task observation of streaming signals in time domain. To eliminate the blocking effects, the underlying correlation between observation tasks is exploited in this paper, and an on-line dynamic recovery algorithm called Stream-TMSBL is proposed based on the lapped orthogonal transform (LOT) and the traditional TMSBL algorithm. In the proposed algorithm, the traditional LOT scheme is extended from single-task case to multi-task case, where a sliding window system is established for the online multi-task observation of the streaming signals in time domain, and by Bayesian probabilistic modeling for the system, the traditional TMSBL algorithm is fused such that the information in both dimensions of tasks and time is utilized to improve the performance of signal reconstruction. Experiments based on engineering measured data showed that the averaged running speeds of the proposed algorithm and the traditional TMSBL algorithm are similar, while the accuracy and success rate of the proposed algorithm is much higher that of the traditional TMSBL algorithm.