A wealth of structural health monitoring (SHM) data has been collected from the grand civil structures instrumented with long-term SHM systems. Nevertheless, the mining of information associated with the structural condition from the big data is still a great challenge. The signal processing is the first and essential step to exploit potentialities of the large SHM data. Collected by sophisticated systems from large-scale civil structures, which operate interacting with intricate loadings and environment, long-term SHM data are usually non-stationary and contaminated by noises, spikes and trends. Moreover, different signal sources are desired to be separated in some specific researches. Consequently, there is an immediate necessity to process the signals, in terms of de-noising, de-spiking and decomposing. Especially, for the large data that has an extraordinarily large volume automation and efficiency are particularly important. With the merit of time-frequency analysis and multi-resolution, the wavelet transform is a promising tool to process the long-term SHM data. A fast and unsupervised signal processing scheme is proposed in this paper to remove the noises, spikes, and trends embedded in the signals, and to separate different signal sources as well. To improve the computational speed, the algorithm is designed to be as simple as possible, so the procedures like constructing templates and searching anomalies from the beginning to end of a long signal have been avoided. The proposed methodology is demonstrated by applying to the long-term strain and displacement data measured from a long-span suspension bridge.