By analyzing shortages of current MSPCA model, an online multivariable statistical process monitoring method is proposed, which uses some concepts from online multiscale filtering and can be applied to sensor fault diagnosis. In the method, wavelet decomposition is employed to the signals using edge correction filter in a fixedlength data window, and then wavelet denoising is conducted with wavelet threshold filtering. Next, an online multiscale model is constructed for data combining wavelet transformation and adaptive PCA in the previous data window. This model avoids time waste in direct signal denoising and reduces time cost in multiscale data with conventional PCA, which eventually increases accuracy in fault diagnosis. Experiments on eight vibration signals of 6135D diesel engine under severe leak condition prove the practicability and feasibility of the proposed method.