The invention discloses a method for monitoring tool wear on line based on multiscale principal component analysis. The method comprises the steps of dividing an original data set into training samples and test samples, decomposing the training samples into a plurality of scales through wavelet decomposition, and carrying out dimension reduction processing on all the scales and the whole by virtue of principal component analysis to establish models; carrying out wavelet decomposition on the test samples according to the layer numbers of the training samples, loading data with all the scales to principal component models with corresponding scales, and calculating the statistics of SPE and T<2> of the data with all the scales; and enabling signals with remarkable scales to form a new test sample by using a wavelet reconstruction method, loading the new test sample on a whole principal component model, and monitoring the tool wear state through calculating the statistics of SPE and T<2> and judging whether the statistics of SPE and T<2> exceed a control limit or not, wherein if the statistics of SPE and T<2> exceed the control limit, the tool wear is shown to be abnormal, and if the statistics of SPE and T<2> do not exceed the control limit, the tool wear is shown to be normal. By using the method, the accuracy rate of online tool wear identification in the machining process is increased; and particularly, the method has high adaptability for complex nonlinear signals existing in difficultly-machined materials.