The invention provides a multivariate time series-based train brake system fault detection method. The method comprises the following steps: sample data relevant to train brake system fault detectionis collected, a multivariate time series matrix for the sample data is built, sample data for time series segments is extracted from the multivariate time series matrix via use of a sliding time window, the extracted sample data for the time series segments and characteristic data of an abnormal mode are subjected to matching detection, and train brake system fault detection results in the time series segments can be obtained according to matching detection results. The method disclosed in the invention is based on data analysis, machine learning is combined with a multivariate time series excavation algorithm, a sliding time window-based abnormal mode matching algorithm is put forward, faults in existing data can be subjected to monitoring and intelligent diagnosis via mode matching, intrinsic reasons behind exception can be accurately found, and the exception can be well positioned.