The invention provides a data fusion method based on a multichannel sensor for metal crack monitoring. The method includes the steps of: extracting a pulse in a manner of sliding time window, performing initial status recognition to the pulse in a manner of support vector data description to determine whether the pulse is a failure pulse or not; and performing evidence synthesis respectively to the recognition results of all failure pulses of different detection points within available time by means of an evidence theory method, thereby finally obtaining a recognition result. The method effectively avoid a problem of false alarm caused by incomplete information due to single detection point and single time point, thereby effectively reducing the rate of false alarm and improving accuracy of metal crack recognition. The method satisfies the on-line diagnosis demand in practical engineering application.