The invention relates to a fabric weaving on-line defect detection method based on a line laser. The method includes the steps that a laser line or plane is shone onto a yarn and a needle, and light curtains formed by the yarn and the needle and the laser are interlaced to form bright spots; a camera collects the formed bright spots; threshold segmentation is conducted on each frame of image in aspecified sampling time to obtain the positions of the bright spots formed when the yarn and the needle are interlaced with the laser plane, and the center coordinates of the bright spots are extracted; the positions of the same bright spot in different images are sequenced to form a sequence, so that the motion trail of each single bright spot in the sampling time is obtained, and bright spot one-dimensional motion signal graphs are obtained; the bright spot one-dimensional motion signal graphs are transformed into two-dimensional spectrograms through wavelet transformation; deep learning isconducted through a CNN, and two-dimensional spectrograms corresponding to normal yarn motion laws and two-dimensional spectrograms corresponding to abnormal conditions are marked respectively for training and learning.