The invention discloses a hydropower unit fault diagnosis method based on convolution neural network information fusion. The method comprises the steps of: obtaining stability data at low, medium andhigh rotation speeds in the start-up process of a unit; respectively pre-processing the unit data at low, medium and high rotation speeds; respectively dividing the unit data at low, medium and high rotation speeds into a training set and a test set of convolution neural networks at corresponding rotation speeds; drawing a unit axis trajectory chart by respectively using the unit data at low, medium and high rotation speeds, furthermore, converting an axis trajectory image into a grey-scale image, and forming a matrix by contrasting the remaining information to be fused with the axis trajectory grey-scale image; respectively establishing the convolution neural networks by using the unit data at low, medium and high rotation speeds; respectively training the corresponding convolution neuralnetworks by using the unit data at low, medium and high rotation speeds; and, performing unit fault diagnosis by using the trained three networks, and fusing diagnosis results, so that the final conclusion is formed. By means of the hydropower unit fault diagnosis method based on convolution neural network information fusion in the invention, the hydropower unit fault diagnosis accuracy can be effectively improved.