The invention discloses a three-dimensional convolution neural network based universal non-reference video quality evaluation algorithm. The method includes two parts comprising a 3DConvNets model training phase and a quality evaluation phase. Compared with the prior art, the algorithm provided by the invention can evaluate multiple distortion types accurately and can keep stable evaluation accuracy and good performance stability when being applied to different video bases. Besides, the quality evaluation calculation complexity is distinctively lower than a current mainstream non-reference video quality evaluation algorithm and the evaluation speed is higher.