Objective To improve the efficiency of cement production and prevent production stoppages caused by major mechanical equipment failures fault diagnosis for roller presses in cement production lines is essential. Methods A convolutional residual Transformer coding network was proposed. The first part of this model was a convolution block containing four convolutional layers. Each convolutional layer contained convolution pooling batch normalization and activation functions to extract local features from the input data. The extracted features were then fed into a Transformer coding module after adding residuals. This Transformer coding module incorporated a multi-head attention mechanism and a feedforward neural network to further extract features from the input. After obtaining the encoded output the residuals were added again. The difference was finally classified through the fully connected layer and softmax layer to achieve fault diagnosis. Results A preliminary verification of the model usability was carried out on the public bearing data set of Case Western Reserve University and then sensors were installed on the roller press to collect vibration data of normal operation and fault operation as an empirical data set. The prediction accuracy on the Case Western Reserve University test set was 99. 75% while the identification accuracy on the roller press data test set was 96. 55% with a fault recall rate of 98. 32%. Conclusion The method of using convolution to extract fault data features and employing an attention mechanism to highlight important features enables effective assessment of the operational status of the roller press in actual production particularly achieving a higher recognition rate for fault states.