The shortcomings of fault diagnosis methods based on deep convolutional neural networks is that， tensor data is easily destroyed when reducing the dimension of high-order input tensors by pooling layers， which results in a loss of data information， and the relatively complex network structure. Therefore， a Deep TensorProjection Networks method is constructed via replacing the pooling layer in the traditional CNN by a TensorProjection Layer. The TensorProjection Layer reduces the dimensionality of in‐ put high-order tensor data without causing damage to the data， thus avoiding the impact of the loss of feature information， and greatly improving the recognition accuracy of the model. The dimensionality of the TensorProjection Layer used for dimensionality reduction is variable， thus simplifying the networks structures. Based on this， combined with the respective advantages of high-or‐ der spectrum and deep TensorProjection networks， a mechanical fault diagnosis method based on deep TensorProjection networks is proposed. In the proposed method， the feature of fault signal is extracted by high-order tensor spectrum， which is input into the constructed model for reducing high-order tensor dimensionality and identifying faults. The proposed method is applied to diagnose gearbox faults. Experimental results show that the proposed method can better retain the original fault information and effectively recognize the different types of faults. And the accuracy is better than traditional deep convolutional neural network fault diagnosis methods