Rolling bearings play a vital role in the overall operation of rotating machinery. In practice, many learning methods for variable-speed fault diagnosis ignore task-specific decision boundaries, making it very difficult to completely match feature distribution between different domains. Therefore, to overcome this problem, an adversarial domain adaptation of asymmetric mapping with CORAL alignment is presented. The asymmetric mapping feature extractor is able to extract more specific-domain features with obvious distinction. Meanwhile, combining the maximum classifier discrepancy of deep transfer to give an adversarial approach and taking the task-specific decision boundaries into account, class-level alignment between the features of the source domain and target domain can be attempted. To prevent degenerate learning, which is possibly caused by asymmetric mapping and adversarial learning, the model is constrained by deep CORAL alignment to extract more domain-invariant features. Experimental results show that the proposed method can solve the variable-speed (a small span of intermediate vehicle speeds) fault diagnosis problem well, with high transfer accuracy and strong generalization.

In recent years, transfer learning has become more and more favored by scholars from all walks of life. At present, although transfer learning has achieved certain results in the field of fault diagnosis, the use of transfer learning alone may lead to poor transfer effects or even negative transfer due to the sample gap being under variable conditions in the same machinery. Therefore, deep domain adaptation with adversarial idea and coral alignment (DAACA) is proposed in this paper in order to solve the problem. DAACA is briefly summarized below. The domain adaptation with adversarial idea is added on the basis of transfer learning. The deep coral is then appended to further reduce the distribution difference between the data from the source and the target domain, which improves the invariant features of adversarial domain adaptation learning. In addition, a gradient reversal layer is introduced in the method to achieve gradient reversion and avoid the adversarial disadvantage of fixing parameters separately. It can be seen from the experimental results that the DAACA can not only solve the problem caused by the sample gap in variable conditions, but also achieve higher diagnosis accuracy and generalization ability.

Deep learning has been widely used in the field of fault diagnosis due to its excellent performance in feature extraction and has been gradually applied to solve various problems in fault diagnosis. Convolutional neural networks and transfer learning networks have been gradually employed to solve the problems of sample imbalance and domain adaptation under variable rotational speeds in fault diagnosis. However, there are still some weaknesses in current research. Firstly, sample imbalance in fault diagnosis is accompanied by the domain-adaptive problem in practice. Secondly, transfer learning cannot extract domain-invariant features that do not change with the rotational speed, which leads to poor diagnosis results when it is applied to other working conditions without transferring the rotating speed. To solve the above problems, an architecture named a convolutional transfer feature discrimination network (CTFDN) is proposed in this paper. It uses the scaled exponential linear unit activation function, and the main body adopts a two-branch network architecture of weight sharing, which mainly includes a deep feature extraction network for fusion feature extraction, a transfer learning network for domain-adaptive problems, and an unbalanced-sample feature discrimination network for feature similarity determination. Finally, two fault datasets are used to test the validity of the proposed model. The results show that the CTFDN model can extract domain-invariant features well in unbalanced fault sample information, and still has good diagnostic accuracy under variable rotational speeds.

As fault diagnosis for motor drive systems enters the era of big data, intelligent fault diagnosis methods exhibit excellent performance because of their learning capabilities. However, the existing methods are strict with the signal size, which reduces the performance of these methods for modern condition monitoring. Besides, most existing intelligent methods have a great limitation: the training data and testing data are under the same working conditions. To overcome these limitations, we propose a novel three-layer model inspired by a recurrent neural network (RNN) and transfer learning, which has the ability to process variable size sequences under different working conditions. In the first layer, the input unit is extended to ensure that there is adequate dimension to store the information of sequences. In the second layer, the main information of the whole sequence is stored, transmitted, transformed and output by gates. In the final layer, softmax is employed to classify the health conditions based on the output of the RNN with a long short-term memory cell. The classification loss based on the whole framework and the domain loss using kernel method are proposed to train the model. Furthermore, a bearing dataset is adopted to verify the effectiveness of the proposed method. The experimental results show that the proposed method can not only break the limitations of existing methods, but also achieve a superior performance compared with related methods.