Accidental failures of rotating machinery components such as rolling bearings may trigger the sudden breakdown of the whole manufacturing system, thus, fault diagnosis is vital in industry to avoid these massive economical costs and casualties. Since convolutional neural networks (CNN) are poor in extracting reliable features from original signal data, the time-frequency analysis method is usually called for to transform 1D signal into a 2D time-frequency coefficient matrix in which richer information could be exposed more easily. However, realistic fault diagnosis applications face a dilemma in that signal time-frequency analysis and fault classification cannot be implemented together, which means manual signal conversion work is also needed, which reduces the integrity and robustness of the fault diagnosis method. In this paper, a novel network named WPT-CNN is proposed for end-to-end intelligent fault diagnosis of rolling bearings. WPT-CNN creatively uses the standard deep neural network structure to realize the wavelet packet transform (WPT) time-frequency analysis function, which seamlessly integrates fault diagnosis domain knowledge into deep learning algorithms. The overall network architecture can be trained with gradient descent backpropagation algorithms, indicating that the time-frequency analysis module of WPT-CNN is also able to learn the dataset characteristics, adaptively representing signal information in the most suitable way. Two experimental rolling bearing fault datasets were used to validate the proposed method. Testing results showed that WPT-CNN obtained the testing accuracies of 99.73% and 99.89%, respectively, in two datasets, which exhibited a better and more reliable diagnosis performance than any other existing deep learning and machine learning methods.

Over the past few years, deep learning–based techniques have been extensively and successfully adopted in the field of fault diagnosis. As the diagnosis tasks become more complicated, the structure of the traditional convolutional neural network (CNN) has to become deeper to deal with them, while the gradient of fault features may vanish within the deep network. In addition, all the features are treated equally in the traditional CNN, which cannot make the most of the representation power of CNN. Here, we proposed a method named dual attention dense convolutional network to handle these issues, which is constructed by the dense network and the dual attention block. On one hand, the dense connections and concatenation layers can reinforce the propagation of fault features among layers and mitigate the vanishing gradient phenomenon in the deep network. On the other hand, as the features flow through the channel attention and spatial attention within the dual attention block, this attention mechanism can learn which feature to emphasize or suppress and then obtain the cross-channel and cross-spatial weights of the features. These weights can make the most of the abundant information, elevating the expressive power of network. After passing through these dense and attention blocks, the generated high-level features are then fed into the final classification layer to obtain diagnosis results. The effectiveness of the dual attention dense convolutional network is validated by eight datasets of spindle bearings under various machinery conditions. Compared with eight other approaches including support vector machines, random forest, and six existing deep learning models, the results indicate that the proposed dual attention dense convolutional network possesses higher accuracy, fewer parameters and computations, and faster convergence under complex operational conditions.

Modern machinery becomes more precious with the advance of science, and fault diagnosis is vital for avoiding economical losses or casualties. Among massive diagnosis methods, deep learning algorithms stand out to open an era of intelligent fault diagnosis. Deep residual networks are the state-of-the-art deep learning models which can continuously improve performance by deepening the network structures. However, in vibration-based fault diagnosis, the transient property instability of vibration signal usually calls for time–frequency analysis methods, and the characters of time–frequency matrices are distinct from standard images, which brings some natural limitations for the diagnosis performance of deep learning algorithms. To handle this issue, an enhanced deep residual network named the multilevel correlation stack-deep residual network is proposed in this article. Wavelet packet transform is used to preprocess the sensor signal, and then the proposed multilevel correlation stack-deep residual network uses kernels with different shapes to fully dig various kinds of useful information from any local regions of the processed input. Experiments on two rolling bearing datasets are carried out. Test results show that the multilevel correlation stack-deep residual network exhibits a more satisfactory classification performance than original deep residual networks and other similar methods, revealing significant potentials for realistic fault diagnosis applications.

Till now, deep learning–based intelligent diagnosis models combined with multisource information have become popular, but tough issues like multisource feature extraction and information redundancy may sacrifice the models’ representational power and result in a degraded performance. Aiming at the above problems, this paper proposed a novel model called multisource heterogeneous information selective fusion network (MHI‐SFN) for rolling bearing fault diagnosis. In MHI‐SFN, multisource heterogeneous signals were stacked together and directly fed into model and grouped convolution was adopted to replace standard convolution throughout the structure, enabling kernels to firstly focus on the feature extraction of every individual signal and then perform efficient feature fusion work as needed. Then, selective kernel modules were designed to adaptively assign suitable kernel sizes and selectively fuse the valuable information between different scales of feature map from different signal sources. Lastly, channel attention was introduced to adaptively alleviate the information correlation and redundancy between the extracted features. Compared with other multisource information–based methods, MHI‐SFH automatically solves the multisource feature fusion and information redundancy problems with its specially designed structure, avoiding complicated hand‐crafted signal processing steps and achieving a powerful end‐to‐end intelligent fault diagnosis. The proposed method was experimentally verified on two rolling bearing datasets, and the results proved the feasibility and superiority of the MHI‐SFN model.

Deep residual network (DRN) is a recently-developed powerful algorithm in the deep learning filed. This paper introduces the superiority of DRN into the fault diagnosis of rotating machinery for simplifying traditional diagnosing process as well as enhancing predicting performance. DRN can not only extract features automatically from raw or processed signals but also benefit from its especially deep architecture to continually improve representation capacity without worrying gradient divergence issues. The functions of DRN result from the unique structure design called residual building block, which will be described clearly with the network overall architecture in this paper. Additionally, the comparison of the DRN with other machine learning-based and neural network-based fault diagnosis methods are presented.