Chen Yang

Deep neural networks, represented by convolutional neural networks (CNNs), have been widely used in fault diagnosis. In recent years, frequency-domain modeling can complement spatial-domain convolution by capturing global spectral structures. In this work, a global feature learning module based on the image Fourier transform, termed deep frequency-domain adaptive filtering (DFAF), is proposed. We first observe that the time–frequency images of mechanical vibration signals tend to exhibit low-frequency components along the time direction, leading to a vertical concentration of spectral energy. The corresponding theoretical basis is given. Then, a frequency filter with a learnable bandwidth is designed to generate a soft, smooth mask that encourages the model to focus on frequency components beneficial to noise robustness diagnosis. The proposed DFAF module is simple, lightweight, and plug-and-play. It provides a nonlocal receptive field and enables multiscale information without substantially increasing the parameter count. By inserting the DFAF module into a CNN, we construct a spatial-frequency fusion convolutional (SFConv) network and evaluate it on two rotating machinery fault datasets. At signal-to-noise ratio = −10 dB, SFConv exceeds the best comparison methods by 6.68 and 3.26 percentage points.

Vibration monitoring using feature parameters like kurtosis is a key early warning indicator for assessing gearbox health in industrial equipment. However, gearboxes in heavy industries, such as port cranes and steel rolling mills, face complex operating conditions with frequent reversing, start-stop cycles, and random external impacts. The vibration signals from these gearboxes contain considerable random impulse interference, impacting the accuracy of condition monitoring based on feature parameters. To address this issue, a strategy for detection and suppression of random impulse interference is proposed. Firstly, the L-kurtosis and variance-weighted interquartile range (LKVIQR) impulse detection indicator is constructed, and then the random impulse interference detection method based on the ratio of short-time statistical characteristics to long-time statistical characteristics (STSC/LTSC) is proposed. Further, a morphological top-hat averaging operation and a method for determining the length parameter of structural elements (SEs) based on maximizing L-kurtosis are proposed to remove the detected random impulse interference. Finally, the validity of the proposed method is verified through laboratory data and vibration data of the ship unloader in the industrial field. The results indicate that the proposed method effectively reduces the kurtosis index of gearbox vibration signals under random impulse interference, which demonstrate that the proposed method can accurately identify and eliminate random pulse interference in gearbox vibration signals, offering a novel approach to filtering out random impacts during the real-world operational processes of industrial equipment.

A gearbox is an indispensable component of a mechanical transmission system, and in the event of an anomaly, it can have a serious impact on the stability of the system. Graph convolutional networks have achieved some results in the study of fault diagnosis, but the way of topology construction from a single metric space is somewhat inexact and one-sided. Meanwhile, graph convolutional networks can only aggregate node features in the spatial dimension, which does not guarantee a comprehensive expression of fault information. Therefore, we propose a gearbox fault diagnosis method based on refined topology and spatio-temporal graph convolutional networks to accurately identify the gearbox health state. First, two topologies are constructed based on the distance metric and the similarity metric, respectively, and the edge structure is pruned to obtain a more lightweight topology based on the analysis of the structural similarity between the two. Then, depending on the number of common neighbors of two nodes, the weights between them are enhanced or weakened to further refine the topological information. Finally, a network model with spatio-temporal feature co-extraction is developed to mine fault features in depth in two dimensions to improve network performance. Case studies, ablation experiments, and robustness validation of the proposed method are carried out, and the results show that the accuracy of the proposed method is as high as 99.41% and 99.93% on the two datasets, respectively, which is much higher than that of other comparison methods, and at the same time, it has good noise robustness.

Nowadays, deep learning has been widely applied to estimate the hard-to-measure quality variables in industrial processes due to its excellent non-linear modeling and feature extraction capabilities. Stacked isomorphic autoencoder (SIAE) is one of deep learning methods, which minimizes the reconstruction error of the original input variables during the pre-training phase, extracting features from low to high layers. Despite this, the features extracted in this manner, in fact, disregard the information associated with the quality variables, which have an impact on the effectiveness of the succeeding supervised fine-tuning phase. Besides, only the features extracted by the last IAE are used to predict the quality variables, which leads to poor prediction results. Therefore, a hybrid stacked isomorphic autoencoder (HSIAE) is developed in this paper to implement gated quality-driven feature extraction. During the hierarchical pre-training phase, HSIAE can extract more features associated with quality variables by incorporating the error of predicting quality variables into the loss function of SIAE. Subsequently, gated units leverage the different hidden features to the fullest and quantifying their contributions. Finally, the feasibility and effectiveness of the proposed HSIAE model are validated with debutanizer column dataset.

The stochastic resonance (SR) in a bistable system driven by nonlinear frequency modulation (NLFM) signal and strong noise is studied. Combined with empirical mode decomposition (EMD) and piecewise idea, an adaptive piecewise re-scaled SR method based on the optimal intrinsic mode function (IMF), is proposed to enhance the weak NLFM signal. At first, considering the advantages of EMD for dealing with non-stationary signals, the segmented NLFM signal is processed by EMD. Meanwhile, the cross-correlation coefficient is used as the measure to select the optimal IMF that contains the NLFM signal feature. Then, the spectral amplification gain indicator is proposed to realize the adaptive SR of the optimal IMF of each sub-segment signal and reconstruct the enhanced NLFM signal. Finally, the effectiveness of the proposed method is highlighted with the analysis of the short-time Fourier transform spectrum of the simulation results. As an application example, the proposed method is verified adaptability in bearing fault diagnosis under the speed-varying condition that represents a typical and complicated NLFM signal in mechanical engineering. The research provides a new way for the enhancement of weak non-stationary signals. This article is part of the theme issue ‘Vibrational and stochastic resonance in driven nonlinear systems (part 1)’.