Bearing incipient fault diagnosis is very crucial to the timely condition-based maintenance (CBM), but it is also relatively difficult because of the faint feature and environmental interference. In this article, an adaptive optimized time-varying filtering-based empirical mode decomposition (TVF-EMD) (AO-TVF-EMD) is put forward to extract more explicit and abundant incipient fault features by optimizing the parameter combination in conventional TVF-EMD. First, a novel integrated indicator termed sparsity-impact measure index (SIMI) constructed from dispersion entropy (DE) and envelope characteristic frequency ratio (ECFR) is designed as the parameter selection criterion. This index fully guarantees the optimal sparsity and impact properties of the obtained modes simultaneously, considering both time- and frequency-domain characteristics. Simulation verified that SIMI can not only sensitively characterize the variation of fault feature but also is robust to the environmental disturbance. Subsequently, we employ the Salp swarm algorithm (SSA) to realize the accurate search of parameter combinations where the minimum SIMI is defined as the fitness function. Moreover, comparative analysis in two runs to failure tests of bearings demonstrate that AO-TVF-EMD is more effective to extract incipient fault features than the maximum weighted Kurtosis (MWK) optimized TVF-EMD, the conventional TVF-EMD with fixed parameter, and some other advanced signal processing methods, which highlights the superiority of the proposed method.

Aiming at the problems of poor self-adaptive ability in traditional feature extraction methods and weak generalization ability in single classifier under big data, an internal parameter-optimized Deep Belief Network (DBN) method based on grasshopper optimization algorithm (GOA) is proposed. First, the minimum Root Mean Square Error (RMSE) in the network training is taken as the fitness function, in which GOA is used to search for the optimal parameter combination of DBN. After that the learning rate and the number of batch learning in DBN which have great influence on the training error would be properly selected. At the same time, the optimal structure distribution of DBN is given through comparison. Then, FFT and linear normalization are introduced to process the original vibration signal of the gearbox, preprocess the data from multiple sensors and construct the input samples for DBN. Finally, combining with deep learning featured by powerful self-adaptive feature extraction and nonlinear mapping capabilities, the obtained samples are input into DBN for training, and the fault diagnosis model for gearbox based on DBN would be established. After several tests with the remaining samples, the diagnosis rate of the model could reach over 99.5%, which is far better than the traditional fault diagnosis method based on feature extraction and pattern recognition. The experimental results show that this method could effectively improve the self-adaptive feature extraction ability of the model as well as its accuracy of fault diagnosis, which has better generalization performance.

On the condition of variable speeds, the blade crack signal generated by the single-mode sensor is an inherent risk of being misled by noise. Besides, it is challenging to reach the required integrity of information through feature extraction in single analysis domain, leading to the loss of detection accuracy for blade cracks. Therefore, a blade crack detection method based on domain adaptation and autoencoder (AE) of multidimensional vibro-acoustic feature fusion is proposed to extract more complete fault information from multi-sensor and multi-analysis domains. This method extracts one-dimensional (1D) and two-dimensional (2D) features from the frequency domain and smoothed pseudo-wigner-ville distribution (SPWVD) spectrum of the vibro-acoustic signal. Firstly, the 1D and 2D features are extracted without supervision through Stacked Denoising Autoencoder (SDAE) and Convolution Denoising Autoencoder (CDAE), respectively. Then, the extracted multidimensional features are fused and classified through the fully connected network. Finally, for insufficient diagnosis accuracy and data imbalance under certain working condition models with variable speed and combined with inductive transfer mechanism and domain adaptation, the stable diagnosis model under specific speed is applied for transferring and learning the unsupervised networks of other models compensating the weak generalization ability and robustness of the model. The experimental results show that this method is characterized by higher accuracy and robustness compared with the diagnosis method of single-domain feature for the single-mode sensor. Moreover, the proposed migration mechanism can well solve the data imbalance under different working conditions and improve the generalization ability of the few-shot model.

The evolution of advanced sensing techniques and intelligent algorithms has significantly underpinned the growth of structural health monitoring and damage identification. Modern industry equipment like compressors, which are indispensable to the petrochemical and other process industries, usually operate under complex conditions including variable speed. The more vulnerable compressor components, such as the blades, are prone to diverse levels of damage over time. Existing research usually discusses the damage identification problem of blades under the Euclidean space, facing challenges in linking multi-source heterogeneous signals. This study introduces a novel approach, employing a graph-structured data-based method for identifying compressor blade cracks. It specifically focuses on variable rotating speed conditions, subsequently proposing an intelligent identification framework based on vibro-acoustic graph-structured data. Firstly, the affinity graphs made of vibro-acoustic damage signal are constructed to express the latent damage information beyond Euclidean space after Fourier transform and residual learning-based feature extraction for one-dimensional data. Then the developed multi-order graph convolutional network and domain discriminator layers are used to extract the domain-invariant damage features, which will be fed into the linear layer for class prediction. Finally, the method's efficacy is cross-verified through experiments with actual measurements on a compressor platform, specifically focusing on variable rotating speed cases.