It is difficult for rolling bearings to realize high-precision fault diagnosis with variable speed. To obtain the features of variable speed fault signal effectively and complete the classification work of high accuracy, robust local mean decomposition (RLMD), fractional hierarchical range entropy (FrHRE), hunter–prey optimization algorithm (HPO) and random forest (RF) are combined. Then the paper advances a model for fault diagnosis based on RLMD, FrHRE and HPO-RF. Firstly, RLMD is selected to reconstruct the signal to eliminate some noise interference in this paper. Secondly, FrHRE is chosen to extract the useful feature. Next step, HPO is used to optimize the important parameters of RF and enhance RF’s classification ability. Finally, these obtained features are imported into the optimized RFmodel to achieve the classification. The experimental data is provided by University of Ottawa. The experiment analysis demonstrates that the proposed method performs very well in classification.

As a critical part of the coal mine ventilation system, the main mine fan ensures fresh air supply and timely removal of harmful gases and dust. Monitoring and predicting its motor bearing vibrations is essential for early fault detection and operational safety. This avoids unplanned shutdowns of the main mine fan while ensuring equipment stability and production safety in coal mines. Unlike other mining equipment, the motor’s operation mode for the main mine fan is periodic, and a high background noise characterizes the working environment. Consequently, the vibration signals exhibit an alternating pattern with complex distribution between operational and shutdown states. To address these challenges, this study proposes a vibration signal prediction model named series decomposition-attention-enhanced temporal convolutional network(SD-ATCN). The model is built upon a temporal convolutional network (TCN) structure, enhanced with series decomposition (SD) and an improved convolutional block attention module (ICBAM). The SD-ATCN model first decomposes the vibration signals into low-frequency signals reflecting long-term trends and high-frequency signals capturing short-term variations, thus enhancing the model’s ability to handle long-term and short-term features. The attention-enhanced TCN model (ATCN) combined with the ICBAM is then utilized to extract features from the vibration time series data of the main mine fan bearings in coal mines. The model employs the ICBAM attention mechanism to locate the key features in the signal more accurately. Finally, temporal patterns and feature dependencies within the vibration signals are extracted by stacking multiple ATCN modules, further improving prediction accuracy. The publicly available intelligent maintenance systems (IMS) bearing dataset from the University of Cincinnati and an industrial field dataset collected from a coal mine fan are adopted for model validation. Experimental results demonstrate the superiority of the proposed SD-ATCN model to other baseline models in terms of MAE, MSE, R², and other evaluation metrics on both the IMS and industrial field datasets.

Deconvolution-related methods are the mainstream choice when it comes to enhancing the pulse impact of bearing fault and reducing noise interference. Kurtogram algorithm is used to optimize the minimum generalized Lp/Lq deconvolution to improve the nonconvexity of other optimization criteria. However, it has low computational efficiency and poor diagnostic accuracy under strong background noise. The paper proposes an optimized method using protrugram algorithm that combines fast iterative filter decomposition (FIFD) with minimum generalized Lp/Lq deconvolution (OMGD) for the 1.5-dimension Teager energy spectrum demodulation. Here is the specific process of the application: Fast iterative filtering (FIF) was used to reduce noise interference before using the maximum kurtosis to obtain the center frequency and frequency band and optimize the filter design, which was for the MGD initialization operation to prevent the result from falling into the local optimal solution and check the interference of impulse noise to a certain extent. The 1.5-dimension Teager energy spectrum was then used for demodulation analysis to extract small fault features of rolling bearings. The verification of simulation signals and actual data showed that this method was better in terms of extraction effect and efficiency than the use of fast kurtogram algorithm to optimize minimum generalized Lp/Lq deconvolution when it comes to extracting microfault features with high interference of background noise.

To address the performance degradation caused by the coupling of fault-sample scarcity and cross-domain distribution shifts in intelligent diagnosis of rotating machinery, scarce-sample cross-domain fault diagnosis for rotating machinery (SSCDFD-RM) has emerged as an important research topic. However, existing reviews are largely organized around individual paradigms, such as few-shot learning, transfer learning, or domain generalization, and a unified analytical framework is still lacking. To bridge this gap, this paper establishes a systematic review framework for SSCDFD-RM from two orthogonal perspectives: training-stage information availability and the source–target label-space relationship. First, according to the accessibility of target-domain information and the sufficiency of source-domain supervision, existing studies are categorized into three representative scenarios: few-sample target-domain cross-domain diagnosis, few-sample source-and-target-domain cross-domain diagnosis, and zero-sample domain-generalization-based cross-domain diagnosis. Closed-set, partial-set, open-set and universal-set settings are further introduced to characterize the source–target label-space relationships under each scenario. On this basis, representative methods are critically reviewed in terms of their core mechanisms, applicable conditions, failure modes, and application boundaries, and their common bottlenecks are further abstracted into several fundamental issues, including the reliability of limited supervision, statistical uncertainty under scarce samples, preservation of class-conditional discriminative structures, and generalization to unseen domains. Subsequently, reproducible methods are quantitatively evaluated on cross-working-condition and cross-machine tasks under unified task definitions, data partitions, sample budgets, and evaluation protocols. The benchmark further incorporates recent technical paradigms, including deep meta-transfer learning, large-scale time-series pretraining, self-supervised representation learning, and diffusion-based representation learning. The results show that cross-machine diagnosis is generally more challenging than cross-working-condition diagnosis, and the difficulty further increases when target-private unknown classes are involved. Moreover, the performance gains achieved by multi-source information fusion and domain generalization are strongly task-dependent, and increasing the number of source samples or source domains does not necessarily lead to monotonic improvement. Finally, based on the methodological analysis and experimental findings, this paper discusses several promising directions, including unified learning under continuously varying information boundaries, collaborative known-class recognition and unknown-class rejection under highly heterogeneous cross-machine conditions, digital-twin-assisted domain expansion, trustworthy and interpretable diagnosis, and lightweight modeling with safe online adaptation. These discussions provide guidance for advancing SSCDFD-RM from controlled experimental studies toward trustworthy deployment in complex industrial environments.