The determination of coal’s geographical origin is pivotal for assessing coal quality and improving import and export inspections. Conventional methods for determining coal’s geographical origin, which primarily involve time-intensive, labor-intensive, and costly chemical experiments, necessitate optimization for greater efficiency. Near-Infrared Spectroscopy (NIRS) offers a viable solution with its accuracy, speed, and non-destructive nature in detecting chemical compositions, which has achieved success in various research fields. This study introduces a non-destructive technique that integrates NIRS with deep learning for identifying the geographical origin of coal, addressing the limitations of conventional approaches. In tackling the issue of abnormal Near-Infrared (NIR) data, the study employs a data cleaning method based on Euclidean distance to identify and eliminate outliers. Standard normal variate transformation is utilized to extract features from the NIR data in the experiment. Furthermore, we introduce an improved U-Net3+ model by integrating residual modules into the encoder and incorporating the concurrent spatial and channel squeeze & excitation attention block to enrich the model’s feature capturing ability. Experimental results demonstrate that the optimized U-Net3+ model attains an impressive 97.45% identification accuracy, outperforming other algorithms. The proposed method stands as a promising solution in the realm of coal quality assessment and inspection.

Gear pitting fault is a common issue in gear systems, affecting transmission efficiency and potentially leading to severe equipment shutdowns. Effective diagnosis enhances reliability, reduces maintenance costs, and extends equipment lifespan. However, existing deep learning based methods often neglect the inherent structure of temporal vibration signals and fail to address domain variations, resulting in poor generalization and performance. To overcome these limitations, we propose a novel approach based on domain-independent features. Vibration signals are mapped to time-frequency representations via short-time Fourier transform, and dependencies between different frequencies are effectively captured using a Transformer encoder. The proposed method incorporates a feature decoupling structure that combines singular value decomposition and Pearson correlation coefficient to extract low-rank approximations of domain-related and pitting-related features, while quantifying their correlation. This approach mitigates feature degradation in constructing domain-independent features. Additionally, the weighted LinSoftmax function is introduced as a replacement for the traditional Softmax, leading to a more stable optimization target and improved model accuracy, with a distance-based penalty weight focusing on significant prediction errors. Experiments on the 2023 PHM Data Challenge dataset demonstrate the effectiveness of the proposed method, achieving a mean absolute error of 0.11, an accuracy of 92.32%, and a fault tolerance accuracy of 98.02%.