By: Kelly Smith
In the roar of industrial manufacturing, rotating machinery serves as the “heart” of production lines. Yet under harsh conditions—such as high temperatures, high speeds, and intense impacts—equipment failures can result in significant operational and financial challenges. Traditional fault diagnosis methods often depend heavily on manual expertise and large volumes of labeled data, which can limit their accuracy, response times, and adaptability in real-world environments.
Addressing these ongoing industry challenges, researcher Chuanyuan Tan has recently published two studies proposing an intelligent fault diagnosis framework that integrates deep learning with transfer learning. These approaches appear to significantly improve fault recognition accuracy and engineering applicability under data-limited conditions, offering promising AI capabilities to the manufacturing sector and marking a notable advancement in the field of intelligent diagnostics.
In his 2025 paper titled “Proposed Damage Detection and Isolation from Limited Experimental Data Based on a Deep Transfer Learning and an Ensemble Learning Classifier,” Tan and his team introduce a fault diagnosis framework that combines deep transfer learning with an ensemble learning classifier. Using a simulated shaft-disc system generating multiple health state datasets, they apply a Retinex-based Wavelet Transform (RT-WT) to convert vibration signals into images.
By integrating the DenseNet architecture with transfer learning strategies, the model achieves a notable degree of generalization capability. Experimental results suggest that this model demonstrates superior accuracy, stability, and computational efficiency compared to mainstream models—particularly when addressing multiple fault types in data-scarce industrial settings.
In a concurrent publication, “Highly Reliable CI-JSO based Densely Connected Convolutional Networks using Transfer Learning for Fault Diagnosis,” Tan advances the framework by employing MinMax Scalar-based Continuous Wavelet Transform (MMS-CWT) for signal preprocessing, coupled with Confidence Interval-based Jellyfish Search Optimization (CI-JSO) for efficient feature selection. Building on this foundation, the DenseNet-based model—leveraging pretrained knowledge from ImageNet—has been shown to deliver excellent diagnostic performance under few-shot learning conditions.
This method excels in both accuracy and generalizability across diverse fault classification tasks, while maintaining a lightweight structure and fast training speed, which makes it potentially well-suited for resource-constrained industrial environments.
Together, these two studies published in 2025 represent a substantial shift from traditional manual feature extraction to end-to-end image-based recognition, addressing long-standing challenges of data dependency and deployment difficulty. They suggest a new technical paradigm for intelligent manufacturing applications.
Tan’s research is grounded not only in theoretical development but also in real-world industrial practice. His experience as a process engineer at Han’s Laser—where he was involved in laser processing, equipment maintenance, and after-sales technical support—has provided him with valuable insights into machinery operation patterns, fault mechanisms, and engineering requirements. This practical experience lends a strong engineering-oriented perspective to his research in intelligent fault diagnostics.
The intelligent fault diagnosis solutions Tan leads are applicable not only to conventional rotating parts such as bearings and gears but also show potential adaptability to more complex systems like industrial robots and composite material structures. These innovations could help facilitate a shift from reactive maintenance to predictive maintenance, enhancing operational reliability and potentially reducing downtime.
At the 2025 Green Materials and Manufacturing Technology International Conference (GMMT 2025), Tan presented a study titled “3D Printer Nozzle and Stepper Motor Temperature Measurement Based on Homemade NiCr/NiSi Thin Film Thermocouples,” which introduces a cross-disciplinary advancement in microscale sensing technology.
This research uses magnetron sputtering to fabricate NiCr/NiSi thin-film thermocouples, achieving a high Seebeck coefficient of 39.80 μV/°C. These custom sensors allow for precise temperature monitoring of 3D printer nozzles (189.62°C) and stepper motors (55.67°C). Compared to conventional K-type thermocouples, these thin-film sensors provide faster response times and improved thermal stability, which could help mitigate printing defects due to temperature fluctuations and improve the reliability of additive manufacturing processes.
From intelligent diagnostics in harsh environments to breakthroughs in microscale sensor design, Chuanyuan Tan has developed a comprehensive research model driven by “problem identification – technological development – engineering application.” His work not only contributes to advancing key technologies in intelligent manufacturing but also injects sustainable innovation into traditional industrial systems.
As the shift toward automation, data transparency, and smart operations continues, Tan’s AI-driven research is emerging as a significant force behind the next phase of manufacturing evolution—enabling machines not only to perceive their own state but also to make autonomous decisions and, in the future, possibly achieve self-directed operation.
