Skip to main navigation Skip to search Skip to main content

Fault detection and localization of petrochemical melt pump gearboxes using lightweight multimodal vision-language model via oil monitoring

  • Guo Yang
  • , Shizhong He
  • , Wei Feng
  • , Hui Tao
  • , Zhibin Zhao
  • , Ruxu Du
  • , Yaohua Deng
  • Guangdong University of Technology
  • Ltd.
  • Guangdong Polytechnic Normal University
  • Xi'an Jiaotong University
  • Guangzhou Janus Biotechnology Company Ltd.

Research output: Contribution to journal › Article › peer-review

1 Scopus citations

Abstract

The large petrochemical melt pump is essential for synthesizing key chemical products in the national economy, with oil monitoring playing a vital role in ensuring the safe operation of its critical gearbox. However, existing deep learning-based intelligent fault detection methods exhibit poor performance under data scarcity conditions, while large language models (LLMs) can identify faults but cannot precisely localize them. To address these problems, we propose a lightweight multimodal vision-language model (LMVLM). Firstly, the integrated oil monitoring systems were used to collect online and offline multimodal data (including viscosity, temperature, and wear particle images, etc.) for data preprocessing. Secondly, a lightweight image decoder was designed for pixel-level fault localization using intermediate patch-level features. Additionally, a prompt learner was employed to ensure semantic consistency between the LLM and decoder outputs. Thirdly, the Poisson editing method was utilized to clone objects between images by solving Poisson partial differential equations, with question-answer content being used for prompt tuning. Finally, a multi-objective optimization framework was designed to jointly enhance the LMVLM's performance in semantic alignment, class-imbalance mitigation, and pixel-wise segmentation accuracy. The method was validated using multimodal monitoring data collected from April 2021 to July 2024 on a large petrochemical melt pump gearbox in China. Experimental results demonstrate that the overall performance of LMVLM surpasses that of conventional deep learning models (CNN, RNN, and ResNet-18) and large language models (GPT-4 and GPT-4 Turbo) in fault detection accuracy. The proposed method provides an effective solution for intelligent fault detection and localization of critical equipment in the petrochemical industry.

Original languageEnglish
Article number100203
JournalChinese Journal of Mechanical Engineering (English Edition)
Volume39
DOIs
StatePublished - Dec 2026
Externally publishedYes

Keywords

  • Fault detection
  • Fault localization
  • Lightweight multimodal vision-language model (LMVLM)
  • Oil monitoring
  • Petrochemical melt pump gearbox

Fingerprint

Dive into the research topics of 'Fault detection and localization of petrochemical melt pump gearboxes using lightweight multimodal vision-language model via oil monitoring'. Together they form a unique fingerprint.

Cite this