Machine learning-enhanced fiber optic sensor system for leakage detection in enclosed devices

Abstract In marine exploration, precision equipment deployed in the deep-sea encounters significant challenges due to extreme high pressure and complex ocean currents, which increase the risk of potential leakage. This paper proposes a fiber optic sensor system for leakage monitoring in sealed devices. The system is based on Optical Time-Domain Reflectometry (OTDR), which achieves real-time leakage monitoring by accurately capturing changes in Fresnel reflection intensity caused by variations in the medium at the fiber end face. Given the complexity of the marine environment, the sensor was designed with an encapsulated probe structure to isolate external interference. Through experimental analysis, this study selected water temperature, environmental humidity, and characteristic peak values as features for machine learning. After comparing five aforementioned machine learning models and one classical statistical model, Support Vector Machine (SVM) was ultimately chosen as the leakage detection algorithm for identifying leakage events. Experimental results indicate that the sensor’s measurement data has a relative standard deviation of only 0.0834, reflecting the stability of the sensor’s measurements. The system achieved a high accuracy rate of 99.51% in detecting leakage events across various environmental conditions. Additionally, it demonstrated strong adaptability and scalability for application in other liquid media, including alcohol and oil.