Integrating Convolutional Neural Network and Broad Learning System for Moisture Detection in Oil via Near-Infrared Spectroscopy

Accurate determination of moisture content in transformer oil is critical for assessing transformer insulation, planning maintenance, extending life expectancy, and reducing failure rates. However, existing detection methods are often limited by high costs, time inefficiencies, susceptibility to human error, and low accuracy. To overcome these challenges, this study proposes an oil moisture content detection method that integrates a convolutional neural network (CNN) with a broad learning system (BLS) using near-infrared spectroscopy (NIR). This approach aims to achieve efficient, precise, rapid, and environmentally friendly quantitative analysis. The research began with the creation of a NIR spectral database for trace water content in transformer oil, followed by the application of the Savitzky-Golay convolution smoothing method for data preprocessing. The CNN was employed to extract features from the preprocessed spectral data, with the output from the penultimate fully connected layer serving as input for the BLS. The model's performance was further enhanced through training and parameter optimization. Validation and testing were conducted using a real NIR spectral dataset of transformer oil with trace water content. Experimental results show that the proposed method outperforms CNN, BLS, and PLSR models, achieving a Root Mean Square Error of Prediction (RMSEP=0.064) and Prediction Coefficient of Determination (R²=0.926). These findings confirm the method's effectiveness and superiority in detecting moisture content in oil.