Light fuel classification based on Raman spectroscopy and region-adaptive convolutional neural networks

In modern warfare, the demands of individual combat and rapid deployment place increasingly stringent requirements on energy security. The safe and reliable supply of light fuels directly impacts combat effectiveness and strategic success. Therefore, there is an urgent need for a rapid, accurate, and field-adaptable detection method to ensure the security of fuel supply. Among the available techniques, portable Raman spectroscopy, owing to its non-destructive nature, rapid analysis capability, and high sensitivity to molecular structures, has been widely applied in the detection of light fuels. However, in Raman spectroscopy, light fuel samples often exhibit highly similar spectral profiles. Moreover, traditional feature extraction methods suffer from limited capability, and under conditions of low signal-to-noise ratios or strong background interference, they show insufficient sensitivity to weak characteristic peaks, which substantially constrains classification performance. To address this challenge, we collected a total of 16 different grades and sources of light fuels, including gasoline, diesel, and jet fuel, and proposed a novel one-dimensional convolutional neural network (1D-CNN) for fuel classification. The proposed model segments each Raman spectrum into four regions, performs region-specific feature extraction, and assigns initial weights based on the relative peak density of each region. During training, these weights are further optimized through a constraint mechanism that adaptively adjusts regional contributions, thereby guiding the network to focus on the relative importance of peaks within each region and enhancing the model' fine-grained discriminative ability. Experimental results demonstrate that the proposed model achieves an average accuracy of 96.16%, a macro-averaged precision of 96.23%, a macro-averaged recall of 96.08%, and a macro-averaged F1 score of 96.12%, all consistently outperforming those of the other models, while maintaining a relatively low parameter complexity of 0.7 M and a short training time of 226.32 s. Further sensitivity analysis validates that the parameters we selected are indeed optimal. It provides a novel and practical solution for rapid battlefield identification of light fuels and offers valuable insights into the design and application of region-aware 1D deep learning models.