A novel deep learning model for spectral analysis: Lightweight ResNet-CNN with adaptive feature compression for oil spill type identification

[Display omitted] •LightResNet-AFC introduces a new CNN paradigm for spectral analysis.•The model achieves 100% accuracy in classifying six types of engine oils.•It offers faster training times, smaller model size, and enhanced stability.•Eliminating the fully connected layer and using GAP enhances model efficiency.•LightResNet-AFC is ideal for deployment in resource-limited environments. This paper presents a novel structural paradigm for convolutional neural networks (CNN), termed Lightweight ResNet-CNN with Adaptive Feature Compression (LightResNet-AFC), which has broad applicability in the field of spectral analysis. The LightResNet-AFC model has the potential to transform spectral analysis applications, particularly in real-time oil spill detection, environmental monitoring, and industrial quality control. The LightResNet-AFC, combined with near-infrared (NIR) spectroscopy technology, accurately classifies six types of engine oils. The performance of the proposed model is benchmarked using accuracy, training time, and model size, showing significant improvements. Following multiple training and validation cycles, the LightResNet-AFC achieves an accuracy rate of 100% by employing a ResNet-based CNN to establish connections between local and global spectral features, using global average pooling (GAP) to eliminate the fully connected layer, a key component in traditional one-dimensional convolutional neural networks (1DCNN). This approach not only outperforms the traditional 1DCNN model in terms of accuracy but also demonstrates advantages in training time, model size, and stability. Additionally, when benchmarked against traditional spectroscopic analysis models, our method shows improved model robustness by reducing the impact of preprocessing steps, enhancing response speed, and enabling rapid end-to-end online spectral analysis. In summary, our proposed method addresses a significant limitation of traditional spectroscopic analysis algorithms − the inability to accurately analyze substances lacking distinct characteristic peaks or exhibiting similar spectral trends − and introduces a rapid, non-destructive spectral analysis technique for oil spill detection. Moreover, it challenges the conventional structure of CNN models, simplifying the design and optimization process of spectral analysis models, offering fresh insights for the advancement of CNN in spectral analysis. Simultaneously, the omission of the fully connected layer results in a smaller p