New AI-Driven Network Achieves 98.5% Accuracy in Diesel Fuel Brand Identification

Recently, a team of scientists at Yanshan University conducted research into diesel fuel brands using a hybrid artificial intelligence (AI) network. This study, which was published in the journal Measurement, examines how spectroscopy and AI can improve industrial inspection (1). Led by researcher Zhiwei Wang, the research team developed a multimodal feature fusion network (MFFN-INIRS) that delivered a 98.5% classification accuracy (1). These results suggest the potential of their method in establishing a new standard for quality control in the energy and chemical sectors.

What is the state of gas-powered vehicles?

Despite the push for more electric vehicles, gas-powered machinery and vehicles still reign supreme in the turbomachinery industry. According to the U.S. Department of Energy, natural gas powers more than 175,000 vehicles in the United States and roughly 23 million vehicles worldwide (2). To power this machinery, it is required that their fuel systems are performing optimally. As a result, much effort has been placed on accurately identifying diesel fuel brands to maintain the efficiency and reliability of fuel systems in vehicles and industrial machinery (1).

Close up view of fuel nozzle refueling a modern car at a gas station. Generated with AI. | Image Credit: © Pungkas - stock.adobe.com

The problem, though, is that the usual techniques are dependent on one-dimensional (1D) spectral data. These methods either model raw spectral data directly, which results in a loss of global structural patterns, or convert spectra into images using encoding methods that risk distorting important information (1).

What did the researchers do in their study?

In their study, the research team looked at resolving these limitations by developing and testing a new dual-branch modeling strategy. In their MFFN-INIRS framework, the researchers relied on three key components. First, the Oversampling and Image Encoding (OIE) Module uses a KMeans-SMOTE algorithm to address class imbalance by synthetically expanding underrepresented sample categories while preserving their structural integrity (1). Then NIR spectral data are then encoded into two-dimensional image representations using a technique called recurrence plotting (RP), effectively transforming time-series data into visual patterns for deeper analysis (1).

Second, the Multimodal Feature Extraction (MFE) Module comes into play. It utilizes ResNet50 to extract high-level semantic features from the RP images (1). Simultaneously, a one-dimensional convolutional neural network (1D-CNN) processes the original NIR spectra to extract localized frequency domain features. This dual-pathway architecture allows for a more holistic understanding of the data by capturing complementary information across different formats (1).

The final component, which was the Dynamic Weighted Gated Feature Fusion (DWGFF) Module, combines the features extracted from the two branches. By assigning adaptive weights to various feature channels based on relevance, the module ensures that the most informative signals are prioritized while redundant or noisy information is filtered out (1). This results in more robust and accurate classification outcomes.

Once the multimodal features are fused, a Particle Swarm Optimization-based Support Vector Machine (PSO-SVM) is employed for the final classification stage. The PSO algorithm fine-tunes the hyperparameters of the SVM model, ensuring peak performance (1). The end result was that this model was capable of identifying diesel fuel brands with 98.5% accuracy.

What are the key takeaways from this study?

There are a few key takeaways from this study. First, this MFFN-INIRS approach is environmentally friendly and non-destructive. Because this approach is completely data-driven, the MFFN-INIRS approach is a more sustainable solution for industrial inspection and quality assurance (1).

Second, this method uses both spectral analysis and deep image processing to improve classification accuracy (1). By showing how multimodal data fusion, combined with advanced deep learning techniques, can be leveraged to solve complex classification problems in the energy sector, the researchers offer a new pathway to creating technologically advanced tools for ensuring fuel quality and safety.

References

1. Xie, B.; Liu, X.; Wang, Z. Multimodal Feature Fusion Network based on Image and NIR Spectroscopy for Diesel Fuel Brand Identification. Measurement 2025, 256 Part C, 118360. DOI: 10.1016/j.measurement.2025.118360
2. U.S. Department of Energy, Natural Gas Vehicles. Energy.gov. Available at: https://afdc.energy.gov/vehicles/natural-gas (accessed 2025-08-05).