Abstract
Driving conditions significantly influence the energy control and fuel economy of hybrid electric vehicles (HEVs). This paper proposes an automated feature extraction scheme using Convolutional Neural Networks (CNNs) and Kernel Principal Component Analysis (KPCA) for real-time driving pattern recognition (RTDPR). By transforming one-dimensional speed sequences into two-dimensional datasets and integrating a KPCA layer to reduce model parameters, the proposed architecture improves generalization on small datasets. Experimental results demonstrate that this hybrid strategy achieves state-of-the-art recognition accuracy, outperforming conventional manual feature extraction methods.
Introduction
For Hybrid Electric Vehicles (HEVs) to optimize energy management, they must accurately identify the current driving environment—such as congested urban traffic versus steady highway cruising. Traditionally, this required engineers to manually define "hand-crafted" features like average speed or maximum acceleration. However, these features often fail to capture the complex, non-linear dynamics of real-world driving.
While deep learning, specifically CNNs, can automate feature extraction, standard models are often too parameter-heavy for automotive applications where labeled datasets are small. This leads to overfitting, where a model performs well on training data but fails in real-world deployment. This paper addresses these challenges by introducing a hybrid CNN-KPCA architecture designed for efficiency and robustness.
Methodology
The proposed system utilizes a multi-stage pipeline to transform raw driving data into actionable insights:
- Dimension Expansion: 1D speed-time sequences are transformed into 2D "images," allowing the CNN to capture spatial-temporal relationships as visual patterns.
- Automated Feature Extraction: Convolution and max-pooling layers extract high-level features from the 2D data.
- The KPCA Layer: In a standard CNN, fully connected layers often contain millions of redundant parameters. The authors replace these with a KPCA layer, which projects high-dimensional features into a much smaller, non-linear subspace.
- Classification: A simple softmax layer classifies the reduced features into one of four patterns: Manhattan bus (congested urban), EPA urban, West Virginia suburban, or Highway fuel economy test cycles.
The model was validated using standard EPA driving cycles and real-world data collected from a 27.2km bus route in Ningbo, China, featuring 41 stops.
Results and Discussion
The hybrid CNN-KPCA approach yielded significant improvements in both efficiency and accuracy:
- Parameter Efficiency: The architecture reduced the parameter count from approximately 1.64 million (standard CNN) to just 3,472. This 99% reduction makes the model far more suitable for embedded vehicle controllers.
- Superior Accuracy: The model achieved 97.40% accuracy on the test set, vastly outperforming standard CNNs (79.78%), K-Nearest Neighbors (81.51%), and Multilayer Neural Networks (85.16%).
- Generalization: Unlike the standard CNN, which suffered from massive overfitting (100% training vs. 79% test accuracy), the CNN-KPCA model maintained high performance across both datasets.
- Real-World Performance: In field tests on the Ningbo bus line, the model correctly identified driving patterns with 95.81% accuracy.
Visual analysis of the feature space confirmed that the KPCA projection made different driving patterns more linearly separable, simplifying the final classification task and enabling the model to learn discriminative patterns directly from raw data.
Conclusion
The integration of CNNs and KPCA provides a powerful, automated alternative to manual feature engineering for HEV energy management. By drastically reducing model complexity while improving recognition accuracy, this approach enables more consistent real-time performance. Future work will focus on integrating this recognition engine into HEV energy management strategies to quantify actual fuel savings in real-world driving scenarios.
Summary updated