Abstract
Advanced combustion technologies such as homogeneous charge compression ignition (HCCI) engines have a narrow stable operating region defined by complex control strategies such as exhaust gas recirculation (EGR) and variable valve timing. Identifying the stable operating envelope is critical for diagnostics and control, yet physics-based modeling remains computationally intractable for real-time applications. This paper employs a machine learning approach to identify the HCCI stable operating boundary directly from experimental data. To address significant class imbalance (17:1 stable to unstable ratio), re-sampling and cost-sensitive learning techniques were applied to Support Vector Machines (SVM) and Extreme Learning Machines (ELM). Results indicate that cost-sensitive versions of these algorithms effectively model the HCCI operating envelope and offer the potential for real-time instability prediction based on sensor history.
Introduction
Homogeneous Charge Compression Ignition (HCCI) is often considered a "holy grail" of internal combustion, combining the high efficiency of diesel engines with the low emissions of gasoline engines. Unlike traditional engines, HCCI relies on spontaneous auto-ignition triggered by heat and pressure rather than a spark plug. This makes the combustion process highly sensitive to transient conditions; minor temperature fluctuations can lead to misfires or high cyclic variability.
Mapping the "stable operating envelope"—the range where the engine runs smoothly—is traditionally the domain of high-fidelity physics simulations. However, these models are too computationally expensive for onboard vehicle computers. Machine learning offers a faster alternative, but it faces a major hurdle: class imbalance. Because engines are rarely operated in unstable modes, experimental datasets are dominated by "stable" points. Standard models often fail to detect rare but critical misfire events, favoring overall accuracy over minority class detection.
Experimental Setup
The researchers utilized experimental data from a 2.0L 4-stroke in-line gasoline HCCI engine equipped with variable valve timing (VVT). The engine was instrumented with a comprehensive sensor suite to capture the complex dynamics of HCCI combustion.
The dataset comprised approximately 17,000 combustion cycles, with a training/test split of 6,400 and 10,200 cycles, respectively. The data exhibited a sharp imbalance ratio of 17:1. Cycles were labeled "unstable" if the Net Indicated Mean Effective Pressure (IMEP) dropped below a 0.1 bar misfire limit or if the combustion phasing (CA50) exhibited excessive variance.
Methodology
To map the stability boundary, the authors compared three algorithms: Logistic Regression (LR), Support Vector Machines (SVM), and Extreme Learning Machines (ELM). Two primary strategies addressed the class imbalance:
- Heuristic Re-sampling: Under-sampling the majority (stable) class or over-sampling the minority (unstable) class to balance the training set.
- Cost-Sensitive Learning: Modifying the algorithms' objective functions to include a penalty factor ($f$), increasing the "cost" of missing an unstable event relative to a false alarm.
A key insight of this work is that HCCI combustion is a dynamic system with "memory." Residual gases from previous cycles significantly influence current combustion. Consequently, the models were fed a 39-dimensional feature vector encompassing the current cycle and the two preceding cycles.
Results and Discussion
The study found that cost-sensitive learning generally outperformed re-sampling by allowing models to find a more nuanced decision boundary in the "fuzzy" transition zone between stable and unstable operation.
- Performance: The cost-sensitive SVM achieved the highest total accuracy at 93.1%, while the cost-sensitive ELM reached 86.4%.
- Efficiency: While SVMs provided superior accuracy, they required 120,900 parameters. In contrast, the ELM required only 3,690 parameters—a reduction of over 96%—making it highly suitable for resource-constrained embedded systems.
- Predictive Capability: Crucially, both models demonstrated the ability to predict misfires one cycle in advance. This allows engine control units to preemptively adjust fuel injection or valve timing to maintain stability.
Conclusion
This research demonstrates that machine learning, particularly cost-sensitive ELM and SVM, can effectively map the HCCI stable operating envelope despite severe data imbalance. While the current study focused on constant engine speeds (2500 RPM), future work will aim to incorporate variable speeds and ambient conditions. The parametric nature of ELMs also suggests significant potential for online adaptation, allowing the model to update as the engine ages and its stability boundaries shift.
Summary updated