Introduction
Modern Cyber-Physical Systems (CPS), such as autonomous vehicles, increasingly integrate Learning-Enabled Components (LECs) like Convolutional Neural Networks (CNNs) for perception and control. While LECs offer high performance, their "black box" nature makes safety verification challenging. The traditional solution is the Simplex Architecture, which employs a high-performance "Advanced Controller" (the AI) and a safety-critical "Baseline Controller" (a simple, trusted algorithm). A decision manager monitors the system and switches control to the baseline if the AI's behavior is deemed unsafe.
However, traditional Simplex systems face two major flaws:
- Jarring Transitions: Sudden binary switches between controllers cause performance "jerks," such as abrupt steering adjustments.
- Resource Constraints: Running complex AI on embedded devices like a Raspberry Pi can cause overheating, leading to CPU throttling and increased latency, which endangers system safety.
The Dynamic-Weighted Simplex Strategy
This paper proposes a novel Dynamic-Weighted Simplex Strategy that replaces abrupt binary switching with a smooth, reinforcement-learned blend of controller outputs. Instead of selecting one controller, the system calculates a weighted average: $Output = W_1 \cdot LEC + W_2 \cdot Baseline$
The weights are dynamically optimized using Q-Learning to balance high speed (performance) with lane centering (safety). This approach treats the transition as a continuous spectrum, avoiding the jarring instability of discrete binary adjustments.
Hardware-Aware Resource Management
To address thermal constraints, the framework includes a Temperature Forecaster (a 2-layer neural network). If the hardware temperature reaches a 70°C "soft limit," the system offloads non-critical tasks to a nearby "fog" device via WiFi, preventing thermal throttling and maintaining control stability.
Experimental Setup
The strategy was evaluated using the DeepNNCar testbed, an RC car platform equipped with a Raspberry Pi 3 and LIDAR. The authors developed several datasets for this purpose:
- CNN Training: 6,000 labeled images from two indoor tracks for steering prediction.
- Thermal Monitoring: 3 hours of CPU and temperature data to train the forecasting model.
- Generalization: A third, unseen track was used to evaluate performance in novel environments.
Key Results and Insights
The dynamic-weighted strategy demonstrated significant improvements over traditional methods:
- Enhanced Safety: Track violations were reduced by 60% compared to using the LEC alone.
- Optimized Performance: The system maintained a speed of 0.4 m/s, outperforming the baseline controller.
- Context-Aware Blending: The RL agent learned to give the LEC nearly total control ($W \approx 0.95$) on straightaways while increasing the baseline's influence to $W \approx 0.80$ on curves to prevent drifting.
- Thermal Stability: The resource manager successfully kept the Raspberry Pi below 70°C. While offloading introduced minor network latency—requiring a speed reduction to ~0.34 m/s—it prevented total system failure from overheating.
Conclusion
The Dynamic-Weighted Simplex Strategy improves CPS safety and stability by transitioning from binary switching to a continuous, learned blending of controllers. Future work will focus on implementing online learning for real-time adaptation and addressing network dependencies in task-offloading scenarios.
Summary updated