Abstract
This paper has made a qualitative and quantitative analysis by establishing the tunnel fault tree and giving the minimal cut sets of the faults in tunnel, and tested the data in tunnel combined with artificial neural network. The fault detection mechanism in this article has been simulated by MatLab and processed a lot of the actual data through the tunnel operating history. Experimental results show that: This fault detection mechanism is effective.
Introduction
Tunnels represent high-stakes, complex environments where sensor reliability is paramount. A single failure or data error can lead to catastrophic traffic accidents or severe fire hazards. Traditional monitoring systems often struggle with the "stochastic" nature of tunnels—variables like weather, human behavior, and mechanical wear create a noisy environment that complicates fault detection.
The primary challenge addressed in this research is identifying when data from sensors, such as vehicle inspectors or CO sensors, drifts from reality. If a sensor incorrectly reports low traffic during peak hours, critical systems like ventilation may fail to engage, leading to dangerous air quality. This paper proposes a dual-layered mechanism to identify both system-level hardware failures (qualitative) and data anomalies (quantitative).
Methodology
The authors employ a two-pronged strategy combining logical rigor with pattern recognition.
Fault Tree Analysis (FTA)
The "logic" layer of the system is built upon a top-down diagram of potential failures within the tunnel monitoring system. This includes power outages, line faults, and communication failures. By identifying Minimal Cut Sets (MCS), the researchers determined the smallest combinations of component failures that would result in a total system exception.
Backpropagation (BP) Neural Network
The "intelligence" layer utilizes a BP neural network designed to detect data anomalies by predicting traffic flow based on environmental proxies. The network architecture includes:
- 7 Input Vectors: CO concentration, wind speed, visibility, number of active fans, fan running time, number of active lights, and light intensity.
- 2 Hidden Layers: Two layers (containing 16–22 nodes) were utilized to handle the non-linear, uncertain variables inherent in real-world tunnel environments.
- Activation Functions: Sigmoid functions were used for hidden layers to ensure efficient convergence, while a Purelin (linear) function was used for the output to provide a continuous range of predicted traffic values.
Results and Discussion
The model was trained and validated using historical data from major tunnels in Wuhan, China, primarily the 1700-meter Wuhan Shuiguohu Tunnel. Data was pre-processed through averaging to reduce noise and ensure sample homogeneity.
The core insight of this approach is using environmental "exhaust" as a proxy for traffic. Because CO levels and visibility are physically tied to vehicle volume in a confined space, the ANN can accurately estimate traffic flow. A significant deviation between the actual sensor data and this neural "guess" flags a data anomaly.
As shown in the contrast figure above, the predicted traffic flow (dashed line) demonstrates a high-fidelity fit with the actual historical data (solid line). This confirms that the neural network successfully learned the complex physical relationships between environmental factors and traffic volume.
Conclusion
The integration of FTA for hardware failures and ANN for data anomalies creates a comprehensive safety net for tunnel management. While the visual fit of the predictions is impressive, future work could benefit from specific statistical error metrics (like RMSE) and the inclusion of external weather data to further refine the model's accuracy. Additionally, evaluating the real-time latency of the two-layer ANN is necessary for practical deployment in live monitoring systems.
Summary updated