Towards automatic smoke detector inspection: Recognition of the smoke detectors in industrial facilities and preparation for future drone integration

1. Problem Statement

Fire safety in industrial facilities relies heavily on smoke detectors, which are often located in hard-to-reach areas (high ceilings, restricted zones). Manual inspection is dangerous, labor-intensive, and costly. While drone-based inspection offers a solution, it requires a robust, real-time object detection system capable of running on embedded hardware with limited computational power, battery life, and memory.

Key Challenges:

• Data Scarcity: Collecting sufficient real-world training data is difficult due to the inaccessibility of sensors and restricted industrial environments.
• Hardware Constraints: The detection model must run efficiently on edge devices (e.g., Raspberry Pi) with low latency and energy consumption.
• Environmental Variability: The system must handle motion blur, low resolution, partial occlusion, and varying lighting conditions typical of drone flights.

2. Methodology

A. Dataset Construction

Due to the difficulty of gathering real-world data, the authors created a hybrid dataset strategy:

• Real Data: 1,672 manually annotated photos from public and industrial environments.
• Semi-Synthetic Data: 3,840 images generated by compositing rendered sensor models and photos of real sensors onto real background images. This included adaptive brightness adjustment and Gaussian blur to reduce domain shift.
• Data Split: The data was split into Training, Validation, and Test sets. Two distinct test sets were created:
  • Test-N (Normal): Standard appearance.
  • Test-D (Difficult): Contains motion blur, defocus, and incomplete objects to simulate drone flight conditions.

B. Training Strategies

The authors evaluated five distinct training strategies to determine the optimal use of synthetic data:

1. R-R: Real training + Real validation.
2. R-G: Real training + Generated validation.
3. G-G: Generated training + Generated validation.
4. G-R: Generated training + Real validation.
5. M-R: Mixed (Real + Generated) training + Real validation.

C. Model Architectures

Three state-of-the-art object detectors were compared, focusing on single-stage detectors for speed and a transformer-based model for accuracy:

1. YOLOv11: Variants N, S, M, and X (Convolutional).
2. SSD (Single Shot MultiBox Detector): Variants with VGG16 and MobileNetV3-L backbones (Convolutional).
3. RT-DETRv2: Variants S, M, and L (Transformer-based).

D. Augmentation and Evaluation

• Augmentation: Compared default framework augmentations against a custom Albumentation pipeline simulating drone-specific distortions (Autocontrast, MotionBlur, Defocus, ChromaticAberration, ISONoise).
• Metrics: Primary metric was mAP@0.5 (mean Average Precision at IoU 0.5).
• Hardware: Models were trained on an AMD Ryzen 7/RTX 4080 workstation and evaluated for inference speed on Raspberry Pi 4B and Raspberry Pi 5.
• Integration: A ROS-2 (Robot Operating System 2) pipeline was developed for future drone integration.

3. Key Contributions

1. Data-Efficient Learning Strategy: Demonstrated that a Mixed (M-R) training strategy (combining real and semi-synthetic data) significantly outperforms using only real data or only synthetic data, particularly for transformer-based models.
2. Comprehensive Model Benchmarking: Provided a rigorous comparison of YOLOv11, SSD, and RT-DETRv2 specifically for the niche task of smoke detector detection on edge devices.
3. Synthetic Data Analysis: Proved that while purely synthetic training (G-G) leads to overfitting and poor generalization, synthetic data acts as a powerful augmentation tool when combined with real data.
4. Edge Deployment Pipeline: Delivered a functional ROS-2 pipeline and analyzed inference times on Raspberry Pi hardware, identifying the feasibility of real-time detection.
5. Open Source: Publicly released the code, pre-trained models, and the custom dataset.

4. Experimental Results

A. Training Strategy Performance

• G-G (All Synthetic): Failed to generalize, resulting in very low mAP on test datasets (e.g., YOLOv11m dropped to ~0.385 mAP).
• R-R (All Real): Performed well but was often outperformed by the mixed strategy.
• M-R (Mixed): Best performing strategy. It improved generalization, especially for RT-DETRv2 and YOLOv11, by providing more diverse training examples without the domain shift of pure synthetic data.

B. Model Comparison (Best Results on M-R Strategy)

Model	Variant	Avg mAP@0.5 (Test)	Test Diff (Normal vs. Difficult)	Notes
YOLOv11	n (Nano)	0.884	0.112	Best overall. High accuracy, low compute, robust to blur.
SSD	VGG16	0.876	0.066	Most consistent between datasets; VGG16 outperformed MobileNet.
RT-DETRv2	L (Large)	0.833	0.128	Best transformer; L variant handled "difficult" data better than M.

• YOLOv11n emerged as the winner, balancing high accuracy (0.884) with the lowest computational requirements.
• SSD with VGG16 showed the most consistent performance (smallest gap between normal and difficult test sets), though it is computationally heavier than MobileNet.
• RT-DETRv2-L was surprisingly robust on difficult data, suggesting larger transformer models can handle complex conditions if trained with mixed data, though they require more data to avoid underfitting.

C. Augmentation Impact

• Custom augmentations simulating motion blur and defocus did not improve results; in fact, they slightly degraded performance for most models.
• The authors hypothesize that the models were already overfitting to standard appearances, and aggressive augmentation disrupted feature learning without providing enough new signal.

D. Edge Inference (Raspberry Pi)

• Raspberry Pi 5 significantly outperformed the Pi 4B (approx. 3.5x faster).
• YOLOv11n on Pi 5: ~6.36 FPS (Total inference time ~157ms). This is sufficient for slow-speed drone navigation.
• SSD and RT-DETRv2: Inference times were in the order of seconds (0.2–0.8 FPS), making them unsuitable for real-time drone navigation on this specific hardware.

5. Significance and Conclusion

This study provides a practical roadmap for deploying automated fire safety inspections using drones.

• Practicality: It confirms that YOLOv11n is the optimal choice for this specific application, offering the best trade-off between accuracy and edge-device feasibility.
• Data Strategy: It validates that for industrial applications where data collection is hard, semi-synthetic data combined with real data (M-R) is a superior training strategy compared to relying solely on real data or purely synthetic data.
• Future Work: The authors propose integrating this detection system with a mechatronic sensor-testing device (using aerosol) and advanced navigation (LiDAR/SLAM) to create a fully autonomous inspection drone.
• Licensing: The paper highlights licensing considerations, noting that while YOLOv11 requires a commercial license for some uses, SSD (MIT) and RT-DETRv2 (Apache-2.0) offer more flexibility for commercial deployment.

In summary, the paper successfully demonstrates that automatic smoke detector inspection is feasible on low-cost embedded hardware, provided that the right model (YOLOv11n) and data strategy (Mixed Real/Synthetic) are selected.