Thermal Brake Disc Fissures

Description

The core hypothesis of this research is that integrating localized object detection (YOLOv8m) with fine-grained semantic segmentation (U-Net with a ResNet50 backbone) on thermal infrared imagery significantly enhances the accuracy and reliability of crack detection in railway brake discs compared to conventional visual inspection or single-model AI approaches. Thermal imaging provides superior contrast for identifying heat dissipation patterns around metallic fissures, which are often invisible to the naked eye under industrial conditions. This dataset contains a comprehensive collection of thermal images (160 x 120 px resolution) captured from railway brake discs during preventive maintenance. The data includes: Original Thermal Frames which are raw captures of brake discs showing various thermal signatures. Annotated Masks: Ground-truth labels for three types of cracks: Penetrating, Incipient, and Superficial. Augmented Samples: Images processed with noise injection and geometric transformations to ensure model robustness.

Steps to reproduce

Data Acquisition and Hardware Setup Thermal Imaging: Images were captured using a Flir Lepton 3.1R thermal sensor, which provides a resolution of 160 x 120 pixels. Environment: Captures were taken under controlled industrial conditions to minimize thermal noise from external sources. Image Pre-processingUpscaling: Due to the sensor's native resolution, bilinear interpolation was applied to resize the images to 160 x 120 pixels, ensuring compatibility with deep learning architectures. Data Augmentation: To improve model generalization, the dataset was expanded using the following techniques: Rotations and flips, LongesMaxSize, PadIfNeeded, GridDistortion, OpticalDistortion, ElasticTransform, GaussNoise, RandomBrightnessContrast. Labeling and Segmentation Masking Annotation: Manual labeling was conducted using professional annotation software. Classification: Fissures were categorized into three classes: Penetrating, Incipient, and Superficial. Ground Truth: Binary masks were generated for each thermal image to train the semantic segmentation branch (U-Net). Deep Learning Workflow (Software and Training) Environment: The experiments were implemented in Python using the PyTorch framework. Detection Phase: YOLOv8m was utilized for real-time object detection to identify the region of interest (ROI) where cracks are located. Segmentation Phase: A U-Net architecture with a pre-trained ResNet50 backbone (Transfer Learning) was integrated to perform pixel-level segmentation within the detected ROIs. Hyperparameters: Training was conducted for 150 epochs with a learning rate of 0.0001 and a batch size of 32. Model Architecture & Fine-Tuning TechniqueTransfer Learning: We utilized a Fine-Tuning approach, loading a U-Net architecture with a ResNet50 backbone pre-trained on the ImageNet dataset. Architecture Layers: ReLU Activation was applied within the ResNet50 encoder and U-Net decoder blocks to process non-linear thermal features and avoid the vanishing gradient problem. SiLU (Swish) Activation: Utilized in the YOLOv8m feature extraction modules to refine the detection of subtle incipient cracks. Sigmoid Activation: Employed in the final output layer to generate a pixel-level probability map, which is then binarized at a 0.5 threshold to produce the final segmentation mask. Evaluation Metrics The reproducibility of the results can be verified using the provided test set by calculating the Mean Intersection over Union (mIoU), Precision, and Recall.

Institutions

• Peruvian University of Applied Sciences

  Lima Province, Lima

Categories

Licence