Two-Stage Deep Learning for Rail Car Coupler Cotter Key Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning algorithms for detecting objects in images, particularly in the context of rail car couplers, face challenges in accurately predicting the presence or absence of cotter keys due to variations in environmental conditions and image quality, leading to potential coupler securement failures.
Innovation Solution
A two-stage system comprising a deep-learning localization algorithm and an ensemble of deep-learning classification models is trained using curated datasets that include raw and secondary images, with the latter generated through adjustments, augmentations, and duplications to simulate real-world conditions, ensuring robustness and resilience against imperfections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single deep-learning algorithm is used for detecting coupler conditions, then the system complexity is low, but the detection accuracy and reliability are insufficient under varied environmental conditions
Solution Approach 1:
The detection system is divided into two independent stages: a localization algorithm that identifies the region of interest (coupler) in the image, and a classification algorithm that detects the presence or absence of cotter keys. This segmentation allows each algorithm to specialize in its specific task, improving overall detection accuracy while maintaining manageable complexity through modular design.
Solution Approach 2:
The localization algorithm serves as an intermediary component between the input image and the classification algorithm. It first identifies and extracts the region containing the coupler, then passes this refined region to the classification algorithm. This intermediary step eliminates the need for the classification algorithm to process entire images, improving both accuracy and computational efficiency.
2Adaptability or versatility
If training data is limited to available raw images, then the training process is simple, but the model performance degrades under diverse real-world conditions
Solution Approach 1:
The system performs preliminary data preparation by generating synthetic training images through image augmentation techniques (rotations, flips, brightness adjustments, noise addition) before the actual training process. This preliminary action ensures the model is pre-adapted to various environmental conditions it may encounter in deployment, improving robustness without requiring extensive real-world data collection for every scenario.
Solution Approach 2:
The system creates synthetic copies of existing raw images by applying various transformations and augmentations. These copied and modified images serve as additional training data, allowing the model to learn from diverse variations without requiring physical access to numerous real-world examples under different conditions. This copying approach efficiently expands the training dataset while maintaining consistency with real-world patterns.
Data Source
AI summary
Systems, devices, media, and methods are presented for training a predictive model to detect objects in digital images, such as detecting whether a cotter key is present or absent in a digital photograph of a rail car coupler. The training system includes curating a plurality of training datasets, each including a number of raw images, together with a number of adjusted, augmented, and duplicate images. The predictive model includes a localization algorithm and an ensemble of models for classification. The localization algorithm is a deep convolutional neural network (CNN) which identifies a region of interest. One of more of the deep CNN classification models generates a plurality of candidate regions associated with each region of interest, thereby generating a large number of additional regions useful for training. In use, the trained predictive model is part of a detection and notification system that processes new images from the field and broadcasts a notice when an anomaly is detected.