Deep Learning of Railway Track Faults using GPUs Defect Detection. Swiss Federal Railways (SBB) Swiss Center for Electronics and Microtechnology (CSEM) Nathalie Rauschmayr (CSEM) Matthias Hoechemer (CSEM) Marcel Zurkirchen (SBB) Stefan Kenzelmann (SBB) Maitre Gilles (SBB) GTC 2018, Santa Clara Fingerprinting.
Monitoring the conditions of the Swiss rail network. Need for automation Some Statistics ▪ • Manual inspection is infeasible 15k trains per day • 1.2 million rides per day • Size of rail network: 4000 miles ▪ Ever increasing traffic leads to faster attrition ▪ New train types (Cargo, High Speed and Tilting Trains…) cause the development of railway faults that have not been observable 10 years ago Multiple specially equipped trains «Diagnosis Trains» ▪ Travelling up to 100 mph ▪ Multiple high resolution cameras and other sensors Big Data Deep Learning of Railway Track Faults using GPUs 2
Monitoring the conditions of the Swiss rail network. Diagnosis train operates since 2007, BUT ▪ Software generates too many false positives/negatives ▪ Railway experts have to filter all of them by hand System has not been of much use until now Deep Learning of Railway Track Faults using GPUs 3
Railcheck project. Our Project Goals: Use Deep Learning to: reduce the time for onsite visual inspections to a minimum minimize the time experts spend on false positives increase network safety Diagnosis train Machine Learning Human Being / Business Unsupervised Learning DFZ since 2006 IBN gDFZ 2018 Coniditon monitoring Supervised Learning Deep Learning of Railway Track Faults using GPUs 4
Railcheck project. Joint project between Swiss Center for Electronics and Microtechnology (CSEM) and Swiss Federal Railways (SBB) CSEM: Swiss R&D Lab specialized in microtechnology, system engineering, robotics, automation and computer vision mainly doing applied R&D for industry (www.csem.ch) Deep Learning of Railway Track Faults using GPUs 5
The Input Data. Challenges: ▪ Changing weather conditions: rain, snow, ice ▪ Artefacts: leaves, dirt ▪ Different forms and shapes Snow In the street Small artefacts Rain Turnouts Deep Learning of Railway Track Faults using GPUs 6
Preprocessing. Tensorflow Object Detection API Transfer Learning on faster R-CNN inception resnet in order to segment railway surfaces and clamps Deep Learning of Railway Track Faults using GPUs 7
Anomaly Detection. Next Step: Clustering of railway components Anomaly: components that do not fit to any cluster Clustering with Generative Adversarial Networks (GAN): Random Fake Data Discriminator Generator Noise Training Data Deep Learning of Railway Track Faults using GPUs 8
Anomaly Detection. Generator and Discriminator learn the underlying structure of the data Search for similar images (Clustering) – Example: Input Output: most similar images Deep Learning of Railway Track Faults using GPUs 9
Anomaly Detection. Pre-Clustering of normal railway components Train Convolutional Autoencoder per Cluster Input Image Difference Reconstructed Anomaly Deep Learning of Railway Track Faults using GPUs 10
Fault Classification. Major Problem: ▪ Very little training data < 100 ▪ Tricky to define fault category: in principal 20 categories Was simplified to 6 categories and later on to 4 ▪ A lot of variation Welding Joint Surface Defect Deep Learning of Railway Track Faults using GPUs 11
Fault Classification. Major Problem: ▪ Very little training data < 100 ▪ Tricky to define fault category: in principal 20 categories Was simplified to 6 categories and later on to 4 ▪ A lot of variation Squat Wheel Slip Not Defect Deep Learning of Railway Track Faults using GPUs 12
Fault Classification. Labelling faults: Tedious, ▪ railway surface does mostly not contain any fault (luckily) ▪ faults can be easily overlooked Speed the labelling process up Transfer Learning & Inference Add to Inspect & Training Correct Data Output Increased training data for faults by a factor 100 within a month Deep Learning of Railway Track Faults using GPUs 13
Fault Classification. Major Problem: bias, wrong categories, different expert opinions The worst faults are sometimes the ones that are nearly invisible (e.g. squats) Surface Defect Squat Surface Defect Deep Learning of Railway Track Faults using GPUs 14
Fault Classification. Solution: Crowd Decision Making Experts can view output of neural network via a website and give feedback Add to Training Data Railway Experts Transfer Statistical Learning & Analysis Inference Inspect & Correct Output Surface Defect Deep Learning of Railway Track Faults using GPUs 15
Fault Classification. Even with little training data network achieves astonishing results ▪ Detection rates: Joint 99%, Welding 57%, Surface Defect 65% No training data: Squat, Wheel Slip Welding Tiny surface defect Surface Defect Surface Defect Deep Learning of Railway Track Faults using GPUs 16
Fingerprinting. «Diagnosis trains» go routes multiple times per year ▪ Which fault is new? ▪ Which fault is already recorded in DB? Challenge: ▪ Faults change over time ▪ Environmental conditions can be different ▪ Faults << GPS accuracy Deep Learning of Railway Track Faults using GPUs 17
Fingerprinting. Are these surface defects identical? 2016 2017 Locality Sensitive Hashing Input faulty image into Reduce size of Check for images within GPS feature vector Range and compares hashes pretrained model Deep Learning of Railway Track Faults using GPUs 18
Fingerprinting. Recording faults over time helps to understand ▪ How do they develop ▪ How to prevent them November February April May June August 2016 2017 2017 2017 2017 2017 Deep Learning of Railway Track Faults using GPUs 19
Summary. Deep Learning of Railway Track Faults using GPUs 20
Summary. Highly parallel problem: ▪ Create more instances of Fault- and Anomaly-Detectors depending on available GPUs Ideally: real-time processing (120-160 km/h) HPC ▪ Deep Learning of Railway Track Faults using GPUs 21
www.csem.ch www.sbb.ch
Recommend
More recommend
