Making Trains See, Think and Talk Back
By Wes Mukai, CTO, GE Transportation Digital Solutions and Dattaraj Rao, Principal Software Architect, GE Transportation Digital Solutions
Specifically, for the Railway Industry, this is happening with advanced sensing capabilities on locomotives, railcars and wayside assets, increased processing power onboard the trains and improved connectivity and processing power in the cloud. The locomotive is turning into a mobile “data-center on wheels” as trains are becoming more “self and network aware”. The locomotive can now connect to the entire rail network and collect useful insights to improve its operating profile. For example, a locomotive can collect information like upcoming weather alerts, rail network congestion and track conditions and make intelligent decisions to adjust notch and braking levels over the journey to optimize fuel consumption. This is made possible through a high processing power computer onboard with the ability to wirelessly connect to the cloud. Data collected onboard is processed in real-time to make immediate decisions, while it is also transferred to the cloud to run analysis at a fleet and history level to enable long-term outcomes.
The next evolution of the self-aware train is the train that can “see” the track infrastructure it runs upon, “think” to assess potential issues and “talk” back to the rail network with actionable insights. We have trains with high-definition video cameras mounted facing the track that continuously collect high quality video data. Almost 300GB of data is collected from these cameras per day and stored on an onboard computer. There is analysis that happen onboard using computer vision models. A computer sees video as a stream of images and an image as an array of pixel intensity values. The analytical models extract geometry information from these images and try to decipher location of the track, ties, and ballast. They try to measure the co-linearity of the rails and distance between them and compare to benchmark numbers.
Analysis that runs in cloud typically involves machine learning, historical trending, and fleet-level comparisons
Any deviation is noted and tracked across frames and if it persists – this is reported as a potential defect with the GPS coordinates of the location.
Not all processing can and should be done onboard. For analysis that doesn’t need to happen in real-time – all this video “BigData” is transferred to the cloud and analyzed using hundreds of computers working in parallel. Analysis that runs in cloud typically involves machine learning, historical trending, and fleet-level comparisons.
Machine learning is all about finding patterns in data that are not obvious enough to be “coded” as rules. They typically need massive amount of data to train – but once trained can work very well on new data and handle unseen video in different lighting conditions and terrains. The next evolution of machine learning is deep learning – which builds many layers of learning models with each layer extracting significant high-level features from low-level features in previous layers. Deep learning is the state-of-the-art technology used by many of the new and popular newer inventions like Google’s self-driving car and Facebook’s face tagging algorithm. By uploading video to cloud we provide the ability to run models at scale. Technologies like Hadoop and Spark enable us to run our deep learning models on hundreds of machines in parallel and produce results in minutes rather than days.
One of the major struggles with deep learning is achieving the right accuracy and eliminating false positives – that is flagging a defect when there is none. We don’t want to send our maintenance crew to fix a track defect when there is none – or worse stop a train for a non-existent rail crack. As videos are analyzed in parallel, comparisons are made to the video from same track location window over time and between video collected from multiple locomotives for same window. These comparisons improve our confidence level of saying that the defect found by the algorithm is indeed real and worth creating a workorders for maintenance crew. This way we enable condition-based maintenance of the railway track and help the railroads increase their coverage of inspection. And we use the locomotives themselves as our eyes in the field – thus getting the true picture of the rail network.
These are just some examples where we use intelligence onboard the locomotive to provide insights for the whole network. This can very easily be extended to other areas of the railroad like service shops, network planning, transport logistics and shipper management systems. The latest and accurate location of the train will help dispatch planning systems to route trains accurately and handle delays. Similarly, the most accurate weight and temperature of the railcars carrying oil is extremely essential for providing accurate updates to shippers. This is driving the rail industry towards a connected railroad and a common operating system for the rail network operators.