| | December 202019CIOReviewPRESENT AND FUTURE OF IOT ANALYTICSBy Andrei Khurshudov, Ph.D., Director, IoT Analytics, CAT DigitalCXO INSIGHTSCaterpillar is synonymous with heavy machinery and, for more than 90 years, our products help our customers build a better world. We've also been driving industry-leading connectivity solutions for over 20 years, introducing our first telematics device in 1999. Nowadays, Caterpillar's growing digital team is supporting the company's digital strategy through advanced data science and IoT solutions.With onboard computers, sensors, and cameras, approximately one million assets are transmitting data to Caterpillar to enable advanced IoT analytics at scale. This data can include time-series data, machine health alerts, fuel usage, GPS, and operator-specific usage.Powered by data, Caterpillar's IoT analytics can provide customers value at a lower cost of ownership, increased productivity, safety, and reduced maintenance costs. We use analytics to tell when a machine or part of a machine needs to be serviced or replaced, how to operate more effectively to increase production, how to reduce operational costs, how to increase service life, and more.So, what is the future of this fast-growing field - IoT analytics?To answer this question, let us review and discuss the main trends and challenges in this field.TREND: The increasing amount of data. CHALLENGE: Quality of available data. The Industrial Internet of Things (IIoT) is a data-generating engine for IoT analytics. For example, a large modern truck can have over 100 IoT sensors, each producing telematics data at 1 Hz frequency or faster. An analyst can face quality problems in working with such a large set of data. Missing batches or messages, missing channels, malfunctioning sensors, buggy extract, transform, load (ETL), and other factors degrade the efficacy of IoT analytics. Channel naming irregularities could be another issue which can result in a data scientist investing more of their time to quality control rather than analytics.ACTION: Invest in data quality monitoring and improvements. This will pay off long term. As the amount of data keeps increasing, tackling this issue later will become incrementally more difficult.TREND: Preference for "supervised" analytical models. CHALLENGE: The lack of quality "ground truth." Most "supervised" machine learning models rely on "ground truth" to separate data into two or more classes. For instance, such classes can be "healthy" or "unhealthy", which is the type of prediction one would want to make for an IoT device. Supervised models are more directly usable, typically come with an estimate of their accuracy, and are thus preferred for many applications. However, until issues with data quality - including the ground truth data quality - are resolved, achieving high accuracy is hardly possible for supervised models. ACTION: Focus on "unsupervised" modeling today, but build an infrastructure that is compatible with "supervised" models,and continue improving the quality of data. Transition to supervised models can and will happen gradually as infrastructure is put in place to automatically label data. TREND: Automated analytics. CHALLENGE: Lack of algorithm transparency limits scalability due to human workload. The ultimate goal for IoTanalytics is to fully automate tasks that can be automated, freeing human experts to focus on the most complex problems. This is the most efficient and cost-effective way. However, most of the machine learning Andrei Khurshudov
< Page 9 | Page 11 >