When it comes to machines, there’s never really a good time for them to breakdown. Machinery breakdowns are inevitably costly. This is true for manufacturers of equipment, for their customers, and for providers of managed machine maintenance services. In response to the challenge, companies have explored a variety of strategies – even preventive maintenance programs - to increase machine uptime. The goals have always been to: ensure high availability of machines, rationalize service and repair costs across the network, and incorporate learning into the manufacturing and application of new models of machines. But every strategy – even a combination of strategies - has had only limited success.
The juxtaposition of data from machines across multiple sites and over months of usage, with near real-time data from individual machines that must be maintained, would help to derive insights that provide warning signals and alert managers to potential machine failure. Here’s how it works:
For example, we curated four million failure ticket records from over 8,500 ATMs to develop, train and test a machine-learning model for predicting ATM failure in North America. Harvest diverse data – past events of dysfunction, maintenance schedules, log data, transaction load, time since the last repair, the age of the machine, and defects reported.
Ingest the data into an Apache Spark data-processing engine for logistic regression. The Spark open source cluster computing framework cleanses and enriches data as quickly as 27 seconds. In the case of the ATM predictive maintenance program, the logistic regression algorithms predicted ATM fault in 60 milliseconds with 80% accuracy.
Use Tableau visualization tools to present reports for interpretation and analysis. The color-coded dashboard helps maintenance teams review notifications, service calls and failure patterns over time periods, cities / states, and model types.