Using search for engineering diagnostics and prognostics Jim Austin Overview Problem domain Drivers - why we need better solutions Example applications Our approach Challenges Slide 2 Prognostics and Diagnostics Find out what is wrong with some thing Find out what may be about to happen Use data to achieve this, but deliver knowledge Wide applicability (not just engineering) Slide 3 Engineering problems Asset monitoring Large numbers of sensors Many types of sensors Distributed sensors and systems Possibly hostile domains Large data rates Slow connections Data incomplete, noisy hard to characterise Slide 4 Engineering problems Response Needs to be rapid Qualified response (i.e. how good) Must include users in the loop, not yet automatic Conclusion must be justified – able to dig into problem Slide 5 Drivers Why now? Data to Knowledge is a prime motivator Sensors are now robust small and reliable Data collection is very cost effective (2Tb < £200) Large computing capability is now possible Most easy wins have been achieved Green agenda is forcing issues Slide 6 Example Applications Aero-engines & fixed assets Rail – track and carrage Roads Slide 7 Gas turbines High speed, rotating systems Typically very reliable Used for air travel as well as pumps and generators (oil and gas, marine, air, power) Slide 8 Gas turbines Typical problem Spot failure in good time (!) Spot maintenance issue ahead of time Data is High frequency Large Complex Slide 9 Rail Monitoring of both track and carriages Over 2000 alerts on a Thomas virgin voyager Aim is to reduce unplanned maintenance Slide 10 Rail Track Look at data from track inspection systems Find if track is bent or broken and needs maintenance Slide 11 Road Monitoring for congestion problems Data from road ‘loops’ (flow and occupancy) Weather Accident reports Adjust Traffic lights Variable message signs Slide 12 Road Hull road bus gate, York 13 Road Slide 14 Our approach Use historic data as a prediction of now and the future Basically search the historic data Use AURA neural network Have a set of systems within Signal Data Explorer Share data and services through portals Building on CARMEN Slide 15 SDE and CARMEN Slide 16 Data compatibility Neural Data Format – NDF Allow interoperability between: Multi-channel systems data (.mcd). Comma delimited (.csv). Alpha map (.map). Neural event (.nev). NeuroShare native (.nsn). Nex (.nex). PC spike2 (.smr). Plexon data (.plx). TDT data format (.stb) Supported in visualisation tool (SDE), soon in services Slide 17 Data entry Slide 18 Services Slide 19 Execution log Slide 20 Examples Slide 21 Search for signals Data converter Correlation matrix Time series data Historical data Compare 22 Known? Fault identification Data converter Time series data Correlation matrix Historical data Compare 23 Known? Challenges Best practice in data collection – build system when you know how to process it! Better tools, for analysis of signals, images and text (three main groups). Better collaborative technologies, new in industry sector User adoption of the technology Slide 24 Summary Data now available in large quantities Real opportunities to improve the systems that are being built Slide 25