In this blog post we are reviewing the webinar Accenture presented at the SigOpt Summit. During this webinar, Accenture detailed how they used SigOpt to train over 200 LSTM models used in Predictive Maintenance for an Oil Platform. By using SigOpt, Accenture was able to minimize the training time of their models (both wall clock and machine time) by over 10x. After successfully deploying the models in 2018, this platform was able to correctly detect numerous failures including detecting six high value events that were averted weeks before failure.
This blog will detail why predictive maintenance is important and how Accenture was able to use SigOpt to quickly and easily training and deploy their models for their real-world predictive maintenance platform.
Predictive maintenance is an aspirational level of maintenance most operations organizations strive to achieve. The main goal of all maintenance is to reduce downtime of operational equipment. According to Accenture, there are 4 levels of maintenance; with Predictive Maintenance being the future of maintenance.
Level 1: Reactive Maintenance
- Fix equipment as it breaks down. Least cost effective as fixing broken equipment is costly. However, this is the simplest operationally
Level 2: Planned Maintenance
- Schedule maintenance events. Less cost effective than level 3 or 4, but hopefully prevents unplanned downtime. However, this level requires performing maintenance on equipment that does not require maintenance, thus introducing extra downtime that could be avoided if the organization knows the machine is in good working order.
Level 3: Proactive Maintenance
- Use sensors to alert when equipment is about to fail. Uses better “just-in-time” maintenance to plan maintenance so you are not performing unneeded maintenance.
Level 4: Predictive Maintenance
- Uses advanced analytics and sensing data to predict machine reliability and suggest maintenance. This level allows smart predictions to pinpoint when equipment may fail and which specific fixes are recommended. This will further reduce unneeded maintenance as well as giving better fixes to get the equipment back to working order faster.
Using predictive maintenance as opposed to other levels of maintenance has been shown to reduce maintenance costs by up to 25%, reduce downtime by up to 50%, and reduce scheduled repairs by up to 12%. Organizations should strive to implement predictive maintenance in their operations to realize these gains.
Why did Accenture use SigOpt?
To deploy a predictive maintenance system, Accenture used a machine learning model to predict when a piece of equipment is outside of nominal behavior. Specifically, Accenture used LSTM type models since they react to multiple types of sensor data (i.e. temperatures, vibrations, sounds, pressures, etc). The first step was to model the nominal behavior for the equipment by training the LSTM models. Training machine learning models is a resource intensive task and also requires optimizing hyperparameters for fast, effective models. Accenture used SigOpt during this stage to narrow the network architecture space as well as effectively plan to handle when unusable data is being used during the training process (not all data is clean when coming from sensors). SigOpt also allows for searching for hyperparameters with respect to multiple metrics: 1) minimizing training time and 2) minimizing validation loss. Additionally, SigOpt can do the parallel search which allows for searching for multiple optimizations at the same time; minimizing the wall clock needed to complete the search and optimizations.
Using SigOpt resulted in a 10x reduction in resources used when training the 200+ LSTM models. So not only was SigOpt able to deliver higher performing LSTM models optimized in a quicker timeframe, but it required less IT infrastructure resources to do so.
Platform in Operation
The predictive maintenance system in operation would take in data from sensors and compare them to nominal values. Any values outside of nominal behavior is scored then aggregated with other sensors to give sub-system and total system scores to give focused health readings of the systems. These results are fed into a dashboard where operators can then make decisions to either schedule maintenance or keep the system running. This dashboard gives details with time-layered outputs of each component, sub-system, and system. These time-layered results allow users to cross-reference events with system failures and prescribe specific fixes to the system, rather than performing unneeded fixes.
Since deploying the production system in 2018, numerous events have been detected, including six high value events that were identified weeks before failure.
Accenture learnings during this process:
- Application of machine learning in maintenance requires embedding domain knowledge to be successful
- Adoption of such a machine learning application requires significant cultural change
- Use of digital accelerators (SigOpt) is critical for developing and maintaining such a tool
To try SigOpt today, you can access it for free at sigopt.com/signup.