What makes ARP different than other predictive maintenance or AI solutions?
Asset Risk Predictor (ARP) is deeply integrated to OT layer and uses a multi-variable model, which includes inputs from multiple sensors, machine status, and recipe information to provide a 360 view into a machine’s health. It also offers fast time to value, as it can be up and running in as little as 2 weeks.
How is ARP different than condition-based monitoring (e.g. SCADA)?
The main differences are:
- Automated learning: Unlike traditional condition-based monitoring where thresholds have to be manually set based on expert knowledge and data analysis, ARP uses machine learning algorithms to learn what is normal for each asset automatically, eliminating the need to manually set up thresholds.
- Intelligent anomaly detection: While traditional condition-based monitoring might trigger false alarms due to small deviations or spikes in sensor readings, the ARP algorithm is intelligent enough to ignore these minor anomalies that are not indicative of a problem.
- Complex problem detection: Traditional condition-based monitoring often relies on simple thresholds that may not be sufficient to detect complex problems that involve a gradual change in condition over time or correlations between multiple data points. ARP, however, can identify these complex patterns, allowing for earlier detection of potential issues.
- Predictive maintenance: Unlike traditional condition-based monitoring, which tends to focus on current conditions, ARP includes next-day prediction capabilities. This feature facilitates predictive maintenance, allowing potential issues to be addressed before they cause a failure.
- Reliability metrics and predictions: Given failure codes from PLCs, ARP is capable of calculating reliability metrics, which are key indicators of an asset's probability of performing its intended function without failure over a specified period of time. By analyzing these metrics, ARP can predict not only the risk of failure but also other aspects of asset performance, such as the expected remaining useful life of the asset.
- Feedback and self-improvement: Unlike traditional condition-based monitoring, ARP is capable of improving with feedback, as models are retrained after set periods of time, and will generate better prescriptive maintenance suggestions on work orders after user correction.
- Convenient alerts: ARP allows for integration with web and mobile, giving you a more accessible alert system.
- Scalability: While traditional systems must be set up from scratch whenever there are changes to scale, ARP is scalable with your operation.
What is the difference between using a threshold-based system and ARP?
Threshold based systems are often used to indicate when a parameter or metric is outside the bounds of their specs. When you go out of spec, you get an alarm. Often these systems are single-dimensional and don't give much warning before failure.
ARP learns the signature of how the sensors behave during normal operation, and also throughout the year. It also is multi-dimensional, so it can indicate risk based on small movements when it finds patterns of more than one sensor movement outside of “normal”. It can also give warning prior to failure, allowing you to schedule maintenance before actual failure happens and in a controlled and planned manner.
How do we know the data and predictions we’re receiving are accurate?
Technicians can verify the accuracy by looking at the raw sensor values and operating statuses during the risk events we specify. If you provide failure logs, accuracy can also be measured automatically.
How are false positives handled?
For now, we need to review the machine to understand if there is a false positive. More often than not, the sensors are in fact drifting, maybe not because the machine itself is failing, but because an upstream process could be causing changes in the behavior of the machine that we are monitoring. If this is not the case and there is in fact a poor model causing false positives, we would trigger a retrain or wait for the monthly relearn to self-correct.
How does ARP's data retrieval work with the prediction schedule?
Every hour, ARP retrieves the plant historian’s aggregated minute-by-minute data for all connected sensors in the past hour. Every eight hours, ARP's prediction models use the collected data from the last eight hours to generate risk predictions for the next eight hours.
This eight-hour cadence lines up with shift changes and allows you to see emerging issues a full shift before they can turn into downtime. It also reduces minute-by-minute noise and minimizes computing costs.