Improving anomaly detection and prevention: What is your data telling you?
By Ruban Phukan, Co-Founder of Progress DataRPM and VP, Product, Progress
The pace of digitisation means manufacturers today are under growing pressure to deliver perfect products in increasingly shorter timeframes, and at a lower cost.
They can’t afford unplanned interruptions, unforeseen failures, or unexpected breakdowns, nor can they afford to wait until the quality check stage to identify issues that could have been avoided during the production line.
According to Vanson Bourne, 82% of companies have experienced at least one unplanned downtime outage over the past three years, which can cost anywhere from $US50k-$150k per hour up to $US2 million for a major outage on an industrial critical asset. Industry research shows more than a third of the manufacturers lose 1-2% of their annual sales to scrap and rework.
Data to the rescue!
In order to reduce downtime, improve operational efficiencies and quality, manufacturers are heavily investing in data-led technologies. The Industrial Internet of Things (IIoT), machine learning and artificial intelligence (AI) for example are helping automate the process of analysing a growing number of datasets to understand and prognose machine health.
Yet, most industrial anomaly detection efforts fail, with research from Capgemini showing almost 60% of organisations do not have the analytics capabilities to take advantage of the data generated from IoT sources.
The issue is, many anomaly detection systems end up identifying either too many anomalies (false positives) or not enough (false negatives). Identifying true anomalies involves scouting for those “unknown unknowns”, amidst a sea of changing industrial data patterns.
Avoiding downtime: Illuminating the dark spots in your industrial data
The key is to detect early signals of future problems, and take proactive actions to prevent them.
There are a few best practices used for anomaly detection and prediction that every manufacturer should look to follow:
- Rule-based/supervised vs unsupervised anomaly detection and prediction
Rule-based systems are designed by defining specific rules, and typically rely on the experience of industry experts detecting “known anomalies.” The thing is, real business scenarios are quite complex and full of uncertainties.
Unsupervised learning can help learn patterns of normal behavior and identify anomalies that are very different from the expected normal behavior.
It is about enabling the production system to constantly learn, update and predict what is likely to happen next in the data stream, providing an intelligent way to detect and predict the “unknown” anomalies with greater accuracy, much before the incident occurs and alert plant operators.
- Top down approach vs bottom up approach
In the traditional top down approach, the same set of features are calculated for each sensor. But all sensors may not exhibit the same characteristics, and even those which do may not do so during all operational stages, making the data much more complex to analyse.
In the bottom up approach, the different stages of each individual sensor are first identified from the data. Then the state space of the machine is developed acknowledging that each sensor stage is part of a dynamic process’s portion that determines the state of the machine at any time.
An ensemble of baseline models for the normal conditions of the machine is then created, which helps identify anomalies based on how much the state of a machine is different from the expected normal state.
3. Manual vs cognitive approach
A manual approach to anomaly detection is useful to detect common outliers or extreme value points, which are commonly occurring across all machines. But it brings in significant human biases, it can only factor in known problems from the past, and assumes anomalies are only outliers.
A cognitive approach to anomaly detection and prediction applies a “machine-first” approach. It creates a mechanism where the algorithms, which can adapt to changing conditions, learn the data domain for each individual machine and transfer learning across similar machines. It then validates the learning with feedback from subject matter experts.
You eventually get a fully automated and cognitively enabled machine learning system, where anomalies are detected and predicted before they occur.
Manufacturers still lack full awareness of when equipment is due for maintenance, upgrade or replacement.
Investing in data-led technologies and taking a cognitive approach can help build up rock solid foundations to design accurate anomaly detection scenarios, and build truly efficient predictive maintenance strategies.