Traditional monitoring and visualization
You need to make sure your systems operate properly and take appropriate action (raise alarms, create work orders, etc.) as soon as there is abnormal behavior or failure in the system or its components.
Many companies are accustomed to monitoring system components and reporting failures. One way around the manual monitoring approach is to implement Business Rules, which are logics defined by domain experts to detect abnormal behavior in a system and take proper actions when needed. A major challenge with the Business Rules approach is that people do not have the analytic capacity to simultaneously take hundreds and thousands of parameters into account while also considering every edge case. Also, these rules might be contradictory and could therefore end up generating numerous false alarms and target misses.
This is where data science, and, more specifically, machine learning can help. Instead of trying to define Business Rules, your data science team or a partner can use a machine learning algorithm that will process the data coming from your system and automatically separate failure operating conditions from normal ones. This process is well known as anomaly or outlier detection. Unsupervised machine learning algorithms such as LOF (Local Outlier Factor), k-NN (k-Nearest Neighbors), PCA (Principle Component Analysis), and unsupervised SVM (Support Vector Machine)can be used to address the anomaly detection problem, among others. The algorithms may produce different accuracy of detection dependent on your specific domain and corresponding data, and data scientists will explore and compare the accuracy to suggest the best approach for your specific needs. Few data science experts, including GE Digital’s Data Science Services team, will also incorporate physical modelling into anomaly detection to further improve the quality of outcomes needed for industrial applications.
By utilizing machine learning to automate the monitoring of processes, you will be able to not only avoid time-consuming business rules creation, but also “catch” failure situations that you’d never thought about before and weren’t captured by the rules.
Once your team has the results of automatic monitoring performed by the algorithm, your subject matter experts can review the outcomes and confirm or reject the failures that were detected. The machine learning algorithm will learn from such operator’s input, and the accuracy of failure detections will be improved automatically.
Prediction and forecasting
Even if you are able to identify and quickly respond to system failures, downtime costs are still a major factor to consider. Is there a way to predict how the system will work under future performance parameters and forecast any potential failures?
Instead of relying on the reactive strategies used in traditional monitoring and visualization methods, consider a predictive approach. By feeding data into a predictive model, you can simulate the relationship between multiple variables and forecast how the system will perform under a variety of conditions. This approach also allows your staff to take proactive action to manage future requirements for incoming events.
For business users, in the majority of such cases, data scientists will create a Health Index for an asset that will show the probability of asset or system health in the prediction interval. This creates an easy-to-understand metric that requires a sophisticated mathematical approach. For example, Linxia Liao from GE Digital’s Data Science Services team recently co-authored a paper for the International Journal of Prognostics and Health Management (of assets) in which he proposed a method to integrate feature extraction and prediction as a single optimization task by stacking a three-layer model as a deep learning structure.
The proposed method was tested on a small dataset collected from a fleet of mining haul trucks. The model resulted in the ``individualized'' failure probability representation for assessing the health condition of each individual asset, which well separates the in-service and failed trucks. The proposed method was also tested on a large open source hard drive dataset and it showed promising results.
A good example of the benefits received with the predictive approach is Intel, who saved $3 million using predictive analytics to prioritize silicon chips inspections in just one year alone.
System reliability and survival analysis
Real-world equipment reliability can be significantly different from the stats provided by manufacturers. Is there a way to predict reliability using actual data?
By analyzing actual field data, machine learning and artificial intelligence can predict the likelihood of equipment failures and warn you ahead of time, regardless of what the manufacturer’s predicted failure rate may be. This allows your team to replace equipment and key system components before they fail, which reduces downtime and revenue loss due to system failure.