Machine Learning in Root Cause Analysis

It’s hard to keep up with the news in the manufacturing industries and not find articles on how artificial intelligence will bring improvements. Machine learning is one area that can help spot patterns to improve reliability, quality, and throughput.

Emerson's Mark Nixon


Emerson's Noel Bell


At this past Emerson Exchange conference in San Antonio, Emerson’s Mark Nixon and Noel Bell teamed with senior leaders from Integration Objects to present, Learning and Root Cause Analysis Application.

They opened describing the traditional control room where operators observe alarm conditions, identify the scenario that is causing these alarms and compare it with their earlier experiences, find an appropriate solution and react accordingly to address the situation.

The operator must digest all the information coming from printer logs, HMI alarm interfaces, operator screens and trends to decide on the best course of action. This is where artificial intelligence can help. From additional data provided by Industrial Internet of Things (IIoT) sensors and other suppliers of big data, software for smart decision support and prescriptive analytics can be applied.

The sources of data can include the control systems, plant data historian, plant asset databases, legacy applications, and laboratory information management systems (LIMS) to name a few. Additional non-process data can include operating manuals, key performance indicator (KPI) calculations, workflows, advanced analytics, rules, best practices and root cause analysis.

The application of machine learning algorithms can process this data into actionable information to support corrective actions to address the abnormal conditions. Other objectives include identifying/predicting performance gaps before unplanned shutdowns occur, isolating the root cause before it affects operating performance, identifying areas for efficiency gains and cost reductions, retaining expertise being lost to retirements, and preventing off-spec products while waiting for lab results.

The methodology to build the machine-learning models is to take input sources of data and first apply offline machine learning algorithms such as cleaning, clustering, principal component analysis (PCA), decision trees, event clustering, regression, sequential rules, and optimization. After this processing has been performed online machine methodologies include recursive density estimation, online clustering, and model retraining. These methods feed additional areas of refinement such as expert rules, workflows and root cause analysis.

The model generated from this processing can help to improve the management of asset health, detection of performance gaps, failure predictions, identification of root causes, asset useful life, etc.

They shared some examples of how this technology has been applied. For refiners, Reid Vapor Pressure (RVP) is a measure of gasoline volatility. Predicting RVP in real time during production helps to prevent variability from infrequent lab samples.

Realtime root cause analysis using the RVP soft sensor (click to enlarge)

Some other applications they discussed included smart heat exchangers and KNet Reasoning integration with Plantweb Optics asset performance platform.

Realtime root cause analysis for the heat exchanger based on FMEA information (click to enlarge)

By incorporating data streams from the process and from work processes and other business functions, machine learning technology can help develop models to avoid abnormal conditions and improve operational performance.

As part of your digital transformation efforts to achieve top quartile performance, visit the Operational Certainty section on Emerson.com. You can also connect and interact with other performance improvement experts in the Services group in the Emerson Exchange 365 community.

Posted Thursday, November 29th, 2018 under Industrial IOT.