The basis for artificial intelligence is the data itself - and connectivity is the first step on the road to Industry 4.0. However, large amounts of data alone are not enough. What is needed is a way to extract meaningful information from it and derive options for action for users and the company. This is exactly where advanced analytics come into play. Analyses that provide the user with information about the type of incidents help to avoid similar incidents.

Advanced analytics goes one step further. It includes autonomous analyses of real-time device data, comparison with historical information, predictions or suggestions for intervention. These techniques are already being used in predictive maintenance applications. They offer direct benefits to manufacturing companies as they prevent unnecessary downtime by avoiding interruptions and carrying out necessary maintenance and repair work within a planned downtime window.

With multiple interconnected machines and devices, the challenge is clearly to analyze the rapidly growing data stream, interpret the data and trigger reports or actions. At Rockwell Automation, advanced analytics that fall into the areas of AI and machine learning are brought together under the LogixAI umbrella. The term goes back to the Logix range of programmable automation controllers and the RSLogix software used to program them. The overarching LogixAI technology platform can be divided into four areas that Rockwell Automation's R&D specialists focus on:

• a stream clustering module,
• an automated modeling module, currently code-named Sherlock,
• a non-linear, solid and computationally efficient 'mixed-integer optimization module'
• and a self-learning module for the self-optimization of the coupled PID controllers.

This article will focus on the first two modules as they relate to performance monitoring and real-time diagnostic applications.

Clustering based on real-time data streams

The clustering algorithms are considered first. Clustering involves bundling data sets based on their similarities. They form the basis for the most commonly used performance monitoring and predictive maintenance solutions in the manufacturing industry. Clustering offers the advantage that employees with little knowledge of data science can carry out complex analyses. Such analyses have traditionally relied on relatively small volumes of static data, but companies now need to systematically record operational data - continuously, discretely and at ever-increasing data rates. This leads to enormous volumes of data that could potentially grow immeasurably over time.

Unlike static data, this streaming data may contain 'unknown' elements that have not yet been assigned to a specific cluster. That's why Rockwell Automation's data science team is developing a stream clustering engine for real-time performance monitoring and diagnostics. This engine - which is not provided as an actual product, but rather as part of the company's service and solutions offering - has already been tested in an oil and gas exploration application.

Example from the oil and gas industry

The most important source of information for monitoring operations in oil production is borehole data from the dynamometer maps, which can be divided into different clusters using stream clustering algorithms.

© Rockwell Automation

Deep rod pumps are usually used for oil production. Oil companies must be able to monitor their operation and diagnose any impending problems. The most important source of information is downhole dynamometer maps, which are widely used in the industry. Therefore, Rockwell Automation researchers have developed stream clustering algorithms that automatically divide continuous well data into different clusters. These clusters correspond to the different operating conditions of the rod pump and the oil well. As mentioned earlier, data streams may contain 'unknown' elements, which is why the stream clustering algorithm used here processes the dynamometer data as unlabeled data. The solution includes a workflow that allows a specialist to link it to an existing cluster or tag it for a new cluster that is currently unknown. Existing clusters may already have a number of actions associated with them (for example, sending an alert, stopping the pump, etc.), while new clusters will send a request to the relevant expert to name the new cluster and define the appropriate actions for that cluster.

This is a hybrid approach to performance monitoring and real-time diagnostics applications that is both data-driven and expert-driven. The main objective is to enable the detection of operational anomalies based on historical data while capturing the unique characteristics of a downhole/wellhead pump unit based solely on live operational data. This approach can be applied to any application that requires continuous monitoring.

Project 'Sherlock'

Sherlock is a plug-in module attached directly to the controller chassis that creates models from predefined controller tags and automatically compares them with real operation to detect anomalies.

© Rockwell Automation

Another AI approach from Rockwell Automation's R&D department is the use of modeling algorithms based on physical rules that support machine learning. Sherlock' will be a plug-in module attached directly to the chassis of the controller. Sherlock's data models recognize the controller's applications and check them for anomalies. The previously assigned controller tags are checked in order to recognize which application is involved or to allow users to determine what is to be modeled by intuitively selecting the inputs and outputs. The data streams running through the controller are then analyzed to create a model.

What would take weeks or months for a human to do, Sherlock does in a matter of minutes. The module does not require a large amount of historical data and the data does not have to be transferred outside the automation level. As soon as the model has been completed, the Sherlock module continuously monitors real operation and compares it with the model of normal operation that has been created. If the module detects anomalies, an alarm is triggered on the operator's HMI screen or dashboard. If the incidents recur, the model goes beyond pure diagnostics and guides the user to rectify the problem - or it automatically adjusts the system parameters accordingly to resolve the problem without human intervention.

Minimization of false alarms

Sherlock's physics-based approach was developed from the ground up for industrial applications and is designed from the outset to minimize false alarms that are otherwise common in such applications. For example, the module detects whether a change in boiler temperature is due to a harmless change in upstream process sections or whether it is a fault that needs to be corrected.

Devices communicating and connecting with each other is the most important first step towards a connected enterprise - and the combination of data and linked analytics uses this connectivity as the basis for faster fact-based decision making. AI takes this a step further, using algorithms to stop or prevent problems that can lead to unexpected downtime and lost production. And that can only be beneficial for any manufacturer.

Authors:
Kadir Liano is Senior Scientist, Analytics, at Rockwell Automation;
Bijan Sayyar-Rodsari is Director, Advanced Analytics, at Rockwell Automation;
Alex Smith is Research Manager, Analytics, at Rockwell Automation.