Perhaps the most common application of machine learning in the process industries is to support predictive maintenance. Anomaly detection is widely used to identify equipment performance degradation much earlier than using traditional condition monitoring systems and data analytics. However, the use of artificial intelligence to improve reliability and reduce maintenance costs is still in its infancy. 

Reliability is critical to operational excellence, it impacts safety, emissions and profitability. More than a third of safety incidents may be related to equipment failure. In the process industry, leading operators can have 10% higher availability than the worst performers. At the same time, their maintenance costs can be 30%  lower. This has a dramatic impact on profitability and return on capital employed.

The recent developments in cloud, big data, IIoT and AI have a particular impact on reliability. New IIoT sensors can be cheaply and easily installed using wireless networks such as LoRaWAN. These can provide vibration and temperature data for the smaller rotating equipment which does not justify expensive condition monitoring systems. This data can be combined with process data in the cloud and aggregated over many assets. The almost unlimited computing resources available in the cloud can then be used to apply the latest machine learning algorithms to detect changes in equipment and process performance. Having more information on equipment condition facilitates the use of a risk based approach to maintenance.

Many operators have implemented sophisticated maintenance management and asset performance management systems.  Often these systems are not adequately integrated with each other or with the operations management systems and real-time process data. In addition the systems are usually not fully configured as it is time consuming to define the risks and failure modes for each asset and the asset policies which dictate the maintenance strategy. Semantic web technologies and knowledge graphs can be built to assist in the integration and the configuration of the maintenance and reliability systems.

Using knowledge graphs also facilitates the use of natural language processing to assist with managing engineering data, structuring operator logs, maintenance records and combining them with condition monitoring information to provide better information to support risk based work selection.  Natural language processing (NLP) can also support root cause analysis by analyzing historic maintenance and failure information. The end goal is to use semantic web and NLP technologies to provide the right information to the maintenance team when they need it and provide a cognitive assistant.

Although a data driven approach is essential to support condition monitoring, it can be enhanced by the use of first principle models of assets and the process.  Asset models can provide additional features for probabilistic machine learning and ensure that the input data obeys the laws of thermodynamics. Process models can be used to determine the impact of the forecast equipment performance and help to determine the best way to manage and maintain the equipment in the context of its impact on production.

As with all digital transformation, knowing where and how to start can be a challenge.  The first step is to assess the current maintenance systems, processes and team to identify the most impactful opportunities for improvement and their value. Then the best technical approach can be assessed overall and for each opportunity. It is often tempting to start with the technology but it is always better to start with the problem and the business process. It is also important not to underestimate the need for change management. Effective maintenance, like safety, is as much a culture as a technology.