The world is in a state of accelerating digital change. In an industrial context machine learning, advanced analytics and artificial intelligence (AI) tools have the potential to deliver improved predictive maintenance strategies, but how these tools are applied is critical to success.

There has been a shift from the initial, somewhat idealistic, objective to monitor everything and anything on a machine. It makes a lot more sense to identify priorities and focus attention on optimizing the machine learning algorithms to spot data pattern anomalies in critical areas first – identifying the areas with the fastest ROI.

Predictive maintenance based on AI offers additional advantages compared to traditional condition monitoring approaches. Increasingly data from the machinery can be merged with process data and evaluated using analysis models and cloud-based solutions.  For example, AI can be used to detect deviations at an early stage from the normal state of production machinery. This reduces unplanned downtime, lowers energy costs and increases efficiency. Reducing unplanned production stoppages directly also increases overall equipment effectiveness.

Delivering this new vision for predictive maintenance is an interesting challenge. Today’s automation manufacturers need to look for new ways to support their customers in achieving predictive maintenance based on real time data mining. For companies within the automation technology sector, this requires the ability to merge mechatronics expertise with digital analytical solutions, which is no simple task.

Experience has taught us it is critical to an AI project’s success that we not only provide software expertise, but we also have the knowledge to integrate it into the production environment and provide the experience to interpret the data in the terminology of the application. All this is achieved by analyzing data in real time. The solution can be integrated into systems – on premises, on edge or in the cloud and the programs can run on edge components directly by the machine. This offers advantages in latency and costs for data transfer are minimized.

We have also observed the benefits of being able to monitor machine output data in parallel to the standard control architecture. Separating these two functions minimizes any overloading or slowing of the standard machine control. This is particularly valued in existing, running installations where replacement of the control system would be costly in installation and programming time.

Where possible, the data is accessed through standard protocols such as OPC UA and message queuing telemetry transport (MQTT). It is processed by algorithms based on AI and machine learning in real time, outputting notifications when the data differs from the ‘healthy state’ of the asset. The notification informs the user of the sensors involved and can also provide a recommended course of action.

Using the operator’s knowledge to classify anomalies, this type of system can be trained so that the software knows how to react next time the same anomaly occurs.  Through this iterative (human-in-the-loop) process, the algorithms improve and unwanted notifications are eliminated. It is a “reward” and “punishment” process where the algorithm is positively reinforced for good notifications and refined for those that aren’t judged by the operator as useful. The more frequently the anomalies are generated, either by repetition across many identical machines or a high reoccurrence rate, the faster the digital model learns.

Once the optimization cycle is underway, anomalies are detected very early on, meaning that unexpected downtime can be avoided, spare parts are at the ready and maintenance can be carried out without having a negative impact on production.

Deeper data dives with AI

As the benefits of data analysis become better understood, the need for greater insight grows. Today a root cause analysis can be automatically generated for every abnormal behavior of the machine, showing which sensors are crucial for the anomaly detection. This additional data visualization enables users to get to the bottom of the anomalies and to recognize important correlations.

For this type of predictive maintenance tool it is important that the data analysis is not limited to components and modules from the solution provider.

The use of AI in predictive maintenance will continue to evolve. This in turn means that learning and experience needs to be acquired as early as possible to take advantage of the technology.  For end-users, it doesn’t make sense to try to justify big, all-encompassing block-buster installations.  Instead, take an agile approach and identify the quick wins with the quick paybacks.  The most successful predictive maintenance projects based on AI and other emerging technologies have taken a staged approach – proposing and testing a hypothesis with a pilot evaluation and then upscaling from the learnings gained.

Essentially, the application of machine learning and data analysis using AI helps lift the fog within large data lakes and bears down on the areas with the fastest ROI.  This enables manufacturers to pursue a prioritization strategy to benefit from the quick wins and to work down the priority list with the payback examples already in-hand. Without a doubt this is an exciting development area and it will be interesting to see how it evolves in the next few years.

– This originally appeared on Control Engineering Europe’s website.

Do you have experience and expertise with the topics mentioned in this content? You should consider contributing to our WTWH Media editorial team and getting the recognition you and your company deserve. Click here to start this process.