Predictive maintenance, predictive quality and automated root cause analysis are industry 4.0 initiatives driven by the power of AI and machine learning. However, many implementations – specifically in the process manufacturing industry – fall short in delivering the promised value of industry 4.0 due to inaccurate insights and too many false positives.

The solution? Production process context.

With process-based machine learning, the specific characteristics of the manufacturing process and its assets are taken into consideration within the algorithms.

Specific sensors in machines and mechanical equipment, production recipes, production process flows, and the facility’s environmental factors – all contribute to the algorithm’s accuracy.

 

Too Many Anomalies, Too Many Alerts

In process manufacturing, data from thousands of tags (sensors) are typically captured in a data historian. This data is heavily influenced by production context: which product is being manufactured, and how – product recipe, machine settings, production flow, and manual interventions.

For this reason, applying machine learning algorithms to raw data, without factoring all the relevant contextual information, will result in an abundance of alerts, most of which will be false positives.

Unfortunately, for many manufacturers, this is where the credibility of their industry 4.0 efforts begins to crumble:

Generic Vs. Process-Based Machine Learning

What about weights?

One method used as an attempt to shape data towards actionable insight is to use formulaic weights. In this method, weights are attributed to certain behaviors (measurements/groups of measurements) in an attempt to define problematic states.

While introducing weights to a process will indeed add a contextual layer to the data collected from a production line, it will also introduce new problems. This is because the logic behind weight allocation is based upon human rationale.

 

Process-based Machine Learning to the Rescue

The good news is that this type of problem is perfectly suited to the application of industrial AI. That’s why machine learning and artificial neural networks are considered a key part of Industry 4.0, especially when it comes to driving predictions as in the cases of predictive maintenance and predictive quality.

Including the Process

With process-based machine learning, the topology of the production floor is modeled precisely to include all the lines, manufacturing stages, machines, and the flow of the product through the system.

This model provides an accurate representation of all the assets via directional graphs and the movement of the product through each stage of the production process.

Process-based Machine Learning

Enter the Algorithms

The production context, critical for closing the gap between the representative data and the real-world manufacturing environment, is added by means of feature engineering. With this contextualized data in place, machine learning algorithms become extremely powerful predictors. While more conventional ML algorithms such as Random Forest and XGBoost may be used, applications designed for industrial process data such as the Seebo platform, use graph-based models such as Hidden Markov (HMM) and Directed Acyclic Graph (DAG).

Since the data fed into the ML algorithms includes the context of the entire operation, analytics produce accurate predictions regarding maintenance, quality, and the supply chain.

These predictions are then presented to personnel and management with enough time to be acted upon strategically, significantly improving the performance of the entire plant.

 

Get in touch for a 1-on-1 demo to see how process-based machine learning can significantly improve your operation’s uptime and output quality.