Machine learning spotlight: Industry 4.0 and predictive maintenance

June 29, 2018

Story

Machine learning spotlight: Industry 4.0 and predictive maintenance

I interviewed Eitan Vesely, CEO of Presenso to discuss Industry 4.0, the Smart Factory, and how companies like Presenso can fill the knowledge gap with Machine Learning for Industry 4.0.

Industry 4.0 is characterized by applying cloud and cognitive computing to current automated and computerized industrial systems resulting in the ability to create smart factories that monitor physical processes, identify issues or optimizations, and perform iterative refinement or proactive maintenance and updates. A recent study was released by Emory University and Presenso called The Future of IIoT Predictive Maintenance.

The study is focused on predictive maintenance current state, implementation, resulting impact, and future needs identified within smart factories. Over 100 operations and maintenance professionals across Europe, North America, and Asia Pacific participated. The results showed that while there was good satisfaction with existing predictive maintenance environments, the modeling and machine learning aspects are lagging behind where spreadsheet based statistical modeling has not been replaced by more advanced capabilities.

One of the biggest issues raised by the study is the skill shortage of Big Data scientists and lack of understanding of machine learning technology and techniques. I interviewed Eitan Vesely, CEO of Presenso to discuss Industry 4.0, the Smart Factory, and how companies like Presenso can fill the knowledge gap with Machine Learning for Industry 4.0.

How did Presenso get started?

VESELY: We first thought of the idea when I was systems support engineer at Applied Materials in 2013.  My job was to troubleshoot equipment after breakdown had already occurred and I was required to sort through masses of data to identify the root cause of the failure. At the same time, Deddy Lavid, the co-ounder and CTO of Presenso, was looking for a research topic for his master’s thesis in machine learning. Only a year or so later, while Deddy was working on an energy consumption prediction algorithms, we jointly decided to found the company.

Eitan and Deddy recognized that hidden within exabytes of sensor data generated by industrial machines were micro-patterns that could warn them in advance of failure. These patterns were undetectable to even the most advanced industrial monitoring tools. Even data scientists working to improve machines’ uptime lacked the tools to find these patterns. They received two million dollars in funding in 2016 and less than two years later, they have released Machine Learning analytical tools for Predictive Maintenance using advances in data science such as Automated Machine Learning (AutoML). Presenso fills the Big Data skill gap by making these tools accessible to maintenance and reliability professionals.

Our client list includes some of the largest industrial producers in Europe and North America including Wien Energy, MAN Diesel, Ansaldo Energy, Total-Eren and others. We have also received positive analyst coverage and were selected as a 2018 Gartner Cool Vendor for AI across the Supply Chain.

What are key predictive maintenance goals and objectives? Are they different between a factory floor versus a vending machine for example?

VESELY: When we started the company, the driver for selecting a predictive maintenance system was to minimize machinery downtime.  While operational goals are still important, today customers speak to us about the economic potential from increased uptime and higher production yields. This reflects the recognition from executives of the strategic value of Predictive Maintenance. At a high level, the business objectives of Predictive Maintenance for a factory floor and a vending machine reflect the requirement to prevent breakdown and the loss of revenue.  The similarity ends there.

A vending machine is a small machine with small data. Let’s assume that a vending machine can be monitored remotely.  For instance, if the thermostat exceeds a certain temperature, a maintenance work order is activated.  

Factory floors contain hundreds and even thousands of sensors. When Machine Learning is applied to the sensor-generated Big Data, algorithms are trained to detect anomalous behavior or patterns of anomalous behavior.  Using this data, the root cause of the breakdown can be traced to the sensor where the anomalous behavior originated.

The vending machine is simply monitored for signs of a breakdown. In most cases, by the time an alert is generated, the breakdown has already occurred. The key difference with Machine Learning applied to the factory floor is that the factory breakdown can be anticipated hours or days before it occurs.

How does predictive maintenance affect IoT architecture? How do you plan for this up front?

VESELY: There are different scenarios for deploying Predictive Maintenance. There are predictive maintenance solutions that are dependent on IoT architecture. To utilize the analytics component of the IoT platform, the factory is required to make an upfront commitment and investment in an IoT architecture at either at a plant or enterprise-wide level. This is the best-case scenario for platform vendors seeking to cement long-term relationships and cash flows.  However, locking into one solution vendor limits the flexibility of the organization itself.

There are Predictive Maintenance software solutions that analyze SCADA data from a plant’s existing historian database. In this way, the Predictive Maintenance solution is not dependent on a specific architecture or vendor. This is the model used by Presenso.

How much of the machine learning and intelligence needs to be at the sensor, edge, and cloud for an effective predictive maintenance solution?

VESELY: There is no one size-fits-all answer to this question. There is a lot of discussion about edge versus cloud today. The selection of edge versus cloud is dependent on the definition of the “real time” requirement. “Real-time” cloud deployment assumes a latency of a few minutes which is acceptable for most industrial scenarios.  For instance, with wind turbines, our systems analyze data every 10 minutes.

There are other applications, such as medical devices or electronics where there is less tolerance for delays and latency. In this case, edge is the optimal analytics processing solution.

It is important to match the solution with the underlying maintenance requirements. In my experience, most processing industries can rely on cloud-based Predictive Maintenance solutions.

What are the key takeaways about predictive maintenance and smart factory applications?

VESELY: The Predictive Maintenance discipline is evolving rapidly. The big change is AutoML. Advancements in Big Data are being applied to predictive maintenance to address the lack of data scientists in the labor market. The data science community has recognized the need to automate many laborious and time-consuming tasks such as data preprocessing and algorithm selection. The next wave of Machine Learning PdM solutions will likely incorporate AutoML.  

Another trend is that OEM’s recognize that hardware is becoming a commoditized item and to stay competitive, they will need to integrate smart technologies into their product offerings. Business models and product offerings will need to change because OEM’s cannot sit on the sidelines while technology vendors provide value added analytics and services.