Machine Learning / Predictive Maintenance

The many flavors of machine learning for manufacturers

Much has been written about machine learning in manufacturing, but it can still be difficult to understand the different approaches.

By Abhinav Khushraj, co-founder and CEO of Petasense

Machine learning (ML) is everywhere. Startups, OEMs, and industrial suppliers are investing heavily in developing technology to collect and analyze manufacturing data. Much has been written about ML in manufacturing, but it can still be difficult to understand the different approaches.

In manufacturing, the application of wireless sensors and ML offers potential to reduce costs and improve efficiency across the entire organization. Companies have created ML-based systems for almost all operational functions, including:

  • Supply chain
  • Quality
  • Production
  • Demand forecasting
  • Predictive maintenance
  • Inventory management
  • Other functions such as finance, sales, and marketing

The IIoT and ML within predictive maintenance


Predictive maintenance (PdM) has widely been seen as one of the most promising applications for ML because it gives reliability teams real-time insight into the condition of physical equipment. Companies can immediately see benefits in terms of better planning and scheduling, reduced downtime, reduced risk, and increased production. According to a January 2019 survey from Tech Pro Research, 79% of respondents currently use or plan to use IoT for predictive maintenance. This was the highest rate for any application.

The IIoT connects sensors on industrial equipment such as pumps, motors, and gearboxes to the internet. As more data is being gathered, ML is needed to make sense of the data. Manufacturers don’t have the resources to analyze all data coming in, meaning that they will rely on algorithms to flag equipment for a human to look at. ML can be thought of as an assistant to the existing reliability team.

Not all ML-based systems for PdM are the same. They vary in terms of domain knowledge, breadth, and complexity. Many suppliers offer technology in this area, and it can be difficult to distinguish among them.

The companies that offer ML-based software for PdM generally fall into the following categories:

  1. OEM
  2. General purpose ML companies
  3. Application specific ML companies
  4. Existing software companies (e.g., EAM, historian)

OEM-provided machine learning


Many OEMs are moving from selling products to a service model. Rolls-Royce has been selling engine hours instead of engines since the 1960s, but only within the past five years have industrial OEMs started shifting to a similar business model. Selling a product as a service allows the OEM to move up the value chain and increase customer entanglement. They can track the installation base and equipment use, as well as gain knowledge about equipment usage that can improve product design.

As OEMs move away from selling parts to selling the function or performance of the equipment, they will need to collect real-time condition information. Several OEMs have partnered with or developed their own IIoT devices that incorporate machine learning algorithms.

The benefit of using an OEM-developed system is that an OEM will have the deepest knowledge about its equipment and often will embed this knowledge into the algorithms. The downside is that it is only relevant for that maker’s equipment, and most industrial facilities have equipment from multiple OEMs. It isn’t practical to maintain and monitor separate systems for each OEM.

General-purpose machine learning software


Many of the analytics software startups provide building blocks to be applied to any manufacturing problem. The benefit of this approach is that you don’t need separate software programs for different functions (e.g., maintenance, operations, finance). Companies with data scientists on staff will benefit from the breadth of these systems. However, data science is a skill that is lacking at many industrial companies. According to Emory University, a skill shortage of big-data scientists is one of the most significant factors negatively affecting Industry 4.0 deployment.

Several analytics companies provide tools so that a customer doesn’t need to have a data scientist on staff, but the systems vary in terms of complexity. They still require technical and statistical expertise to build models that will accurately predict failures. Many vibration analysis practitioners have pointed out that it is difficult to predict failures by algorithm alone without knowing the physics behind how and why failures occur.

Application-specific machine learning


A third category of companies offers pre-built applications of machine learning—these are applications that are already built for specific manufacturing problems such as quality improvement or predictive maintenance.

The benefit of pre-built applications is that they are optimized to solve a particular problem. An algorithm that is trying to answer predictive maintenance questions such as “What equipment is at risk of failing and what is the root cause?” does not have the same variables as an algorithm that is answering quality questions such as, “What can I do to reduce off-spec paper?” They are taking different inputs, using a different frequency of data collection, and making different types of recommendations. An algorithm that has already been used hundreds of times for the same application will perform better than one created for a different purpose.

For PdM, the best-performing algorithms are going to be the ones that combine physics-based knowledge with machine learning. Users can go beyond anomaly detection to understanding the root cause of problems.

The downside of these algorithms is that they are built to solve a specific problem and cannot be scaled across departments.

Evolution of existing plant systems


EAMs, data historians, and other plant systems provide deeper insights about equipment using machine learning. The benefit of using the ML algorithms within these systems is that it leverages existing investments. For example, both SAP and Maximo have advanced ML capabilities to predict downtime, and it is built from data that is already collected. The downside is that it is based off historical data rather than actual condition, and it is only as good as the data within it.

Another example of the evolution in existing plant systems is data historians. There are advanced algorithms built from process data that can provide information about asset condition. The main drawback of these systems is that they are taking information only on assets that are already instrumented. Many facilities don’t have continuous sensors on balance-of-plant (BOP) equipment, so they won’t have any visibility into asset condition unless they add new sensors.

Important considerations when setting up the program


Each form of machine-learning offering has merits, depending on the resources and goals of the company implementing it. When evaluating what type of ML-based program makes sense, industrial facilities should focus on the following three criteria:

  1. Time to implementation
  2. Industry focus
  3. Scope

Time to implementation


The companies that offer more of a platform or “do-it-yourself” ML tools will require more expertise and typically a longer payback period. This is balanced by the scope of these systems, which can often be applied to several different problems or applications rather than on one specific need. An ML algorithm that already has been tested and tuned specifically for predictive maintenance (as opposed to general tools) will be much easier to implement, require fewer technical resources, and generate returns faster, but it will be more limited in scope.

Industry focus


The variables (or features) that make sense for an automotive manufacturer are very different from the ones most relevant for an oil refinery. You cannot just apply a generic model across different situations; you need to have the industry and process knowledge that makes a given model work better in specific circumstances. One of the most important steps in implementing an ML-based system is feature selection. Companies that take the approach of throwing all data into the algorithm without applying asset knowledge risk confounding the algorithm or making connections that don’t really exist.

Scope


If a company has a digital transformation strategy in place, it should select a vendor that fits with the company strategy. With the proliferation of options, manufacturers need to give consideration to factors such as security, breadth of application, and interoperability with existing systems. The most successful programs are ones that have collaboration between corporate IT/OT and plant-level reliability teams. There needs to be a match in terms of culture and capabilities given that digital transformations are major undertakings.

Conclusion


In the 2018 Gartner Hype Cycle of Emerging Technologies, IoT platforms and digital-twin models were labeled as at the peak of inflated expectations. This means they have received a lot of media attention; there have been some early successes; and there have been a lot of failures. The benefits have yet to be fully realized, and many companies don’t take action because they don’t see the business value.

For companies looking to get started, the first step is to look at internal resources, specific business needs, available data, and budget considerations. They should evaluate each vendor in terms of how well they align with company goals and consider the three dimensions of industry expertise, time to implementation, and scope.