Advertisement

The limitations of applying machine learning to industrial machinery

August 21, 2018

Eitan_Vesely-August 21, 2018

Most industry analysts expect that Artificial Intelligence and Machine Learning will have a transformative impact on the industry.  Machine Learning features prominently in research reports including Gartner’s Top 10 Technology Trends for 2018 and McKinsey & Company’s Artificial intelligence: The time to act is now. However, practical considerations for the real-world applications of Machine Learning are often missing from these analyses.

Even if executive decision makers recognize the economic potential from the digitalization of assets, engineers tasked with the design of Industry 4.0 compatible equipment lack direct experience and a formal education in the data science discipline. It is not the physical integration of sensor or hardware which needs to be addressed, but the dynamic use of Big Data that is generated from these sensors.

In my experience, industrial organizations are struggling to build Machine Learning competencies and that the ultimate responsibility lies with senior management.  

This issue is relevant for both the industrial plants and the OEM manufacturers of the industrial equipment. In a research study conducted by Emory University and Presenso, Operations and Maintenance employees were cautious about the rate of adoption of Machine Learning based Predictive Maintenance practices.  Fred Schenkelberg, a reliability engineer who is a lecturer at the University of Maryland summarized the sentiments of many O&M professionals: “Gathering more data does not solve anything. We already don’t use the data we have nor know what to do to analyze and use the data today.”

Skill Shortages are Preventing Machine Learning from Widespread Adoption

The noise about the potential of benefits from Machine Learning hides some basic realities.

First, there are simply not enough qualified big data scientists and engineers in the labor market.   There are too few high caliber institutions of higher learning that can produce sufficient quantity of qualified graduates in the field.  Deep-pocketed tech and financial services companies are attracting talent with highly competitive salary packages and the opportunity to gain cutting edge work experience.  The industrial sector lacks the resources and the glamour to effectively compete against these entities.

Second, an alternative solution to labor shortage proposed by Gartner -  Citizen Data Scientist – is not practical on a large scale.  This role has been defined as a person “who creates or generates models that use advanced diagnostic analytics or predictive and prescriptive capabilities, but whose primary job function is outside the field of statistics and analytics.” The problem is that we still lack the tools to empower Citizen Data Scientist to perform impactful roles.  

Is Automated Machine Learning the Silver Bullet?

There is a lag between the advances in data science and its application to industry. Is Automated Machine Learning (Auto-ML) the solution?

Auto-ML is creating a lot of interest within the data science community because it automates many of the time consuming and repetitive data science tasks such as data pre-processing, model selection and hyperparameter optimization. Many view Auto-ML as a game changer.

In fact, both Google’s Cloud AutoML and Amazon’s Machine Learning are positioned to help non-data scientists apply Machine Learning algorithms to their data.  Unfortunately, these tools are too generic to be applied to most industrial Big Data scenarios.

The challenge with innovations such as Auto-ML is that they are developed primarily as tools for data scientists. Within the data science community these innovations can accelerate Machine Learning processes.  Generic versions in the hands of Citizen Data Scientists are of limited value.

Machine Learning is not a Standalone Solution

Another factor that is often overlooked is that Machine Learning is a dynamic discipline compared with traditional industrial hardware equipment which can be built to last decades or even software which has a lifespan of as little as three years.  

Furthermore, for Machine Learning to be applied to large data sets on an ongoing basis, it needs to be integrated into a scalable (software) solution.  The graph below depicts the development of a Machine Learning solution that is based on the Software Development Life Cycle or SDLC.

click for larger image

The Software Development Life Cycle. (Source: Presenso)

It is important to recognize that building an application based on Machine Learning requires experience with a software development process.  These are two different skillsets as Machine Learning engineers are unlikely to have experience with most aspects of software development including design, coding, testing and ongoing upgrades. 

Without significant resources, these two disciplines – Machine Learning and Software Development - cannot be incorporated into industrial machinery.

No Quick Fix

The same way that organizations cannot build new competencies overnight by hiring data scientists, engineers without formal training in the discipline cannot become data scientists.  

The options to embed a Machine Learning solution into equipment are limited to outsourcing the development of custom Machine Learning applications or partnering with solution vendors with this expertise.  This is not a one-time purchase of an off-the-shelf software package because by its nature, the Machine Learning discipline is evolving at a rapid pace.  Processes for algorithm selection, calibration and optimization that were common in 2017 will be antiquated by 2019.

Conclusion

The first step in addressing the challenges for applying Machine Learning to industrial machinery   is to acknowledge them.  They cannot be solved incrementally by adding a small number of Machine Learning experts to the R&D organization. 

Whether it is forming alliance agreements with third-party entities who have Machine Learning capabilities or even acquiring organizations with these resources, it is incumbent upon senior management – and not engineers in the field – to find and fund a long-term strategic solution.


Eitan Vesely is the CEO of Presenso. He was previously a hardware specialist and a support engineer for Applied Materials, where he specialized in software-hardware-mechanics interfaces and system overview. Mr. Vesely holds a BSc degree in mechanical engineering.


 

Loading comments...