512-800-6031 editor@ramreview.com

Do a quick search on the accuracy of machine learning (ML), artificial intelligence (AI), or Internet of Things (IoT) predictive maintenance (PdM) to understand what the potential implications are. Most of what you’ll typically find are white papers and other material describing how great AI is with condition-based maintenance (CBM). However, with little information or success stories to back it up, much of that information is related to simulation.

While there are some excellent systems out there, primarily designed by experts in IoT-based CBM systems, there also are a lot of “textbook’ systems developed by sources with little to no experience. In fact, we’ve seen quite a rise in field projects related to “Why is my AI telling me to do something, but it’s wrong?’

According to a paper titled, “Establishing the Right Analytics-Based Maintenance Strategy,” by McKinsey and Company (which I found buried among multiple pages of online marketing material), a majority of companies were not seeing a return on investment (ROI) from AI-driven predictive-maintenance applications. The primary reason appears to be that most AI systems are built based upon limited knowledge or substandard capabilities that consider the best result to be a significant number of false positives. With a false positive, systems may be shut down or maintenance may be performed when there are no issues due to things such as noisy data.

Most of the articles on the benefits of IoT-based maintenance identify elimination of the technician or expert as a benefit. This is equally concerning as the primary purpose of ML systems is to reduce repetitive work and move non-routine work and decision-making to a human being (re: IBM Watson). The concept behind AI surpassed the idea of “Industry 4.0” (elimination of workforce) to cobotics, or “Industry 5.0.” Cobotics involves the human-machine interface as AI is rules-based and explicit while complex decision- making is not rules-based and implicit. The concepts behind Industry 5.0, due to the failures of Industry 4.0, have simply been to keep the human involved but let AI assist in performing simplistic tasks, thus making the human work more rewarding.

One of the significant challenges as we move through this new world is that of education associated with understanding the principles behind the systems that are being studied. A lack of understanding of the systems where Machine-Learning tools will be applied, the bias of programmers based upon what experience they may have, and the reliance to fall back on what they were taught, is an extremely present issue.

For instance, in one current-signature-analysis IoT application, we investigated the hardware which was only capable of 1,500 samples per second.  That rate results in a Nyquist limit (FMAX) of 750Hz (45,000 CPM), on an 1,785 rpm motor with 28 rotor bars. At 1,790 rpm (29.833Hz) calls of “mixed eccentricity: were made, which would require a frequency range of (29.833Hz * 28 rb then +/- 1, 3, 5 times line frequency + running speed frequency).  That would be 835Hz plus the sidebands, or up to 1,165Hz, for a sample rate of at least 2,400Hz. How a call of eccentricity was made is questionable. In the end, the problem was not what was called when a 12kHz analyzer was used, and there were no issues with the machine after multiple repair attempts had been made following the original IoT-device findings. This was one of many cases across multiple technologies, and those were only the ones that we’ve dealt with.

Another condition that we’ve had experience with is application of “unsupervised” learning systems. During the learning period, the present condition of the machine is developed and then monitored for deviation. If a bearing is in bad shape, there’s misalignment or some other condition impacting reliability and energy consumption that is now considered “normal.”  In many cases, the condition in question is near failure and the IoT device continues to monitor the system as green right to failure. The drift away from expert systems in which we set band alarms has been, well, alarming. Now, anyone with a laptop and access to “AutoML” (yes, literally a method that does everything in the ML pipeline) can develop a system based on real or simulated data, i.e., basically, an edge-based system, in mere hours. However, it doesn’t end there.

Most of those systems are cloud-based and encouraged to be so by a small handful of cloud-based AI companies. Each of those companies has its own catalog of solutions based on specific models that are preferred within the organization. When applied in the cloud, the model development must fit exactly into that specific model, or it will not perform. If a system is primarily a “deep learning” neural network, which is great with images and facial recognition, but not always with CBM (or the mere suggestion of true remaining life), the developed system must bend to fit.

Ultimately, the most accurate systems have been those developed by experts in their field as means for automating or transferring knowledge. The challenges in such cases still relate to personal bias of the data scientists and programmers, should the expert rely on assistance. All of the marketing hype associated with IoT/AI/ML systems requires the end user to be more vigilant on the true capabilities of the systems in question. This means that decision-makers must become more knowledgeable in how these systems are developed and be able to come up with some tough questions.  As is often stated in the ML/AI industry: not every application calls for AI/ML.

Based upon the aforementioned McKinsey report, it was estimated in 2019 that 25% of businesses were using IoT to connect their equipment. Improperly selected strategies resulted in excessive cost increases in the worst cases due to false positives and did not generate expected ROI in others. Organizations that properly applied strategies to meet needs and understood the capabilities of the ML systems did see an improvement, but rarely near the hype suggested when solely used for CBM.TRR


ABOUT THE AUTHOR
Howard Penrose, Ph.D., CMRP, is Founder and President of Motor Doc LLC, Lombard, IL and, among other things, a Past Chair of the Society for Maintenance and Reliability Professionals, Atlanta (smrp.org). Email him at howard@motordoc.com, or info@motordoc.com, and/or visit motordoc.com.


Tags: reliability, availability, maintenance, RAM, asset management, electric motors, Internet of Things, IoT, predictive maintenance, PdM, condition-based maintenance, CBM, machine learning, ML, artificial intelligence, AI, cobotics