An interview with Lou Zhang, Chief Data Scientist, MachineMetrics
While buzzwords such as predictive maintenance, artificial intelligence, digital twin and augmented reality have promised to enable the fabled digital transformation of manufacturing, when it comes to Industry 4.0, most practical applications start and end with machine connectivity. And when it comes to driving value, look no further than answering these questions; “What’s happening?” and “Why is it happening?”
Simply put, most manufacturers are unable to see what’s actually happening on the shop floor in real time because their machines are not connected to any sort of data collection or data visualization system. This inability to both see and use data to drive continuous improvement leads to massive inefficiencies that affect every component of a company’s operations, from the shop floor all the way to the C-Suite.
That said, as the excitement around the opportunity presented by AI continues to grow, we conducted an interview with our very own Lou Zhang, Chief Data Scientist at MachineMetrics, so he could give us his perspective on where AI lands within the Analytics Journey and its relationship to technologies such as machine monitoring and data collection.
The industry is progressing in the right direction, but we’re tackling a number of very difficult challenges at the moment. Manufacturing as a whole struggles with modernizing effectively in the face of antiquated and ingrained cultural traditions. There may not be as much of a bias for action as other industries – when you’re producing physical parts, often times the metric is your tangible end product. As long as the end product looks good and sells, you sometimes don’t want to bother with the rest of the process. This makes it hard to drive change and modernization. We find that in general, manufacturing is 5-7 years behind other industries in terms of adopting new technologies like machine learning. In short, we’re beginning to use AI but it’s only in the nascent stage.
For machine monitoring specifically, it gets even harder. Manufacturers can be hesitant to have their data leave the factory – and most machine monitoring companies therefore submit to doing an on-premise implementation. You’re looking at the problem of siloed data, which makes it tough to aggregate enough data and across sufficiently diverse domains to actually train an AI model. Imagine if Netflix tried to build a recommendation engine but only had data from one household.
So, what’s happened in the last few years is that you get all these one-off, single-purpose models that aren’t all that helpful outside their respective domain or company. MachineMetrics aims to solve this problem by pushing a pure-cloud solution – we do not offer on-premise solutions and are therefore able to pull together a representative sample of manufacturing processes across the discrete manufacturing space. This has helped us begin to develop more general-purpose AI algorithms for detecting failure on multiple different types of processes and machines. We’ve seen some success in this and have published papers and received patents for our work, but there’s a long way to go as this is totally greenfield territory.
Even after tackling the issue of siloed data, there still remain many outstanding issues both technical and cultural – the data can be very messy and inadequately labeled, shop employees may resist any implementation of AI in the factory, seeing it as an affront to their job security, and there can be a general lack of understanding of the capabilities and limitations of AI in its current stage. The industry is slowly working through these problems though, and as time goes on, adoption of AI will only become greater.
The two general types I’m most aware of are the low-cost/free types that are specifically for small manufacturers, and the huge, enterprise-level applications that need to be customized for different manufacturers like the ones sold by giants like IBM.
On the low-cost end, these can be quick to implement but have extremely limited capabilities – perhaps only being able to give you a display of when the machine is up or not and what percentage of the time it’s been utilized today. They can also be buggy and unmaintained. Generally they’re known as lightweight solutions that solve one very specific problem, like either tracking parts or monitoring uptime, but not both.
On the higher-cost, enterprise end, these require long implementation timelines and a huge upfront cost. According to the SBA, over 50% of manufacturing in the US and over 90% of exports are from small and medium size companies. This is a huge market that doesn’t have the resources or time to deal with corporate behemoths that need to do extensive customization to their software for it to work with each different machine shop.
I would also add that Cisco reports that 76% of industrial IoT implementations have failed on the enterprise end so even manufacturers with more resources struggle to achieve value with generic IoT platforms.
MachineMetrics was started recognizing that there’s a huge, untapped market for a vertically integrated IoT platform that’s plug-and-play but also feature-rich and extensible to create continuous value and innovation.
We should seek to remember that we’re only at the beginning of a paradigm shift here, so overconfidence and exuberance for AI that’s disproportionate with the actual capabilities of it is a huge downside we see. Yielding decisions purely to AI without human context is extremely damaging – smart managers will always seek to understand why the AI recommendation occurred and supplement it with human input. Just like with relying on AI for anything, like self-driving cars – you need to stay awake still or you’re going to crash and burn.
Yes, computer vision is a huge area that’s maturing in manufacturing. Installing cameras inside the machine and around the factory floor is helping manufacturers automatically identify product defects, track inventory, and generally make the factory a more automated experience. Again, this is in its infant stages so impact is limited, but as time goes on this type of technology will only get better.
The ultimate goal of AI for machine monitoring is not to replace humans, but to supplement their expertise with additional, computer-guided capabilities to make the factory as a whole run smoother. We find that shop culture actually gets better with more visibility – people can reference data when driving decisions and the transparency helps foster an environment that’s healthier in the long run.
As to replacing jobs- in our experience, when a piece of software is able to automatically track the number of parts created and how often the machine is up or down, workers’ time is freed up to do higher-value tasks like running initiatives to improve efficiency across the factory floor with all that data that’s collected by the computer program. The higher revenue that the company realizes can be reinvested into worker-training programs, so that blue-collar workers doing rote tasks can be retrained into white-collar workers with a keen mind for management and strategy.
Speaking broadly, my prediction is that AI will be able to help manufacturers drive down costs through preventative maintenance and increase revenue through increased production and efficiency gains. As to the next paradigm shift – it’s hard to tell these things. If people knew the next big thing to invest in getting rich would be easy. I think continually pushing on transparency and visibility in the manufacturing workplace will be critical. I’m excited to see where that brings us.
My advice is that they should start with the basics – walk before you run. If you really have no digitization or automated visibility into your processes, start with descriptive analytics. Just knowing when your machines are running or not can help you uncover obvious things about your business. For example, we’ve had customers not realize that some of their machines were sitting idle for two hours at the beginning of their first shift – something like this can help save you thousands of dollars right off the bat without any fancy machine learning algorithms.
As to costs and considerations, it’s important to get buy-in from not just one party, but multiple key stakeholders in the organization before embarking on a machine-monitoring project, even without the AI. It’s a complicated venture into new, pioneering territory for most organizations and can easily fail without organizational maturity and a steadfast eagerness to see change and improve.
Besides that, evaluating if your company is the right fit for machine monitoring and advanced analytics is important – managers should be considering what types of efficiencies they’re looking to drive and realize it’s not a magic box that will solve all their problems without effort on their side. The journey into AI and machine monitoring is a partnership which requires constant learning and a thirst for knowledge on all sides. If this isn’t the company’s attitude, then benefits will be limited. But if there’s organizational buy-in and a desire to be a long-term partner, we’ve seen tremendous successes that have transformed the way companies do business.