Setting your organization up for predictive maintenance success - with ISPE and Johnathan Bonner
Show notes
Welcome to the Trend Detection podcast, brought to you by Senseye Predictive Maintenance – the platform which enables predictive maintenance at scale across all of your assets, across all of your plants.
This session was originally recorded for ISPE and their Technical Tuesday session.
Watch the video here: Technical Tuesday: Boosting Pharmaceutical Efficiency Through Scalable Predictive Maintenance
Find out more information here: Homepage | ISPE | International Society for Pharmaceutical Engineering
Transcript
Welcome to the trend detection podcast powered by Siemens. Thank you very much and hello to everyone online. So, let's start off with what is predictive maintenance. And I always do like to start off at this point because whenever I engage with new people or a new audience, some of you would have heard about predictive maintenance, potentially even investigated predictive maintenance. But quite often, the understanding of what predictive maintenance is falls in line with machine magic rather than machine learning. So predictive maintenance is an evolution from traditional condition monitoring. But instead of having a specialist go around the facility, taking manual measurements, then uploading them to the computer, analyzing them and reporting back on them, this function is now automated and online, so continuous. Also, with traditional condition monitoring, it was a snapshot in time. When I used to run uh condition monitoring programs, it would be once a month, you'd manage to get to that machine again on your route-based, time-based schedule, and you would take a snapshot to identify the condition of that asset at that time. However, with online continuous monitoring and an understanding of an uh an operation, a machine's behavior, you can really start to get amazing insight into the condition of that asset, as well as highlighting signs of degradation. But like with any good condition monitoring engineer, a predictive maintenance solution should evolve and learn. Learn about failure modes, failure patterns, maintenance events, and give you recommended actions based off of that. It's also important to understand that predictive maintenance is not a technology, it's a maintenance strategy. Predictive maintenance cannot be installed in a site. Um, really, it's the usage of the information provided by a system that enables a predictive maintenance strategy in your environment. It's also not about predicting when a machine will fail. Quite often I get asked, um, as a requirement, you know, we want to know a year in advance of exactly what the failure is and exactly the day it will fail. And the further out you're trying to predict a negative event like that, the less specific you can be. Really what is of use to industries, maintenance teams, to project teams, is an understanding of the early signs of degradation with enough information that they can go and perform the corrective action when it best suits their, their teams, their production requirements and their maintenance availability. It's also not a big data problem, uh, especially when you're looking at cloud solutions and having to send large volumes of data to the cloud continuously. That can put a huge strain on your network requirements. So it's not about big data, really looking at the, the fine details inside of it and using machine learning to understand inherently noisy data. It's actually about providing the system with the good quality information. Um, making it, uh, a solid baseline to learn from and highlight results. Predictive maintenance should also not be restricted to your high criticality assets or only your high criticality assets. The real value from a predictive maintenance program in your environment is to cover the entirety of its plant. And probably the most important factor, at least for myself, is that this isn't about replacing expert knowledge or skills you have. It's about making them more efficient. It's about scaling their efforts and availability in your environment. And to look at predictive maintenance, at the core of it is condition monitoring data. And this is data which can come from PLCs, drives, inverters, um, even your process data is used for predictive maintenance. Things like flows, pressures, temperatures, um, understanding of batch requirements and so on. You know, condition-based maintenance is in maintenance, not done on a reactive or planned approach. However, it is done on condition when the machine or asset requires the maintenance to be done. Predictive maintenance sits around all of those and really enables a best practice approach to maintenance in an environment. And with technology being what it is these days. There is now the advent of uh conversational predictive maintenance. So this is now making the information accessible to everyone, and we will touch on that a bit more in a little while. But what does predictive maintenance give? What is the point? So I'm sure you would have all seen a PF curve with within your time. So, what predictive maintenance enables from the very early stages is when you start to see the signs of degradation. It will give multiple sets of early warning. So an indication that something is degrading with enough information that you can perform maintenance before your existing systems will set off an alarm before the the process production machine has failed. But as I said, it continues to learn. So once it does learn and starts to understand what these signs of degradation mean, it can start to be more specific. So from the, the stage of early degradation, from those early signs, you may get some anomalies or trends detected. Then insight descriptions, letting the user know what is happening, um, interpreting the, the data. You may even get some prognostic, uh, forecasting and a lead time to repair indications. Again, long before any conditions or thresholds have been violated on your environment and your existing systems would have notified you that something has gone bang. Whoops, it's all gone wrong. So it's about giving insight early on that is actionable with enough time for users to, to make use of this and to plan in maintenance when it best suits the company and the environment they're working in. It's also important to note that predictive maintenance is not about thresholds. And this is something I've personally have had to try and combat quite a lot in the eastern part of the world, there are hundreds if not thousands of companies with hardware, sensors. They have potentially developed their own software solution, which is rule or threshold-based. Um, and they're selling that as predictive maintenance, and it just isn't. So if we look at the graph on the right-hand side, let's say that is uh the current profile being amps for an electric motor over a week period. So considering your, your production cycles, Monday morning, there's an idle period and then you start producing production uh Y or X. Then you may go down for commissioning, uh, pre-setup and an idle phase, and then you start in production for product Y. You may go down for a little bit of maintenance midweek, um, something that which was planned in. You carry on with production of P Y or product Y, and then you tail off for the rest of the week. So that is a, an example of a typical production requirements and profile of an electric motor for a week's period. So with the, the system in question, the threshold based, the company will come in, install sensors, and possibly even set up some thresholds or limits for you. These could be based off of ISO standards, the size of the motor, the bearing type, the uh rigidity of the, the base and mounting. But then you have a simple threshold limit. Well, it is missing all of the important detail, especially around the operational complexity that you have. So, maybe someone in your team needs to go and reset the threshold. Well, maybe the 2nd time, they have gotten it a little bit more rights, a little bit more fit for purpose for um an understanding of the production requirements for product X. However, it's missed the detail for product Y. So, then the threshold needs to be reset again. And potentially this time, they've captured or managed to get the right limit to capture the event that you're looking for. However, by this time, the teams uh on site that need to use this have lost trust in it. They've been overalerted. They've uh performed maintenance and essentially wasted money because there is limited understanding of the complexity of the asset, um, and the production requirements. So really, it's not fit for purpose. What is needed, and especially within the pharmaceutical industries, is an understanding of normal production complexity, an understanding of your machine state. What is it trying to do? What speed is it running at? What product does it have or batch is it running through, um, what are the, the output capacity requirements? And with that understanding, a system which can highlight the specific signs of degradation with an understanding of machine state. And really, I think this is gonna be a universal requirement, but also very specifically within the phar pharmaceutical industry, is a system which is applicable to every single type of machine that you have. You know, from your process environments, having your fermenters, uh, chromatography columns, uh, Lilysers, all the way through to your warehousing and distribution. If you were to implement a solution or a system that is specialist in one of these areas, fantastic. I'm sure it will be very good at that one area. But when you look at the full range of production um and environments that you have, you could now end up with 5, 10, even 20 individual solutions. Therefore, that is problematic because you now have multiple logins, people that need multiple sets of training. It's really hard to I truly understand the differences in capability, layouts, user interfaces between these different systems. What is needed uh in industry and in the and in the pharmaceutical industry is one solution which is fit for purpose for everything. Predictive maintenance is evolving and these uh requirements are very much uh in your grasp. So predictive maintenance, really, if we think about it as uh online condition monitoring with uh automated knowledge and understanding, that could fall under diagnostics and statistical forecasting. Essentially, how is my machine, my process, my line behaving now, and how is it going to be behaving in the future? In a few days, in a few weeks, in a few months. Uh, the solutions that you are assessing should be able to prioritize maintenance. And again, this is not based off of machine or asset criticality. It shouldn't just be, oh, my very critical, uh, machine is degrading. I need to run off to that immediately at the first sign of, uh, an increase in current or vibration. No, it should be about understanding how Concerned the user should be about an event, about an insight, and to be able to prioritize this so that your maintenance teams and uh project teams can decide which assets they want to work on, which is going to have the highest impact or which is at risk of failing first. And we, with technology being where it is and the um the huge boom in generative AI functionalities, generative AI and large language models have really revolutionized the predictive maintenance space, really making that information accessible to everyone and also giving The ability to ask questions to the system which is doing the monitoring for you. Oh, why would geneative AI even be a thing? Why would it be used for predictive maintenance? Well, what we've found is that there is an inherent lack of data literacy. And this is at no fault, uh, for the maintenance teams. They are extremely experienced, extremely knowledgeable in maintenance, in engineering, in their field. However, Expecting them to have any sort of data analytics skills is in many cases unreasonable, and what we found is they just don't. They also don't have time to interpret graphs. Most of the, the customers that we, we assist, their maintenance teams are stretched thin already. There have been a huge amount of budget constraints limiting their, their team's size, the, the team's availability, and really having to make do with what they have on site. So time is a huge thing. Also, whenever using a new system, who do you ask questions to? If it's just been installed and preconfigured and then if you have to potentially pay the supplier for any bits of uh consultancy afterwards, you are very limited with the ability to ask questions and really get use from a, from a system. Also, language barriers for global organizations, you know, you will have a wide range of languages spoken from your skilled expertise. So being able to speak to someone or experts in your native language so you get a true understanding is difficult. There's also the aspect of lost knowledge. I'm sure we are all facing and experiencing these challenges. In your environment, you'll have maintenance engineers who have been working with you or in your environment or in your field for decades, and now the up and comers, the apprentices coming through. So anything that your experienced engineers cannot impart onto the, the new up and comers, that knowledge is lost. And we all know of uh, that person on site who can walk around and he knows when, uh, when the homogenizer isn't sounding right. He knows that there may be something going wrong and can go and fix it just based off of that knowledge. That is experience which, yes, can be taught, but isn't always imparted on. So that, that type of knowledge and understanding of an environment, of a process, of a, a problem, more importantly, does get lost. It's also it's also difficult to share knowledge and collaborate with global organizations. Quite often you'll find a company doing exactly the same thing. Maybe they have a facility in India and one in the UK as an example. Their processes are pretty much the same. Um, the products are identical. The machinery, may be some variation on OEMs. However, they are like for like and extremely comparable. But how do they collaborate? How do they share knowledge between each other? Often we find these organizational silos in place where there isn't collaboration, there isn't best practice sharing between them. And this is where generative AI comes in. First of all, it's about getting instant answers to your questions. So generative AI is able to interpret the data, the graphs, the, the data analytics for you and present it to the user in the natural language. It also provides the ability to ask questions to, to the, the system. Um, the answers you'll get will be from an inherent knowledge, you know, trained off of the knowledge on the internet. But for companies that are, are feeding their large language models with expert knowledge, it starts to become extremely powerful. So this expert knowledge could be in the form of uh condition monitoring techniques, maintenance and reliability, uh, practices, as well as user manuals, maintenance manuals, decision trees. Um, and having instant answers to that expert knowledge and having instant reference to that knowledge is extremely powerful. Generative AI retains the knowledge forever, and I know forever is a definite term, but for the foreseeable future we could say. Because it is going to continuously learn. It's going to have the storm of Uh, failure patterns of maintenance events. It has reference to the expert knowledge, your user manuals, maintenance manuals, and it can really start to build a deep, deep understanding of your environment and what is important to you and make this knowledge accessible to all of you new up and comers, the, the apprentices who don't necessarily know what this components for or uh what the part numbers of that um asset are. And it makes it extremely easy to share knowledge across all locations because as you would expect, large language models work in pretty much every language available and because the users can interact with it in their native language and get results back in their native language, it means that your, your cross-country and uh cross geography collaboration is heavily enhanced. Really tying in all of this human knowledge into one space and making it accessible to every single one that uses the system. Really, predictive maintenance should be the convergence of the human-machine interface. It should be having a true understanding of your machines, of your processes, using your OT data, IT data, but also to learn from the human element, the expertise that you have on site. But predictive maintenance or any digitalization program should be focused on business outcomes. And we'll discuss some of those in a while. An example of deployments within the, the, the farmer industry. Um, this this specific customer, uh, they make insulin, um, and the machines that we're monitoring for them specifically is the vial filling line, so little vials. One of the key factors though is that we had to use existing data from their existing infrastructure. This wasn't a a project to come in and install huge amounts of hardware, sensors, data collection devices. It was about can we derive value from their existing infrastructure, their existing investment. And the answer is yes. Um, some of their success criteria and the pain points was on the silicone valves specifically and the degradation that these caused. The fact that the, this company didn't have a centralized visibility, these were happening unexpectedly and halting production until they were able to inspect 20 or 30 silicone valves to find which one had actually failed. So we started working with this company in January 2023. Thankfully for us and I would say for them as well, they actually had tried to develop their own system in-house. Unfortunately, that did fail. However, fortunately for us, it meant that they had done a huge amount of investigation. They weren't starting at a zero knowledge base. They knew what the complexity of predictive maintenance and had done a lot of the data work beforehand. So, quite rapidly, um, you know, a month and a half later, we had the system integration with data being pushed into the application. Their first success was roughly 6 months later, from there. And This first success was just the first interesting event that was detected. And this also goes to show why going for your most high criticality assets isn't always the best approach, because high criticality assets are typically over-maintained. Um, a huge amount of maintenance, effort, time, and money is spent on them. So when you're looking at trying to achieve value and a return on investment, in a twelve-month period, what are, what is the likelihood that something interesting is gonna happen, that something is actually going to degrade, that you can show results and show something working? So that is why we go for value. And again, I will discuss value a little bit later on in this presentation and how we target where the value is. But for this customer, even though the silicone valves were the main cause of frustration and loss in production, it still took over 6 months before anything happened for us to show results. However, with one notification, they were able to achieve a 3 times return on investment. By the end of that year, over 300,000 units or additional units were able to be produced because of reduction in breakdowns. So, early indication that something was degrading with enough time for the maintenance and project teams to assess when they wanted to do the production, to fit it in with their planned stops, and to still maintain production through all of those um additional units. Huge amounts of value in a relatively short space of time. Let's talk about some of the challenges faced. Again, I would say these are quite universal, the maintenance of uptime. For maintenance teams, again, they are very good, very specialized, very uh experienced in performing maintenance. However, when do you perform maintenance? Well, when something fails. You know, the process is something fails, potentially an operator, uh, pick up the phone, call the maintenance department and say, hey, my, uh, chromatography column has stopped, uh, something's wrong. …