How to use production data in a dairy facility – and actually trust it | Ep 8

Henrik Kæmpe (00:00):
That’s the thing about collecting data, if you don’t use them, if you don’t use them for decisions, they are worthless. When we make cheese, the number of measurements we can do per day is very limited.

(00:12):
We can’t make any statistical process control on it. For me, it’s like working with the gold. Cheese production, we use 10 liters of milk for one kilo of cheese. Even small deviation in fat and protein has a huge impact of cost.

Mikkel Svold (00:26):
So if you’ve got a new job at a new cheese production facility, what would be the first data point that you would introduce to that factory?

(00:51):
Welcome to The Food Tech Podcast. Today, we are zooming in on the use of data in dairy productions. We want to talk about how to go from cradle to grave and how to catch and use data in modern dairy facilities.

(01:06):
And today with me, I have Henrik Kæmpe, who is the Operations Development Manager at Mammen Mejerierne, which is the second-largest dairy in Denmark, with an annual turnover of just around 270 million euros yearly across three production sites.

(01:22):
Welcome to you, Henrik. And I think we had a good little talk before we turned on the microphones, but one of the things that I want to ask you is day-to-day, how do you personally use production data? What metrics are you looking at every day?

Henrik Kæmpe (01:44):
See, I’m not looking at them every day. That’s not my role, but I’m really still working with the data and building up the data structure so the organization can work with the data. Because when you ask what do we use the data for, I think we have to see who is looking at the data.

(02:07):
And if we take the different decisions loops, so we have the production, we have the operator, which in a short loop, it could be let’s say minutes, control the production. They look at the data, the hands-on data. Then it could be the department leader typically look into the daily production and look into deviations from normal, look into adjusting targets and so on.

(02:48):
And then if you look at the management, it could be weekly, monthly data to control the production on a higher level, but also be the backbone in taking strategic decisions on what kind of products to make and what should be the cost of the product and so on. Yeah.

Mikkel Svold (03:14):
If we zoom all the way down to the actual production floor at sensor level, what sensors are we talking about? Where are they placed and what are they, I’m going to say counting, what are they measuring?

Henrik Kæmpe (03:30):
It depends. If we take direct process data, it’s temperature, it’s flows, it’s pressure, it’s pH meters. It’s also, let’s say, time that we use to control the production. Then, of course, we have to validate the quality of our products. So at that point of view, it could be measuring fat and protein and moisture, microbiology and so on. And these data, we aggregate them and use them for performance data. So, to look into it could be yield, it could be losses, downtime, energy.

(04:27):
So, we also collect data from our equipment to energy to look into energy consumption, but also to look into how often we have to have maintenance, how long time our equipment is running, and if it’s running stable and so on. So there’s huge amount of data we collect. And I guess we also collect more data than we use. And that’s the thing about collecting data, if you don’t use them, if you don’t use them for decisions, they are worthless.

Mikkel Svold (05:08):
I wonder, when you say you have these three different levels, and that’s something that we will talk about more in another episode as well, but just to get it clear, when delivering data to different levels in the organization, does that data come from the same sensors or is it different sensors? And if they come from the same sensors, I assume that it’s just because you aggregate them or you dice the numbers in different ways to present other results, or how does it work?

Henrik Kæmpe (05:43):
Both. Because we could take a measurement, a pH. If we measure pH in the cheese production, it’s part of controlling, “Is the production going right. Do we expect that we can see if the production is running right, if the bacteria grows that we assume they should grow and so on?”

(06:10):
And then on a later stage, pH, some of it could be from the same data. It could also be another pH measurement is used to part of the quality control. And that, combined with sensory analysis is a way to… When we look at cheese data and taste the cheese and say, “Okay, how is this cheese? Is it as expected?”

(06:45):
Then we can use it to the overall adjustment and the overall, “How do we want our cheese to taste?” So say, “This is more flat tasting,” then, “Okay, we need some more acid in the cheese.” And in that way of course we use an adjust pH, and then as a follow-up loop on that, control the pH again, and see, “Do we get the right target and do we get the new normal for the cheese?”

Mikkel Svold (07:21):
And in your work, you work with developing the operational facility or the operations, what troubles or what issues in production will basically catch your attention and start you thinking and working on new solutions? And maybe you have an example.

Henrik Kæmpe (07:48):
We are in a process now where we started to collect more data than we used to do, and one of the first steps in that is, of course, to validate if we are measuring the right things.

(08:06):
If we are measuring a volume, is it actually the right volume? When we shift from one batch to another, what about the milk and the pipes? Do they belong to one batch or the other batch?

Mikkel Svold (08:25):
So the waste milk or what happens?

Henrik Kæmpe (08:28):
No, it’s not wasted, but does it belong to one batch or the other batch? When we make the shift, for a traceability point of view, that part of milk, where does that belongs? Have we measured correct? And if we have a sensor that breaks down, what do we do? How do we secure that we put in some numbers that are correct? Because if we don’t put in the numbers, we can’t calculate yield or we can’t make a traceability and so on. So, all that, if everything go well and we have set up everything, then it goes well.

(09:14):
But the number of possibilities that things are not normal are actually quite big. Do we measure right? Is sensor the calibrated? Do we measure it right? What do we do if we measure wrong and so on? So, a lot of things to be sure that the numbers that we actually collect, and the values that we collect are validated. So that’s the first step in order to be able to use the data, because if we can’t trust the data, can’t use them. And we also will see the operators that have to use on the management. If we don’t trust data, then we have a huge chance to get that in place. Yeah.

Mikkel Svold (10:11):
Yeah. We’ve circled around that a little bit, but maybe we should get quite precise in what kinds of data would you recommend that a dairy your size should collect? What kind of data is important?

Henrik Kæmpe (10:37):
It’s directly data related to the production. Yeah, I mean volumes, the temperatures, pH, flow, and then of course all the energy consumption data. And then if we take it into a broader perspective. Then because we use that data to calculate yield and the amount of cheese, if take cheese as an example, amount of cheese or amount of product we have produce, then all the planning data is actually part of the production too has to be used in the production.

(11:26):
So the sales forecast, which it’s of course not from sensors, but the planning forecast, and convert that to production forecast, and then setting up a production plan so we produce the right amount of products that we have the demand.

Mikkel Svold (11:52):
Have you any examples of sales forecasts that were completely wrong, and were you either overproduced or under produced significantly?

Henrik Kæmpe (12:04):
Yeah. Yeah, yeah. That’s, of course, the thing. That could be wrong numbers. There could be strategic, it could be in our system, but our customers also sometimes they have a campaign, and suddenly something else become interesting and the campaign is canceled.

(12:25):
And of course, we have agreements and so on, but sometimes we are also ahead of planning the production and so on. So, there can be incidents where we have to navigate and find another way.

Mikkel Svold (12:39):
Yeah. Yeah. When you collect all this different data, so you have supply chain data, you have… Not supply chain. You have sales data coming in, you have planning data forecasting all those numbers, combined with the production data.

(12:56):
How does that live inside of Mammen, the organization? Because that’s a lot of data points. So you need them split out. And I imagine that if you put them all in one big dashboard, even if you visualize them very nicely, it’s still completely information overload. What do you do?

Henrik Kæmpe (13:18):
Yeah, basically if you take all the data, it’s from outside in data, which is the planning part, and then from inside out, which basically, and also where we started the dialogue, basically starting with the milk and our measurements on the milk, and our forecast on how much milk we receive. So basically, coming from outside in and inside out, and then we have to balance it in the middle.

Mikkel Svold (13:50):
Do you work with organizational=level dashboards? So people in operations, they have one dashboard that they can look at and then you go up the ladder, and are they like… I have an image in my head where you have this screen behind you with all the metrics you want to look at. Is it like a dashboard that’s in your face or do you need to go in and find it?

Henrik Kæmpe (14:16):
Both ways, but we have both. The most important about both KPIs and dashboard is that they are showing data KPIs that the people working with them has influence on. Although it has to be relevant data for the right people. So if we start in the production, it has to be KPIs or production data that the operator needs, and the operators need to take decision upon.

(14:59):
So look at it, say, “Are we in control? Yes, we are in control. Don’t spend time with them. Are we out of control or are they out of limit?” Then adjust and get on the target again and get inside the limits. If we take one step up at, let’s say, department manager-level, middle management, it would be basically the same.

(15:29):
They should not waste time of looking at data that the operators can handle, and they should not spend time of data that it’s actually management decisions, “What kind of production should we set up, how should we do it on a strategic level?” So, the task when we set up dashboards and we set up KPIs and so on is that they are relevant for the relevant people.

Mikkel Svold (15:57):
How do you make sure they are?

Henrik Kæmpe (16:01):
Try to, and then sometimes we implement the KPIs and keep them on the dashboard and say, “Okay,” we put them on because I believe, or we have a working group that think they are relevant, and think, “This must be relevant for site manager.” And then we use them and have it on the KPI board, and then we find out, “Okay, this is not relevant for us to look into. It has to be, let’s say, one closer to the shop floor.”

(16:39):
So out of our dashboard and to someone else’s dashboard, it could also be the other way around, that we’re looking at KPI, that we don’t understand because we don’t understand the underlying things happening. We see a dashboard and a KPI, and we have a feeling we don’t own it and we don’t know how it becomes how it is, it’s color, why is it yellow?

(17:10):
Well, it’s yellow this week. So sometimes we have to change it to the underlying KPIs in order to react on it, because if we don’t own the KPIs and we don’t are capable of or able of reacting on them, and don’t tell us things that we have to act on, then it’s not relevant.

Mikkel Svold (17:32):
I think that’s a very good point. And I would suspect that this is a place where a lot of people go wrong. And again, I can just even look at my own day-to-day and looking at it’s very tempting when you have a lot of data to look at all kinds of metrics, but making them actionable and making… That, I think, is the hard part. Do you find that it’s easy to implement actionable insights? Because you come, of course, both with experience from Mammen, but also your work before that.

Henrik Kæmpe (18:15):
I think it goes on both ways, that we have KPIs that we can’t control and we can’t discuss, and we are not the right forum to discuss it, but we as a site management look at it and are not able to discuss it because we don’t have the information. And we also see that we end up working from KPIs, from decision on tasks that we should actually let people closer to the production floor work.

(18:56):
Because they are the right people to solve it, but also sometimes let them grow in their role, taking decisions on a higher level than they are used to, which they grow into the, let’s say, future career by the management letting them work with things that are a bit more complex than they used to.

(19:31):
So, it’s actually quite interesting, too. It’s a lot about people, and it’s a lot about involvement. And it is easily to, from a management point of view, to think, “Okay, this must be an operator, this must be a team lead, this must be the management.” But in reality, have to go and look and validate if that’s correct, and if the right discussions comes out of the right meetings and of these KPIs.

Mikkel Svold (20:05):
That sounds like a really hard task when you’re sitting back office and trying to do adjustments?

Henrik Kæmpe (20:15):
Yeah, yeah, but it’s quite satisfying when it works, when the decisions loops are right when you attend a meeting. And then there could be a production problem that people solve. And you join a meeting and you can hear that you have a couple of operators and a team lead, and they discuss something.

(20:40):
And you can know if you hear, “Ah, the KPI was like that and it was an out liner.” And we have seen it a couple of times. And when you can hear that some of the job we’ve gone, the system we have built, the KPI set up we’ve built is actually used and used for decision at the right levels, then that’s quite satisfying.

Mikkel Svold (21:06):
Yes. Yeah, maybe the numbers, they will turn out to become better profitability or higher yields from your production, but having that emotional connection with someone whose job you’ve actually helped make easier, I can understand why that is also quite appealing.

Henrik Kæmpe (21:37):
I’m much a people manager. And without people, we do nothing. But connecting, getting the people connected in a way, and get the discussion right. And then getting the right technical discussion, where we set up production, set up measurement, end up involving our technical department, buying the sensors that’s needed in production in order to enable operators.

(22:16):
And then at the end, to collect data. For me, it’s really satisfying because I know end of the day I’m also driven of driving business. So at the end of the day, when we on the shop floor, set up things better than we had before, then it will end at the basically better bottom line of the company.

(22:44):
And that dependence from sensors to the SCADA net, collecting data from the sensors, presenting them for the operators, the operators taking decision, management taking decisions, that whole change is fully connected and fully dependent of each other.

(23:15):
And it’s one of the things I think a lot about and use in my work, is what kind of things that goes from sensors and that goes from task from operators, maintenance zone, do we have to perform on each level in the organization.

(23:39):
In that way, really thinking about the structure in the organization in order to support what do we have to do where in the organization? Because when we do it right, and involve the right people, we can build rockets and fly to the moon again.

Mikkel Svold (24:00):
Fly around the moon.

Henrik Kæmpe (24:01):
Fly around the moon. Exactly.

Mikkel Svold (24:06):
I want to come back down to some of the more nitty-gritty stuff on working with the data that you do. You collect a lot of data on all kinds of different sources. Can you try and take me through the data processing? How do you do that? What happens, what’s important, what’s aggregated, how does the processing look like?

Henrik Kæmpe (24:32):
I have started slightly different place, but getting into it. I’ve worked many, many years with statistical process control in order to improve cheese production. And cheese production is mostly a batch process, at least in the cheese vat, cheese production that’s not in cheese vat, but mostly cheese production is batch related.

(25:05):
And SPC statistics process control is common used tool in order to control production, and in order to adjust production and keep it on track. And if we were producing nuts and bolts, the output was significant and we could measure a lot of data. “Is this not the right size?”

(25:35):
And we can measure this one, and the next and the next. And the collection of data will be huge, and the processing of data and then information we’ll get from the data is based on a statistical good, solid, based… We have solid ground for this decision.

(25:54):
But when we make cheese, the number of vat and the number of measurements we can do per day is very limited, and it could be, let’s say, 10 batch, it could be someplace, maybe 30 batch.

(26:10):
And then most dairies produce different kind of cheese. So let’s say it’s a 20 batch per day and then it’s three different kind of cheese. So we end up dividing the production into such small amount of data, that we can’t make any statistical process control on it because it’s milk and then the variation of milk.

(26:38):
Today, we collect milk from these farmers and tomorrow it’s from other farmers. It could be mixture, where we have organic milk. Organic milk is cheese. Bacteria grows different in organic milk than they do in conventional milk, so that the milk from different cows has different fat content and protein contents, so that’s a huge amount of variables.

(27:14):
And then with the sensors we have today, measure one number, we say, “Ah, this has a pH of 6.3.” We have one number based off so many different variables, so we can’t really use it. So that’s the thing, that’s difficult in that kind of production we have here.

(27:46):
If we had a flow production of… Not to say it’s everything else than what we do are easier, because then they just have other problems, but other challenges other optimization possibilities, but within batch cheese production is really hard to get a high number of measurements that we can really use for that kind of optimization.

(28:18):
Then, of course, we can measure the variation between batches, and we can measure pH between batches and water contents, and fat contents and so on, and still work by getting the variation lower. We could get the variation of fat lower, we could get the variation of protein lower, the variation of the value of melt lower, the variation in temperature.

(28:49):
We could add, let’s say, better temperature control, better sensors for that, better regulations for temperature. So as much as all the variables, as much as we can stabilize them by measuring how they are and then adjust them, the more stable we can of course make our production.

(29:12):
The more stable we have a production, the more closer to limits we can go. If it’s the regulation, there’s a lot of regulation that we can get closer to regulation limits, and that way optimize our production, that way optimize yield.

Mikkel Svold (29:31):
Do you think we will ever see a production where you’ll be able to sell the product with variations in them? Because of course the whole issue here is that the raw milk that you get in varies maybe not in quality, but it has different levels of different components. And it varies from day to day, from batch to batch, the input. But the output that you need to produce must be very close to the exact same thing every day.

(30:07):
So you can’t have taste differences, you can’t have smell differences, color differences. It has to be the same product. And the reason why I’m asking is because I just came to think about it. I saw I think it was an apple juice where on the label they said, “We use fresh apples from last week” or something like that.

(30:30):
So basically coming down to it, so the color varies and the taste can vary because we use whatever is fresh right now. Do you think that we could move to a similar place with at least some of this, maybe specialty cheeses, where you could kind of, I’m going to say get away with just stating on the label, “This is basically a living product”?

Henrik Kæmpe (30:54):
A living product, yeah.

Mikkel Svold (30:55):
So, it’s going to be different depending on the season of the year, depending on what cows we got the milk from. Do you think that would be plausible? And would it even help anything in production?

Henrik Kæmpe (31:13):
Yeah. We are caught in the middle, because if you go to the grocery store and you look then cucumbers or you try to feel an avocado, is it good? Or the mango, is it as you expect it to be? And when you come home and eat the mango, “This one was really good,” or, “This one, it was two acid this time compared to what I like.”

(31:46):
And there with fruit we accept that they vary during the season. And then on your other hand, if you buy cookies, if you buy candy, then you totally expect that they are the same. And then if you take a cheese, then you expect it that it’s the same.

(32:15):
And I really understand, because if you have from the last cheese you bought from that company and that kind of cheese, then you say, “Okay, this is really what I like.” And cheese is also quite expensive. It’s 50 day, 100 Danish krone for a piece of cheese. And it will last in the fridge for also sometimes some weeks.

(32:47):
So, you have a high expectation that last time you bought it, and it was exactly as acid and soft as you like it, then you expect it to be the same. But if you just take the variation, if you take some mature cheeses that are, let’s say, 40 weeks old, then just hitting the demand that the same customer can buy some cheese that one week is 50 weeks old and some months later it’s… Because the demand has really increased and then it could be 20 tweaks younger if they don’t have had a steady production.

(33:33):
And so the variation within buying the same cheese, it’s quite huge, and that’s a big challenge. And so goes with our production. Then when we salt the cheese, we do that in brine. We know it when we bake cookies in the oven, if they’re placed in one corner, they often get a bit more brown than the other corner.

(34:06):
And if it’s a big brining vat, it’s also how do the brine circulate, and that could be small variation in salt content. And if the small ways in the salt content. And not just the taste of salt, it’s also the texture of the cheese.

(34:22):
Same goes to the pH. pH is not just how sour you think the cheese is, but it comes from breaking down the protein. So if you have a sour cheese, the protein is typically more broke down and then you’ll get a softer cheese and a bitter cheese, a softer cheese. And when the cheese is softer it can obtain more salt, and then it has more salt. So it’s really trying to standardize all parameters in the cheese production.

Mikkel Svold (35:02):
I actually like your analogy with baking in the oven, because that’s very clear that you can see the differences. And it’s down to not just how is it baked and what is the oven representing the hardware here, but it’s also how well have you mixed it? How much did you actually pour in of flour this time? Was it five deciliters or was it 5.2 deciliters? And that will vary the outcome.

(35:37):
And then with cheese, of course, you just have, and especially large production batches, you have to control these variations both at a very high level, but also at a very detailed level, I guess.

Henrik Kæmpe (35:54):
Yeah.

Mikkel Svold (35:54):
Yeah. I think that’s very interesting.

Henrik Kæmpe (35:57):
So somehow all parameters come into play. Right?

Mikkel Svold (36:00):
Yeah.

Henrik Kæmpe (36:00):
And it’s-

Mikkel Svold (36:01):
And they also interact with each other.

Henrik Kæmpe (36:03):
Yeah.

Mikkel Svold (36:03):
That I think is very interesting. Right?

Henrik Kæmpe (36:04):
Yes, exactly. Exactly.

Mikkel Svold (36:05):
So if you have more broken down protein, you would obtain more salt, and that will then, again, start a new process.

Henrik Kæmpe (36:12):
Exactly. Yeah, exactly.

Mikkel Svold (36:14):
Yeah. One thing I want to ask you before we round this episode off, is if you came to a new place, so if you got a new job at a new cheese production facility, what would be the first data point that you would introduce to that factory?

Henrik Kæmpe (36:45):
Good question. Yeah, quite good.

Mikkel Svold (37:00):
Or the first few couple of data points may be a bit easier.

Henrik Kæmpe (37:06):
Yeah. They will come from different level because we have to be in control of our production. On the basic in production, at the shop floor, are we producing the right quality and do they live up to food standards and so on? We have to have that.

(37:31):
But on top of that, I really look into, are we in control of yield? Because if we are not in control of yield, we don’t know the cost of our products and we don’t know if we set the right prices, and we don’t know actually if we have the right prices and so on.

(38:03):
That really could be, let’s say, company critical if we don’t have the right price point. And the value of the milk and the cream is so high. For me, it’s like working with the gold, because the raw material is so expensive. And cheese production, we use 10 liters of milk for 1 kilo of cheese.

(38:31):
And so even small deviations in fat and protein that is not needed, and adding too much fat that we would not have to do, that has a huge impact of costs. So the pH, the fat and protein measurement is really key. We also have to measure flow right and so on in order to get stable production. So it’s quite complex question.

Mikkel Svold (39:08):
Yeah. No, I think there are very good points, and also it’s a beautiful segue into our next talk, which is going to be on how production data can change, not just operations, but also tactics and all the way up to the business strategy, how we can use production data here.

(39:25):
So I think it was actually a perfect answer without being on purpose. But Henrik Kæmpe, thank you so much for joining into this discussion on what kinds of data to capture and how we process them.

(39:39):
And also what limitations, especially the batch limitation, that kind of stuck with me, that it’s actually although you have infinite amounts of data, because you do batches, that data is broken up into so small, sometimes those small segments that is even hard to use the actual data.

(39:59):
And yeah, that’s an interesting problem that of course you’ll have to somehow solve, and I know that you are solving. Henrik Kæmpe, thank you so much for joining us for this episode.

Henrik Kæmpe (40:12):
Yeah, welcome. Thank you.

Mikkel Svold (40:14):
And to the listeners out there, thank you also to you for joining for this show. Again, if you have any questions or if you have any topics that you would want us to pick up on this podcast, do let us know. It’s always interesting to hear what you guys have to say. All I have left to say now is thank you so much for listening.