Sensor data analytics in manufacturing: the ‘why’, the ‘when’ and the ‘how’
Pick any piece of electronics in front of you – chances are, it has sensors in it. And not only electronics: now sensors are even in trees and buildings. And they are there for a reason: people extract value from their readings. So if you feel that sensors on your factories’ equipment are also potentially a profit gainer, you’re right. But the mere feeling is not enough. To extract real value from sensor data, before turning to big data consulting, you need to understand the basics of sensor analytics. And this is what you will gain down below, if you look closely at:
Why you should analyze sensor data depends on what your enterprise needs. Usually, manufacturers analyze sensor data to monitor/optimize processes or design products.
Monitoring usually presupposes the following course of action: you create a fault model to determine parameters defining failure, then you analyze sensor data, spot faulty patterns based on the model and immediately correct errors to restore the normal flow of your processes.
On a particular example, it would look like this. A plastic goods manufacturer that has sensors on their equipment can monitor all stages of their production process. So, if the temperature of the melted plastic reaches a maximum admissible value, they can lower the temperature of the plastic and avoid splay on the final product. This allows to minimize the expenses connected with defected goods and enjoy better quality assurance.
Also, process monitoring is a good chance to set up predictive maintenance within a factory (through predictive analytics). Sensor data analytics can warn your maintenance team about any faulty patterns on the production floor. And this will give them the opportunity to mend the soon-to-be-malfunctioning equipment and avoid expensive downtimes.
Manufacturing processes optimization means analyzing the whole production cycle (or a certain part of it), identifying key influencing parameters and tuning them to increase the yield, stabilize the output or improve quality.
Now, let’s look how it can work for a tire manufacturer. After analyzing sensor data, their big data analytics tool reveals a process peculiarity. If a tire contains 6% more sulfur and the temperature of the steam pumped into the tire mold balloon is 9% lower, the output’s quality is 12% higher than average. The analytics team also finds that saving in steam-heating costs outweighs additional expenses connected with sulfur. Besides, the lower the temperate of the steam, the longer it takes for tire molds to wear out. And using this insight, the manufacturer tunes their key manufacturing parameters accordingly and manages to economize and improve quality at the same time.
3. Product design
Taking into account sensor analytics results, manufacturers can design better products allowing for the performance characteristics of the products in use.
Let’s consider smartphone manufacturers. Besides other design insights, to come up with new models, they use data from sensors deployed on tested prototypes and smartphones already in use. The prototype data shows whether the manufacturer is designing something good technically. And the in-use sensors show what gaps of previous models should be fixed and what not-too-popular technical features can be done without.
When you analyze your sensor data must also be based on the specific tasks you need to accomplish. Maybe you think it’ll be enough to do only once, like in case with one-time process optimization. But having tasted the sweet big data analytics pie together with the profit it brings, you will probably want more. And there are two models that describe when your sensor data is analyzed: ad hoc and in real time.
1. Ad hoc sensor analytics
Ad hoc sensor data analysis means looking into your sensor data on demand, only when you need to. Usually, it is performed by a team of data scientists or analysts.
Here’s an example. A chemical manufacturer is unsure about how frequently they need to change their industrial air and water filters. They do it every 3 months, as the instructions say. But they don’t know whether the manufacturer’s calculations allow for the specifics of their chemical production, which presupposes emitting toxic fumes.
Analyzing the data from equipment and filter sensors shows that the chemical plant should in fact change filters every month. Moreover, they should choose a different filter type, since the one they use isn’t good enough at handling toxic waste. This allows the plant to avoid a huge ecological scandal and a government fine as well as make a contribution to the cause of fighting air and water pollution.
2. (Near) real-time sensor analytics
In contrast to ad hoc, the mere name of real-time analytics gives off the air of rapidness. And it does so with no intention to mislead: real-time sensor analytics does provide a constant flow of analysis results.
However, according to this article, different companies understand the word ‘real-time’ differently. For some, it can mean a 40-millisecond interval of collecting and analyzing data. While for others, a 30-minute interval would be pretty real-time as well. And as many problems as sensor data analytics may have, choosing the right interval for real-time analysis is definitely one of them.
A perfect example of the 40-millisecond interval would be real-time analytics on a wind turbine. While it collects and analyzes data at a 40-millesecond rate, the analytics tool uses it, for instance, to find the optimal way to adjust the pitch of the blades. And it’s easy to justify such a high speed of data collection: it depends on the very nature of wind’s changeability.
Whereas for the 30-minute interval, the example could be a solar park. It would definitely be too much to transmit solar energy conversion data every 40 milliseconds. As sunrays angle changes with time, the solar panel can adjust its position to the Sun to convert more energy. Taking into account this fact, a 30-minute data collection and analysis interval could come in handy.
How you analyze your sensor data also, guess what, depends on your particular needs, tasks and context. But there are some best practices. And over the years, our big data consulting experts came up with an architecture that works for pretty much any sensor data analytics solution. It may resemble the IoT architecture shown here.
The starting point is a sensor. When it collects data from its ‘host’ and transmits it to gateways, the data gets filtered and moves to a data lake.
The data lake is a reservoir that stores data in its natural format until you need to analyze it. Then, the data is extracted, transformed (‘dressed and groomed’) and loaded into a big data warehouse.
The big data warehouse is the place that stores the cleaned, dressed and groomed data that is then used for analysis. Besides a data lake, a big data warehouse gets data from the control applications that govern actuators. It also stores data on your machinery configurations, the places where certain sensors are deployed and all other info that puts your sensor data into context. This way, the warehouse is ‘aware’ of not only what sensors are transmitting but also where they are and what your system told their actuators to do.
And there’s obviously a data analytics segment. This is where analysis itself happens. It is the true source of all the valuable business insights you can get.
The last and ‘smartest’ form of sensor data analysis is machine learning. It watches your sensor data, notices new patterns, makes new models for control apps and sends them into action. This way, your sensor analytics is always updated.
A bit more on data lake vs. big data warehouse
The key difference between a data lake and a big data warehouse is the approach to storing data:
- With a big data warehouse, it’s all difficult: before you load it, you need to filter, process, integrate, model it. You need to give the data a proper ‘look’ – shape and structure. This approach is called schema-on-write.
- With a data lake, it’s all easier: you don’t need much fuss preparing and staging data. You just take it and load it into the lake. Just like that. This approach is called schema-on-read.
Now, you can see why storing all your data in a big data warehouse is costly: imagine how much sensor data you’ll need to ‘refine’ and how much resources it’ll take.
The differences between a data lake and a big data warehouse are all in this handy table.
A bit more on machine learning
Here’s how it works. A data scientist, together with an engineering technologist, say, at an aircraft engine plant, pick out a set of influential process parameters. Then, the ML algorithm goes through a huge set of sensor data for these parameters and creates models. In terms of aircraft engines, the result could mean something like this: if the pressure is 18% lower than average and the alloy contains 7% more aluminum, then with 78% probability it leads to an increase in end product quality. After that, the team manually tests the model and, if the model deserves to, it gets applied by control apps.
The main benefit of machine learning models is accuracy. This is the key reason for the rivalry between experts and ML. If they say the temperature should be 15°C, an ML algorithm can find that it should in fact be 15.4°C. And a big data analytics tool will tell you that the 0.4°C difference will make a substantial financial difference.
Technologies used for machine learning are Spark Machine Learning Library (MLlib), Amazon Machine Learning, Azure ML Studio, TensorFlow, Theano, Torch, etc.
So, to recap it all, remember that you can analyze your sensor data:
- At regular intervals or on demand.
- Using data lakes, big data warehouses and machine learning algorithms.
- In order to monitor, optimize your processes and design new products.
But most importantly, remember that sensor-based big data analytics can let your business grow into something fascinating. And this is not an empty promise: there is quite a number of sensor-based manufacturing big data use cases that prove it.