With the introduction of Industry 4.0 by the German Government and the designated workgroup, a roadmap was created to introduce Artificial Intelligence, Smart Controls, Smart Automation and Industrial Internet of Things (IIoT) in combination with Big Data into the manufacturing industries in Germany and beyond. The roadmap for Industry 4.0 focuses on introducing information and communication technology to combine the growing demand for individualized production and the benefits of the growing digitization of society and industry. In combination with the LEAN principle of flexibility over complexity, Industry 4.0 lays the foundation for flexible smart factories.

In reality, there is however a lot of manufacturing equipment out there that barely made it through the previous industrial revolution. Big and inflexible, build for mass production and still going strong, despite not being from this modern industrial age. Even companies that own hyper modern facilities, in many cases also own and operate equipment that was installed decades ago, and given the investments made in the past and the investments needed to replace them, have no intention of shutting these facilities down in the near future. With the gap between modern Industry 4.0 and aging Industry 2.0/3.0 equipment growing annually, many facilities operating aged equipment consider the benefits of modernization and Industry 4.0 as unrealistic. Or as one Plant Manager told me recently “These machines were build before you and I were born in a country that no longer exists, and still run stable without robots and internet. How could I improve anything with modern technology when all they have is panel with ON/OFF switches?”.

As cynical as this comment might sound, it does reflect reality for many production facilities. However, this doesn’t mean that the benefits of Industry 4.0 are completely lost for legacy production equipment. What follows is an example of own recent  experience, in which aged equipment was upgraded with smart sensors, smart controls, (limited) artificial intelligence/machine learning and maximized flexibility.

Teaching the old paint shop new tricks

A paint shop with 4 processes is linked with a single chain on which the parts are transported through the equipment. Experience over the years has shown that the combination of speed of the chain and total weight of parts on the chain have a significant impact on stoppages and parts falling from the chain. Parts can vary up to 100% in weight between the lightest and heaviest parts and the mix is very dynamic based on customer demand. When the chain is fully loaded with heavy parts, the chain starts to vibrate leading to parts falling of, and the drives and chain itself suffer from the vibrations, leading to additional stoppages and wearing of material. To complicate the process even further, the robots for on- and offloading of the chain and the spray guns and bells need to manually synchronized with the actual speed of the chain. So the team had set the speed of the chain at less than 50% of the available speed, to eliminate vibrations and stoppages as much as possible. And with that, also parting from more than 50% of the equipment capacity!

During our initial solution session, we made several significant discoveries:

  • All robots were able to take input from the chain drive on the actual speed and synchronize automatically. Although each drive had sensors for the actual speed available and even provided an available signal block for it, these were not connected to the robot controls and all equipment was set to manual speed control.
  • “Time under the gun”, meaning the actual time that the parts are being painted by the spray guns or bells, has significant impact on the quality and costs. When the time is too long, the amount of used paint is much higher than required and has possible negative impact on the quality. Is the time too short, not enough paint is applied to the parts. The spray gun and bell assemblies could be upgraded with the ability to measure the size of the part and calculate the optimal pressure/speed automatically where these were now run in manual mode, even the 2 assemblies which already had this capability. By setting the speed of the chain and spray equipment manually, the amount of paint used was much too high, especially for the smaller and lighter parts. Investigating further, we found that the controls even lacked data to set pressure and time based on speed and size, clearly indicating that these features had never been used.
  • The supplier of the chain provided an upgrade set to create 4 segments matching the 4 paint processes to upgrade the current single segment chain and was able to demonstrate that a segmented chain not only increased flexibility but also significantly reduced maintenance costs in combination with extended lifetime of the chain segments compared to a single chain.
  • Based on data provided by the supplier, we also established that additional maintenance costs of running the chain at maximum speed were slightly less than 15% of the costs of running the paint shop at below 50% of the capacity.

A few Design Thinking sessions later, we started by feeding the spray gun and bell controls with initial master data to properly set the pressure and time in relation to speed and size of the parts on the chain. We monitored consumed paint and paint coverage quality as our KPI’s for this modification. Even when the speed of the chain was still fixed and 6 out of the 8 assemblies were not yet able to measure the seize of the parts, we already reduced the consumed paint with 13% and decreased paint coverage quality issued with almost 10% (mainly in the area of over-painting the parts). Showing the Plant Manager the financial benefits of this modification, we got permission to install the upgrades to measure the seize of parts in the remaining paint gun and bell assemblies, and activated the automated pressure optimization on all assemblies. Now cutting paint consumption to 25% and further decreasing paint coverage quality issues, the paint gun supplier and local technicians were convinced that these already impressive results would even improve further if we would be able to vary the speed of the chain.

Although we were all very keen on creating segmented chains, this would require significant CAPEX which wasn’t budgeted so we had to postpone this upgrade until the next year, assuming the Plant Manager would get approval. Over a beer and pizza, the Plant Manager said “demonstrate on the current setup how important that is and I will fight for the CAPEX”. Challenge accepted and we locked ourselves in the Design Thinking mode again. We found that we could read the actual dimension of the gripper on the robots when they were loading and offloading the part on the chain, giving us a rough 2 dimensional indication of the size of the part. Knowing roughly the size of the part, we could estimate the weight of the part. Experiments showed that we could reach +/- 10% accuracy of the actual weight of the part, which was good enough to start with.

Using a flip chart and sticky notes, we designed a control circuit that would take the estimated weight of all parts loaded and offloaded on the chain from the robots and by doing so could calculate the total weight of parts on the chain. Collection and analyzing data from 2 weeks production, we started to compare the estimated loaded weight with the reference tables of the supplier and found that theoretically, we could increase the speed of the chain at least 30% without causing vibration of the chain. Back to the drawing board because we had to figure out how to control the speed of the chain and feed this to the robot and paint controls. We included experts from the various suppliers in our sessions and learned that the controls of the paint guns and bells could adjust to the actual speed but were also able to set the optimal speed based on measuring the size of the part. The robots would simply follow the set speed of the chain within their technical limits.

Because none of these smart controls were ever utilized in the past, and the supplier of the paint shop builds highly customized equipment for its customers, there was no usable reference data available to work with so we had to build settings from scratch. We starting by using the standard reference data for the chain drives to set the maximum and optimal speed of the chain based on the estimated weight loaded. These limits were supplied to the paint assembly controls as upper limits for the speed variations. Experts from the paint gun and bell supplier helped us to develop high, low and best speed settings based on the size of the parts and the type of paint used. We had 8 paint gun and bell assemblies, aligned in 4 pairs (top and bottom painting units) which would use the same optimal speed for the part currently being painted but the actual size of the parts in each of these 4 pairs could vary. We solved this by taking the lowest optimal speed of each pair against the limitation of the maximal speed for the total loaded weight on the chain. As KPI’s we decided to measure speed related stoppages, actual speed of the chain, and again the amount of used paint.

We could have spend months making complicated measurements and calculations to find the most optimal settings based on countless variables but we decided to go with what we had. Rather learn from mistakes and learning by doing than invest time in complicated designs of which there is no guarantee that it will actually pay off. Starting this control with the estimated optimal settings, the chain picked up 37% speed and in the first 30 minutes, we couldn’t detect any vibration or quality issues. During the corporate meetings that followed, I was rather impatient and wanted to return to the paint shop as quick as possible to see the results. When I finally got out of the meetings, I found that the automatic speed control was switched off. In the following weeks, it kept happening that the operators switched of the control with every stoppage or quality issue. We kept pointing out the benefits of this control and that we had to learn from the actual results to be able to improve even further. It took a lot of coaching and mentoring but the mindset and acceptance kept improving. A month later, the paint shop was running completely on automated speed control, providing us with valuable data on which settings gave the best results and which settings lead to increased stoppages and quality issues. Through it all, the consumed paint decreased significantly, the availability and output of the paint shop increased and the quality improved. And this was still only the simplified version of what we wanted to do!

Taking it to the next level

As promised, that Plant Manager used the achieved results to get permission for CAPEX investment in a segmented chain for the paint shop and further implementation of smart controls. The segmented chain was installed during the winter shutdown and we added small conveyors between the chain segments to allow us to vary speed between the segments. In addition, we upgraded the robotized on- and offloading units so we could in the future load and offload optimal mixes of different parts. The actual modifications of the robots were simpler than expected. Instead of loading parts from pallets in the sequence as they were placed in front of the robot, we created 5 pallet locations and trained the robot to find the position of the pallet. With the capability of loading and offloading the chains with different parts, and the capability of varying the speed of the chains in segments, we started to design smart controls.

The first step was to improve data on the actual weight of the parts and feed it to the controls based on parts handled by the robots. This was solved by taking the master data directly from the ERP system and scanning the part number by the operator when placing the pallets at the robot. With that, we changed the method of the control to establish the maximum speed of the chain based on the weight. Instead of calculating the maximum speed based on the load, we created a control that tried to find the optimal mix of speed and weight based on the loaded weight and available parts to be loaded on the chain. Now we attacked the speed of the segments in combination with the created buffers between the segments. Each segment was controlled by an algorithm that established the optimal speed based on weight for that segment, available parts on the buffer conveyor and optimal settings for the spray guns and bells connected to that segment.

What bothered us was that we kept working at fixed settings for different scenarios based on assumptions we made in the past. Although the team kept improving those assumptions, the amount of variations was simply to big to find plausible ways of correlating them and at the same time find the optimal speed and capacity of the paint shop. Building on the demonstrated concepts of ReVista – Ressourcen- und verfügbarkeitsorientierte Instandhaltungsstrategien, see section 8.5 of the project publication, we analyzed the available sensors, stoppages, minimum requirements for “time under the gun” and many more variables. Even after several months visiting the plant regularly and doing many improvement projects, I was still surprised about the amount of data available which simply wasn’t actively used and especially not set in relation to each other. With the success of the previous project as our motivation and my experience in ReVista and other projects, we made an ambitious roadmap of how we wanted to create an intelligent control that would actively manage the speed of the paint shop and learn from the actual results. The success of the previous project on that paint shop, and the significant ongoing cost reductions we established with that project, made that we also had more buy-in from the management team and the operators.

The process started with a control that would monitor all known and measured variables, and with that would increase the speed of the chains in small steps between Optimal and Maximum. A second control would take over when issues were detected, which could for example be speed related stoppages but also the signal that the buffer between the segments would be depleted at the current speed. This second control would decrease the speed until the detected issue was solved, after which the first control would start to increase speed again. This in combination with the control that constantly calculated the optimal load and speed based on the weight of the parts. These automatic fluctuations of the speed based on the measured conditions accumulated to an improvement of 11% of the capacity. On a 8-hour shift, that means almost an additional hour of production output!

All data from sensors, the actual loaded weight, stoppages, speed, etc. but also the “recipes” of presets for minimal, optimal and maximum speed, were loaded directly into a database and constantly updated in real-time. With this, a fourth control was introduced: artificial intelligence to develop best demonstrated recipes for presets. Now we were able stop using static presets based on assumptions. Over 500 data points and a virtual model that kept track of each part on the chains and conveyors, fed this algorithm with live data. Current and past results were analyzed in real-time and used to set the 3 controls of the chain segments, the robots and the paint gun and bell controls.

In our last review of the settings being used by this AI based control, the team found relations between data which they had never identified as significant. Other combinations and setting had been considered impossible but now proved to be the optimal setting for certain scenarios. And although the frequency kept decreasing, the AI control every now and then created a loop for itself from which it could not recover. The 3 controls directly connected through the paint shop chain segments and paint assemblies identified the invalid instruction and simply continued with the previous valid setting so no real harm was done there. What we called the “No thanks” signal from the controls would trigger the AI master control to step back to the previous scenario and restart the analyzes and optimizing process. Most charming was how the responsible paint shop engineer responded to this:

It took me 10 years to learn half of what this computer learned in 3 months. When there is a minor problem in the program, it makes up for that immediately in the next run. It takes me longer to understand that something is wrong than this program needs to solve the issue and get back to work again!