Portfolio Optimization Series: Driving Economic Value with Industrial Computer Vision
Tldr;
Following investment in a new business private equity companies will often execute on a playbook to realize the revenue and cost synergies identified during the deal process. We believe that building machine learning (ML) capabilities into the acquired business where appropriate should be part of that playbook going forward.
A portfolio business does not have to operate in a high-tech environment to benefit from ML, but the business case to implement ML technologies must be sound.
We worked with an industrial inspection business to implement a computer vision solution that, from the outside, would not appear to be the perfect candidate to adopt an ML system.
However, the business case was sound and the implementation produced 20% more profit over a 7 year period and a 250% ROI
Want to evaluate the business case for applying ML in your business? Get in touch
Welcome to the first perspective in our series on portfolio optimization. Our objective, like with all our work, is to demonstrate the genuine business value that implementing machine learning (ML) can generate.
In this perspective, we explore a use case in which an industrial inspection business created economic value by implementing computer vision technology. Some details in this perspective have been altered to protect the client's confidentiality. The underlying methodology and value drivers remain unchanged. Through this case study, we hope to provide a tangible example of how ML can be used by Private Equity firms and Management Teams to optimize portfolio businesses and do so in way that helps to draw parallels between ML implementation and other common portfolio optimization techniques.
Background
This use case is centered on a business that inspects brake pads for major airplane manufacturers. These inspections are critical for ensuring the reliability of aircraft and due to regulatory requirements, manufacturers are required to outsource the inspection of them to a third-party company like this one. The relationships with manufacturers are formalized through long-term contracts (5+ years) where the company agrees to perform inspection for a certain volume of parts that the manufacturer ships to them. If manufacturing throughput accelerates, new contracts are drafted to pay for the additional volume.
At a high level, the inspection process operates as follows:
A brake pad is manually loaded onto a conveyor belt and is carried down the inspection line.
It passes through four different stations where a trained person will inspect it and identify defects. Each station is responsible for inspecting a different area on the brake pad.
When a defective pad has been identified, the person at that station removes it from the line and places it into a defect bin so it can be sent back to the manufacturer.
If no defects are identified, the part is manually removed from the conveyor belt, placed into a bin, and trucked out to the assembly facility.
This process requires six people to operate; one to load the part on the conveyor belt, four to inspect the brake pad, and one to remove pads from the conveyor belt.
Opportunity & Challenge
The company faced an opportunity when one of their top customers informed them that they were about to double their throughput of brake pads to meet new demand. The new contract was theirs if they could handle the additional capacity. Historically, the response to this situation was to build a new inspection line, double the number of inspectors on staff, train everyone on the inspection process, and take on the new contract as soon as possible. However, the CEO wanted to use this opportunity to re-think their process and how their business could scale more efficiently. In recent months, additional inspectors who were available to start with little notice were in short supply and the business did not envision that changing in the coming years. As the company thought about the opportunity, the CEO suggested that they explore computer vision and decided to kick off a project to explore the value it could provide.
Solution & Approach
The objective of the project was to implement a ML system that could automate as much of the inspection process as possible. The solution would:
Capture multiple images of the brake pad simultaneously using multiple cameras.
Pass the images to computer vision models that could analyze each image and determine whether the pad was defective or not.
Indicate the part status (i.e. defective or not defective) to an operator who would subsequently triage parts; removing them if they are defective or loading them into a container to be sent to the assembly facility if they were not defective.
The ML-based process required two people instead of six; one person to load the system with a new brake pad and another person to triage the pads to different spots in the facility based on whether they were defective or not.
This specific project was separated into two phases. Phase one focused on building a system that could replace the existing inspection line. Once successful and value was proven, phase two could begin. The second phase focused on how to deliver additional systems to improve the efficiency at which the business could serve new contracts.
The technical challenges, learnings, and solution that were implemented was a fascinating combination of industrial hardware, modern development frameworks, and ML techniques. We will return to these topics in a subsequent technology-focused perspective. The remainder of this perspective will dive int the levers driving economic value.
Phase one: Automating the existing inspection process
During phase one, an ML system was designed and implemented to replace the role of the manual inspectors in the existing process. Rather than needing multiple inspection stations, the new system had a single station, with multiple cameras that captured specific images of the brake pads. Those images we passed along to ML models that were trained to identify whether there was a defect or not.
The primary value driver for this phase was expanding gross profit by reducing the business’ cost of goods sold. We have modeled the benefits from reducing the number of inspectors required in the process from six to two against the investment required to build the initial ML system. For the purposes of this analysis, we have excluded any additional benefits from increased throughput or reduced system downtime.
The key question to answer in this phase was whether implementing an ML system would lead to higher profits within a short enough time frame to generate more profit than the existing manual process. The key inputs and analysis are summarized in the table and chart below:
The results from the implementation of the system in phase one was a clear success:
Profit generated from the line increased 20% over the 7-year contract
Return on invested capital was 250%
Investment payback period was 2 years
It is important to highlight how short the payback period was in this case. Like many new custom ML initiatives there is a learning curve, challenges that are impossible to foresee, and many approaches that need to be compared against one another. A short payback period provides sufficient time to learn, iterate, and adjust to maximize the long-term financial benefit to the business.
With phase one complete, it was clear that automating the inspection process reduced labor costs and increased gross profit of the line. Next, we turned our attention to how much incremental value ML systems could generate if they were installed and used to support new contracts.
Phase two: More efficient scaling using ML systems
As mentioned earlier in the perspective, it is common for manufacturers to suddenly need to increase throughput and therefore inspection volume. This phase was meant to explore the financial impact that incremental ML systems could have on new demand compared to the existing manual inspection process. The primary value driver in this phase remained the same as phase one; gross profit increases due to reduced labour costs. However, the returns on incremental ML systems improve because the cost of implementing additional ML systems benefits from economies of scale.
To explain why this is the case, it is useful to understand some of the activities required in developing the first ML system and how they are different after you build one system:
Hardware & software selection – The number of possible hardware and software configurations for computer vision is enormous. Deciphering technical documentation and performing the engineering analysis required to make the right choice is time consuming yet critical to the project’s success. After it has been completed however, it becomes a much simpler activity. Instead of designing from scratch, you can verify the conditions of the new system against the parameters of the pre-selected configuration, shaving weeks of research, analysis, evaluation, and sourcing off the project timeline.
Integration design & build – Computer vision systems like this require many different components that can communicate with one another and if you are not purchasing an off-the-shelf solution (which we believe should only be done in a subset of cases), getting industrial hardware to communicate with ML models is a complex task. The standard communication protocols for each domain are different which requires middleware to be written that translates information from the industrial hardware and sends it to the ML models and vice versa. Once built and tested however, it can be implemented into a new system almost immediately.
Data collection & model training – Defect detection models must be trained on many samples of defective and non-defective pads before a reliable model can be produced. It is also important for the conditions that the sample data is captured in to resemble the actual inspection process as much as possible. To collect training data, we used the ML system’s camera hardware to take pictures of defective and non-defective pads. Although time consuming, it is a process that only had to be done to build the first system and does not have to be done again to create a new system that is inspecting the same type of brake pads.
We modeled the economics of implementing additional systems to explore the financial impact that the above dynamics have. We compared the existing manual process against incremental ML systems in the analysis below:
From the analysis we concluded:
Using an ML system to handle new contracts generates 28% more gross profit than the manual process
Return on invested capital for each incremental line is 900%
The bar chart above paints a clear picture of how incremental ML systems deliver more value as the company adds new inspection lines. Returns of this magnitude should give any company facing a similar use case the confidence that an ML system can be a tremendous investment.
Even more than a high-return investment, ML systems in this environment can be thought of as a means to improve the underlying cost structure of a business. Not only does it generate better returns on higher gross margins in the short term, but it makes the business a more attractive asset for investors or acquirers in the long-term.
Qualitative Benefits
While the economic benefits of implementing an ML system in this environment became clear, there are also qualitative benefits that should not be overlooked because they provide incredible flexibility and strategic value.
Firstly, it provides the company with an opportunity to increase their market share. This system gives them a unique value proposition that they can use to compete against other manual inspection companies and win new business. Additionally, the improved gross margins give the company stronger pricing flexibility. It enables them to reduce their price to win a contract from an important new logo, provide discounts for longer-term contracts, or win competitive bids on the basis of price if they so choose.
Another important benefit is that scaling the business is less dependent on finding available talent. The use of computer vision reduces the need for human inspection, in turn reducing the risk of having to turn down or delay new contracts until new inspectors can be hired and trained.
Finally, these projects developed intellectual property (IP) for the business that should ultimately be accretive to their valuation. Should the owners and/or investors choose to sell the business in the future, they can demand a premium over similar businesses that did not choose to develop a similar solution.
Conclusion
We hope that this perspective provides a better understanding of how ML systems can be used to drive value in ways that are similar to traditional company optimization projects. These systems are not a case of implementing “technology for the sake of technology”. They are all about using ML to drive changes in the cost structure of a business and make it more scalable and valuable. The industrial sector in particular has many opportunities like the one described above and the market has an increasing appreciation for how industrial processes and ML are coming together. In future perspectives, we will continue to explore how ML technologies can be used in different sectors to spearhead growth, efficiency, and portfolio company value.
Does this situation sound familiar? Get in touch