Stanley Black & Decker is a company known for its consumer and professional tools, storage solutions and other products. It has quite a few well-known brands under its umbrella, including DeWalt, Proto, Lista, and Craftsman. However, tools and storage make up only half of their business. At Stanley, they also offer various solutions in consumer and residential security and in healthcare, as well as fastening solutions, infrastructure products, and pipeline services.
The company faces lots of technological challenges resulting from its size. Specifically, the company is growing by acquisition, which results in different information systems, different levels of technological advances, and different cultures.
Do you like this in-depth educational content on applied machine learning? Subscribe to our Enterprise AI mailing list to be alerted when we release new material.
Building an AI-Supporting Culture
Implementing AI in a manufacturing company with dozens of different business units and 175 years of history sounds like a very challenging task. At Stanley, they have developed their own approach to creating a business culture that supports technological developments, including AI-driven solutions. Their approach is based on three key factors:
- Support from top management.
- Driving progress within business units by educating and engaging business people in the development of new technological solutions.
- Driving progress across the enterprise by developing the “AI Center of Excellence” and helping business units in acquiring their own AI specialists so that this AI Center doesn’t become a bottleneck for technological advancements.
The AI Center of Excellence at Stanley acts a lot like a sales team. They introduce themselves to business leaders, and explain what they can do and what they can’t do. They also do a lot of proof-of-concept developments, where they propose their solution to problems defined in conjunction with the specific business unit or identified by the AI Center itself. Finally, the center identifies ways to replace AI vendors that often do not deliver the results the company expects but always charge high prices for their services.
Defining a Problem that AI Can Potentially Solve
One of the business units at Stanley discovered that an important part of their manufacturing process might be inefficient. Specifically, they have a complex process that requires the selection of part combinations. For example, if you need to produce a drill, there is a variety of motors, power sources, bushings, bearings and other parts that can be used, and the combination of these parts has an ultimate impact on the quality and lifetime of the final product.
Here is how the part combination process looked before the implementation of an AI-driven solution:
- The customer sends an order, which usually includes multiple products and hundreds of items for each product.
- The engineers for each product make an educated selection of parts based on their prior experience.
- The selected parts are used to assemble the product.
- The final product is tested, and then disassembled for detailed lab tests.
- If the product passes the testing, mass production begins. If not, the part selection process starts all over again.
The problem is that this process is very time-consuming. It takes on average over an hour to test one combination. Since there are usually multiple combinations that need to be tested per product and multiple products in each order, it took the company up to 850 hours just to identify the good part combinations for a single customer order. Much less time was spent on actually producing the products after this selection was made.
Solving the Part Combination Problem with AI
When choosing a solution for the identified problem, Stanley considered the cost of the solution, the speed of implementation, and the strategic advantage. Additionally, in their evaluation process, Stanley relied heavily on another indicator: accuracy per development-hour. Basically, you can always improve the accuracy of your model, but the question is how much time it will take to improve the model’s accuracy by another percentage point.
They had three options under consideration:
- An external provider, who could potentially solve the problem without much effort on Stanley’s side but would charge over $175K just for the initial system development. This option also has additional disadvantages, including external dependencies and a lack of strategic advantage as Stanley’s enterprise capabilities would not increase.
- Open-source Python stack (scikit-learn, SciPy, etc.), which is basically free, assumes no external dependencies and would lead to internal skill improvement, but would take lots of time for model development and deployment, and thus would have quite a low accuracy per development-hour.
- Driverless AI by H2O, which would give the company a strategic advantage by providing a reusable tool, would simplify the development and deployment process and would have a very high accuracy per development-hour. The disadvantages of this solution for Stanley include cost and external dependencies on the vendor.
After considering all the pros and cons, the company decided to address its part combination problem with the help of a Driverless AI solution by H2O. Here is how machine learning is now embedded in the manufacturing process:
- Production requirements are fed into the machine learning model.
- The model estimates the test results for each combination of materials (up to 50,000 combinations) – will it pass or fail the quality inspection?
- The model provides such estimations for several properties including reliability, durability, etc.
- The combinations are rank-ordered based on the model’s estimations of the testing results.
- This ranking is presented to an engineer, who builds the product based on the most promising part combination.
- The product goes through testing and the results are fed back into the system for batch retraining later.
Evaluating the Results from the Deployed Solution
This project started with the need to evaluate the quote from the external AI solution provider, and the company actually had only about two weeks for this evaluation. Within this short time period, they were able to develop the full proof-of-concept model with Driverless AI that outperformed the accuracy quoted by the external vendor. So speed to solution was a critical factor in this particular case. Furthermore, accuracy per development-hour was a key metric considered when evaluating several machine learning solutions.
Every company is looking for an approach that will be good, fast, and cheap, but this is an almost impossible combination in the real world. Stanley evaluated Driverless AI as good and fast but not cheap. At the end of the day, they were happy with their choice and now use Driverless AI as a preferred tool for rapidly validating project ideas.
Luckily, H2O also has a fully open-source machine-learning platform that shares lots of functionality with the Driverless AI platform and might be suitable for some projects.
If you are interested in more details of how Stanley implemented the Driverless AI solution to optimize their manufacturing process, please check out the video below that this article is based on:
Enjoy this article? Sign up for more updates on applied ML.
We’ll let you know when we release more technical education.