Data – The Secret to Successful AI Deployment in Manufacturing

AI implementation in manufacturing

There is a secret to a successful AI deployment in manufacturing and it is not cool algorithms – it’s data. In fact, 80% of the effort of an average AI implementation is getting the data needed in the right format so you can start training the algorithms.

Generating, collecting and cleansing data is a decidedly unsexy aspect of AI implementation but absolutely critical. Without the data, any AI deployment will fail.

In this blog, we discuss the data challenges we have seen working with several large manufacturing customers deploying AI-based visual inspection and predictive maintenance applications and share how to launch a successful project.

Where is the data?


AI deployment in manufacturing

For an AI algorithm to learn, it needs to “see” a large amount of data. In quality inspection, for example, images are used to teach the algorithm to differentiate between a good and a defective product as well as to categorize defects.

Plenty of cleansed training images are needed before algorithm training can start. The critical question is “where is the data?” We have seen two scenarios play out repeatedly:

  • The data exists but in disjointed places and formats. It can take significant programming time to collect, reformat and combine that disparate data.
  • The data does not exist. This is a very common problem and not entirely unexpected. When the system was originally set up, no future mining of the data was anticipated and so there was no reason to collect the data. For an AI implementation, however, the lack of data results in significant delays. Sensors, such as cameras, need to be installed and hooked up. Only then can one begin gathering data to train the algorithm.

In all our years in business, we have never encountered the third, ideal scenario, namely that the customer had all the data needed, nicely cleansed and easily accessible.

The delays data issues cause can be significant. In the case of disjointed data, there might be a way to brute force collection and cleansing, such as by throwing more resources at the problem. However, in the scenario of the missing data, the only path forward is to actually collect it. Depending on cycle time or maintenance intervals, this can take a long time. If, for example, a defect or quality issue only surfaces once every month, it will take months to collect enough data for the algorithm to learn.

These are common problems that everybody suffers from – but they don’t have to jeopardize the success of your AI deployment.

Planning for a successful AI deployment in manufacturing

Before you even start thinking about algorithms get your data ready by following this four-step process:

  1. Get insights into the critical variables in your manufacturing process by speaking to in-house subject matter experts such as process engineers, technicians and operators. Their feedback allows you to answer the critical question: “Improvement to which processes will have the biggest impact on the bottom line?”
  2. Next, ask them what the critical (process or quality) data are that describe the process – this answers what exactly you have to measure and how.
  3. Find out which data are currently being collected and perform a gap analysis to determine whether the data you need exists somewhere or whether it needs to be generated from scratch.
  4. With the gaps identified, decide how to best generate and collect the missing data and then install the appropriate sensors to do so.

Step 1 through 3 of the analysis will also provide insights into which application will be most promising for an initial proof of concept. Good candidates are processes where

  • ample data is available, or
  • cycle times or maintenance intervals are short so missing data can be collected quickly

These initial applications may not be the most high-value ones but will allow the organization to quickly go up the learning curve.

Once your data strategy is underway, it is time to start thinking about algorithms, deployment and integration and how to make sure you build an AI platform that flexibly allows you to add new capabilities and makes sure that the different algorithms you will deploy over time work together.

If you are exploring how to implement AI on your shop floor and have questions, we are happy to discuss them with you. Contact us here.

To learn why predictive maintenance is a good way of getting started with AI implementation in manufacturing check out our blog on that topic.