Bringing Predictive Analytics to the Shop Floor: OK, but How?

By

Predictive Analytics can bring a lot of value to shop floor operations, especially to improve quality, yield or process sustainability, for example in composite manufacturing.

Machine learning algorithms allow to extract patterns from past production data. These patterns, which make up a model, can in turn be used to obtain predictions (“what is the risk to have a defective part?”) or even recommendations (“what can I do to reduce the risk?”).

Data Scientists wanted!

I have intentionally used the expression “machine learning algorithm” and you may think that companies that want to go in this direction need to hire a team of data scientists.

Indeed, many open source or commercial solutions require the availability of data scientist skills along with good programming skills in order to:

  • Identify the proper algorithms to use
  • Fine tune algorithms to get good results
  • Ensure scalability and performance

So it is no surprise that with the explosion of Big Data and Predictive Analytics, job postings in this field have skyrocketed since early 2012:

data_scientist_job_trends
Percentage of job offers with words “Data Scientist” or “Data Science” ©  Indeed.com.

And, as a result, salaries have soared and positions are hard to fill, which slows down the adoption of Predictive Analytics solutions.

Empowering Quality Managers and Process Experts

In order to overcome this difficulty, the DELMIA Operations Intelligence  solution for shop floor optimization (DELMIA OI) has been designed from the start for Quality Managers and Process Experts. There is no need to select or fine-tune algorithms and “correlation” is probably the most complex word used in the User Interface. Training is achieved in a few days.

shop_floor_quality
A failure analysis engineer prepares boards for corrosion testing. © Intel.

We also think that expertise is essential to obtain reliable models in the manufacturing field. For example, a process expert may identify irrelevant parameters, add relevant durations between operations, spot errors in data… And, last but not least, he may get inspiration from the model, which in the case of DELMIA OI comes in the form of human-readable rules.

Does this mean that data scientists are out of the picture? No, if you are lucky enough to have such resources, you will realize that best results are actually obtained by the collaboration between all profiles. Data scientists bring their experience on how to prepare and handle data, while quality managers and process experts can make informed decisions using their process knowledge.

The need for a Method

Even simple concepts and an intuitive user interface will not guarantee best results. You need a method to avoid pitfalls when you have to deal with potentially erroneous or incomplete data and different ways to address the problem.

Using the experience of DELMIA Operations Intelligence past projects, we have been able to build such a method, which has been recently shared in the DELMIA Enterprise Intelligence community.

The method consists in 8 steps:

understand_process
Understand process		 import_curve_data
Leverage curves		  cleanup
Clean data		  prepare
Prepare data
 target
Define output		 build_model
Build model		  validate
Validate model		  assess_value
Assess value

The method answers questions such as:

  • How to leverage curve data (hint: you may need BIOVIA Pipeline Pilot)?
  • Where should I put the frontier between a good and a bad yield?
  • How can I measure the reliability of the model (its ability to predict)?
  • How can I improve my model?
  • How can I evaluate the number of defective parts that could be spared if DELMIA OI recommendations were applied on the shop floor?

Discover more about how to build reliable predictive models to optimize your manufacturing operations by joining our free DELMIA Enterprise Intelligence community.

Once you are registered, it all starts with this post!

0

Christian