Using HTM technology to predict failure points in optical manufacturing of eyeglasses

Engineering Applications
1

I am interested in any thoughts on using HTM to identify failure points, based on probability. This prior to the start of production. I believe that specific steps/processes could be avoided before the job is started, avoiding breakage of the lens, resulting in a re due and doubling the production cost.

1 Like
2

Hey @terrypickett1, welcome!

What kind of data do you have access to prior to production? Any small sample we could see?

HTM has been successfully used to detect oncoming failures in servers, which may be similar to your objective. The HTM models monitor streaming metrics coming from the machines and detect anomalies when the sequential behavior of the metrics change.

3

Thank you for your email.

The information on hand prior to the start of the manufacturing of glasses includes the following:

Prescription details (O.D., O.S., PD, etc)

Special handling (coatings, safety bevel, tinting, mirroring, etc.)

Manufacture of the lens material (Zeiss, Essilor, Hoya, Varilux, etc.)

Special handling (coatings, safety bevel, tinting, etc.)

Additional information available:

All human/machine steps are recorded in chronicle order as well.

Each step is monitored to report where and when the failure occurred.

It takes both machine and human steps to complete the final object.

Together this forms a pattern.

My goal is preventing a repeat of a failed pattern.

Terry Pickett

4

Could we see an example of this? Could be toy data, just to get the data types & structure.

The other info you mentioned like Prescription & Manufacturer sounds like metadata, which may not be directly useful since HTM only learns transitions from sequential data.

If failed patterns repeat an HTM model could be trained to recognize known failed patterns. So when the model anomaly scores are low it means the model is recognizing a familiar sequence. If this model is only familiar with known failed patterns and is recognizing the current input, we can surmise the current input resembles a known failed pattern.

If however the failed patterns aren’t known like this, a model could be trained on only successful patterns. In this case high anomaly score would mean the current input isn’t a known successful pattern — implying it may be a failed pattern.

1 Like