Final Project for ME 597 IIoT: Machine Learning for Manufacturing Failure Detection

As a semester-long project and part of a team of 3, we worked with a local manufacturing company, Kirby-Risk, to develop an ML algorithm for detecting when a tool has failed during a manufacturing process and is now creating a part that will be later considered scrap.

The current method to identify a scrap part is primarily a post-manufacturing visual inspection and tolerance check.

With our solution, the machine operator would have the ability to stop the machining of a part before it would be considered scrap. Alternatively, if the part can’t be saved, it at least reduces the amount of time spent machining a bad part. Additionally, parts were made in groups of 4, so it potentially saves the remaining parts from the same issues.

To summarize the work done (and my contribution):

The system was installed into the CNC machine and data collected over January through March.

I primarily focused on pre-processing. I took in raw MTConnect data with timestamped auto data, and the output was an organized fileset consisting of audio files sliced into per-tool segments with the accompanying mt data. It was then run through a feature extraction process, where I wrote the code to process STFT (Short Time Fourier Transform). We ultimately used this in later ML development.

My final contribution was per-tool CNN models that proved to be semi-successful, although we were working with a relatively small dataset and were unable to obtain more data as the process had been changed since we had started, so we couldn’t test on more recent data.

Although further testing was needed to confirm, our dataset showed two parts that were pulled in 1 month for not meeting standards (No specific data on what day/time/batch etc),but my 2D model (After much tuning) was able to detect 2 instances in March where abnormal audio was recorded!

Other team members attempted various forms of feature extraction or data manipulation, but none seemed to yield significant results in the end. They built other models (Convolutional AutoEncoders, Random Forest Classifier, etc) that sometimes worked, sometimes didn’t. They were very helpful in tuning the CNN in later stages.

A visual of how our system worked:

⁠

An example spectrogram from one tooling process: Each frame is a different part being machined, but by the same tool.

⁠

Want to print your doc?
This is not the way.

Try clicking the ⋯ next to your doc name or using a keyboard shortcut (

CtrlP

) instead.