Anomaly Detection

Introduction

Anomaly detection is a useful tool to determine if your machine is experiencing unusual behavior. The algorithm looks at each part being created and generates a profile of what's going on in the machine while the part is being machined. It will send an alert whenever a new part being created is very different from all the other parts.

What it does

Anomaly detection can detect when a new part deviates from the norm. The “norm” is established by looking at the activity of your machine for the last 20 parts. It looks like speeds, feeds, and loads, and sets a baseline based off of these metrics.

The limitations

Anomaly detection does not predict 100% of tool failures or otherwise consequential issues with your machine. This is because tool failures sometimes do not have preceding indicators. For example, a chip re-cut, which is when a chip accidentally gets thrown back into the machine, will have minimal, if any signal beforehand to indicate that it will happen.

Similarly, not every anomaly will indicate that something bad is about to happen. This is because machines can do strange things which do not necessarily cause anything of interest. This may be because the anomaly occurs due to something like an e-stop being hit or a door being opened in the middle of the part cycle, or the abnormal behavior is within the stress tolerance of the machine, as in the case of a load spike that doesn't bring down the machine.

 

The requirements


Anomaly detection will work if your machine fulfills the following conditions:

1. The machine has 3 or more numeric variables. You can check this on your machine diagnostics page, but generally all non-I/O integrations will fulfill this condition.

2. The machine has part cycles that are greater than 30 seconds long and less than 1 hour long.

3. The machine must count parts consistently. For example, if the part counter increments by 2, then 3, then 2, the machine cannot qualify for anomaly detection. An occasional inconsistent count is acceptable and can be worked around, but generally 80%+ of part count increments must be the same for the algorithm to work.

 

Activating anomaly detection

Anomaly detection can be activated and deactivated on the monitors page. You can simply create or delete a monitor with the monitor type "anomaly".



Examples

Examples of interest can be found here. This page is updated periodically as more data is gathered.

https://medium.com/machinemetrics-techblog/using-unsupervised-learning-to-detect-machine-anomalies-part-4-92bdf0b9079a

Have more questions? Submit a request

0 Comments

Please sign in to leave a comment.
Powered by Zendesk