Anomaly detection, warning system and real time simulation in predictive maintenance using LSTM and CNN

I am currently working on a project regarding predictive maintenance with LSTM and CNN. The algorithms are good to run. However, my task now is to add few features to the project. The features are anomaly detection, warning system and deployment of the algorithm to perform live data simulation.

[Arkadiy Turevskiy]

Sounds like a cool project. Is that for school or work?

What are your already developed algorithms doing: fault classification, remaining useful life estimation?

You might be interested to take a look at this video where we went through a cse study of using autoencoder for anomaly detection.

What is the question you wanted to get an answer to?

[Ng Yong Jie]

Hi, thanks for your reply!

It is a project for my university.

The developed algorithms is doing remaining useful life prediction using LSTM and CNN.

Thanks for the link for autoencoder. Will check it out!

My question is how should I proceed to enable the algorithm to be fed with live data for real time simulation? I am currently reading about digital twin but I am not sure if that is a right direction.

And the next question, is there any suggestion on how to implement an early warning system to the algorithm?

Thanks.

Answer 1

[Prateek Rai]

To my understanding, you have created a deep learning model for predictive maintenance with LSTM and CNN and want to add anomaly detection, warning system, and real time simulation functionalities.

For anomaly detection, you have to first set the criteria as to what output values will be distinguished as an anomaly. Based on that you have to develop your warning system.

After you are done with both the steps, then you can proceed with deployment to enable the algorithm to be fed with live data for real time simulation.

For deployment, you can refer to Prototype Deep Learning Networks on FPGA MathWorks Documentation page to learn more on deploying deep learning networks onto target FPGA and SoC boards.

[Ng Yong Jie]

Do you have any suggestions on what approach to take for the anomaly detection and warning system?

[Prateek Rai]

It depends on your model. Anomaly detection depends on what activity you think anomaly is. For example, if you are reconstructing the input back as output then anomaly depends on the mean square error between input and output. Another example is what I stated in my answer, if you want your prediction to be less than a threshold then anomaly would be the case when output is more than that threshold.