Some Ideas about Predictive Maintenance
Today Blog will be without code, but with many different ideas. Before describing the Paper from www.arxiv.org : Tie Luo and Sai G. Nagarajan, “Distributed Anomaly Detection using Autoencoder Neural Networks in WSN for IoT”, as I promised last time, I want to discuss the situation in Predictive maintenance with machine learning in general.
After some Internet surfing about this topic I found out that there are quite a lot of StartUps which work in this direction. For example, here the most successful: https://www.nanalyze.com/2017/12/7-industrial-iot-startups-using-ai/ . Almost all of them have quite the same general idea – use only existing data from customers machines. I have found only one in this list which propose to use own sensors for anomaly detection. I found this approach not quite write. Using existing data is ok, but in most cases it is not enough for predictive maintenance. I think that each machine should be approached individually and should be analyzed which sensors this machine needs.
I think that both ways (supervised and unsupervised learning) should be used for predictive maintenance. Maybe in the future we will have such nice models which could learn by themselves, but they are not available yet. What I also think is a good idea is to start monitoring each machine from its lifestart (from installation), as it is supposed that the machine is in best condition in this time and after that it starts degrading. It is, of course, not always a situation in real life, as a new machine could be made not perfectly, but it is a good way to start.
Now let’s talk about the paper about anomaly detection. The idea which is described in paper is quite simple: use simple autoencoder directly on sensor and then send information to server when anomaly detected or just send data ones a day to server to retrain model and get new parameters. I found this idea interesting. As I tested my wireless vibration analyses system I found out that most problems come from data sending from sensor. Often occurs interruption in data transfer and data comes corrupted. It would be better to make data analysis direct on sensor and then just send the result to server for further analysis. The problem is only that the controller which comes with sensor has very limited resources and can work with very simple model, like 3 layers autoencoder in the Paper. For example, using FFT directly on sensor would be problematic. But I think that soon we will get powerful and cheap enough controller which we can use with sensors.