I've designed a machine learning model for the predictive maintenance of machines. The data used for training and testing the ML model is the data from various sensors connected to various parts of the machines. Now, I'm searching for a good approach for deploying the model in the real-time environment as explained here. I did some research and found some information about using real-time data for prediction such as using Kafka. I have some questions unanswered regarding the deployment of the ML model. Following are the details of my system:
- The sensors (pressure, temperature, flow, vibration, etc) are deployed across the parts of the machines.
- The ML model is trained with historical data.
- For predictive maintenance (anomaly detection), streams of data will be available via MQTT. As there are 3000 machines, the volume of data will be very high.
My questions are:
- Where will be the best place to perform prediction operation, at the factory premice where machines are located (edge computing), at our office (that designs ML model), or at cloud server? I want to know it in regard to operational cost.
- Is there any way to estimate the effectiveness of the complete system (full-stack ML architecture)?