These units are used to predict the failure event at base transceiver stations applying 4 ML models.
The description of raw data is in bts data and a small pre-processing data file is within this directory. The pre-processing data was transformed from the raw data for the ML.
Many months of raw data are not shared here, available based on the discussion with the author.
- DNN Multi Regression
- DNN Single Regression
- Multi Var LR
- Single Var RL
- LSTM single series
The ML unit relies on ML models trained with the data mentioned above.
TODO: a short info to indicate where is the code for training
This model is
- applied to a single station for a single parameter (e.g., Load of Power Grid)
Further trained information can be found in the QoA4ML paper.
A simple test case includes:
- A message broker, using RabbitMQ, are used for sending and receiving requests and results
- A client-v1 sends normalized data for prediction and gets results
- A ML Unit as a service loads trained models from exported format (in TensorFlowLite), obtains requests from the the broker, performs the prediction and returns the predicted value
TODO:
- A client-v2 just sends raw sensoring data to a ML service v2 which performs data preprocessing and other data processing tasks and serving
- Python3
- Pika
- Docker
- TensorFlow/TensorFlowLite
- Start RabbitMQ using docker composed by running the script (run.sh) in server folder
- Start server/bts_prediction_server.py
- Use the example client code with our data