Smart Intent Guarantee based on Closed-loop - ML MS Enhancements

Refer Intelligent Slicing flow for complete closed loop sequence. Existing ML MS (not officially exposed to ONAP) requires certain enhancements to satisfy the intent guarantee. PM data generated for the closed loop realization is based on the RAN network elements - GNBCUCPFunction .

Overview

Prerequisites:

1. Configuration of E2E network slices  (~ 5 slices)
2. RAN Simulator, VES Collector, PM Mapper, Data File Collector, Slice Analysis MS, ML Prediction MS, SO, Policy, SDNR, Config DB components should be up and running
3. SFTP setup to store the PM messages from RAN Simulator
4. Manual configurations are detailed at Closed Loop for Network Slicing
5. To Train and predict Machine learning model, following ML frameworks are used Tensorflow, Keras and sklearn.

Functionality of ML block is divided into two parts. 

1. Model Training (Offline) and validation
2. Prediction

Step 1: ML Offline Training

This step requires a huge amount of historical data for the analysis. It requires both the input (PM metrics) and output data (intent). 

Configure E2E Slice (shared or non-shared) that need to be trained.

Training is performed by first acquiring the data. The data are acquired from following topics.

• The training modules trigger PM data generation for slices using the below RAN Simulator API
• The process typically waits for more data to be generated. Currently it waits for 2 hours.

curl -X POST -H "Content-Type:  text/plain" http://localhost:8081/ransim/api/GenerateIntelligentSlicingPmData -i

• After the time elapse, the training modules perform stop PM data generated for slices using following API.

curl -X POST -H "Content-Type:  text/plain" http://localhost:8081/ransim/api/stopIntelligentSlicingPmData -i 

• The Training module start to get the acquired data for all the slices and cells using PM Mapper topic

curl -X POST -H "Content-Type:  text/plain" https://localhost:8081/events/org.onap.dmaap.mr.PERFORMANCE_MEASUREMENTS/mlms-cg/mlms-cid 

• While acquiring each slice and cell data, the training module also acquire the intent configuration for each cell and slices. using following Config DB API. Config DB will be replaced by CPS in the next release.

curl -X POST -H "Content-Type:  text/plain" https://localhost:8081/api/sdnc-config-db/v4/nrcellcu-configdata/1/snssai/01-B989BD

• Following are the processes involved in the training the model.

• Once the all the training data are acquired, the training module performs preprocessing to convert the data from JSON format into Dictionary format
• In Preprocessing the training Data is normalized and Transformed.
• The preprocessed data (PM data Parameters data content) for all slices and cells which are in time series are converted into forecast series.
  • The forecast series is cells PM data parameters for current time instance, Plus the previous past or last four time instance data.
  • Similar forecast series data is prepared for all the cells in the slice.
• Data are split into training and validation datasets
• The expected results for each time instance cell data are synthetically generated based on rule conditions and the cell intends.
  •  Sample training data referred at following source code location components/ml-prediction-ms/train/ExampleSample_train_data_s1.xlsx
  • P.S: This fill is only for reference, on how the training data are prepared. In actual training process, the data are fetched  RANSim topic.
  • Following wiki attached excel file contains sample cell data from a slice. Here each of the cell's timeseries data will appended from first cells to the last cell its time series successive to that  slice. This append timeseries data from all the cells from a slice will create a large training data block..
  • a
• The training module perform Training against an LSTM model.
• The Training model file are archived.
• The results are manually verified .

PM file format

Step2: Prediction 

The prediction is performed by first by acquiring the current time instance data from slices and cells. The data are acquired from following APIs.

• The Prediction modules trigger PM data generation for slices using the below RAN Simulator API
• The process typically waits for more data to be generated. Currently it waits for 2 hours.

curl -X POST -H "Content-Type:  text/plain" http://localhost:8081/ransim/api/GenerateIntelligentSlicingPmData -i

• After the time elapse, the Prediction modules perform stop PM data generated for slices using following API.

curl -X POST -H "Content-Type:  text/plain" http://localhost:8081/ransim/api/stopIntelligentSlicingPmData -i 

• The Prediction module start to get the acquired data for all the slices and cells using the PM Mapper topic

• At start of the prediction process, we have a cold start condition for the model to make accurate predictions, The actual prediction starting from the 5th time instance, while each time instance data are generated for every 15 minutes, the 5th time instance data will be retrieved after 1 hour 15 minutes, till then the model generated synthetic data for the first 4 time instant we are generating and then performs prediction. This is  done as a softmax correction, essential to get better accuracy. After the first 4 time instance the predicted values are used and and taken forward.

curl -X POST -H "Content-Type:  text/plain" https://localhost:8081/events/org.onap.dmaap.mr.PERFORMANCE_MEASUREMENTS/mlms-cg/mlms-cid 

• While acquiring each slice and cell data, the Prediction module also acquire the intent configuration for each cell and slices. using following Config DB API. Config DB will be replaced by CPS in the next release.

curl -X POST -H "Content-Type:  text/plain" https://localhost:8081/api/sdnc-config-db/v4/nrcellcu-configdata/1/snssai/01-B989BD

The response will be in the form of:

{

    "dLThptPerSlice": 1,

    "uLThptPerSlice": 2,

    "maxNumberOfConns":300

}

• Following are the processes involved in the Prediction the model. 

• The data from above topics are used to acquire the current time instance and also the previous last four time instance data for all cells and slice. 
• The collected data are parsed, normalized, transformed and converted into forecast series, the steps followed here are almost the same to the input preprocessing steps used in Training module.
• The preprocess forecast series along with the intend are fed to the ML model to perform prediction for the resource allocation for each cells. The predicted results are further optimised by the 25% +/- on cell config and 10% +/- on slice config thresholds.
• Recommendation for each cell can be maximum of 5120 and the aggregation of maximumNoOfConnections recommended for a slice (sNSSAI) can be upto 110% (10 % buffer)  of the intent.
• Now predict recommend suggestions for the cell level configurations (maxNumberOfConns)  for a slice. A new micro service will be introduced to do the data processing and prediction for the dynamic data.

Suggested Configuration from the ML MS should be in the form of below format

CPS API:

The CPS integration for RAN slice allocation is done in J-release, Config DB API will be replaced with the CPS API in the next release.