Federated Learning in Automotive

Use case 1: EV battery range prediction model update

Model inputs (features extracted from a time window)

Feature description	Label
1. Initial State of Charge (SoC) (%)	soc_percent
2. Delta State of Charge (DSoC) (%)	delta_soc_percent
3. Total battery energy spend (kWh)	battery_energy_spend_kwh
4. Regen energy gain (kWh)	regen_energy_gain_kwh
5. Average speed (km/h)	avg_speed_kmph
6. Vehicle mass (kg)	vehicle_mass_kg
7. Distance travelled in time window (km)	distance_step_km
Model output - Remaining vehicle range (km)	remaining_range_km

Using data from EV Ramnge Simulator for training/validation

Using EV battery range simulator form here. Features are typically avilable or can be calculated from raw data published in the vehicle CAN bus.

Model training (without Federated Learning)

python3 model_trainer.py

It is expected that model weights will not converge during training as randomized data is used.

Model training using flower FL library with pytorch model

Following steps are required to run the demo for EV range model training using Federated Learning

Code changes

1. Use flwr new to create a template director
2. Update model architecture and data source for the use case in ./ev_range_pred/task.py
   • Replace the CIFAR dataset with random data
   • Replace flower_datasets with torch DataLoader

Above steps are completed in ./ev_range_pred directory.

Running the demo on localhost

Run below commands in different terminals as given in deployment tutorial.

Terminal 1 - Run superlink

flower-superlink --insecure

Terminal 2 - Run supernode

flower-supernode \
     --insecure \
    --superlink 127.0.0.1:9092 \
    --clientappio-api-address 127.0.0.1:9094 \
    --node-config "partition-id=0 num-partitions=2"

Terminal 3 - Run client (to be ported to run on the vehicle edge ECU)

flower-superexec \
    --insecure \
    --plugin-type clientapp \
    --appio-api-address zupernode-1:9094