7.9.22. Torque Measurement Regression

Estimate the electromagnetic torque of a permanent magnet synchronous motor (PMSM) from electrical and thermal measurements, as a virtual (sensorless) torque sensor.

Example folder: examples/torque_measurement_regression (GitHub)

7.9.22.1. Overview

Task: timeseries regression (generic_timeseries_regression)
Application: motor control, sensorless torque estimation
Device: F28P55 (config.yaml)
Model: REGR_10k
Target: torque

7.9.22.2. The Dataset

The data are measurements from a PMSM prototype on a test bench (German OEM motor, recorded by the LEA department at Paderborn University), sampled at 2 Hz. It is the Electric Motor Temperature dataset on Kaggle.

Column	Description
`i_d`, `i_q`	Current d- and q-components
`u_d`, `u_q`	Voltage d- and q-components
`motor_speed`	Motor speed
`ambient`, `coolant`	Ambient and coolant temperature
`pm`	Permanent magnet surface temperature
`stator_winding`, `stator_tooth`, `stator_yoke`	Stator temperatures
`torque`	Torque induced by current (the target)

The example uses the first 10 columns (i_d to stator_tooth) as inputs (variables: 10) and predicts torque. The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/torque_measurement_dataset.zip. The dataset follows the regression layout described in Regression Dataset Format.

7.9.22.3. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/torque_measurement_regression/config.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\torque_measurement_regression\config.yaml

7.9.22.4. Configuration

common:
  task_type: generic_timeseries_regression
  target_device: F28P55
dataset:
  dataset_name: torque_measurement
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/torque_measurement_dataset.zip
data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  variables: 10
  frame_size: 128
  stride_size: 0.25
training:
  model_name: REGR_10k
  batch_size: 512
  training_epochs: 50
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 1
  learning_rate: 0.001
  output_int: false
  auto_quantization: true
testing: {}
compilation: {}

num_gpus: 1 expects a GPU; set it to 0 to train on CPU. Try other frame_size and stride_size values to see how they affect training. In the readme experiment the smaller frame (128) gave the better R2 score.

7.9.22.5. Results

The example readme reports these results for two frame sizes (stride size 0.25):

Frame size	Float R2 (MSE)	Quantized R2 (MSE)	Cycles
128	0.994 (12.98)	0.963 (83.59)	114102
256	0.983 (36.82)	0.948 (113.36)	187445

For frame size 128 the readme also reports a test-set RMSE of 9.40, MSE of 88.36 and R2 of 0.98. The cycle counts are taken from the readme and were measured on the device for the quantized model.

7.9.22.6. Output Location

Results are written to ./data/projects/torque_measurement/run/<date-time>/REGR_10k/. The four files needed on the device are compilation/artifacts/mod.a, compilation/artifacts/tvmgen_default.h, training/quantization/golden_vectors/test_vector.c and training/quantization/golden_vectors/user_input_config.h. See Understanding the Run Output.

7.9.22.7. Running on Device

The C2000Ware (or tinyml-sdk) example f28p55x_generic_timeseries_regression takes these four files. Follow Deploying Regression Models from ModelMaker to Device and Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check). Check the test_result variable for the cases in test_vector.c.