7.9.27. PMSM Rotor Temperature Forecasting

Forecast the permanent magnet (rotor) temperature of a permanent magnet synchronous motor (PMSM) one step ahead, without a temperature sensor inside the motor.

Example folder: examples/forecasting_pmsm_rotor_temp (GitHub)

7.9.27.1. Overview

Task: timeseries forecasting (generic_timeseries_forecasting)
Application: motor protection and derating. Magnets can lose strength permanently above about 150 C, and motor life shortens as temperature rises, but sensors deep inside a running motor are costly and fragile.
Configs and devices:
- config.yaml: F28P55
- config_MSPM0.yaml: MSPM0G5187
Model: FCST_LSTM8 (an LSTM, 393 parameters according to the readme)
Target: pm (permanent magnet surface temperature), one step ahead

7.9.27.2. 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: Electric Motor Temperature (Kaggle).

From the original columns the example keeps signals that are easy to measure: ambient temperature (ambient), coolant temperature (coolant), the voltage components (u_d, u_q) and the phase current magnitude i_a = sqrt(i_d^2 + i_q^2). Using past values of these and of the permanent magnet temperature pm, the model forecasts pm for the next time step. The config uses variables: 6 (i_a, u_d, u_q, ambient, coolant, pm) and target_variables: [5] (pm).

The dataset is downloaded from https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/pmsm_rotor_temp.zip. It follows the forecasting layout (files/ plus annotations/), see Forecasting Dataset Format.

7.9.27.3. Running the Example

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

./run_tinyml_modelzoo.sh examples/forecasting_pmsm_rotor_temp/config.yaml

# MSPM0G5187 variant
./run_tinyml_modelzoo.sh examples/forecasting_pmsm_rotor_temp/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\forecasting_pmsm_rotor_temp\config.yaml

7.9.27.4. Configuration

common:
  task_type: generic_timeseries_forecasting
  target_device: F28P55
dataset:
  dataset_name: forecasting_pmsm_rotor_temp
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/pmsm_rotor_temp.zip
data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 3
  stride_size: 0.4
  forecast_horizon: 1
  variables: 6
  target_variables:
  - 5
training:
  model_name: FCST_LSTM8
  model_config: ''
  batch_size: 256
  training_epochs: 20
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 0
  quantization: 2
  optimizer: adam
  learning_rate: 0.005
  output_int: false
testing: {}
compilation: {}

SimpleWindow and frame_size are mandatory for forecasting; feature extraction is not supported.
output_int must be false.
config_MSPM0.yaml differs only in the target device (MSPM0G5187) and in having no early stopping.

7.9.27.5. Evaluation

The metrics are SMAPE (0 to 200%, lower is better) and R2 (higher is better). The epoch with the lowest SMAPE across all targets and predicted time steps is the best epoch. Per-epoch plots and CSV files are written to the best_epoch_<n>_results and test_results folders (see below).

7.9.27.6. Results

Results of the example readme (best epoch statistics, SMAPE and R2 of pm, time step 1):

Stage	SMAPE	R2
Float train, best epoch 19	0.09%	1.0000
Quantized train, best epoch 10	0.26%	0.9999
Test set	0.25%	0.9998

The automated 1.5.0 config test run of config.yaml gave a quantized best epoch SMAPE of 0.11% (R2 0.9999) and a test SMAPE of 0.13% (R2 0.9999), consistent with the readme.

PMSM rotor temperature predictions — Predicted versus actual permanent magnet temperature (black dotted line is the ideal x = y)

7.9.27.7. Output Location

Results are written to ./data/projects/forecasting_pmsm_rotor_temp/run/<date-time>/FCST_LSTM8/:

training/base/best_epoch_<n>_results: float model
training/quantization/best_epoch_<n>_results and training/quantization/test_results: quantized model, each with prediction_plots and predictions_csv
compilation/artifacts/: mod.a and tvmgen_default.h
training/quantization/golden_vectors/: test_vector.c and user_input_config.h

7.9.27.8. Running on Device

The readme compiled the example with the TI-NPU software preset for F28P55 and ran it on a LAUNCHXL-F28P55X (C2000Ware 6.0.1, CCS 20.2.0). Follow Deploying Forecasting Models from ModelMaker to Device. After flashing, the model output matches the golden vectors.

On-device results — On-device output matching the golden vectors

Measured on the device (example readme, written for 1.3):

Metric	Value
Cycles	102307
Inference time	682.05 us
Code size	4375 bytes
Results match golden vectors	TRUE