7.9.23. Induction Motor Speed Prediction

Predict the speed (RPM) of a three-phase induction motor from supply, motor and operating parameters.

Example folder: examples/induction_motor_speed_prediction (GitHub)

7.9.23.1. Overview

Task: timeseries regression (generic_timeseries_regression)
Application: motor monitoring and sensorless speed estimation
Device: F29H85 (config.yaml)
Model: REGR_1k (two BatchNorm + Conv + ReLU layers and two light linear layers)
Target: motor speed in RPM

7.9.23.2. The Dataset

The dataset is simulated: 15,000 observations of three-phase induction motors under varied electrical and mechanical conditions. It combines physics-based relations (frequency determines synchronous speed) with non-linear interactions (load, voltage and temperature affect slip).

Supply: voltage (200 to 460 V), current, frequency (50 or 60 Hz), power factor (0.65 to 0.95)
Motor: number of poles (2, 4, 6, 8), stator resistance
Operating conditions: load torque, ambient temperature
Target: motor speed (RPM), derived from synchronous speed and slip

Electromechanical relationships of the induction motor dataset — Electromechanical relationships used to generate the dataset

The key relations are Sync = 120 * Frequency / Poles, Slip = (Sync - Actual) / Sync and Actual = Sync * (1 - Slip). Active power P = V * I * PF and apparent power S = V * I are also included.

7.9.23.3. Preparing the Dataset

config.yaml downloads induction_motor_speed_prediction_dataset.zip from https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/. To regenerate the zip locally, run the helper script and point dataset.input_data_path at the output:

cd examples/induction_motor_speed_prediction
python induction_motor_speed_prediction.py

7.9.23.4. Running the Example

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

./run_tinyml_modelzoo.sh examples/induction_motor_speed_prediction/config.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\induction_motor_speed_prediction\config.yaml

7.9.23.5. Configuration

common:
  task_type: generic_timeseries_regression
  target_device: F29H85
dataset:
  dataset_name: induction_motor_speed_prediction
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/induction_motor_speed_prediction_dataset.zip
data_processing_feature_extraction:
  frame_size: 8
  data_proc_transforms: ['SimpleWindow']
  stride_size: 1
  feature_extraction_name: Generic_8Input_ABS_8Feature_1Frame
  variables: 11
training:
  model_name: REGR_1k
  batch_size: 256
  training_epochs: 50
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 0
  learning_rate: 0.000001
  output_int: false
testing: {}
compilation: {}

The dataset is not a true timeseries, so a small frame size works better. The shipped config uses frame_size: 8 together with the Generic_8Input_ABS_8Feature_1Frame feature extraction preset (absolute value of 8 inputs). The readme also reports frame sizes 1 and 4; these need a frame_size and feature extraction setting that match each other.

7.9.23.6. Results

The example readme reports results for three frame sizes:

Frame size	Float R2 (MSE)	Quantized R2 (MSE)	Exported model R2 (RMSE)
1	1.000 (185.94)	0.999 (451.44)	0.990 (87.46)
4	0.997 (645.68)	0.997 (785.67)	0.990 (46.36)
8	0.978 (2347.63)	0.978 (2341.76)	0.970 (59.88)

Device measurements from the readme (F29H85x, flash and SRAM from the map file, cycles from the profiling API):

Frame size	Flash (B)	SRAM (B)	Cycles	Inference time (us)
1	4306	128	1889	9.45
4	4582	130	3519	17.59
8	5292	256	7288	36.44

These numbers come from the example readme (written for 1.3.0) and were not regenerated for 1.5.0.

7.9.23.7. Output Location

Results are written to ./data/projects/induction_motor_speed_prediction/run/<date-time>/REGR_1k/. The files for 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.

7.9.23.8. Running on Device

Copy the four files into the generic_timeseries_regression example of the F29H85x SDK (examples/ai/generic_timeseries_regression/f29h85x), build, flash and check the test_result variable. See Deploying Regression Models from ModelMaker to Device and Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check).