7.9.24. Washing Machine Load Weighing

Estimate the weight of the clothes in a washing machine from motor voltage, current and speed, which removes the need for a mechanical weight sensor.

In 1.4.0 the example folder was named reg_washing_machine. In 1.5.0 it is washing_machine_load_weighing. The docname of this page stays washing_machine_regression.

Example folder: examples/washing_machine_load_weighing (GitHub)

7.9.24.1. Overview

Task: timeseries regression (generic_timeseries_regression)
Application: sensorless load weighing for automatic water level adjustment, lowering manufacturing cost
Configs and devices:
- config.yaml: F28P55, frame_size: 512
- config_MSPM0.yaml: MSPM0G5187, frame_size: 1024
Model: REGR_13k (12,477 trainable parameters according to the readme); REGR_4k (4,093 parameters) can be used as a smaller alternative

7.9.24.2. The Dataset

The data comes from TI Kilby Labs: 6 voltage, current and speed based signals from a washing machine running in constant torque mode, with loads from 0 g to 900 g in 100 g steps.

Variable	Description
`Vd`	Voltage component along the d-axis
`Id`	Current component along the d-axis
`Vq`	Voltage component along the q-axis
`Iq`	Current component along the q-axis
`Iqref`	Reference current
`Speed`	Speed of the washing machine motor

The dataset has 391 files of about 12,300 data points each: 276 for training, 51 for validation and 64 for testing (70% / 14% / 16%). The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/washing_machine_loading_data.zip. It follows the regression layout (files/ plus annotations/) described in Regression Dataset Format.

7.9.24.3. Running the Example

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

./run_tinyml_modelzoo.sh examples/washing_machine_load_weighing/config.yaml

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

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\washing_machine_load_weighing\config.yaml

7.9.24.4. Configuration

common:
  task_type: generic_timeseries_regression
  target_device: F28P55
dataset:
  dataset_name: washing_machine_load_weighing
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/washing_machine_loading_data.zip
data_processing_feature_extraction:
  stride_size: 0.1
  data_proc_transforms:
  - SimpleWindow
  frame_size: 512
  variables: 6
training:
  optimizer: adam
  lr_scheduler: cosineannealinglr
  model_name: REGR_13k
  batch_size: 128
  training_epochs: 100      # use 500 epochs for a better R2 score and RMSE
  early_stopping: true
  early_stopping_patience: 5
  lambda_reg: 0.01
  num_gpus: 1
  quantization: 2
  auto_quantization: True
testing: {}
compilation: {}

Feature extraction such as FFT, binning or ABS did not improve the RMSE for this task, so only SimpleWindow with a suitable frame size is used. variables: 6 is the number of input signals.
num_gpus: 1 expects a GPU; use 0 for CPU.
config_MSPM0.yaml targets MSPM0G5187, uses frame_size: 1024 and has no early stopping.
The training loss is MSE and the reported metric is RMSE. The readme recommends about 500 epochs for a good RMSE.
With quantization: 2 and auto_quantization: True the model is partially quantized. The input BatchNorm, the first conv or linear layer and the last linear layer stay in float, and all other layers are INT8. quantization: 0 gives the float model.

7.9.24.5. Results

Test-set RMSE in grams from the example readme, after 500 training epochs each:

Frame size	`REGR_4k`	`REGR_13k`	Partially quantized `REGR_4k`	Partially quantized `REGR_13k`
256	63.00	35.77	62.32	33.37
512	46.91	32.58	52.42	31.23
1024	59.01	25.78	64.59	36.78

For the 13k model the RMSE falls as the frame size grows, because the model sees more context. The 4k model has too few parameters to exploit larger frames. The readme quotes a best RMSE of 25.78 g for the float model and 31.23 g for the partially quantized model. These numbers are from the example readme; they were not regenerated for 1.5.0. The automated 1.5.0 config test run (short, early stopped) completed for both config.yaml and config_MSPM0.yaml.

Float REGR_13k predicted versus actual weight — Float `REGR_13k`, frame size 512 (from the example readme)

Partially quantized REGR_13k predicted versus actual weight — Partially quantized `REGR_13k`, frame size 512 (from the example readme)

7.9.24.6. Output Location

Results are written to ./data/projects/washing_machine_load_weighing/run/<date-time>/REGR_13k/. The four device files 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.

7.9.24.7. Running on Device

See Deploying Regression Models from ModelMaker to Device and Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check).