7.7.7. Motor Bearing Fault

Motor bearing fault classification detects and identifies different bearing failure modes from vibration sensor data.

7.7.7.1. Overview

  • Task: Multi-class classification (6 fault types)

  • Application: Predictive maintenance for motors

  • Dataset: 3-axis vibration data from bearing experiments

  • Model: MotorFault_model_1_t or CLS_4k_NPU

7.7.7.2. Fault Classes

The dataset includes 6 bearing conditions:

  1. Normal - Healthy bearing operation

  2. Contaminated - Foreign particles in lubricant

  3. Erosion - Surface wear

  4. Flaking - Material flaking from bearing surface

  5. No Lubrication - Dry bearing operation

  6. Localized Fault - Point defect on bearing

7.7.7.3. Running the Example

cd tinyml-modelzoo
./run_tinyml_modelzoo.sh examples/motor_bearing_fault/config.yaml

7.7.7.4. Configuration

common:
  task_type: 'generic_timeseries_classification'
  target_device: 'F28P55'

dataset:
  dataset_name: 'motor_fault_classification_dsk'
  input_data_path: 'https://software-dl.ti.com/C2000/esd/mcu_ai/01_03_00/datasets/motor_fault_classification_dsk.zip'

data_processing_feature_extraction:
  feature_extraction_name: 'Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1'
  variables: 3

training:
  model_name: 'MotorFault_model_1_t'
  training_epochs: 20

7.7.7.5. Dataset Description

  • Sensor: 3-axis accelerometer mounted on motor bearing

  • Variables: X, Y, Z acceleration (variables: 3)

  • Sampling: Multiple sampling frequencies available

The vibration patterns differ between fault types due to:

  • Different impact frequencies

  • Varying severity levels

  • Distinct frequency signatures

7.7.7.6. Feature Extraction Presets

Several presets are available for this dataset:

Preset

Features

Accuracy

Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_1D

384

~99.99%

Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1

384

~100%

Input256_FFT_128Feature_1Frame_3InputChannel_removeDC_2D1

384

~98%

Input128_RAW_128Feature_1Frame_3InputChannel_removeDC_2D1

384

~92%

FFT-based binning performs best due to the frequency-domain nature of bearing faults.

7.7.7.7. Available Models

Model

Parameters

Description

MotorFault_model_1_t

~1,000

Baseline

MotorFault_model_2_t

~2,000

Improved

MotorFault_model_3_t

~4,000

Best accuracy

CLS_4k_NPU

~4,000

Generic NPU model

7.7.7.8. Expected Results

With default settings:

Float32 Model:
Accuracy: 99-100%
F1-Score: ~1.0

Quantized Model:
Accuracy: 98-100%

7.7.7.9. Multi-Class Evaluation

For 6-class classification, examine:

Confusion Matrix

Shows which fault types are confused with each other. Common confusion pairs may indicate similar vibration signatures.

Per-Class ROC Curves

One-vs-Rest ROC shows how well each class separates from others.

ROC Curves for Motor Bearing Fault

One-vs-Rest Multi-class ROC curves for motor bearing fault detection showing excellent separation

Class Score Distributions

Histograms show classification confidence for each class.

Class Score Histogram

Distribution of class score differences showing model confidence across different fault types

7.7.7.10. Dataset Quality Analysis

Use the Goodness of Fit (GoF) test to visualize class separability:

data_processing_feature_extraction:
  gof_test: True
  frame_size: 256

training:
  enable: True

This generates 8 plots showing cluster separation using different transformation combinations.

7.7.7.11. Practical Considerations

Sensor Placement

Vibration patterns depend heavily on sensor location. Train with data from the same mounting position as deployment.

Operating Conditions

Include data from different:

  • Motor speeds

  • Load conditions

  • Temperature ranges

Fault Severity

Early-stage faults have subtler signatures. Include samples from different severity levels if available.

7.7.7.12. Anomaly Detection Alternative

If you only have normal data, use anomaly detection instead:

common:
  task_type: 'generic_timeseries_anomalydetection'

training:
  model_name: 'AD_4k_NPU'

This detects any deviation from normal without needing fault labels.

Reconstruction Error Analysis:

Reconstruction Error Log Scale

Reconstruction error distribution showing separation between normal and fault conditions

7.7.7.13. Next Steps