8.3. Feature Extraction

Feature extraction transforms raw sensor data into a representation that helps the neural network learn patterns more effectively.

8.3.1. Overview

Why use feature extraction?

Reduced input size: Compress long time series
Better patterns: Transform to domain where patterns are clearer
Faster inference: Smaller inputs mean faster models
Domain knowledge: Incorporate signal processing expertise

8.3.2. Feature Extraction Pipeline

Raw data flows through two stages: data processing transforms, then feature extraction transforms:

Raw Signal → Data Processing → Feature Extraction → Model Input
             (data_proc_transforms)   (feat_ext_transform)
             e.g. SimpleWindow,       e.g. FFT_FE, BINNING,
                  Downsample               ABS, LOG_DB, CONCAT

8.3.3. Configuration Parameters

The data_processing_feature_extraction section supports the following parameters. There are two usage modes: using a preset name, or defining a custom pipeline.

Core Parameters:

Option	Description
`feature_extraction_name`	Preset name (e.g., `'Generic_1024Input_FFTBIN_64Feature_8Frame'`) or a custom name starting with `'Custom_'` (e.g., `'Custom_Default'`, `'Custom_ArcFault'`). When using a preset, the transform pipeline is predefined. When using a `Custom_*` name, you must specify `feat_ext_transform` and related parameters.
`data_proc_transforms`	List of data processing transforms applied before feature extraction. Common values: `['SimpleWindow']`, `['Downsample']`, `['SimpleWindow', 'Downsample']`, `['Downsample', 'SimpleWindow']`, or `[]` (empty).
`feat_ext_transform`	List of feature extraction transforms applied in order. Common values include: `'FFT_FE'`, `'FFT_POS_HALF'`, `'WINDOWING'`, `'BINNING'`, `'NORMALIZE'`, `'ABS'`, `'LOG_DB'`, `'DC_REMOVE'`, `'CONCAT'`, `'FFT_Q15'`, `'Q15_SCALE'`, `'Q15_MAG'`, `'BIN_Q15'`, `'ECG_NORMALIZE'`.
`variables`	Number of input channels/variables. Supports three formats: an integer (select first N columns), a list of column indices (e.g., `[0, 2, 4]`), or a list of column names (e.g., `['accel_x', 'accel_y', 'accel_z']`).
`frame_size`	Number of samples per frame (e.g., `128`, `256`, `512`, `1024`).
`feature_size_per_frame`	Number of output features per frame after transform (e.g., `8`, `16`, `32`, `64`, `128`).
`num_frame_concat`	Number of frames to concatenate (e.g., `1`, `4`, `8`). Total features = `feature_size_per_frame` x `num_frame_concat`.
`stride_size`	Stride between frames as a fraction (e.g., `0.01`, `0.1`, `0.25`, `0.5`, `1`).

Signal Processing Parameters:

Option	Description
`sampling_rate`	Original sampling rate of the input signal. Used with the `Downsample` data processing transform.
`new_sr`	Target sampling rate after downsampling. Used with the `Downsample` data processing transform.
`scale`	Scaling factor applied to input data (e.g., `0.00390625` for 1/256).
`offset`	Offset added to input data.
`frame_skip`	Number of frames to skip between selected frames (e.g., `1`, `8`).
`normalize_bin`	Enable bin normalization (`True`/`1`).
`stacking`	Feature stacking mode: `'2D1'` or `'1D'`.
`min_bin`	Minimum frequency bin index to include.
`analysis_bandwidth`	Fraction of bandwidth to analyse (e.g., `1` for full bandwidth).

Logarithmic Transform Parameters:

Option	Description
`log_mul`	Multiplier for logarithmic scaling (e.g., `20` for dB).
`log_base`	Base for logarithm (e.g., `10`).
`log_threshold`	Minimum threshold to avoid log(0) (e.g., `1e-100`).

Fixed-Point (Q15) Parameters:

Option	Description
`q15_scale_factor`	Scale factor for Q15 fixed-point quantization (e.g., `4`, `5`).

Data Augmentation and Testing:

Option	Description
`gain_variations`	Dictionary mapping class names to `[min_gain, max_gain]` ranges for data augmentation. Example: `{arc: [0.9, 1.1], normal: [0.8, 1.2]}`.
`gof_test`	Run Goodness of Fit test on extracted features (`True`/`False`).

Output Control:

Option	Description
`store_feat_ext_data`	Store extracted feature data to disk (`True`/`False`).
`nn_for_feature_extraction`	Use neural network for feature extraction (`True`/`False`).

Forecasting-Specific Parameters:

Option	Description
`forecast_horizon`	Number of future timesteps to predict (e.g., `1`, `2`).
`target_variables`	List of column indices or names for the target variable(s) to forecast (e.g., `[0]`, `[5]`, `['temperature']`).

8.3.4. Preset System

Tiny ML Tensorlab provides predefined feature extraction presets. When using a preset, simply specify the feature_extraction_name and variables:

data_processing_feature_extraction:
  feature_extraction_name: 'Generic_1024Input_FFTBIN_64Feature_8Frame'
  variables: 1

Preset Naming Convention:

Generic_<InputSize>Input_<Transform>_<Features>Feature_<Frames>Frame

Example: Generic_1024Input_FFTBIN_64Feature_8Frame
- Input: 1024 samples
- Transform: FFT with binning
- Features: 64 frequency bins
- Frames: 8 temporal frames
- Total: 64 x 8 = 512 features to model

8.3.5. Available Presets

FFT-Based Presets:

Best for frequency-domain patterns (vibration, arc faults):

Preset	Features	Use Case
`Generic_1024Input_FFTBIN_64Feature_8Frame`	512	General purpose
`Generic_512Input_FFTBIN_32Feature_8Frame`	256	Smaller input
`FFT1024Input_256Feature_1Frame_Full_Bandwidth`	256	Full spectrum

Raw Time-Domain Presets:

Best for waveform shape patterns:

Preset	Features	Use Case
`Generic_512Input_RAW_512Feature_1Frame`	512	Full waveform
`Generic_256Input_RAW_256Feature_1Frame`	256	Shorter window
`Generic_128Input_RAW_128Feature_1Frame`	128	Compact input

Application-Specific Presets:

Preset	Application
`Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1`	Motor fault (3-axis)
`FFT1024Input_256Feature_1Frame_Full_Bandwidth`	Arc fault detection
`PIRDetection_125Input_25Feature_25Frame_1InputChannel_2D`	PIR detection

8.3.6. Data Processing Transforms

The data_proc_transforms parameter specifies preprocessing steps applied to raw data before feature extraction:

SimpleWindow

Segments continuous data into fixed-size windows:

data_processing_feature_extraction:
  data_proc_transforms: ['SimpleWindow']
  frame_size: 256
  stride_size: 0.01
  variables: 1

Downsample

Reduces the sampling rate of input data:

data_processing_feature_extraction:
  data_proc_transforms: ['Downsample', 'SimpleWindow']
  sampling_rate: 313000
  new_sr: 3130
  frame_size: 256
  stride_size: 0.01
  variables: 1

Multiple transforms can be chained in order:

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  - Downsample
  frame_size: 256
  sampling_rate: 100
  new_sr: 1
  variables: 1

8.3.7. Feature Extraction Transforms

The feat_ext_transform parameter defines the feature extraction pipeline as an ordered list of transforms. Each step processes the output of the previous step.

Common Transform Steps:

Transform	Description
`FFT_FE`	Compute FFT (Fast Fourier Transform)
`FFT_POS_HALF`	Keep only positive frequency half of FFT
`WINDOWING`	Apply windowing function
`BINNING`	Group frequency bins to reduce feature count
`NORMALIZE`	Normalize features
`ABS`	Take absolute value
`LOG_DB`	Convert to logarithmic (dB) scale
`DC_REMOVE`	Remove DC component
`CONCAT`	Concatenate frames into final feature vector
`FFT_Q15`	Fixed-point Q15 FFT (for MCU deployment)
`Q15_SCALE`	Q15 scaling
`Q15_MAG`	Q15 magnitude computation
`BIN_Q15`	Q15 binning
`ECG_NORMALIZE`	ECG-specific normalization

Example: FFT with Binning Pipeline:

data_processing_feature_extraction:
  feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'ABS', 'BINNING', 'LOG_DB', 'CONCAT']
  frame_size: 1024
  feature_size_per_frame: 64
  num_frame_concat: 4
  variables: 1

Example: FFT without Binning Pipeline:

data_processing_feature_extraction:
  feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'ABS', 'LOG_DB', 'CONCAT']
  frame_size: 256
  feature_size_per_frame: 128
  num_frame_concat: 1
  variables: 6

Example: Fixed-Point Q15 Pipeline (for MCU deployment):

data_processing_feature_extraction:
  feat_ext_transform: ['FFT_Q15', 'Q15_SCALE', 'Q15_MAG', 'DC_REMOVE', 'BIN_Q15', 'CONCAT']
  frame_size: 256
  feature_size_per_frame: 16
  num_frame_concat: 8
  q15_scale_factor: 5
  normalize_bin: True
  variables: 1

8.3.8. Custom Feature Extraction

For advanced use cases, use a Custom_* feature extraction name and specify the transform pipeline manually:

data_processing_feature_extraction:
  data_proc_transforms: ['SimpleWindow']
  feature_extraction_name: 'Custom_Default'
  feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'WINDOWING', 'BINNING', 'NORMALIZE', 'ABS', 'LOG_DB', 'CONCAT']
  frame_size: 32
  feature_size_per_frame: 8
  num_frame_concat: 8
  variables: 5

You can also configure additional parameters for fine-grained control:

data_processing_feature_extraction:
  data_proc_transforms: []
  feature_extraction_name: 'Custom_MotorFault'
  feat_ext_transform: ['FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'ABS', 'BINNING', 'LOG_DB', 'CONCAT']
  frame_size: 1024
  feature_size_per_frame: 64
  num_frame_concat: 4
  normalize_bin: 1
  stacking: '1D'
  offset: 0
  scale: 1
  frame_skip: 1
  log_mul: 20
  log_base: 10
  log_threshold: 1e-100
  variables: 3

8.3.9. Multi-Channel Data

For sensors with multiple axes (e.g., 3-axis accelerometer), set variables to the number of channels:

data_processing_feature_extraction:
  feature_extraction_name: 'Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1'
  variables: 3

The variables parameter supports three formats:

Integer: Select first N columns (e.g., variables: 3)
List of indices: Select specific columns (e.g., variables: [0, 2, 4])
List of names: Select columns by name (e.g., variables: ['accel_x', 'accel_y', 'accel_z'])

8.3.10. Forecasting Configuration

Forecasting tasks require specific additional parameters:

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 32
  stride_size: 0.1
  forecast_horizon: 2
  variables: 1
  target_variables:
  - 0

Note

SimpleWindow must be specified in data_proc_transforms for forecasting tasks.

8.3.11. Data Augmentation

Use gain_variations to augment training data with gain variations per class:

data_processing_feature_extraction:
  data_proc_transforms:
  - Downsample
  - SimpleWindow
  gain_variations:
    arc: [0.9, 1.1]
    normal: [0.8, 1.2]
  sampling_rate: 313000
  new_sr: 3130
  frame_size: 256
  stride_size: 0.01
  variables: 1

8.3.12. Choosing the Right Preset

Decision Tree:

Is the pattern in frequency content?
|-- Yes --> Use FFT-based preset
|   |-- Need full spectrum? --> FFT_FullBandwidth
|   |-- Reduce features? --> FFTBIN
|-- No --> Use RAW preset
    |-- Need temporal context? --> Multi-frame
    |-- Single snapshot? --> 1Frame

Common Choices by Application:

Application	Recommended Preset
Arc fault detection	`FFT1024Input_256Feature_1Frame_Full_Bandwidth`
Motor bearing fault	`Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1`
ECG classification	`ECG2500Input_Roundoff_1Frame`
Vibration anomaly	`Generic_1024Input_FFTBIN_64Feature_8Frame`
Simple waveforms	`Generic_512Input_FFTBIN_32Feature_8Frame`
PIR detection	`PIRDetection_125Input_25Feature_25Frame_1InputChannel_2D`

8.3.13. Performance Impact

Feature extraction affects model size and speed:

Features	Model Input	Model Size	Inference Time
128	Small	Smaller	Faster
256	Medium	Medium	Medium
512	Large	Larger	Slower

Trade-off:

More features = more information = potentially better accuracy
Fewer features = faster inference = fits smaller devices

8.3.14. On-Device Feature Extraction

Feature extraction runs on the MCU before inference. The compilation process generates C code for the feature extraction pipeline configured in your YAML.

Memory Usage:

Feature extraction buffers add to memory requirements:

Input buffer:  frame_size x variables x sizeof(data_type)
FFT buffer:    frame_size x sizeof(data_type)
Output buffer: feature_size_per_frame x num_frame_concat x sizeof(data_type)

8.3.15. Example Configurations

Arc Fault Classification (using preset):

data_processing_feature_extraction:
  feature_extraction_name: 'FFT1024Input_256Feature_1Frame_Full_Bandwidth'
  variables: 1

Motor Bearing Fault (using preset with override):

data_processing_feature_extraction:
  feature_extraction_name: 'Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1'
  variables: 3
  feature_size_per_frame: 4

Anomaly Detection with Downsampling:

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  - Downsample
  frame_size: 1024
  sampling_rate: 100
  new_sr: 1
  variables: 1

Regression with Simple Windowing:

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 512
  stride_size: 0.1
  variables: 6

Forecasting (PMSM Rotor Temperature):

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 3
  stride_size: 0.4
  forecast_horizon: 1
  variables: 6
  target_variables:
  - 5

Goodness of Fit Testing:

Enable the gof_test parameter to run Goodness of Fit analysis on extracted features:

data_processing_feature_extraction:
  feature_extraction_name: 'Input256_FFTBIN_16Feature_8Frame_3InputChannel_removeDC_2D1'
  gof_test: True
  variables: 3

PCA Visualization of Extracted Features:

PCA (Principal Component Analysis) helps visualize how well the extracted features separate your classes. Well-separated clusters indicate good feature extraction.

PCA on Training Data — PCA visualization of extracted features on training data

PCA on Validation Data — PCA visualization of extracted features on validation data

Interpreting PCA plots:

Tight clusters: Features represent the class well
Well-separated clusters: Good class separability
Overlapping clusters: May need different feature extraction
Scattered points: High variance, potentially noisy data

8.3.16. Best Practices

Match to signal characteristics: FFT for periodic, raw for transient
Start with standard presets: Customize only if needed
Consider device constraints: Fewer features for smaller devices
Test multiple options: Compare accuracy with different presets
Use domain knowledge: Understand what patterns you’re looking for

8.3.17. Next Steps

See Goodness of Fit to analyze dataset quality
Learn about Quantization for model compression
Explore Time Series Classification for classification