7.9.33. Glass Break Detection

Detect glass breaking events from their acoustic signature with FFT features and a compact DSCNN on the MSPM0G5187 with NPU. New in 1.5.0.

Example folder: examples/glass_break_detection (GitHub)

7.9.33.1. Overview

Task: audio classification (audio_classification, target_module: audio), 3 classes
Application: security systems, intrusion detection and smart home safety. Threshold-based detectors have high false positive rates, while a learned acoustic pattern model is more robust. Processing is on the device, so there is no cloud dependency.
Device: MSPM0G5187 (config_MSPM0.yaml, the only config in the folder)
Model: DSCNN_GB_NPU
Targets in the readme: response time under 100 ms, detection accuracy above 95%, false positive rate below 5%, under 32 KB SRAM and 128 KB flash

7.9.33.2. The Dataset

The data covers glass breaking (tempered, laminated and plate glass), background noise (household sounds, music, voices) and similar sounds (dishes breaking, metal impacts, door slams), with 100+ captures per class. Before augmentation all data comes from public datasets: DEMAND (noise), FSD50K, ESC-50 and Google Speech Commands.

The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/glass_break_detection.zip. In the 1.5.0 test run the three classes are glass_break, human_speech and others (glass breaking, speech, and all other sounds).

7.9.33.3. Feature Extraction

The config uses the preset GlassBreakDetection_512FFT_Default on 1-second windows sampled at 44.1 kHz:

The window is split into frames of 512 samples.
A 512-point real FFT is computed per frame (FFT_Q15 in the preset transform chain).
The complex magnitude is calculated, the DC component is removed and the Q15 values are scaled.
Adjacent bins are averaged (binning) into 32 features per frame.
86 frames are concatenated, giving an input of 86 x 32 features.

The preset transform chain is ["FFT_Q15", "Q15_SCALE", "Q15_MAG", "BINNING", "CONCAT"] with frame_size: 512, feature_size_per_frame: 32, num_frame_concat: 86 and q15_scale_factor: 8 (shown as comments in the config). See Feature Extraction.

7.9.33.4. Model

DSCNN_GB_NPU is a lightweight depthwise separable CNN. In the 1.5.0 test run the NPU compiler reported code 2.03 KB, read-only data 9.09 KB (flash) and read-write data 15.86 KB (SRAM):

Input (N, 1, 86, 32): 86 time frames x 32 FFT features
BatchNorm, Conv 8x8 stride 4, two depthwise 4x4 plus pointwise 1x1 blocks, Dropout, GlobalAvgPool, FC
Output: 3 classes
Mixed precision: depthwise and pointwise convolutions at 4 bits, activations at 8 bits (4-8x smaller than FP32)

7.9.33.5. Running the Example

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

./run_tinyml_modelzoo.sh examples/glass_break_detection/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\glass_break_detection\config_MSPM0.yaml

7.9.33.6. Configuration

common:
  target_module: audio
  task_type: audio_classification
  target_device: MSPM0G5187
dataset:
  dataset_name: glass_break_detection
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/glass_break_detection.zip
data_processing_feature_extraction:
  feature_extraction_name: GlassBreakDetection_512FFT_Default
  sampling_rate: 44100
  audio_duration_ms: 1000
training:
  model_name: DSCNN_GB_NPU
  batch_size: 32
  training_epochs: 20
  num_gpus: 0
  quantization: 2
  learning_rate: 0.001
  weight_decay: 1e-5
  auto_quantization: False
testing: {}
compilation: {}

quantization: 2 with auto_quantization: False runs quantization-aware training with the mixed-precision scheme. For the audio task type see Audio Classification.

7.9.33.7. Results

The example readme lists targets (above 95% accuracy) but no measured accuracy. In the automated 1.5.0 config test run (test set of 1200 samples: 600 glass break, 300 human speech, 300 others) the quantized model reached:

Metric	Value
Best quantized epoch	6
Validation accuracy / F1 score (best epoch)	92.72% / 0.927
Test accuracy	92.42%
Test AUC ROC	0.984

This short run did not reach the 95% target. A longer training run, more data, or the augmentation options may improve it.

7.9.33.8. Output Location

Results are written to ./data/projects/glass_break_detection/run/<date-time>/DSCNN_GB_NPU/. See Understanding the Run Output and Project Compilation Guide: Timeseries Classification Examples for device deployment with the MSPM0 SDK (2.08.00 or later) and a microphone input.