Business Context
A smart city project deploys thousands of sensors to monitor urban noise pollution. The objective is to classify sound events (sirens, drilling, traffic) using a Deep Neural Network. The model performs better with Filter Bank (FBank) features that include temporal context (frames before/after) and are standardized to handle varying volume levels.
Data Preparation
Simulation of a high-volume sensor data table 'SENSOR_STREAM' with varying recording locations.
Copied!
| 1 | |
| 2 | DATA casuser.sensor_stream; |
| 3 | LENGTH sensor_loc $20 raw_sound $1000; |
| 4 | DO i=1 to 1000; |
| 5 | sensor_loc=cats('LOC_', i); |
| 6 | raw_sound='<simulated_noise_bytes>'; |
| 7 | OUTPUT; |
| 8 | END; |
| 9 | |
| 10 | RUN; |
| 11 |
Étapes de réalisation
1
Computing FBank features with Log-Filterbank values, adding 5 frames of context on each side, and applying standardization.
Copied!
| 1 | PROC CAS; |
| 2 | audio.computeFeatures / |
| 3 | TABLE={name='sensor_stream', caslib='casuser'} |
| 4 | audioColumn='raw_sound' |
| 5 | fbankOptions={useLogFbank=true, usePower=true} |
| 6 | nContextFrames=5 |
| 7 | featureScalingMethod='STANDARDIZATION' |
| 8 | casOut={name='noise_features', caslib='casuser', replace=true}; |
| 9 | RUN; |
Expected Result
The 'noise_features' table contains FBank features where each frame vector includes data from the 5 preceding and 5 succeeding frames. Values are standardized (mean 0, variance 1) across frames.