Feature representation is where signal knowledge meets ML

For acoustic and sensor ML, the choice of representation affects robustness, latency, and interpretability. A common pattern is: raw signal → windowing → time–frequency transform → model.

Common representations (and when they help)

• STFT / spectrograms: strong default for many acoustic tasks; good balance of information and compute
• Mel spectrograms: compact; often effective when frequency resolution requirements are moderate
• Wavelets: useful when multi-scale structure matters or transient events dominate
• Learned embeddings: can boost performance when pretraining aligns with your domain — but require careful validation

Industrial constraints to include early

• Latency: how quickly must a decision be produced?
• Compute: edge device vs. cloud; power and memory limits
• Sensor mismatch: new microphones/sensors and mounting changes are normal in real deployments
• Noise & drift: new machines, materials, and environmental conditions change distributions

A good feasibility plan tests representation choices against these constraints, not only against a benchmark metric.