Introduction

Notch filters are commonly used to remove narrowband interference from signals.

Typical applications include removing mains hum, switching noise, or mechanical vibration tones.

In theory, designing a notch filter is straightforward. Once the interference frequency is known, a filter can be placed precisely at that frequency.

However, in real engineering systems notch filters often perform poorly.

The filter may fail to remove the interference, distort nearby signals, or even introduce numerical instability.

Understanding why requires examining how real signals differ from textbook assumptions.


Textbook Expectation

Most DSP textbooks assume ideal conditions.

The interference tone is assumed to have:

  • a fixed frequency
  • high signal-to-noise ratio
  • constant amplitude

Under these assumptions, a narrowband notch filter can eliminate the tone with minimal impact on the rest of the signal.


Real Signals Are Not Stationary

In practical systems, interference sources rarely remain perfectly stationary.

Frequency drift is common.

For example, a switching converter may produce interference that moves slightly over time:

1000 Hz → 1002 Hz → 998 Hz

Temperature changes, supply variation, and oscillator instability all contribute to this drift.

When the notch filter is too narrow, even small frequency shifts allow the tone to escape the notch.


Spectral Peak Detection Problems

Before designing a notch filter, engineers must estimate the interference frequency.

This is usually done by identifying peaks in the power spectral density.

However, PSD peaks are not always reliable indicators of true tones.

Noise fluctuations, spectral leakage, and estimator variance can produce misleading peaks.

If the wrong frequency is chosen, the notch filter will target the wrong location.


High-Q Notch Instability

Narrow notches correspond to filters with poles close to the unit circle.

While this improves frequency selectivity, it also increases sensitivity to coefficient errors.

In fixed-point implementations, small quantization errors can significantly shift the filter response.

High-Q filters may also exhibit limit cycles or stability issues in embedded systems.


Drift vs Bandwidth Tradeoff

Engineers face an unavoidable trade-off.

A very narrow notch preserves nearby signal content but fails when interference frequency drifts.

A wider notch tolerates drift but removes more useful signal energy.

Choosing the correct bandwidth requires understanding both signal characteristics and system constraints.


Why Trial-and-Error DSP Fails

Many DSP pipelines rely on manual trial and error.

Engineers inspect the spectrum, guess a notch frequency, and test the filter repeatedly.

This approach is time-consuming and often produces inconsistent results.

Small changes in signal conditions can invalidate the chosen filter parameters.


Deterministic DSP Pipelines

A more reliable approach uses deterministic signal characterization.

Instead of guessing filter parameters, the pipeline performs:

  1. spectral characterization
  2. tone detection
  3. filter synthesis
  4. quantitative verification

This structured process ensures repeatable results.


Engineering Constraints

Real filter design must consider multiple constraints.

These include:

  • acceptable signal distortion
  • tolerance to frequency drift
  • numerical stability
  • computational cost

Balancing these constraints requires more than simply selecting a notch frequency.


Conclusion

Notch filters do not fail because the underlying mathematics is incorrect.

They fail because real signals violate the assumptions used during filter design.

Understanding signal drift, spectral estimation uncertainty, and numerical implementation constraints allows engineers to design more reliable filtering systems.


  • How to Detect Tonal Interference in Real-World Signals
  • Why PSD Peak Detection Fails in Low-SNR Signals
  • Spectral Leakage Explained for Real Engineering Signals