Signals: Wavelet Transform Basics

Understanding how to analyze signals that change over time is fundamental in engineering, from diagnosing machine vibrations to processing medical images. While the Fourier Transform is a powerful tool, it struggles with signals containing sudden bursts or localized features. The Wavelet Transform addresses this by decomposing a signal into scaled and shifted versions of a base function, providing a multi-resolution analysis that reveals both frequency content and when those frequencies occur. This makes it indispensable for analyzing transient phenomena, compressing data, and cleaning noisy measurements.

From Fourier's Limitation to Wavelet's Solution

The classical Fourier Transform represents a signal as a sum of infinite sine and cosine waves. This is perfect for identifying all the frequencies present in a stationary signal—one whose statistical properties don't change over time. However, it provides no information about when a specific frequency component occurs. For a signal with a brief spike or a changing rhythm, the Fourier spectrum becomes smeared and difficult to interpret.

The wavelet transform solves this by using a localized, finite-duration function called the mother wavelet. Unlike a sine wave that oscillates forever, a mother wavelet is a small wave that starts at zero, oscillates, and decays back to zero. Think of it as a mathematical "microscope" that you can move along the signal (translation) and zoom in or out (dilation or scaling). By correlating the signal with this scaled and shifted wavelet, you build a time-scale map showing how well the signal matches the wavelet's shape at different locations and resolutions.

Continuous and Discrete Wavelet Transforms

The Continuous Wavelet Transform (CWT) is defined for a continuous signal $x(t)$ using a mother wavelet $\psi (t)$. The CWT coefficient at a scale $a$ (which is inversely related to frequency) and translation $b$ is calculated as:

$$CWT(a,b)=\frac{1}{\sqrt{|a|}}{\int }_{-\infty }^{\infty }x(t){\psi }^{\ast }\left(\frac{t-b}{a}\right)dt$$

Here, $a$ controls the stretch: a large $a$ corresponds to a stretched wavelet (low frequency, broad view), and a small $a$ corresponds to a compressed wavelet (high frequency, detailed view). The term $1/\sqrt{|a|}$ ensures energy normalization across scales. The CWT produces a dense, highly redundant representation useful for detailed time-scale analysis.

For practical computation and applications like compression, the Discrete Wavelet Transform (DWT) is used. The DWT discretizes the scale and translation parameters in a dyadic (powers of two) grid: $a=2^j$ and $b=k\cdot 2^j$, where $j$ and $k$ are integers. This critical sampling leads to an efficient, non-redundant decomposition implemented using a filter bank—a series of high-pass and low-pass filters.

Multi-Resolution Analysis and Filter Bank Implementation

Multi-resolution analysis (MRA) is the conceptual framework behind the DWT. It views a signal at different resolution levels. At each level, the signal is split into two components: approximations and details. The approximations are the low-frequency, broad-stroke components of the signal. The details are the high-frequency, fine-detail components.

This is implemented using a filter bank tree. Imagine you have a discrete signal. You pass it through two complementary filters:

1. A low-pass filter outputs the approximation coefficients (A1).
2. A high-pass filter outputs the detail coefficients (D1).

You then take the approximation coefficients (A1) and feed them into another identical pair of filters, creating a second level of approximations (A2) and details (D2). This process can be repeated, decomposing the signal into a hierarchy of increasingly coarse approximations and their associated details. This tree structure allows you to analyze the signal's features at the resolution most appropriate for them.

Wavelet vs. Fourier Analysis for Transient Detection

The choice between wavelet and Fourier analysis hinges on the signal's characteristics. The Fourier Transform (and its windowed version, the Short-Time Fourier Transform or STFT) is ideal for identifying precise frequency components in steady-state signals, like the harmonic tones in an audio recording of a sustained musical note.

However, for transient signal detection—locating a fault spike in a power line or a heart murmur in an ECG—wavelets are superior. The STFT uses a fixed-duration window, creating a trade-off: a short window gives good time resolution but poor frequency resolution, and vice-versa. Wavelets adapt their window size naturally: they use short windows (high scale) for high frequencies to pin down transient events in time, and long windows (low scale) for low frequencies to accurately determine the frequency. This adaptive resolution makes wavelets exceptionally good at isolating and characterizing short-lived events without smearing.

Application to Signal Denoising

A prime application of wavelet analysis is signal denoising. The core idea is that noise is often concentrated in the high-frequency detail coefficients, while the true signal's energy is concentrated in a few large coefficients across scales.

The denoising workflow follows these steps:

1. Decomposition: Choose a suitable mother wavelet (e.g., Daubechies) and decomposition level, and perform a DWT on the noisy signal.
2. Thresholding: Apply a threshold to the detail coefficients at each level. Small coefficients (likely noise) are set to zero or shrunk. Common thresholding rules are "hard" (keep or zero) and "soft" (shrink toward zero).
3. Reconstruction: Perform the inverse DWT using the original approximation coefficients and the modified (thresholded) detail coefficients to reconstruct the denoised signal.

This method is often more effective than standard frequency-domain filtering because it can remove noise while preserving sharp edges and discontinuities in the signal, which linear filters tend to blur.

Common Pitfalls

Selecting an Inappropriate Mother Wavelet: The mother wavelet must match the signal's features. Using a smooth wavelet to analyze a signal with sharp jumps will perform poorly. Always visualize the wavelet and consider its properties (symmetry, vanishing moments) in relation to your signal.

Misinterpreting the Scale Axis: In a wavelet scalogram (a plot of CWT coefficients), remember that scale is inversely related to frequency: low scale = high frequency, high scale = low frequency. Confusing this relationship leads to incorrect conclusions about the signal's frequency content.

Over-Decomposing with the DWT: When using a filter bank, each level of decomposition halves the number of samples. Decomposing too many times can leave you with too few approximation coefficients to represent the signal meaningfully, especially for short signals. A good rule is not to exceed ${\log }_2(N)$ levels for a signal of length $N$.

Treating Thresholding as a Black Box: Simply applying a default threshold can remove important signal details. The choice of threshold rule (hard vs. soft) and threshold value (universal, SURE, minimax) should be informed by the nature of the noise and the signal.

Summary

• The Wavelet Transform decomposes a signal using scalable and translatable mother wavelets, providing a joint time-scale representation that is superior to Fourier methods for analyzing non-stationary and transient signals.
• The Continuous Wavelet Transform (CWT) offers a detailed, redundant analysis, while the Discrete Wavelet Transform (DWT) uses a dyadic grid and filter banks to achieve an efficient, non-redundant multi-resolution analysis.
• Unlike the fixed-resolution Short-Time Fourier Transform, wavelets provide adaptive resolution—short windows for high frequencies and long windows for low frequencies—making them exceptionally powerful for transient signal detection.
• A key application is signal denoising, which involves decomposing a signal, thresholding the detail coefficients to suppress noise, and reconstructing a cleaner signal.
• Successful application requires careful selection of the mother wavelet, correct interpretation of the scale-frequency relationship, and thoughtful application of decomposition levels and thresholding parameters.