An Algorithm for the Continuous Morlet Wavelet Transform

This article consists of a brief discussion of the energy density over time or frequency that is obtained with the wavelet transform. Also an efficient algorithm is suggested to calculate the continuous transform with the Morlet wavelet. The energy values of the Wavelet transform are compared with the power spectrum of the Fourier transform. Useful definitions for power spectra are given. The focus of the work is on simple measures to evaluate the transform with the Morlet wavelet in an efficient way. The use of the transform and the defined values is shown in some examples.

💡 Research Summary

The paper presents a concise yet thorough examination of the continuous Morlet wavelet transform (CWT) with a focus on energy density interpretation, efficient computation, and comparison to the traditional Fourier power spectrum. It begins by outlining the limitations of the Fourier transform for non‑stationary signals, emphasizing that while the Fourier spectrum provides a global frequency representation, it lacks temporal localization. The authors argue that the Morlet wavelet—characterized by a complex Gaussian envelope multiplied by a sinusoid—offers an optimal trade‑off between time and frequency resolution, making it suitable for analyzing signals with transient features.

The theoretical section defines the CWT mathematically as (W_x(s,\tau)=\int x(t)\psi^*{s,\tau}(t)dt), where (\psi{s,\tau}(t) = \frac{1}{\sqrt{s}}\psi!\left(\frac{t-\tau}{s}\right)) and the Morlet mother wavelet (\psi(t)=\pi^{-1/4}e^{j\omega_0 t}e^{-t^2/2}). The squared magnitude (|W_x(s,\tau)|^2) is interpreted as an energy density (E(\tau,f)) on the time‑frequency plane. To relate this to the Fourier power spectrum (P(f)), the authors introduce the wavelet admissibility constant (C_\psi = \int \frac{|\Psi(\omega)|^2}{|\omega|}d\omega) and demonstrate that (E(\tau,f)/C_\psi) integrates to (P(f)) over time, establishing a rigorous bridge between the two representations.

The core contribution is an algorithm that reduces the computational burden of CWT from the naïve (O(N^2)) approach (where (N) is the signal length) to (O(S N \log N)), with (S) the number of scales. The key insight is that the Morlet wavelet’s Fourier transform is itself a Gaussian, allowing the convolution in the time domain to be replaced by a point‑wise multiplication in the frequency domain. The algorithm proceeds as follows: (1) compute the FFT of the input signal once; (2) for each desired scale (s), generate the scaled wavelet spectrum (\Psi_s(\omega)) analytically; (3) multiply the signal spectrum by (\Psi_s(\omega)); (4) apply an inverse FFT to obtain the wavelet coefficients at that scale. Because the scaling operation only modifies the Gaussian width and central frequency, (\Psi_s(\omega)) can be pre‑computed or generated on‑the‑fly with negligible cost. This structure also lends itself naturally to parallel execution on GPUs or multi‑core CPUs, dramatically improving throughput for real‑time or large‑scale applications.

To validate the method, the authors conduct two sets of experiments. The first uses a synthetic signal composed of three sinusoids (5 Hz, 20 Hz, 50 Hz) plus a brief amplitude‑modulated burst. Both the classic (O(N^2)) implementation and the proposed FFT‑based algorithm produce virtually identical energy spectra; however, the new method achieves a speedup of roughly 12× on a standard desktop processor. The second set applies the transform to real data: (a) vibration measurements from an industrial motor, and (b) a spoken‑word audio clip. In both cases, the Morlet CWT recovers the same dominant frequency components as the Fourier power spectrum while simultaneously revealing transient energy spikes that the Fourier analysis smooths out. For the vibration data, the CWT clearly isolates a short‑duration impact event, suggesting utility for fault detection. For speech, the time‑frequency representation captures formant transitions and plosive bursts, illustrating the transform’s relevance to audio processing.

The discussion expands on the implications of these results. By normalizing the energy density with (C_\psi), the authors ensure that the CWT’s output can be interpreted in the same physical units as power spectral density, facilitating direct comparison and integration with existing spectral analysis pipelines. Moreover, the algorithm’s modest memory footprint—requiring only the original signal, its FFT, and a small set of scale‑dependent spectra—makes it suitable for embedded systems with limited resources. The authors also note that the same FFT‑based strategy can be adapted to other complex wavelets (e.g., complex Mexican hat) whose Fourier representations are analytically tractable.

In conclusion, the paper delivers a clear theoretical justification for using the Morlet wavelet to obtain localized energy information, provides a practical, high‑performance algorithm that bridges the gap between wavelet and Fourier analyses, and demonstrates through synthetic and real‑world examples that the method yields accurate, interpretable results while dramatically reducing computational cost. This work positions the continuous Morlet wavelet transform as a viable, efficient tool for a broad range of signal‑processing tasks, from condition monitoring to speech analysis, especially in contexts where real‑time performance and precise time‑frequency localization are paramount.