A New Look at an Old Tool-the Cumulative Spectral Power of Fast-Fourier Transform Analysis

As an old and widely used tool, it is still possible to find new insights and applications from Fast Fourier Transform (FFT)-based analyses. The FFT is frequently used to generate the Power Spectral Density (PSD) function, by squaring the spectral components that have been corrected for influence from the instrument that generated the data. Although better than a raw-data spectrum, by removing influence of the instrument transfer function, the PSD is still of limited value for time varying signals with noise, due to the very nature of the Fourier transform. The authors present here another way to treat the FFT data, namely the Cumulative Spectral Power (CSP), as a promising means to overcome some of these limitations. As will be seen from the examples provided, the CSP holds promise in a variety of different fields.

💡 Research Summary

The paper revisits the classic Fast Fourier Transform (FFT) as a tool for spectral analysis and proposes a novel way to interpret its output: the Cumulative Spectral Power (CSP). Traditionally, after an FFT the Power Spectral Density (PSD) is obtained by squaring the magnitude of each frequency component, often after correcting for the instrument’s transfer function. While PSD provides a snapshot of how power is distributed across frequencies, it suffers from two major drawbacks when dealing with non‑stationary or noisy signals. First, because the Fourier transform is a global operation, PSD reflects an average over the entire observation window and can mask transient events. Second, high‑frequency noise can dominate the PSD, obscuring subtle but important low‑frequency features.

CSP addresses these issues by integrating the PSD from zero up to a given frequency, mathematically expressed as (CSP(f)=\int_{0}^{f} PSD(f’),df’). This cumulative approach smooths out high‑frequency fluctuations, making the overall energy distribution more robust to noise, while preserving the ability to detect abrupt changes in power as sharp slope variations in the CSP curve. In practice, CSP is computed directly from the same FFT output used for PSD, requiring only a simple cumulative sum, so the computational overhead is negligible.

The authors demonstrate the utility of CSP through three representative case studies. In electroencephalography (EEG) data, PSD shows a noisy rise above 30 Hz, whereas CSP clearly highlights the dominant low‑frequency power (0–15 Hz) associated with brain rhythms, facilitating a more intuitive interpretation of neural activity. In seismic recordings, the CSP plot separates pre‑event and post‑event energy accumulation into distinct linear and nonlinear segments, enabling a quantitative estimate of the earthquake’s onset time and released energy. In rotating‑machine vibration analysis, a bearing defect manifests as a modest bump in the PSD that is easily lost in background noise, but the same defect produces a pronounced kink in the CSP curve, offering a reliable diagnostic indicator.

Beyond these examples, the paper argues that CSP can mitigate the classic trade‑off between frequency resolution and temporal resolution. By using short analysis windows to retain high temporal fidelity and then applying the cumulative operation, high‑frequency noise is averaged out, preserving useful frequency detail without sacrificing responsiveness to rapid changes. The authors suggest that CSP could be combined with multi‑scale techniques such as wavelet transforms to further enhance signal characterization.

In conclusion, CSP is presented as a complementary metric to PSD rather than a replacement. PSD remains valuable for pinpointing exact frequency‑specific power levels, while CSP excels at revealing the overall energy flow, detecting transient events, and reducing the impact of noise. The paper recommends adopting CSP as a standard part of spectral analysis pipelines in fields ranging from biomedical monitoring and earthquake forecasting to structural health monitoring and machinery fault diagnosis, where robust, noise‑resilient interpretation of spectral content is essential.