S-Transform and HHT: A Comparison of Two Time-Frequency Analysis and Localization Approaches

Nov 07, 2026

Leave a message

When you're choosing between the S-transform and HHT, the real question isn't which one sounds more advanced. It's whether the box you're building can handle it in the field. Both methods can locate disturbances in time and frequency. But they make different assumptions about the signal, cost different amounts of hardware, and fail in different ways. That difference decides whether you get usable start and stop times or a pile of results that still need a human to sort out.

How can the accuracy of fault location devices for low-current grounding faults exceed 99%?

The S-transform comes from combining the short-time Fourier transform and the wavelet transform. Its Gaussian window gets narrower as frequency goes up and wider as frequency goes down. So it gives good frequency resolution at low frequencies and naturally tightens time resolution at high frequencies. One practical plus: the output is stable. Feed the same disturbance in again, and the time-frequency matrix looks basically the same. Engineers don't have to relearn how the algorithm behaves every time. For pulling out the start and end of a single-frequency event like a voltage sag or swell, studies keep showing good accuracy and noise tolerance. The cost is computation. A full S-transform throws off a lot of redundant time-frequency data. Run the full version on an embedded device and response time can stretch out fast. That's why real products often use a fast S-transform or an "incomplete" version that keeps only key frequency points. You trade some information for real-time performance, and you keep that trade controlled.

HHT takes another road. It first uses empirical mode decomposition to break the signal into intrinsic mode functions, then runs a Hilbert transform on each one. It doesn't preset a basis function. That adaptability is its main selling point. When the field signal has mixed frequency content or more than one disturbance type, HHT doesn't need you to decide upfront how wide the window should be or which basis to pick. That can be attractive for mixed disturbances. The trouble is exactly in that "adaptive" part. EMD's output depends heavily on the signal shape and noise level. Mode mixing and endpoint effects aren't rare glitches; they're built-in tendencies of the method. Once the IMFs mix modes, the physical meaning of the instantaneous amplitude and frequency gets fuzzy, and the localization results become less consistent. Common fixes include noise-assisted decomposition, swapping EMD for VMD, or using only the first few IMFs for a local call. But those are patches on the method, not a removal of its limits.

Hardware and real-time behavior make the trade-off more concrete. The S-transform is heavy, but its parallel structure is fairly regular, so it fits fixed-point acceleration on an FPGA or DSP, and fast variants have mature hardware paths. HHT may not look heavier on paper, but EMD is an iterative sifting process. The number of iterations and decomposition levels moves with the signal, so real-time behavior doesn't come with a stable upper bound. For a power quality monitor that has to promise a certain response delay, that uncertainty is harder to deal with than simply being slow.

Long-Distance Overhead Lines in Rural Distribution Networks: Challenges in Deploying Small-Current Grounding Fault Location Devices

So the selection logic comes down to this. If your main objectives are voltage sags, swells, and breaches-single or quasi-single-frequency occurrences-and precision in locating and repeatability is paramount, and the equipment is capable of performing a speedy S-transform or main frequency extraction, the S-transform is the more reliable alternative. If one is faced with combined frequency content and disturbance types that are not entirely understood ahead of time, the flexibility offered by HHT can be used. However, one must accept that noise and mode mixing may obfuscate the limits and it would be worth putting decomposition quality safeguards or fusion approach in place. No method wins across every disturbance type and every hardware limit. The engineering call is to figure out what you're actually trying to locate first, then match the algorithm to that job-not the other way around.

Send Inquiry