---
kind: concept
slug: hilbert-huang-transform
title: Hilbert-Huang transform
summary: >-
  Turns a window of a 1-D signal into a time-frequency image from its first intrinsic mode functions. One card on file: 32x32 images of shaft-mounted bearing vibration feed a CNN. The paper gives few HHT details, so the image cannot be rebuilt from the text.
tags: [signal-processing, time-frequency, bearing-fault-diagnosis]
papers: ["2112.08673"]
related: [convolutional-neural-network, multimodal-fusion, wavelet-transform]
status: draft
updated_at: 2026-11-22
---

## Overview

The Hilbert-Huang transform (HHT) is used here as a way to turn a window of a
1-D sensor signal into a 2-D time-frequency picture. The paper applies it to
linear acceleration, keeps only the first three intrinsic mode functions
(IMFs), and renders the result as a 32x32 image [arxiv:2112.08673 §2.1.1].
The paper calls the result "time-frequency images" and feeds them to a CNN
[arxiv:2112.08673 §2].

The paper relies on the original HHT reference, which it cites as ref. 36
[arxiv:2112.08673 §2.1.1] [§4]. That reference is not on file, so this page
does not restate its definitions. Only the choices the card's paper reports
are described here.

The wiki holds one card for this concept. It is an application to bearing
fault diagnosis, not a study of the transform, so no variant comparison is
possible yet.

## Variants

### Windowed HHT images for a CNN (bearing vibration)

Each window is 3,897 samples long, and consecutive windows start 1,559 samples
apart [arxiv:2112.08673 §2.1]. The HHT is applied to the linear-acceleration
window, using the first three IMFs. The result is a 32x32 image with three
channels [arxiv:2112.08673 §2.1.1] [§3.2]. The image is one input of a hybrid
model. The other input is two spectral band powers, N1 and N2, from the
angular-acceleration signal, which is not an HHT feature [arxiv:2112.08673 §2.1.2].

The authors report 98% test accuracy for the CNN branch on these images
alone, and 99.6% for the hybrid. These are author-reported numbers on one rig
at one shaft speed [arxiv:2112.08673 tab. 3] [§4]. The paper also says the
first three IMFs were the "most effective" of the choices it tried
[arxiv:2112.08673 §2.1.1]. The card for this variant is
[Intelligent Bearing Fault Diagnosis Method Combining Mixed Input and Hybrid CNN-MLP model](/papers/2112.08673).

## When to use which

There is one variant on file, so the table compares what the card supports
against the constraints an implementer is likely to meet.

| Situation | Use | Why |
|---|---|---|
| Sensor on a rotating shaft; single known speed; a CNN is available | The windowed HHT image in the variant above, with the same window and IMF count [arxiv:2112.08673 §2.1] [§2.1.1] | This is the only setup on file with reported accuracy. The paper reports no other speed [§4]. |
| Need a reproducible image from the text alone | Not yet: fix the EMD stop rule, envelope method, grid mapping and colour map yourself, and report them | The paper does not state these choices [arxiv:2112.08673 §2.1.1]. Each one changes the image. |
| Want a non-HHT time-frequency input | See [wavelet-transform](/concepts/wavelet-transform) | Not compared in this paper. |

## Implementation sketch

Only steps the card's paper states are included. Items the paper leaves open
are marked "not stated".

```pseudocode
# HHT image of one window  [§2.1] [§2.1.1]
# Input:  linear: 1-D window of 3,897 samples  [§2.1]
# Output: image, 32 x 32 with three channels  [§2.1.1] [§3.2]

def hht_image(window, num_imfs=3, size=32):
    # Decompose the window into intrinsic mode functions  [§2.1.1]
    # The decomposition method and its stop rule are not stated.
    imfs = decompose_into_imfs(window)
    # Keep the first three IMFs only  [§2.1.1]
    kept = imfs[:num_imfs]
    # Map the time-frequency content to a size x size grid  [§2.1.1]
    # Bin counts, envelope method and colour map are not stated.
    return render_time_frequency(kept, size)
```

## Contradictions

None known. The wiki holds one card on this concept, so there is no second
paper to disagree with.

## Papers

- [Intelligent Bearing Fault Diagnosis Method Combining Mixed Input and Hybrid CNN-MLP model](/papers/2112.08673): converts linear acceleration windows to 32x32 HHT images for a CNN; reports 99.6% for the hybrid, 98% for the CNN branch alone, on one shaft speed.
