⚡ electrical engineering

Non-negative tensor factorization-based dependence map analysis for local damage detection in presence of non-Gaussian noise

This paper proposes a novel method for local damage detection in rolling element bearings by using non-negative tensor factorization (NTF) to analyze dependence maps, enabling effective informative frequency band selection even in the presence of non-Gaussian noise and complex spectral structures.

Original authors: Anna Michalak, Justyna Hebda-Sobkowicz, Anil Kumar, Radoslaw Zimroz, Rafal Zdunek, Agnieszka Wylomanska

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anna Michalak, Justyna Hebda-Sobkowicz, Anil Kumar, Radoslaw Zimroz, Rafal Zdunek, Agnieszka Wylomanska

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a single, rhythmic heartbeat in the middle of a heavy metal concert. The drummer is hitting the cymbals randomly and loudly (that’s the non-Gaussian noise), and the person with the heartbeat is only tapping their chest once every few seconds (that’s the local damage in the bearing).

If you just listen to the whole room, you’ll never hear the heartbeat. You might mistake a random drum hit for a heartbeat, or the noise might be so loud that the heartbeat disappears entirely.

This paper presents a new way to "tune in" to that specific heartbeat so engineers can fix a machine before it explodes. Here is how they do it, broken down into simple steps.

1. The Problem: The "Clumsy" Listeners

Most current tools used to monitor machines are like people who only look for loudness. They say, "Hey, I heard a loud bang! That must be the problem!"

But in a heavy industrial machine (like a copper ore crusher), things bang loudly all the time just because rocks are falling. These "fake" bangs confuse the old tools. The old tools can't tell the difference between a random rock hitting the machine and the tiny, rhythmic "click" of a cracked bearing.

2. The Solution: The "Similarity Detective" (Dependence Maps)

Instead of just looking for loud sounds, the researchers decided to look for patterns of similarity.

Imagine you have 30 different microphones placed around the concert hall. Instead of asking, "How loud is it at Microphone A?", the researchers ask, "Does the sound at Microphone A happen at the exact same time and in the same way as the sound at Microphone B?"

If a random rock hits the machine, it’s a one-time event—it won't show up as a consistent pattern across all the microphones. But the damage in a bearing is rhythmic; it happens over and over again. By comparing how different frequencies "dance" together, they create a Dependence Map. This map ignores the random, chaotic noise and highlights the parts of the sound that are "dancing in sync."

3. The Secret Sauce: The "Digital Sorter" (NTF)

Once they have these maps, they have a massive pile of data. To make sense of it, they use something called Non-negative Tensor Factorization (NTF).

Think of NTF as a highly advanced color sorter. Imagine you have a giant bucket filled with a mixture of red, blue, and green marbles, but they are all coated in gray mud. NTF is a machine that can reach into that muddy bucket and pull out the pure red marbles, the pure blue marbles, and the pure green marbles separately.

In this paper, the "mud" is the heavy industrial noise, and the "marbles" are different types of information:

Marble 1: The rhythmic heartbeat of the damaged bearing (the Signal of Interest).
Marble 2: The random, loud bangs of the rocks (the Noise).
Marble 3: The general hum of the machine.

Because the researchers used a specific mathematical setting (called $\beta = -1$ ), their "sorter" is extra good at ignoring the "muddy" outliers and finding the pure, rhythmic signal.

4. The Result: Finding the Needle in the Haystack

By using this "Similarity Detective" and "Digital Sorter" combo, the researchers were able to:

Filter out the chaos: They stripped away the loud, random bangs.
Isolate the rhythm: They found the exact frequency where the tiny "click" of the damage was hiding.
Predict failure: They could see the damage clearly, even when it was almost invisible to traditional methods.

In short: They stopped listening for loudness and started listening for repetition and similarity, allowing them to hear the "whisper" of a failing machine through the "roar" of an industrial factory.

Technical Summary: Non-negative Tensor Factorization-based Dependence Map Analysis for Local Damage Detection

1. Problem Statement

The paper addresses the challenge of detecting local damage (e.g., defects in specific regions of bearing races or rolling elements) in rotating machinery. This task is exceptionally difficult in industrial environments due to several factors:

Weak Signal of Interest (SOI): Early-stage damage produces weak, cyclic, and impulsive signals.
Non-Gaussian Noise: Industrial settings (like copper ore crushers) generate heavy-tailed, impulsive, and non-stationary noise caused by fluctuating loads and random impacts (e.g., falling ore).
Spectral Overlap: Fault signatures are often masked or confused with non-cyclic impulsive disturbances and high-energy resonances.
Limitations of Existing Methods: Traditional selectors like Spectral Kurtosis are highly sensitive to outliers and often mistake non-cyclic impulsive noise for actual faults. Other methods (Alpha-stable, Conditional Variance) focus on impulsiveness but fail to account for the cyclicity of the fault.

2. Methodology

The authors propose a novel, unsupervised framework that combines dependence map analysis with Non-negative Tensor Factorization (NTF). The process follows three main stages:

A. Pre-processing (Tensor Construction):

Segmentation: The raw vibration signal is divided into $M$ segments.
Time-Frequency Representation: A Short-Time Fourier Transform (STFT) is applied to each segment to generate a spectrogram.
Dependence Map Construction: For each segment, the Pearson correlation coefficient is calculated between all pairs of frequency bins. This creates a symmetric Correlation Map (CM), which represents the auto-similarity of spectral content.
Tensorization: The $M$ correlation maps are stacked into a 3rd-order tensor $\mathcal{C} \in \mathbb{R}^{F \times F \times M}$ , where the first two dimensions represent frequency pairs and the third represents the temporal segments.

B. Analysis (NTF Decomposition):
The tensor is decomposed using NTF into three non-negative component matrices ( $W, H, Q$ ). The authors utilize the $\beta$ -divergence objective function, which offers flexibility in handling heavy-tailed residual errors.

They specifically identify that $\beta = -1$ (Itakura-Saito distance) is optimal for separating the SOI from non-Gaussian noise.
The decomposition separates the data into distinct "classes" (e.g., damage, non-cyclic pulses, noise).

C. Decision (Band Selection & Detection):

IFB Extraction: One of the factor matrices ( $H$ ) is used to extract the Informative Frequency Band (IFB).
Filtering: The raw signal is filtered using the extracted IFB characteristic.
Damage Assessment: The effectiveness is measured using the Envelope Spectrum Based Indicator (ENVSI), which quantifies the energy of the fault frequency and its harmonics relative to the total envelope spectrum energy.

3. Key Contributions

Novel Data Structure: Instead of factorizing energy distributions (like standard NMF on spectrograms), this method factorizes dependence maps (similarity structures), which are more robust to amplitude fluctuations.
Robustness to Non-Gaussianity: By using $\beta$ -divergence (specifically $\beta = -1$ ), the method effectively distinguishes between periodic fault impulses and random, high-amplitude impulsive noise.
Unsupervised & Blind: The method requires no prior knowledge of fault frequencies or machine operating conditions, making it highly adaptable to non-stationary industrial environments.

4. Results

The method was validated using synthetic signals and real-world data (a copper ore crusher and a laboratory test rig).

Simulated Data: The NTF approach achieved ENVSI values $> 0.98$ , significantly outperforming Spectral Kurtosis, Alpha-selectors, and Pearson-based selectors, which failed to identify the correct band or were overwhelmed by non-cyclic noise.
Real-World Case 1 (Crusher): In a highly noisy environment where the SOI was completely masked, the NTF ( $\beta = -1$ ) achieved an ENVSI of 0.678, whereas traditional methods yielded values near zero ( $< 0.0012$ ).
Real-World Case 2 (Test Rig): The method successfully identified the outer race damage in an acoustic signal, with results comparable to or better than specialized methods like CFFsgram.
Efficiency: While $\beta = -1$ is computationally more intensive than $\beta = 0$ or $1$, it is the only setting that consistently provides high accuracy in the presence of heavy-tailed noise.

5. Significance

This research provides a powerful tool for predictive maintenance in harsh industrial sectors (mining, crushing, etc.). By enabling the detection of local damage at an early stage—even when the signal is buried in non-Gaussian noise—the method allows for planned repairs, thereby preventing catastrophic machine failures and reducing unplanned production downtime.