💻 computer science

Evolutionary Kernel-based T-Spherical Fuzzy C-Means Clustering, application to EMG signal

This paper proposes the Evolutionary Kernel-based T-Spherical Fuzzy C-Means (EKTSFCM) algorithm, which integrates T-Spherical Fuzzy Sets and a dual-kernel distance metric optimized by evolutionary techniques to achieve superior clustering accuracy and noise robustness in complex datasets like EMG signals.

Original authors: Seyyed Hassan Ziafati Bagherzadeh, Seyed Ehsan Razavi, Gelareh Veisi

Published 2026-08-05

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Seyyed Hassan Ziafati Bagherzadeh, Seyed Ehsan Razavi, Gelareh Veisi

Original paper licensed under CC BY 4.0 (https://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 sort a giant, messy box of LEGOs. Some pieces are clearly red, some are clearly blue, but many are purple, or half-red/half-blue, or so dusty and broken that you aren't even sure if they belong in the box at all. In the world of data science, this is the daily struggle of "clustering." It's the art of teaching computers to find hidden patterns and group similar things together without a teacher telling them what to look for. This is crucial for everything from organizing photos to diagnosing diseases. But real-world data is messy. It's full of noise (static), outliers (weird glitches), and "gray areas" where a piece of data doesn't fit neatly into one category. Traditional methods often force these messy pieces into a box they don't quite belong in, or they get confused by the static and give up. To solve this, scientists use "fuzzy logic," which lets a piece of data belong to multiple groups at once with different strengths, rather than forcing a strict "yes" or "no" answer.

Now, enter the researchers from Islamic Azad University, who have built a new, super-smart sorting machine called EKTSFCM. Think of this as a next-generation LEGO sorter that doesn't just ask, "Is this piece red or blue?" Instead, it asks a much more detailed set of questions: "How much does it look like red? How much does it not look like red? Is it just too dusty to tell (hesitation)? Or is it so weird and broken that it should be thrown in the trash (refusal)?" By adding these extra layers of questioning, the machine can handle the messiest, noisiest data with incredible grace. They tested this new sorter on EMG signals—which are essentially the electrical whispers of your muscles when you move. Since muscle signals are notoriously noisy and overlap with each other, they are the perfect "stress test" for a new sorting algorithm. The researchers found that their new method, which uses a special mix of mathematical tools and a "genetic algorithm" (a computer program that evolves its own best settings like nature does), sorted the muscle signals much better than any previous method.

The Problem: The Messy Muscle Signal

Imagine you are trying to recognize hand gestures using a wristband that reads your muscle electricity. When you make a fist, your muscles send a specific electrical signal. When you open your hand, they send another. But in the real world, these signals are never clean. There is static from the air, interference from your phone, and sometimes your muscles just get confused during the transition between movements.

Older sorting algorithms (like the standard Fuzzy C-Means) are like a strict librarian who insists every book must be on exactly one shelf. If a book is half-fiction and half-mystery, the librarian has to guess which shelf it goes on, often making a mistake. Even "Intuitionistic" versions, which admit some uncertainty, still struggle because they can't distinguish between "I'm not sure" (hesitation) and "This doesn't belong here at all" (refusal). They also get easily confused by the complex, non-linear shapes of real data, much like trying to draw a straight line through a cloud.

The Solution: A Four-Dimensional Detective

The authors propose a new framework called Evolutionary Kernel T-Spherical Fuzzy C-Means (EKTSFCM). Let's break down the three cool gadgets they put inside this machine:

1. The T-Spherical Fuzzy Detective (TSFS)
Instead of just asking "Is it red?" or "Is it not red?", this detective asks four questions for every single data point:

Membership: How much does it belong here?
Non-membership: How much does it not belong here?
Hesitation: Is it just too fuzzy to tell?
Refusal: Is this data point so weird or noisy that it refuses to join any group?

This "refusal" degree is the game-changer. In previous methods, a noisy, garbage data point was forced to join a group, dragging the whole group's average down. In EKTSFCM, the algorithm can say, "This point is too messy; I'm going to refuse to assign it to any cluster." This keeps the groups clean and accurate.

2. The Dual-Kernel Lens
To see the patterns clearly, the algorithm uses two different "lenses" (kernels) at the same time.

The Gaussian Lens is smooth and gentle, good for seeing general shapes.
The Bell-Shaped Lens is sharper; it stays high and steady for a while before dropping off quickly.
By combining them, the algorithm can see both the smooth curves of the data and the sharp boundaries where one gesture ends and another begins. It's like using both a wide-angle lens and a zoom lens simultaneously to get the perfect picture.

3. The Evolutionary Tuner
The machine has a lot of knobs and dials to turn (like how fuzzy the groups should be, how sharp the lenses are, etc.). If you turn them by hand, you might get it wrong. So, the authors added a Genetic Algorithm. Think of this as a digital evolution lab. The computer creates 50 different versions of the machine, each with random settings. It tests them all, keeps the best ones, mixes their settings together (crossover), and adds a little random change (mutation). Over 150 generations, the machine "evolves" to find the perfect combination of settings for the specific data it is looking at.

What They Found

The team tested this new machine on three different sets of muscle signal data, including the famous NinaPro DB1 (with 27 people doing 52 different hand movements) and the high-resolution CapgMyo DB-a (with 128 sensors).

The results were impressive. The new EKTSFCM method consistently outperformed the old standard methods.

On the NinaPro DB1 dataset, it achieved a clustering accuracy of 0.745 (compared to 0.612 for the standard method).
On the CapgMyo DB-a dataset, it hit 0.818 accuracy (beating the next best method, which was 0.795).
It also proved to be much more robust against noise. When the researchers added static to the signals (simulating a bad connection), EKTSFCM held its ground much better than the others.

The researchers also ran a "taste test" (ablation study) to see which part of the machine was doing the heavy lifting. They found that removing any single part—like taking away the "refusal" feature or the "dual-kernel" lens—made the machine worse. This suggests that the magic isn't in just one trick, but in how all the parts work together.

Why It Matters

This isn't just a math puzzle; it's a step toward better human-computer interaction. If we can sort muscle signals more accurately, even when they are noisy or the user is moving strangely, we can build better prosthetic arms, more responsive video game controllers, and medical devices that can detect muscle issues earlier. The authors suggest that by understanding why a signal was refused or hesitated, we can make systems that are safer and more reliable, knowing exactly when to trust the data and when to say, "I'm not sure, let's try again."

In short, the paper shows that by giving computers a more nuanced way to handle uncertainty—allowing them to say "no" to bad data and "maybe" to confusing data—we can solve some of the messiest sorting problems in science.

Technical Summary: Evolutionary Kernel-based T-Spherical Fuzzy C-Means Clustering (EKTSFCM)

Problem Statement
Clustering complex, real-world datasets characterized by highly overlapping classes, severe noise, and outliers remains a significant challenge. Traditional Fuzzy C-Means (FCM) algorithms rely on single-valued membership degrees and Euclidean distance, which fail to capture deep uncertainty in ambiguous transitional states, handle nonlinear feature interactions, or distinguish between genuine ambiguity and noise-induced uncertainty. While Intuitionistic Fuzzy C-Means (IFCM) introduces non-membership and hesitation degrees, it remains constrained by rigid linear restrictions between membership and non-membership and lacks the capacity to model "refusal" or non-participation independently. Furthermore, existing kernel-based methods often rely on single kernel types (e.g., Gaussian) that may not fully exploit the structural dynamics of dynamic data streams, such as Electromyography (EMG) signals.

Methodology
The paper proposes the Evolutionary Kernel T-Spherical Fuzzy C-Means (EKTSFCM) framework, which integrates three core components to address these limitations:

T-Spherical Fuzzy Sets (TSFS) Framework:
The method extends fuzzy clustering by embedding TSFS, which models four orthogonal degrees of uncertainty for each data point: membership ( $\mu$ ), non-membership ( $\nu$ ), hesitation ( $\eta$ ), and refusal ( $r$ ). These are subject to a relaxed non-linear constraint: $\mu^\psi + \nu^\psi + \eta^\psi + r^\psi = 1$ .
- Refusal Degree: A novel contribution where the refusal degree explicitly quantifies the incompatibility of a sample with the discovered cluster structure (e.g., outliers or noise), rather than forcing assignment.
- Augmented Degrees: Hesitation and refusal are integrated into the objective function via augmented membership ( $\mu^*$ ) and non-membership ( $\nu^*$ ) degrees, allowing the algorithm to penalize uncertainty and non-participation.
Dual-Kernel Distance Metric:
To handle nonlinear geometries and complex boundaries, EKTSFCM employs a hybrid distance metric combining:
- Gaussian Kernel: Models smoothly decaying dissimilarity based on Euclidean distance.
- Generalized Bell-shaped Kernel: Provides a robust similarity measure that maintains high validity over a specific parametric interval before sharply declining.
  This dual approach captures both global and local geometric structures, offering better separability for non-convex clusters than single-kernel methods.
Evolutionary Optimization:
To mitigate sensitivity to initialization and premature convergence to local minima, a Genetic Algorithm (GA) is employed to globally optimize the hyperparameter space. The GA dynamically tunes:
- TSFS parameters (fuzzification exponents, T-spherical exponent $\psi$ , Yager complement parameter $\alpha$ ).
- Dual-kernel parameters (Gaussian width $\sigma$ , Bell-kernel span $P_{span}$ , and hardness $P_{hardness}$ ).
- Objective function weights ( $\gamma_1, \gamma_2, \gamma_3$ ).

The algorithm operates via an alternating optimization scheme: the inner loop updates membership matrices and cluster centroids based on the current hyperparameters, while the outer GA loop evolves the hyperparameters to maximize clustering efficacy (minimizing the objective function).

Key Contributions

Pioneering TSFS Integration: This work presents the first formulation of T-Spherical Fuzzy Sets within an unsupervised clustering paradigm, introducing a mathematically rigorous mechanism for modeling multi-dimensional uncertainty and intrinsic outlier rejection via the refusal degree.
Hybrid Dual-Kernel Metric: The development of a novel distance formulation combining Gaussian and Generalized Bell-shaped kernels to enhance nonlinear feature separability and noise resilience in high-dimensional spaces.
Adaptive Evolutionary Strategy: The integration of a genetic algorithm to jointly optimize dual-kernel and TSFS parameters, circumventing the initialization sensitivity and local minima issues inherent in traditional fuzzy clustering.

Experimental Results
The framework was validated on three public EMG/sEMG benchmarks: NinaPro DB1, CapgMyo DB-a (high-density), and a Lightweight EMG dataset. The method was compared against FCM, IFCM, Kernel FCM (KFCM), and Evolutionary Kernel IFCM (EKIFCM).

Performance: EKTSFCM consistently outperformed all baselines across external metrics (Clustering Accuracy, Adjusted Rand Index, Normalized Mutual Information, and Macro-F1) and internal metrics (Silhouette Coefficient, Davies-Bouldin Index, Calinski-Harabasz Index).
- On NinaPro DB1, EKTSFCM achieved 74.5% accuracy, surpassing EKIFCM (72.0%) and KFCM (65.0%).
- On the high-dimensional CapgMyo DB-a, it reached 81.8% accuracy, demonstrating particular efficacy in complex feature spaces.
Robustness: Under additive white Gaussian noise (SNR levels of 5–20 dB), EKTSFCM exhibited smaller accuracy degradation compared to baseline methods, confirming its resilience to severe noise.
Convergence: The evolutionary search allowed the algorithm to converge in fewer iterations (e.g., 28 iterations on NinaPro DB1 vs. 48 for FCM) with lower final objective values.
Ablation Study: Removing components (dual-kernel, refusal degree, or GA tuning) resulted in performance degradation, confirming that the gains are due to the synergistic combination of all proposed elements.

Significance and Claims
The paper claims that EKTSFCM provides a highly adaptable, mathematically rigorous solution for complex clustering scenarios where data uncertainty exceeds the capacity of classical or intuitionistic fuzzy models. By explicitly modeling refusal, the method offers a more interpretable description of uncertainty, distinguishing between genuine ambiguity (hesitation) and signal incompatibility (refusal). This is particularly significant for neuromuscular diagnostics and myoelectric control, where distinguishing between transitional gestures, noise, and valid signals is critical. The authors assert that the framework successfully bridges the gap between advanced fuzzy logic theory and practical, robust application in high-dimensional, noisy biomedical signal analysis.

The Problem: The Messy Muscle Signal

The Solution: A Four-Dimensional Detective

What They Found

Why It Matters

Technical Summary: Evolutionary Kernel-based T-Spherical Fuzzy C-Means Clustering (EKTSFCM)

More like this