Inner-Solver Initialization Noise in Metaheuristic Tuning of Mode Decomposition

Engineers who denoise vibration, motor-current, or partial-discharge signals often split each signal into a few narrow-band components using Variational Mode Decomposition (VMD). VMD is an iterative method that fits $K$ band-limited "modes," each centered on its own frequency, with a penalty $\alpha$ that controls how narrow each band is. Choosing $K$ and $\alpha$ by hand works poorly. A large literature therefore wraps VMD inside a swarm or evolutionary optimizer such as particle swarm, whale, or hunter-prey optimization. That optimizer reruns VMD hundreds or thousands of times and keeps the $(K,\alpha)$ that minimizes a sparsity score such as envelope entropy. These papers treat VMD as a deterministic black box. It is not. VMD is a non-convex solver that starts from initial guesses of the mode center frequencies, and its output depends on those guesses. The fitness value the swarm sees for a given $(K,\alpha)$ is therefore partly a property of the parameters and partly an accident of the inner solver's starting point.

This note proposes to measure that contamination and then remove it. Our hypothesis is that a large share of what these papers describe as a rugged fitness landscape is actually inner-initialization noise. The same is true of swarm "premature convergence" and of "modal mixing" (two physical components collapsing into one mode). If that is true, two fixes should make a cheap, deterministic sweep over $(K,\alpha)$ match or beat the swarms at a small fraction of their cost. The first fix seeds the inner solver from spectral peaks. The second warm-starts each evaluation from the solution at neighboring parameters, which is a continuation or homotopy path. If the hypothesis is false, and the landscape stays rugged even with a controlled inner solver, then we will have the first evidence that global search over these parameters is actually needed. Either outcome changes how this community should spend its compute.

Core Claim

The decisive structural fact comes from Density-based Adaptive Mode Decomposition. That paper shows that VMD's center-frequency updates behave like $K$-means clustering in the frequency domain. $K$-means is the textbook case of a non-convex problem whose result depends on initialization. Two centers can settle on one spectral peak while one center sits between two peaks, and that second failure looks exactly like modal mixing. The VMD tuning papers attribute modal mixing to a wrong $K$ or $\alpha$, including the WOA-VMD partial-discharge work and Tongguang Yang et al.. We claim that at moderate noise a substantial fraction of these failures occurs at the correct $K$ and is caused by a bad inner start. We also claim that this inner randomness is what makes the outer fitness surface look multimodal.

To make the claim testable, write $F(\theta,\omega_0)$ for the envelope-entropy fitness. Here $\theta=(K,\alpha)$ are the outer parameters and $\omega_0$ is the vector of initial center frequencies given to VMD. We define the landscape noise ratio

$$ R=\frac{\mathbb{E}_{\theta}\big[\operatorname{Var}_{\omega_0}F(\theta,\omega_0)\big]}{\operatorname{Var}_{\theta}\big[\mathbb{E}_{\omega_0}F(\theta,\omega_0)\big]}, $$

which is the average spread of the fitness caused by inner starts, divided by the spread caused by actually changing the parameters. When $R$ is near or above one, a swarm is mostly chasing noise. Our prediction is that $R$ is large under random or default inner starts and falls by an order of magnitude under spectral seeding with continuation.

Relation to Prior Work

Three papers motivate the proposal by showing the cost and fragility of black-box swarm tuning without ever asking about the inner solver. HPO-VMD with improved wavelet thresholding initializes hunter-prey optimization uniformly over $[K,\alpha]$ and minimizes envelope entropy. WOA-VMD-NLMS-WT does the same with a whale optimizer on transformer partial-discharge data. PSO-VMD-LSSVM does the same with particle swarm on roadheader motor current. All three report large gains over manual settings. All three re-solve VMD from scratch at every evaluation. None of them reports how VMD's own initialization was set or how much the fitness varies across inner runs. Our proposal keeps their fitness functions and benchmarks but changes how each fitness evaluation is computed.

Two lines of work supply the method. The first is the two-stage recipe from non-convex signal recovery: start inside the region that local descent can handle, then refine locally. Regularized gradient descent for joint blind deconvolution and demixing and PhaseMax both show that a cheap spectral estimate can replace random starts. In PhaseMax this cuts required measurements from $O(n^{1.5})$ to $O(n)$. We borrow the principle but not the guarantees. Their proofs assume Gaussian measurement models that vibration signals violate, and both assume the structural sizes are known. We supply those sizes by an outer sweep. The second line is continuation. Outlier Path tracks a solution as a parameter changes rather than re-solving at each value. We apply the same idea along $\alpha$ and across $K$.

Two papers constrain the design. DAMD (full paper, extraction and runtime results; Density-based Adaptive Mode Decomposition) already uses fast direct extraction to initialize a VMD-style refinement (its "hybrid variational refinement"). It is therefore the closest prior work. However, DAMD picks the number of modes from density peaks in the time-frequency plane and concedes that it "will partition whatever spectral content is present, including noise." It thus abandons the noise-robust outer criterion that makes swarm tuning work at low signal-to-noise ratio (SNR). We keep that outer criterion and use spectral seeding only to make the inner solve deterministic and cheap. Li et al.'s accelerated low-rank method proves that the region where local descent succeeds shrinks as conditioning worsens. In VMD terms, conditioning worsens when spectral peaks sit close together or SNR falls. This predicts where our approach should break, and we test it there rather than assuming it away.

Research Program

The work has three parts.

First, we build an instrumented VMD evaluator that records the inner initialization, the iteration count, and the final center frequencies. With it we measure $R$ over the full grid $K\in[2,12]$ and $\alpha\in[100,5000]$, the bounds used in this literature. We compare three inner starts: a fixed default, uniform random, and spectral seeding. Spectral seeding places the $K$ initial centers at the $K$ largest well-separated peaks of a smoothed power spectrum, chosen greedily in the style of $K$-means++.

Second, we build the continuation sweep. For each $K$ we solve VMD once at the smallest $\alpha$. We then step $\alpha$ geometrically upward, warm-starting each solve from the previous modes and centers. Moving from $K$ to $K+1$ splits the mode with the widest residual bandwidth into two seeded centers instead of restarting. Each step should need far fewer ADMM iterations than a cold solve; ADMM is the alternating-update method VMD runs internally. The whole grid becomes a single path whose total cost we report in ADMM iterations and in wall-clock time.

Third, on synthetic signals with known components, we attribute every modal-mixing failure to its cause. Each failure is labeled wrong $K$, wrong $\alpha$, or correct $(K,\alpha)$ with a bad inner start. The third category is new to this literature. If it is small, the core claim is wrong.

As a secondary testbed we repeat the measurement of $R$ for blind deconvolution tuned by golden jackal optimization (GJO-CYCBD). This method, called CYCBD, recovers a deconvolution filter by maximizing cyclostationarity, and its solver also iterates from an initial filter. Here we seed the cyclic frequency from the known bearing fault frequency. This tests whether the phenomenon is specific to VMD or general to "swarm wraps a non-convex solver" pipelines.

Baselines

The naive reference point is fixed rule-of-thumb VMD with $K=4$ or $K=6$ and $\alpha=2000$, run with default inner initialization. Prior work reports how far this falls short. On roadheader current, manually tuned VMD reached 87.5% lithology accuracy against 97.5% for PSO-tuned VMD (PSO-VMD-LSSVM). On simulated partial discharge, unoptimized VMD with wavelet thresholding gave SNR 5.3207 against 9.6404 for the whale-optimized pipeline (WOA-VMD).

The strongest existing competitor is swarm-tuned VMD with envelope-entropy fitness. Our main instance is hunter-prey optimization as in HPO-VMD, which reports output SNR of 17.1 dB and MSE of 0.90 at 10 dB input noise, and a 60% SNR gain over wavelet soft thresholding at 5 dB. We will re-implement HPO, PSO, and WOA with a population of 30 and 50 iterations, the 1,500-evaluation budget the field uses. They will run on the same signals, the same parameter bounds, and the same downstream wavelet thresholding and classifier. Because the published numbers come from the authors' own signal sets, we will first reproduce their protocol on CWRU-derived signals. We will compare against our reproduction and report the published figures only as a sanity check. To beat this baseline, our sweep must reach equal or lower final fitness, equal or lower mode-recovery error on synthetic signals, and equal downstream accuracy, using at most one tenth of the ADMM iterations. An outcome where we match quality at the same cost would not count as a win.

For low-SNR deconvolution, the competitor is GJO-CYCBD with a BiLSTM classifier, which reports 98.20% diagnostic accuracy on CWRU at −9 dB (CWRU evaluation; An Optimized Maximum Second-Order Cyclostationary Blind Deconvolution and Bidirectional Long Short-Term Memory Network Model for Rolling Bearing Fault Diagnosis). We will rerun it under the same noise injection and classifier.

For speed, the competitor is DAMD's direct extraction and hybrid refinement (DAMD), which reports 2.4 seconds against 14.0 seconds for standard VMD on a 1024-sample signal. Beating it means equal speed order with clearly better mode recovery below 0 dB.

Two ablations isolate the key ingredient. The first is a plain grid over the same $(K,\alpha)$ points with cold, default inner starts at every point. It separates the value of the deterministic sweep from the value of warm starting. The second is the swarm itself run with spectrally seeded inner solves. It asks whether the swarm becomes unnecessary, or simply better, once the inner noise is removed.

Why not just…?

Why not just seed the swarm's initial population with DAMD's estimate of $K$ and the spectral peaks? This is the obvious hybrid, and it does not address our mechanism. Seeding the outer population still leaves each fitness evaluation dependent on VMD's inner start, so the swarm still sees a noisy landscape. Our second ablation, a swarm with seeded inner solves, compared with a swarm with a seeded outer population, separates the two effects directly. If outer seeding alone recovers most of the gain, our claim about inner noise is weakened, and we will report that.

Why not just use DAMD and skip outer tuning entirely? DAMD's authors state that its noise handling depends entirely on the time-frequency representation and that it will partition noise into modes. At the −9 dB to 0 dB levels where swarm tuning earns its keep, density peaks should over-count modes. Our noise sweep tests this. If DAMD's hybrid refinement matches envelope-entropy selection at low SNR, the outer criterion is unnecessary. That would be a real threat, and the experiment will settle it.

Why not just grid-search $(K,\alpha)$, or average fitness over several random VMD restarts? We take this seriously; the integer range of $K$ is small and $\alpha$ is one-dimensional. A plain grid is our first ablation. Restart-averaging, run with 3 to 10 restarts per point, reduces the noise term in $R$ but multiplies cost. Our prediction is that continuation gives the noise reduction of many restarts at less than the cost of one cold solve per point. If restart-averaged grid search matches us at comparable cost, the contribution reduces to the diagnosis, which would still be a useful result for this literature.

Why not just train an end-to-end network on raw signals? This does not compete on the decomposition question. The CWRU evidence suggests raw classifiers degrade badly at −9 dB without an optimized front-end, roughly 9.89 points behind GJO-CYCBD (CWRU evaluation; An Optimized Maximum Second-Order Cyclostationary Blind Deconvolution and Bidirectional Long Short-Term Memory Network Model for Rolling Bearing Fault Diagnosis). We will include one raw 1D convolutional classifier only as a downstream reference.

Proposed Experiments

Ground truth comes from synthetic multi-component signals. These mix two to eight amplitude- and frequency-modulated tones, with controlled spacing between peaks, plus periodic bearing-style impulse trains. Gaussian, pink, and brown noise is added at 10, 5, 0, −5, and −9 dB, matching the noise colors used in HPO-VMD. Simulated single- and double-exponential partial-discharge pulses follow the WOA-VMD setup. Real data come from CWRU bearings (0.007 to 0.021 inch faults at 1797 rpm with injected noise), Paderborn vibration, and XJTU-SY for robustness across operating conditions. For XJTU-SY no reported result found on Lacuna for VMD tuning.

On synthetic signals the metrics are mode-recovery error after optimal matching, accuracy of the selected $K$, output SNR and MSE, and the attribution of modal-mixing failures. On real data the metrics are the envelope-spectrum ratio at the fault characteristic frequency and downstream fault-classification accuracy with a fixed classifier. Compute is reported in total ADMM iterations and wall-clock time.

Three checks decide the outcome. The first is whether $R$ under default or random inner starts exceeds roughly one on a meaningful share of signals and drops sharply under spectral seeding. If it does not, the core claim fails. The second is whether, at correct $K$, random inner starts produce modal mixing that seeding removes. The third is whether the continuation sweep matches the 1,500-evaluation swarms on fitness, recovery error, and accuracy with at least ten times less inner computation.

We expect the approach to degrade as spectral peaks crowd together and SNR approaches −9 dB, as the shrinking-basin result of Li et al. predicts. At that point envelope entropy itself may develop deceptive minima. We will map where seeded continuation stops matching the swarm as a function of peak spacing and SNR, and that map is the practical guidance this note aims to deliver.