Sharper Signal Trick Catches Faint Bearing Faults Before They Turn Catastrophic

Deep inside a spinning wind turbine, a bearing can begin to fail almost silently. A microscopic pit in a raceway, a hairline spall on a rolling element, a tiny imperfection forged during manufacture: each of these defects strikes the surrounding metal with every rotation, producing a faint, rhythmic impulse buried far beneath the roar of gears, shafts and aerodynamic noise. By the time those impulses grow loud enough for conventional monitoring tools to flag them, the damage is often already advanced, and the countdown to an expensive, sometimes dangerous failure has already begun. A new signal-processing method published in PLOS One by Delong Zhang, Yubo Ma and Hongan Wu aims to catch those whispers much earlier, by combining two ideas that rarely appear together: a deliberately non-convex form of sparse optimization and an aggressively refined time-frequency transform.

The core problem the researchers tackle is one that plagues every vibration-based diagnostic system. Early-stage faults generate impulsive signatures, sharp transient spikes that repeat at characteristic frequencies determined by the geometry of the bearing and the speed of rotation. In a clean laboratory signal, those spikes stand out clearly. In a real machine, however, they compete with gear meshing tones, shaft harmonics, electromagnetic interference and broadband random noise, all of which can be orders of magnitude stronger than the fault signature itself. The result is a signal in which the diagnostic information is present in principle but effectively invisible in practice, like a single candle flame photographed against the surface of the sun.

The team’s answer, which they call the concave sparsity-assisted time-reassigned multisynchrosqueezing transform, or CSA-TMSST, works in two stages. The first stage is a denoising and reconstruction step built on sparse optimization, a mathematical framework that exploits the fact that fault impulses are sparse: they occupy only a small fraction of the signal’s total duration, appearing as brief bursts separated by relatively quiet intervals. Noise, by contrast, tends to spread its energy thinly and continuously across time. If a model can be built that rewards sparsity, it can separate the bursty fault component from the diffuse background, reconstructing a cleaner version of the impulsive signal before any frequency analysis is attempted.

What makes the new approach distinctive is the specific penalty function used to enforce that sparsity. Many classical methods rely on convex penalties, most famously the l1 norm, which are mathematically convenient because they lead to optimization problems with a single, guaranteed global minimum. The cost of that convenience is a systematic bias: convex penalties tend to shrink large signal components more than they should, flattening the very peaks that carry diagnostic information. Zhang, Ma and Wu instead adopt the generalized minimax-concave penalty, known as the GMC penalty, a non-convex regularizer that approximates sparsity more faithfully. Because the GMC penalty penalizes small coefficients heavily while treating large coefficients more gently, it preserves the amplitude of genuine fault impulses while suppressing the countless small fluctuations produced by noise. The resulting optimization model is harder to solve than a convex one, but the payoff is a reconstructed signal in which the impulsive fault structure survives with its shape and strength intact.

Once the impulsive component has been recovered, the second stage of the method takes over: characterizing it in the time-frequency domain. Time-frequency representations are the workhorse of vibration analysis, mapping how a signal’s energy is distributed across both time and frequency so that transient events can be localized. The trouble is that standard transforms face a fundamental trade-off, rooted in the uncertainty principle, between resolution in time and resolution in frequency. A short analysis window captures when an event happens but blurs its frequency content; a long window resolves frequency precisely but smears events across time. For bearing diagnosis, where both the timing of impacts and the spacing of their frequency harmonics matter, this trade-off can obscure the very features an engineer needs to see.

The multisynchrosqueezing transform, and its time-reassigned variant employed in this study, attack that trade-off directly through iterative reassignment. The idea is conceptually elegant: after computing an initial, blurry time-frequency representation, the algorithm estimates where each parcel of energy truly belongs in the time-frequency plane and squeezes it toward that location. Rather than performing this correction once, the multisynchrosqueezing framework repeats it across multiple iterations, each pass sharpening the concentration of the energy distribution further. The time-reassigned version extends this sharpening along the time axis as well as the frequency axis, producing representations in which transient impulses appear as tight, well-localized ridges rather than diffuse smears. Applied to the sparsity-reconstructed signal, the transform yields a time-frequency picture in which the periodic structure of a bearing fault, the regular train of impacts and their harmonic families, becomes dramatically easier to read.

The final step of the pipeline is deliberately conventional: spectral analysis of the processed signal to identify characteristic fault frequencies. Every element of a rolling bearing, the outer race, the inner race, the rolling elements and the cage, produces impacts at a calculable rate that depends on the bearing’s geometry and rotational speed. When the processed spectrum shows a peak, or a harmonic series of peaks, at one of these theoretical frequencies, the diagnosis follows directly. The value of the preceding stages is that they make this final step reliable even when the raw signal would have buried the fault frequencies beneath the noise floor entirely.

The researchers report that CSA-TMSST succeeds in extracting weak fault-related signatures and producing high-quality time-frequency representations under strong noise conditions, confirming both its effectiveness and its superiority over existing approaches. That combination matters because the two failure modes of diagnostic methods are complementary: a denoising step that destroys the fault’s impulsive character leaves nothing for the transform to visualize, while a high-resolution transform applied to a noisy signal simply produces a sharper picture of the noise. By pairing a reconstruction stage that preserves impulses with a transform stage that concentrates them, the method addresses both weaknesses at once. The authors frame the work explicitly in the context of bearing fault diagnosis, and wind turbines are a natural target: their drivetrains combine heavy, slowly rotating bearings with highly variable loads and a noisy acoustic environment, making early fault detection particularly difficult and particularly valuable.

The broader significance of the work lies in what it suggests about the trajectory of condition-monitoring technology. As wind energy, aerospace and heavy industry push toward longer maintenance intervals and predictive, rather than reactive, servicing, the economic case for detecting faults at the earliest possible stage grows steadily stronger. A bearing defect caught weeks before failure can be scheduled for replacement during planned downtime; the same defect caught days before failure can force an emergency shutdown, a crane mobilization and, in the case of offshore turbines, a costly and weather-dependent intervention. Signal-processing advances like CSA-TMSST operate entirely on the vibration data that machines already produce, requiring no new sensors or hardware, which means they can in principle be layered onto existing monitoring infrastructure as a software upgrade.

There are, of course, the usual caveats that accompany any new method moving from paper to practice. Non-convex optimization problems can be sensitive to parameter choices, and the GMC penalty introduces regularization parameters that must be tuned for the signal characteristics of a given machine. Iterative reassignment schemes add computational cost, which matters for continuous, real-time monitoring on embedded hardware. And laboratory validation, however rigorous, always differs from the messier reality of field data, where speed fluctuations, load variations and multiple simultaneous faults complicate the picture. Yet the underlying principles, that sparsity in the impulse domain and concentration in the time-frequency domain are two views of the same physical reality, and that enforcing both yields better diagnostics than enforcing either alone, are likely to outlast any particular implementation. For the engineers tasked with keeping thousands of rotating machines alive, a method that turns an inaudible whisper into a clear, readable signature is not an academic curiosity. It is the difference between a maintenance ticket and a catastrophe.

Subject of Research: Weak fault feature enhancement for early-stage bearing fault diagnosis using a concave sparsity-assisted time-reassigned multisynchrosqueezing transform

Article Title: Concave sparsity-assisted time-reassigned multisynchrosqueezing transform for early-stage weak fault feature enhancement with applications to bearing fault diagnosis

Article References: Zhang, D., Ma, Y., & Wu, H. (2026). Concave sparsity-assisted time-reassigned multisynchrosqueezing transform for early-stage weak fault feature enhancement with applications to bearing fault diagnosis. PLOS One, 21(10), e0360733. https://doi.org/10.1371/journal.pone.0360733

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360733

Keywords: bearing fault diagnosis, time-frequency analysis, sparse optimization, GMC penalty, multisynchrosqueezing transform, vibration monitoring, wind turbines, fault detection, signal processing, condition monitoring, impulsive signatures, non-convex regularization