Stochastic resonance (SR) - the enhancement of a weak signal's detectability by noise of an optimal intensity in a nonlinear system - is widely used for weak-signal detection, but published signal-to-noise ratio (SNR) gains rest on hand-tuned, setting-specific choices of the nonlinear system, thresholds, added-noise scale, noise model, and preprocessing transforms. Three recent SR strands - multiscale double-threshold SR in the wavelet domain, two-coupled tri-stable SR, and asymmetric time-delayed bistable SR under correlated multiplicative and additive noise - each state in their own limitation sections that noise-model generality, systematic parameter optimization, and broader validation remain open. Together the three strands pose one problem: establish a principled, tuning-free methodology for SR-based weak-signal enhancement - an adaptation rule that sets all design degrees of freedom automatically from observed data - and demonstrate, by reproducible evaluation, that its gains persist across signal classes, noise statistics, and real measurement systems. Existing results cover single applications, single noise models, or specific architectures, and none closes this general core.

Background

Stochastic resonance (SR) is a phenomenon of noisy nonlinear systems in which adding noise of a suitable intensity improves, rather than degrades, the detectability of a weak signal: noise drives the system across detection thresholds or energy barriers at times modulated by the signal, so that a performance measure such as the output signal-to-noise ratio (SNR) - the ratio of spectral power at the signal frequency to background spectral power - passes through a maximum at an optimal noise level, the SR point (Gammaitoni et al. 1998). In engineering weak-signal detection, the measured signal, typically with added noise, is passed through a nonlinear system and the output is inspected for the recovered signal. The classical architecture is a bistable potential, meaning a double-well energy landscape U(x)U(x) whose dynamics dx/dt=−U′(x)+s(t)+n(t)\mathrm{d}x/\mathrm{d}t=-U'(x)+s(t)+n(t) are driven by the signal s(t)s(t) and noise n(t)n(t); proposed variants include tri-stable potentials (three wells), coupled systems of several units, asymmetric potentials, and systems with time-delayed feedback, in which the force at time tt depends on the state at t−τt-\tau. Noise may be additive (state-independent) or multiplicative (scaled by the state), and the two components may be correlated. Output quality is usually quantified as an SNR computed from the peak of the Fourier spectrum at the driving frequency, and the improvement as the SNR gain, the difference in decibels between output and input SNR. A complementary statistical formulation uses threshold detectors that record only whether the observation exceeds fixed thresholds aa or bb: a signal is sub-threshold when it never crosses a threshold without noise, and SR then exploits noise-induced exceedance events to estimate the signal; theoretical analyses typically assume a known noise distribution (usually Gaussian) and rely on adiabatic approximations, meaning the signal varies slowly compared with the inter-well transition rates.

Three recent papers, one from each main strand, report SNR improvements while explicitly leaving generality open. Vimalajeewa, Muller and Vidakovic (2025) propose a double-threshold detector that combines a sub-threshold and a sup-threshold estimator - one whose signal crosses the threshold even without noise - and performs SR in the wavelet (multiscale) domain: the discrete wavelet transform decomposes a signal into smoothing (low-frequency) and detail components at successive levels jj, and applying SR to wavelet coefficients reduces the noise required for detection. Their evaluations use simulated and standard benchmark signals under Gaussian noise; they state that analyzing performance across noise distributions other than the normal distribution 'remains a potential extension for future research', that the decomposition level must be chosen by 'empirical testing', and that 'future work should investigate the impact of different noise distributions and optimize parameters accordingly'. Li et al. (2018) study a two-coupled tri-stable SR system, in which two tri-stable subsystems are coupled by a linear term; in fourth-order Runge–Kutta simulations they report higher SNR gain than bistable, coupled-bistable, and single tri-stable variants, but select the four design parameters by single-variable sweeps (joint optimization being too costly) and present no experimental data. He, Hu, Zhang and Lu (2019) derive, via a small-time-delay approximation and time-delayed Fokker–Planck equations (the evolution equation for the state's probability density), a closed-form output SNR for an asymmetric time-delayed bistable system under correlated multiplicative and additive Gaussian noise; they verify delay- and asymmetry-enhanced SR numerically and apply the method to real bearing vibration records - but the analytic results hold only for small delay and slow, weak forcing, and the experimental application requires per-case rescaling and parameter choices.

Across these strands the reported gains share one structural limitation: every degree of freedom - potential shape and parameters, thresholds, added-noise scale, noise model, transform settings - is fixed by hand for the specific signal class, noise condition, and application at hand. Two classical results show that tuning is not intrinsic to SR: Collins, Chow and Imhoff (1995) demonstrated 'stochastic resonance without tuning' in parallel arrays of noisy neuron models, where the collective response improves over a broad range of noise intensities, and Mitaim and Kosko (1998) studied adaptive SR in which the optimal noise level is learned from data. A large modern literature automates parameter choice per application with metaheuristic optimizers - for example adaptive SR for planetary gearbox diagnosis (Lei et al. 2013) and synthetic-index adaptive SR for bearing diagnosis (Zhou et al. 2017) - and bespoke treatments exist for particular non-Gaussian noise models, for example alpha-stable impulsive noise (noise drawn from a family of heavy-tailed distributions producing frequent large-amplitude spikes; Liu et al. 2025). Yet each of these covers one application and one noise model; no adaptation rule is known that carries over, with validated performance, across signal classes, noise statistics, and real measurement systems. That is the unresolved target.

Problem Statement

Establish a principled, tuning-free methodology for SR-based weak-signal enhancement, together with a demonstration of its generality. Concretely: give a rule or algorithm that determines, from the observed data alone and without per-instance manual adjustment, all design degrees of freedom of an SR-based detector - the nonlinear system (potential or threshold structure) and its parameters, the thresholds, the added-noise scale and noise model, and any multiscale transform settings such as the wavelet decomposition level - and demonstrate, by reproducible evaluation, that the resulting SNR gain, recovery accuracy, or detection performance persists across (i) signal classes beyond constant or single-frequency sinusoidal signals, such as time-varying, aperiodic, transient, and multi-frequency signals; (ii) noise statistics beyond known Gaussian white noise, including unknown, heavy-tailed or impulsive, and colored noise; and (iii) real measurement systems with sensor and detector nonidealities. The three axes are the dimensions of one common objective - removing the hand-tuned, setting-specific character of SR-based enhancement - so a complete answer must address them jointly rather than as isolated applications; a solution confined to one axis, one signal class, or one application does not close the problem. The signal classes and noise statistics named in (i) and (ii) are illustrative of the required breadth rather than an exhaustive checklist; the demonstration must nonetheless span at least three signal classes drawn from the kinds named in (i) and at least two noise statistics drawn from the kinds named in (ii). Axis (iii) must be met on real, publicly available measurement records from at least two structurally distinct measurement systems - for example open rotating-machine test-rig vibration records carrying sensor and detector nonidealities - rather than on explicitly simulated nonidealities alone. The demonstration must rest on a single performance measure - SNR gain, recovery accuracy, or detection performance - that is fixed in advance, defined on every mandated case, and applied consistently across the three axes rather than switched per axis or per case to a more favourable measure; 'persists' means that, under that measure, the tuning-free method is no worse on the mandated cases of each axis than a per-case-tuned baseline, namely the tuning-free method's own detector architecture for that case with its design degrees of freedom tuned separately for that case by a reported per-case parameter search, and that search protocol must be reported with the results. In addition, the tuning-free method must exhibit a positive gain over the no-SR input under that measure on the mandated cases, and the per-case-tuned baseline must itself be shown to achieve a positive gain. The cases compared on each axis must be named and fixed before results are produced and reported with them. The operative sense of 'without hand tuning' is: no design degree of freedom - potential or threshold structure and its parameters, thresholds, added-noise scale and noise model, or transform settings such as the wavelet decomposition level - may be set by hand for a specific instance, nor by any per-case search or optimization of the chosen performance measure, including selecting among candidate designs or parameter sets by measured outcome; the degrees of freedom must instead be computed by one fixed formula from statistics of the observed data, identical for every case, not re-optimized case by case as they are today through empirical parameter sweeps, hand-picked noise models, and case-by-case rescaling.

References

  1. Dixon Vimalajeewa, Ursula U. Muller, Brani Vidakovic, 'A Multiscale Approach for Enhancing Weak Signal Detection', arXiv:2510.20828 (2025), https://arxiv.org/abs/2510.20828
  2. Quan-Quan Li, Xue-Mei Xu, Lin-Zi Yin, Yi-Peng Ding, Jia-Feng Ding, Ke-Hui Sun, 'Implication of two-coupled tri-stable stochastic resonance in weak signal detection', Chinese Physics B 27, 034203 (2018), DOI 10.1088/1674-1056/27/3/034203, https://doi.org/10.1088/1674-1056/27/3/034203
  3. Lifang He, Dayun Hu, Gang Zhang, Siliang Lu, 'Stochastic resonance in asymmetric time-delayed bistable system under multiplicative and additive noise and its applications in bearing fault detection', Modern Physics Letters B 33, 1950341 (2019), DOI 10.1142/S021798491950341X, https://doi.org/10.1142/S021798491950341X
  4. Luca Gammaitoni, Peter Hänggi, Peter Jung, Fabio Marchesoni, 'Stochastic resonance', Reviews of Modern Physics 70, 223 (1998), DOI 10.1103/RevModPhys.70.223, https://doi.org/10.1103/RevModPhys.70.223
  5. J. J. Collins, Carson C. Chow, Thomas T. Imhoff, 'Stochastic resonance without tuning', Nature 376, 236-238 (1995), DOI 10.1038/376236a0, https://doi.org/10.1038/376236a0
  6. S. Mitaim, B. Kosko, 'Adaptive stochastic resonance', Proceedings of the IEEE 86, 2152-2183 (1998), DOI 10.1109/5.726785, https://doi.org/10.1109/5.726785
  7. Yaguo Lei, Dong Han, Jing Lin, Zhengjia He, 'Planetary gearbox fault diagnosis using an adaptive stochastic resonance method', Mechanical Systems and Signal Processing 38, 113-124 (2013), DOI 10.1016/j.ymssp.2012.06.021, https://doi.org/10.1016/j.ymssp.2012.06.021
  8. Peng Zhou, Siliang Lu, Fang Liu, Yongbin Liu, Guihua Li, Jiwen Zhao, 'Novel synthetic index-based adaptive stochastic resonance method and its application in bearing fault diagnosis', Journal of Sound and Vibration 391, 194-210 (2017), DOI 10.1016/j.jsv.2016.12.017, https://doi.org/10.1016/j.jsv.2016.12.017
  9. Jin Liu, Zan Li, Yue Gao, Li Yang, Danyang Wang, Peihan Qi, 'A novel weak signal reception algorithm integrating a soft-limiter with adaptive stochastic resonance in alpha-stable-distributed noise environments', Measurement Science and Technology 36, 036117 (2025), DOI 10.1088/1361-6501/adb063, https://doi.org/10.1088/1361-6501/adb063
  10. Jinfeng Zhang, Junling Peng, Jimeng Li, Zong Meng, 'Stochastic resonance in an underdamped Duffing bistable system embedded with overdamped feedback control for rolling bearing fault detection', Measurement Science and Technology 36, 036128 (2025), DOI 10.1088/1361-6501/adb5ab, https://doi.org/10.1088/1361-6501/adb5ab