Source-linked AI summary
Bio-inspired efficient cyclostationary analysis in machine and underwater acoustic recordings
Thura Pyae Sone, Abishek Soti, Henry Zhong, Gregory Cohen, RunChun Mark Wang, Ying Xu
TL;DR
Vessel cyclostationary signatures can be weak and obscured, while FAM/ICMC is computationally costly and artifact-prone. The paper evaluates CARFAC-IHC and CARFAC-BM routes; synthetic validation recovers theoretical cyclic frequencies, while CARFAC-IHC has a higher noise floor on ShipsEar than on CWRU.
Problem
Vessel machinery signatures in real radiated noise can be weak, broadband-distributed, and obscured by ambient noise and non-stationary interference, while FAM/ICMC is computationally costly and susceptible to decomposition artefacts.
Source: §1 · p. 1Method
The paper extracts cyclic modulation from CARFAC cochlear channels using IHC energy-envelope spectra, and compares this route with BM filtering followed by FAM/ICMC.
Source: §1 · p. 1, §2.2 · p. 2Results
CARFAC-BM recovered all theoretical cyclic frequencies in synthetic validation, while CARFAC-IHC had a visibly higher noise floor on ShipsEar than on CWRU.
Source: §3 · p. 3, §4.2 · p. 4Takeaways & Limitations
CARFAC-IHC is not tied to FAM’s periodic block advance or channeliser size, whereas the CARFAC-BM route inherits FAM/ICMC’s computational cost.
Source: §5.1 · p. 4On ShipsEar, broadband ambient and flow noise shared across most of the 64 channels raised the CARFAC-IHC noise floor and yielded a comparatively poorer result than on CWRU.
Source: §4.2 · p. 4
Abstract
from arXiv · showhide
We propose a bio-inspired approach that uses the inner-hair-cell (IHC) response of the Cascade of Asymmetric Resonators with Fast-Acting Compression (CARFAC) model to efficiently extract cyclic modulation from acoustic signals. We further investigate the contribution of IHC processing by comparing the CARFAC-IHC response with the CARFAC basilar-membrane (BM) filtering. Furthermore, the CARFAC-IHC and CARFAC-BM approach are benchmarked against conventional FFT Accumulation Method (FAM), Integrated Cyclic Modulation Coherence (ICMC), and Detection of Envelope Modulation On Noise (DEMON) approaches using the Case Western Reserve University (CWRU) bearing dataset and a real ShipsEar work-vessel recording dataset. The results demonstrate reliable recovery of characteristic cyclic components while substantially reducing the computational burden of conventional cyclostationary analysis.
1. INTRODUCTION
Vessel machinery imprints cyclostationary signatures on radiated noise, but these can be weak, broadband, and obscured. The paper proposes CARFAC cochlear processing to extract cyclic modulation while addressing limitations of conventional methods.
1. INTRODUCTION
- Vessel machinery imprints cyclostationary modulation on radiated noise, but real ship signatures can be weak, spread across broad frequency bands, and obscured by ambient noise and interference.The cited passage specifies non-stationary interference as another obscuring factor.Source: §1 · p. 1
- DEMON extracts envelope modulation through band-pass filtering and envelope detection, but does not fully exploit frequency-dependent structure.Spectral-kurtosis-guided envelope analysis sharpens DEMON band selection, within the same cyclostationary theory.Source: §1 · p. 1
- FAM computes a spectral correlation surface that ICMC converts to a one-dimensional detection profile, but the pipeline is computationally costly and susceptible to time-frequency decomposition artefacts.The supplied passages describe ICMC as a normalization step.Source: §1 · p. 1
- The paper proposes CARFAC cochlear processing for cyclostationary estimation, using frequency-localized channels and IHC outputs to extract and integrate envelope modulation into cyclic signatures.It investigates CARFAC-IHC feasibility and compares performance with the FAM/ICMC pipeline.Source: §1 · p. 1
2. METHOD
The method uses CARFAC’s cochlear channels to derive cyclic signatures from IHC envelopes, and also tests a BM-based route coupled to FAM/ICMC. The IHC route uses block-energy spectra and robust local normalization; the BM route retunes and combines channels before conventional analysis.
2.1. CARFAC cochlear processing
- The CAR cascade produces BM outputs, while per-channel AGC feedback sets damping; the DOHC feedback path adjusts BM damping, and the DIHC stage rectifies and integrates to extract channel modulation.The CAR cascade’s centre frequencies create a log-frequency decomposition; DIHC also uses multiplicative transmitter depletion.Source: §2.1 · pp. 1–2, Figure 1 · p. 1
2.2. CARFAC-IHC envelope spectrum
- For cyclic frequencies below the IHC sampling rate, the method averages squared IHC samples in nonoverlapping blocks to form downsampled frame-energy samples.The energy trace is mean-subtracted and high-pass filtered at 5 Hz to remove DC or near-DC offset.Source: §2.2 · p. 2
- Per-channel spectra are normalized by the local median around each alpha bin, then combined across channels by a median to reduce square-law self-noise.The passage motivates local alpha-bin normalization as preserving a weak phase-locked line when a stronger unrelated line dominates the channel.Source: §2.2 · p. 2
- Unlike FAM’s conjugate cross-multiplication of frequency-shifted spectra, the IHC route computes a plain magnitude spectrum from the real-valued energy signal without a spectral correlation surface.Source: §2.2 · p. 2
2.3. CARFAC-BM into FAM/ICMC
- For the BM route, CARFAC is retuned so all summed Greenwood channels lie in the analysis band; each channel is unit-RMS scaled before summation, removing relative channel amplitudes to lessen centre-frequency-dependent gain effects.The summed signal is then mean-removed, RMS-normalized, decimated, and sent through standard FAM, spectral coherence, and ICMC processing.Source: §2.3 · p. 2
- The BM pipeline omits fixed delay calibration because operating delay varies with input level through AGC feedback and amplitude-dependent damping.The described calibration is a linear, small-signal measurement, so the pipeline retains only the amplitude fix.Source: §2.3 · p. 2
3. RESULTS: SYNTHETIC DATA
Synthetic double-sideband AM tests validate both CARFAC routes against known cyclic-frequency structure. CARFAC-IHC estimates stay close to theory, while CARFAC-BM recovers all theoretical cyclic frequencies.
3. RESULTS: SYNTHETIC DATA
- At fc = 1000 Hz, fm = 100 Hz, and m = 0.8, the full-carrier DSB-AM test checks envelope and carrier-related cyclic components.Source: §3 · pp. 2–3
- Table 2 lists the FAM/ICMC channeliser parameters for the baseline pipeline.Source: §3 · p. 2, Table 2 · p. 3
- 0.063 Hz is the maximum CARFAC-IHC error; CARFAC-BM recovers all theoretical cyclic frequencies, showing the cochlear front end preserves second-order structure.Source: Table 3 · p. 3, Table 4 · p. 3, §3 · p. 3
- For CARFAC-IHC, per-channel rectification generates sum or difference terms for every pair of tones a channel responds to, beyond Gardner’s simplified treatment.Source: §3 · p. 3
4. RESULTS: REAL DATA
On real recordings, CARFAC-IHC recovered shaft harmonics and a quarter-order component on CWRU, while ShipsEar comparisons showed agreement among estimators but lacked an independent theoretical line reference.
4.1. CWRU bearing dataset
- For a healthy, unfaulted CWRU bearing (0 hp, 1797 rpm), CARFAC-IHC extracts 30, 60, and 90 Hz shaft harmonics and noisily resolves the 7.5 Hz quarter-order modulation that DEMON misses.CARFAC-BM and raw ICMC also show the shaft harmonics and resolve the 7.5 Hz component and its overtones.Source: Figure 2 · p. 3, §4.1 · p. 3
4.2. ShipsEar dataset
- On ShipsEar, processed at 16 kHz over 100–1900 Hz, CARFAC-BM tracks raw ICMC on 9/10 compared peaks, agreeing with DEMON on 7/10 and CARFAC-IHC on 8/10.Source: Figure 3 · p. 3, §4.2 · p. 4
- ShipsEar has no documented shaft speed or gear geometry for independent theoretical line checking.Source: §4.2 · p. 4
- Broadband ambient and flow noise shared across most of the 64 channels raises CARFAC-IHC’s noise floor, producing poorer results than on CWRU.Source: §4.2 · p. 4
5. DISCUSSION
CARFAC-IHC recovers most low-frequency cyclic structure across the reported recordings while using less compute time than FAM/ICMC. The discussion also considers the method’s computational constraints, broader detector applicability, and possible future extensions.
5.1. Computational considerations and generalisability
- FAM/ICMC’s accumulator exceeds 1 GB at Np=1024, while CARFAC-BM inherits its backend cost and adds a per-sample cochlear pass.Source: §5.1 · p. 4
- Against FAM/ICMC, DEMON and CARFAC-IHC take 13–102× and 2–37× less compute time, respectively, and scale more gently with segment resolution and Np.Source: §5.1 · p. 4, Figure 4 · p. 4
- CARFAC’s cochleagram could feed downstream cyclostationary detectors beyond the demonstrated propeller and bearing-fault applications.Source: §5.1 · p. 4
- FAM’s block advance adds signal-independent spurious correlations at α=kfs/L and α=kfs/Np; DEMON and CARFAC-IHC are not tied to those parameters.Source: §5.1 · p. 4
- Future directions include asynchronous, transient-triggered auditory analysis and AUV homing on detected signatures.The asynchronous approach is projected to cut power by orders of magnitude; AUV homing is proposed as an extension of existing sound-source localization and FPGA-based edge perception.Source: §5.1 · p. 4
5.2. Conclusion
- Across synthetic, CWRU, and ShipsEar recordings, CARFAC-IHC recovers the majority of low-frequency cyclic structure up to 37× faster than FAM/ICMC.Source: §5.2 · p. 4