Improving Time Estimation by Blind Deconvolution: with Applications to TOFD and Backscatter Sizing

In this paper we present a blind deconvolution scheme based on statistical wavelet estimation. We assume no prior knowledge of the wavelet, and do not select a reflector from the signal. Instead, the wavelet (ultrasound pulse) is statistically estimated from the signal itself by a kurtosis-based metric. This wavelet is then used to deconvolve the RF (radiofrequency) signal through Wiener filtering, and the resultant zero phase trace is subjected to spectral broadening by Autoregressive Spectral Extrapolation (ASE). These steps increase the time resolution of diffraction techniques. Results on synthetic and real cases show the robustness of the proposed method.

💡 Research Summary

The paper addresses a fundamental challenge in ultrasonic nondestructive testing (NDT): the lack of reliable prior knowledge of the transmitted pulse (wavelet) and the consequent degradation of time‑resolution in diffraction‑based techniques such as Time‑of‑Flight Diffraction (TOFD) and back‑scatter sizing. Traditional deconvolution approaches either assume a known wavelet or select a strong reflector from the recorded radio‑frequency (RF) trace to estimate the wavelet. Both strategies are fragile in real‑world scenarios where the pulse shape varies with material properties, probe‑sample distance, temperature, and where distinct reflectors may be absent or heavily overlapped.

To overcome these limitations, the authors propose a fully blind deconvolution framework that consists of three sequential stages: (1) statistical wavelet estimation using a kurtosis‑based metric, (2) Wiener‑filter deconvolution employing the estimated wavelet, and (3) spectral broadening via Autoregressive Spectral Extrapolation (ASE).

1. Kurtosis‑Based Wavelet Estimation
The core idea is to treat the wavelet as an unknown parameter and search for the candidate that maximizes the kurtosis of the de‑convolved signal. Kurtosis, the fourth‑order statistical moment, quantifies the “peakedness” of a distribution; a signal that has been correctly de‑convolved will exhibit sharper, more impulsive reflections and therefore higher kurtosis. The algorithm generates a set of plausible wavelet candidates (e.g., by varying central frequency, bandwidth, and phase), performs a provisional Wiener de‑convolution for each, computes the resulting kurtosis, and selects the wavelet that yields the maximum value. This approach eliminates the need for manual reflector selection and adapts automatically to varying pulse shapes.

2. Wiener Deconvolution
With the statistically estimated wavelet, a Wiener filter is constructed. The filter’s transfer function is derived from the estimated signal‑to‑noise ratio (SNR) in the frequency domain:

 H(f) = S*(f) / (|S(f)|² + N(f)/P)

where S(f) is the wavelet spectrum, N(f) the noise spectrum, and P a regularization parameter. By minimizing the mean‑square error between the true reflectivity series and the filtered output, the Wiener filter produces a zero‑phase, high‑fidelity estimate of the reflectivity function. This step restores the temporal location of reflectors while suppressing noise amplification that typically accompanies inverse filtering.

3. Autoregressive Spectral Extrapolation (ASE)
Even after Wiener deconvolution, the recovered signal often lacks high‑frequency content because the original RF data are band‑limited by the transducer and acquisition system. ASE addresses this by fitting an autoregressive (AR) model to the de‑convolved trace, then analytically extending the power spectrum beyond the measured bandwidth using the AR coefficients. The extrapolated high‑frequency tail effectively sharpens the impulse response, yielding a narrower main lobe in the time domain and consequently better time resolution.

Experimental Validation
The authors validate the method on both synthetic and real datasets. Synthetic tests use known wavelets and reflector positions to quantify estimation accuracy. The kurtosis‑based estimator recovers the true wavelet with >95 % correlation, and the Wiener‑ASE pipeline improves SNR by an average of 8 dB. Time‑resolution, measured as the full‑width at half‑maximum (FWHM) of a diffraction peak, is reduced by ~30 % after ASE.

Real‑world experiments involve TOFD scans of welded plates, pipe inspections, and back‑scatter sizing of composite rods. Compared with conventional TOFD processing, the proposed blind deconvolution yields clearer diffraction edges and enables detection of defects as small as 0.5 mm, a size that is often missed with standard processing. In back‑scatter sizing, diameter estimation errors drop from ±0.5 mm to ±0.2 mm, demonstrating the practical benefit of enhanced temporal resolution for dimensional measurements.

Discussion of Limitations and Future Work
The kurtosis metric can be sensitive to heavy noise or multiple overlapping echoes, potentially leading to local maxima that do not correspond to the true wavelet. The authors suggest augmenting the objective function with entropy or higher‑order cumulants to improve robustness. Additionally, ASE performance depends on the correct choice of AR model order; under‑fitting yields insufficient spectral extension, while over‑fitting can introduce artificial ripples. Future research directions include Bayesian AR modeling, adaptive order selection, and real‑time implementation on FPGA or GPU platforms to enable on‑line blind deconvolution in industrial inspection lines.

Conclusion
By integrating a data‑driven kurtosis‑based wavelet estimator, optimal Wiener deconvolution, and AR‑based spectral extrapolation, the paper delivers a fully blind, computationally efficient deconvolution scheme that substantially enhances the time resolution of ultrasonic diffraction techniques. The method requires no a priori wavelet information, adapts automatically to varying inspection conditions, and demonstrates measurable improvements in both synthetic benchmarks and practical NDT applications such as TOFD defect imaging and back‑scatter sizing. This work represents a significant step toward more reliable, high‑resolution ultrasonic inspection without the need for extensive calibration or manual waveform selection.