Apartsin, Cooper, and Intrator (Journal of the Acoustical Society of America, 2012) proposed a biosonar-inspired method for time-of-arrival (ToA) estimation in the semi-coherent regime, where power- and spectrum-constrained narrowband pulses force receivers to operate below the signal-to-noise ratio at which the matched-filter maximum-likelihood estimator breaks down (the threshold effect). A bank of phase-shifted unmatched filters yields biased but only partially correlated delay estimators, and a regressor trained on their relative peak locations corrects the matched-filter estimate, outperforming the maximum-likelihood estimator near threshold in simulation. The paper's closing discussion named four remaining directions: improving the fusion with machine learning, explicit classification of outlier events, adapting the feature-space idea to signal detection in heavy noise, and evaluating behavior beyond the simulated narrowband setting. The first two directions were subsequently resolved by the same authors' 2013–2014 trilogy (IEEE Transactions on Geoscience and Remote Sensing, Parts I–II, and IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing), which introduced machine-learning fusion of the partially correlated estimators, explicit outlier-class classification, and energy-adaptive multi-pulse estimation. The surviving open core is the remaining two directions: whether the phase-shifted unmatched-filter feature space can support semi-coherent signal detection in heavy noise, and establishing the method's behavior under other waveform families, more complex noise conditions, and alternative feature constructions. No published work addresses either direction.

Background

Time-of-arrival (ToA) estimation — recovering the delay t0t_0 of a reflected pulse in a received signal — is the basis of ranging in radar, sonar, ultrasound, and underground exploration (the delay is often called time of flight, ToF, in those communities). The standard estimator correlates the received signal with a replica of the transmitted waveform — a matched filter, which implements the maximum-likelihood estimator (MLE) for the delay in white Gaussian noise — and takes the time of the correlation peak. For narrowband (weakly modulated sinusoidal) pulses, the correlation output has a main lobe surrounded by side lobes spaced by one carrier period. When the signal-to-noise ratio (SNR) falls below a critical level, noise occasionally promotes a side lobe of the waveform's autocorrelation (the correlation of the waveform with itself) to the global maximum of the filter output, so the peak-pick lands on a systematically displaced wrong value (an outlier) and the estimator's mean-squared error jumps abruptly. This threshold effect is classical in time-delay estimation (Zeira & Schultheiss, 1994). Because spectrum occupancy and power limits fix the pulse in several applications, receivers must sometimes operate near or below the threshold; the source paper calls this the semi-coherent state — the receiver still processes the known waveform coherently, but the estimate is dominated by noise-driven outlier events.

The source paper (Apartsin, Cooper & Intrator, 2012) approached this regime with a biosonar motivation: echolocating animals achieve fine ranging with low-power signals, and the group's earlier work studied multiple-ping data fusion and SNR-dependent filtering in low SNR (Apartsin & Intrator, 2008; Apartsin, Cooper & Intrator, 2010). Instead of a single matched filter, the method correlates the received signal with a bank of phase-shifted unmatched filters — replicas of the waveform whose carrier phase is rotated by ±φ\pm\varphi, constructed with the Hilbert transform (the quadrature component of a signal, used here to shift the carrier phase while leaving the envelope unchanged). Each filter's peak-pick is a biased ToA estimator. Under a three-point model that keeps only the main lobe and the two adjacent side lobes and describes the estimate by the probabilities p+1p_{+1}, p−1p_{-1} of picking each side lobe, the bias of the ±φ\pm\varphi estimator is

b±φ=2πω(p+1±φ−p−1±φ)±φω,b_{\pm\varphi}=\frac{2\pi}{\omega}\left(p_{+1}^{\pm\varphi}-p_{-1}^{\pm\varphi}\right)\pm\frac{\varphi}{\omega},

with ω\omega the carrier angular frequency, so opposite phase shifts give opposite biases. The relative peak locations t~i=t^i−t^0\tilde{t}_i=\hat{t}_i-\hat{t}_0 across the bank (the matched filter plus opposite phase-shifted pairs) form a feature vector; a regressor (a neural network with Gaussian radial-basis-function units, RBF, or support vector regression, SVR) trained on simulated data across near-threshold SNRs learns a corrective offset for the matched-filter estimate. In the paper's simulation — one narrowband waveform family in additive white Gaussian noise (AWGN), roughly 20,000 samples spanning ten SNR values and a 21-filter bank — the fused estimator outperforms the MLE in the presence of the threshold effect (a gain of up to about 10% in root-mean-square error is reported).

The same authors then published a trilogy that developed two of the directions named in the source paper's discussion. Part I (Apartsin, Cooper & Intrator, 2014, IEEE Transactions on Geoscience and Remote Sensing, TGRS) replaced the fixed regression with machine-learning combination of the "multiple experts" — the biased, partially correlated estimators produced by the phase-shifted unmatched-filter bank — for single-echo ToF estimation in the presence of outliers. Part II (2014, IEEE TGRS) added explicit outlier classification: a weak classifier estimates the probabilities of appearance of specific outlier classes in a single-echo estimate, and an optimal weighted averaging scheme combines bias-corrected estimates from multiple echoes while accounting for inlier and outlier uncertainties. A companion paper (2014, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, JSTARS) used the classifier information to adaptively control the number of transmitted pulses, reaching MLE-level accuracy with significantly fewer pulses, and thus less energy, on average. All of these remain estimation results; the program's summary report (Cooper & Intrator, 2013) likewise covers single-echo estimation, multi-pulse estimation, and energy efficiency only.

Two directions from the source paper's discussion were never picked up. First, the suggested adaptation of the feature-space idea from ToA estimation to semi-coherent signal detection in heavy noise — deciding whether an echo or target is present, not how far away it is. Second, the method's behavior outside the original demonstration: every paper in the line is a simulation study on a narrowband waveform family in AWGN, and no systematic evaluation under other waveform families, more complex noise conditions, or alternative feature constructions exists. Later citing work applies related machine-learning and kernel ideas in neighboring measurement settings (e.g., kernel-based outlier rejection for underwater acoustic localization, Pinheiro et al., 2017; maximum-length-sequence ultrasonic temperature measurement, Zhou et al., 2016) without addressing either question.

Problem Statement

The source paper's discussion suggested extending its phase-shifted unmatched-filter feature space beyond the demonstrated simulation setting. Since the authors' 2013–2014 trilogy resolved the machine-learning-fusion and outlier-classification directions, the following derived problem addresses the two directions that remain open. The common objective is to establish whether the unmatched-filter-bank feature space is a viable general basis for semi-coherent echo processing — for deciding signal presence as well as estimating delay, and under conditions beyond the single simulated narrowband demonstration. Concretely:

(1) Detection extension. Construct and characterize a signal detector built on the phase-shifted unmatched-filter feature space (the pattern of biased peak locations relative to the matched-filter peak, and/or statistics of the filter-bank outputs) for deciding whether a narrowband echo is present in heavy noise, in the semi-coherent regime that power- and spectrum-constrained transmission imposes. Determine whether such a feature-space detector improves detection performance — for example, detection probability at a fixed false-alarm rate across an SNR sweep through the threshold region — over the coherent matched-filter detector and over two conventional baselines — one non-coherent, using signal energy and discarding carrier phase (for example, an energy or square-law detector), and one using the known carrier phase but not the echo delay — under those constraints, or establish that it cannot. Here 'improves detection performance over the baselines' means attaining detection probability at least as high as that of every baseline throughout a common SNR interval and strictly higher on a non-degenerate sub-interval of that interval.

(2) Generalization beyond the original setting. Establish the behavior of the unmatched-filter-bank estimation methodology (regression- or classifier-based fusion) across all three of the following axes — waveform families other than the single simulated narrowband family, noise conditions beyond additive white Gaussian noise (for example, colored, reverberant, or impulsive noise), and alternative feature constructions — examining at least one case per axis and quantifying, in each regime examined, the accuracy relative to the matched-filter MLE across the threshold SNR sweep.

References

  1. Apartsin, A., Cooper, L. N., & Intrator, N. (2012). Semi-coherent time of arrival estimation using regression. The Journal of the Acoustical Society of America, 132(2), 832–837. DOI: 10.1121/1.4730885. https://doi.org/10.1121/1.4730885
  2. Apartsin, A., & Intrator, N. (2008). A data fusion and multiple ping method for improving the resolution of low-power acoustic and seismic sensing (Acoustical Society of America meeting abstract). The Journal of the Acoustical Society of America, 124(4, Supplement), 2597. DOI: 10.1121/1.4783265. https://doi.org/10.1121/1.4783265
  3. Apartsin, A., Cooper, L. N., & Intrator, N. (2010). SNR-dependent filtering for Time Of Arrival estimation in high noise. 2010 IEEE International Workshop on Machine Learning for Signal Processing, 427–431. DOI: 10.1109/MLSP.2010.5588848. https://doi.org/10.1109/MLSP.2010.5588848
  4. Apartsin, A., Cooper, L. N., & Intrator, N. (2014). Time-of-Flight Estimation in the Presence of Outliers Part I—Single Echo Processing. IEEE Transactions on Geoscience and Remote Sensing, 52(6), 3382–3392. DOI: 10.1109/TGRS.2013.2272737. https://doi.org/10.1109/TGRS.2013.2272737
  5. Apartsin, A., Cooper, L. N., & Intrator, N. (2014). Time-of-Flight Estimation in the Presence of Outliers. Part II—Multiple Echo Processing. IEEE Transactions on Geoscience and Remote Sensing, 52(7), 3843–3850. DOI: 10.1109/TGRS.2013.2276919. https://doi.org/10.1109/TGRS.2013.2276919
  6. Apartsin, A., Cooper, L. N., & Intrator, N. (2014). Energy-Efficient Time-of-Flight Estimation in the Presence of Outliers: A Machine Learning Approach. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(4), 1306–1313. DOI: 10.1109/JSTARS.2013.2295324. https://doi.org/10.1109/JSTARS.2013.2295324
  7. Cooper, L. N., & Intrator, N. (2013). Time of Flight Estimation in the Presence of Outliers: A Biosonar-Inspired Machine Learning Approach (Defense Technical Information Center, DTIC, report ADA597037). DOI: 10.21236/ada597037. https://doi.org/10.21236/ada597037
  8. Zeira, A., & Schultheiss, P. M. (1994). Realizable lower bounds for time delay estimation. 2. Threshold phenomena. IEEE Transactions on Signal Processing, 42(5), 1001–1007. DOI: 10.1109/78.295217. https://doi.org/10.1109/78.295217
  9. Pinheiro, B. C., Moreno, U. F., de Sousa, J. T. B., & Rodriguez, O. C. (2017). Kernel-Function-Based Models for Acoustic Localization of Underwater Vehicles. IEEE Journal of Oceanic Engineering, 42(3), 603–618. DOI: 10.1109/JOE.2016.2578218. https://doi.org/10.1109/JOE.2016.2578218
  10. Zhou, C., Wang, Y., Qiao, C., Zhao, S., & Huang, Z. (2016). High-accuracy ultrasonic temperature measurement based on MLS-modulated continuous wave. Measurement, 88, 1–8. DOI: 10.1016/j.measurement.2016.03.037. https://doi.org/10.1016/j.measurement.2016.03.037
  11. Aiordăchioaie, D. (2025, online; print issue 2026). Some Aspects of Time Delay Estimation Based on Cross-Correlation. Circuits, Systems, and Signal Processing, 45(3), 2139–2163. DOI: 10.1007/s00034-025-03323-2. https://doi.org/10.1007/s00034-025-03323-2