Used parts returned for remanufacturing ("cores") arrive worn, bent, pitted and coated in grease. Cameras miss defects hidden under contamination. Ultrasound penetrates it, but several probes firing at once interfere with each other, and each probe sees its own pattern of frequencies that the worn part rings at or swallows. Coded excitation can handle this: each probe transmits a long, specially designed pulse pattern, and the echoes are separated by correlating against the known code. Algebraic codes built from cyclotomy (splitting the numbers modulo a prime into a few interleaved classes using powers of a generator) come with provable guarantees: their echoes overlap with each other by at most about $\sqrt{N}$ for a code of length $N$, even when certain frequencies are forbidden. Those guarantees, however, assume every probe is forbidden from the same frequencies. On a worn core that assumption fails, because each probe position sees different resonances.
We propose to build and test code sets in which every probe gets its own forbidden-frequency mask. The masks are chosen per core from a short calibration ping, and the codes are synthesized algebraically in microseconds rather than by iterative optimization. Choosing the masks forces a trade-off across probes: no assignment makes every probe's interference small at once. We therefore make the aggregation rule across probes an explicit, one-parameter power mean. It ranges from "minimize the average interference" to "minimize the worst probe," and its exponent is fitted to downstream accept/reject outcomes rather than fixed by convention. In two sentences: give each ultrasonic probe a cyclotomic code that avoids that probe's own bad frequencies on this particular worn part, with provable interference bounds; then let the inspection outcome, not a designer's preference, decide how to trade one probe's interference against another's.
Core Claim
The open problem is stated directly in the closest existing work. Low Ambiguity Zone: Theoretical Bounds and Doppler-Resilient Sequence Design in Integrated Sensing and Communication Systems builds sequence sets from finite-field characters and cyclic difference sets that reach a maximum ambiguity magnitude of exactly $\sqrt{N}$ under an arbitrary spectral mask. That mask is identical for all users, and heterogeneous masks are listed as unsolved. Ultrasonic inspection of worn cores is a regime where heterogeneity is unavoidable. The role of an ontology-based knowledge backbone in a circular factory documents that incoming cores have unknown, non-uniform wear and deformation. The Non-linear Reflection Relationship Model of Remanufacturing Crankshaft Radial Runout Based on the Uncertainty Measure shows that geometric defects such as roundness and bending interact non-linearly, so the acoustic response of each probe location differs from part to part.
Our claim has three parts. First, a code set whose per-probe frequency supports are built from unions of cyclotomic classes, then punctured to avoid each probe's measured bad bands, keeps cross-probe interference within a small, provable factor of $\sqrt{N}$. Second, this preserves substantially more usable bandwidth, and therefore depth resolution, than the obvious fix of giving every probe the union of all probes' forbidden bands. Third, the cross-probe trade-off is better set by a power mean whose exponent is calibrated to inspection decisions than by the worst-case objective sequence designers default to.
The baseline guarantee is easy to state. For a flat-spectrum code supported on frequency set $S_k$, removing $r$ bins changes every correlation sidelobe by at most $r$, by the triangle inequality on the sum of unit-magnitude Fourier terms. The theoretical goal is to beat this trivial bound when the removed bins are chosen with respect to the cyclotomic class structure. The idea is that removing whole or partial classes changes sidelobes by Gauss-period sums, which are typically of order $\sqrt{r}$ rather than $r$.
Research Program
The construction works over a prime length $N = p$. The nonzero residues split into $e$ cyclotomic classes, which are cosets of the subgroup of $e$-th powers. Each class is an almost-difference set, and the cross-correlation between codes supported on different class unions is governed by cyclotomic numbers of order $e$. These are exactly computable, and the same machinery yields the linear-complexity results in Linear Complexity of Two Classes of Binary Generalized Cyclotomic Sequences of Any Order and the order-eight quaternary designs in Some new balanced and almost balanced quaternary sequences with low autocorrelation.
The per-core procedure has three steps. A short calibration ping estimates each probe's attenuation and resonance spectrum. Each probe's forbidden bins are then marked. Finally, each probe is assigned a class union that overlaps its forbidden bins as little as possible, and the remaining overlap is punctured. The phase function on the surviving support follows the character-based recipe of Ye et al. and of New Sequences Design From Weil Representation With Low Two-Dimensional Correlation in Both Time and Phase Shifts. The latter shows that combining multiplicative and additive characters controls correlation without the full Weil-representation machinery, which keeps synthesis to a closed-form evaluation.
The theory work proves a sidelobe bound for punctured class unions in terms of the number of removed bins and their distribution across classes, and characterizes when it improves on the trivial bound. We also bound the peak-to-mean power ratio, since piezoelectric drivers distort high-peak signals. For reference, New Construction of Optimal Type-II Binary Z-Complementary Pairs proves ratios between 2.57 and 4 for its pairs.
Assigning class unions to probes is a small combinatorial problem: $K$ probes, $e$ classes, and a few thousand candidate assignments. Its objective is where aggregation enters. Let $M_k$ be the peak normalized correlation sidelobe that probe $k$ suffers from all other probes and from its own delayed echoes. The assignment minimizes
$$P_\alpha = \Big(\frac{1}{K}\sum_{k=1}^{K} M_k^{\alpha}\Big)^{1/\alpha},$$
where $\alpha \ge 1$ interpolates between average interference ($\alpha = 1$) and worst-probe interference ($\alpha \to \infty$). This is the power mean used in Power mean pyramid scores for summarization evaluation to trade precision against recall. We reuse its one-parameter form deliberately.
The richer operators of Pythagorean Fuzzy Interaction Muirhead Means with Their Application to Multi-Attribute Group Decision-Making and Some q-Rung Dual Hesitant Fuzzy Heronian Mean Operators with Their Application to Multiple Attribute Group Decision-Making model multi-way interactions among expert-scored criteria. Their parameters are set subjectively, and the Muirhead form sums over $n!$ permutations. Their own example shows instability: switching operator variant at fixed $R=(1,1,1,1)$ flips the top airline from Transasia to Daily Air. With a few dozen labeled cores, a single exponent can be calibrated, whereas a Muirhead vector cannot. That is why we take the simplest member of the family.
The exponent is calibrated by grid search against the misrouting rate of the downstream gate on held-out cores. This follows the spirit of feedback-tuned aggregation in FALCON: Feedback Adaptive Loop for Content-Based Retrieval, where a continuous parameter moves the combination rule between fuzzy AND and fuzzy OR in response to relevance feedback. Here the feedback is whether the core was correctly accepted or rejected. The same family, with its own fitted exponent, fuses per-probe defect evidence into the accept/reject score. This addresses the gap that The Non-linear Reflection Relationship Model of Remanufacturing Crankshaft Radial Runout Based on the Uncertainty Measure leaves with its single-property cubic curve fit.
We do not claim the aggregation layer is essential. If $\alpha \to \infty$ is always best, we will report that and drop it.
Baselines
The naive reference is time-division firing with a standard complementary pair. Each probe fires alone, one after another, so cross-probe interference is zero by construction. The cost is inspection time, which scales with $K$. It is run on the same probes, cores and total acquisition budget. We beat it if our simultaneous scheme reaches equal flaw-detection rate at roughly $1/K$ of its acquisition time, or better detection at equal time through longer integration.
The strongest existing method is the common-mask construction of Ye et al., with the mask set to the union of all probes' forbidden bins on that core. It is the only cited construction with a proven bound under arbitrary spectral constraints, with maximum ambiguity magnitude $\sqrt{N}$ (Ye et al.). It is run at the same $N$, number of probes and calibration ping. We beat it if per-probe usable bandwidth, and hence measured depth resolution, is significantly larger while peak sidelobes stay within a stated constant factor of $\sqrt{N}$, and if this translates into higher detection of sub-surface flaws. For ultrasonic flaw detection on worn cores specifically, no reported result was found on Lacuna, so all numbers for this baseline will be our own.
A numerical baseline uses iterative sequence optimization: gradient descent on sidelobe energy under each probe's mask, given a matched compute budget per core. It is the natural competitor on raw sidelobe level. No cited paper in the file reports such a method for this setting.
Two ablations isolate the key ingredients. The first ablation removes the cyclotomic alignment by puncturing random flat-spectrum supports. This tests whether the class structure, rather than puncturing per se, gives the improved bound. The second fixes $\alpha \to \infty$, which tests whether calibrated aggregation matters.
For context on why acoustic sensing is worth the effort, the visual baseline on contaminated ball-screw spindles, Mask R-CNN on Industrial Machine Tool Component Surface Defect Dataset, reaches mIoU 0.316, with failures concentrated on small, early pitting and grease-covered regions.
Why not just…?
Why not just fire the probes one at a time? This is a real threat when throughput does not matter, and it is our naive baseline. The experiment that settles it is detection rate versus acquisition time. If remanufacturing lines can tolerate $K$-fold slower inspection, the heterogeneous construction buys nothing, and we will say so. The argument for simultaneity rests on inline gate speed, which Remanufacturing production planning and control: Conceptual framework for requirement definition frames as already stressed by volatile return volumes.
Why not just use the union of everyone's bad bands as one shared mask? This applies Ye et al. directly and inherits its $\sqrt{N}$ guarantee. It is also the most likely competitor to win if worn-core resonances turn out to overlap heavily across probe positions. The bandwidth-loss measurement decides it. We will report the overlap statistics of measured masks first. If the union removes only a few percent more bins than individual masks, this simpler method is adequate and the proposal narrows to the theory result.
Why not just optimize each probe's code numerically? Numerical design may reach lower sidelobes than any algebraic family. Its weaknesses are the lack of a guarantee for an unseen mask and per-core iterative cost. We will compare sidelobe level, peak-to-mean ratio and synthesis latency at matched compute. If numerical codes match ours within the latency budget of an inline gate, the algebraic route's remaining advantage is only the certificate, and we will present it that way.
Why not just feed raw echoes to a neural network and skip hand-built aggregation? With a few dozen labeled cores, a constrained model is safer. Schlagenhauf et al. note that labeled industrial defect data is scarce and costly. We will nonetheless include a small 1-D convolutional classifier on the same echoes. A fitted power mean that loses to it on held-out misrouting is not worth keeping.
Proposed Experiments
The first stage is purely numerical. We sweep primes $N$ from a few hundred to a few thousand, $K$ from 2 to 8 probes and class orders $e \in \{2,4,8\}$. Masks are drawn from resonance spectra simulated for shafts with varying bending and pitting depth. We report peak sidelobe divided by $\sqrt{N}$, retained bandwidth fraction, peak-to-mean ratio and synthesis time against the trivial puncturing bound and all baselines.
The second stage is a bench study. Shafts are worn on an accelerated-load rig modeled on the 12 kN multi-spindle bench of Schlagenhauf et al., and greased and soiled as in that dataset. Ground truth for surface pitting comes from cleaned camera images. Sub-surface ground truth comes from sectioning a subset after inspection. Bent and out-of-round shafts are measured with CMM roundness and dial-indicator bending, following the protocol of Liu. We cannot obtain his 20-batch data, so we will replicate the measurement scheme on our own parts and use downstream re-machining outcome as the accept/reject label.
Three checks decide the outcome. Heterogeneous masks must retain meaningfully more bandwidth than the union mask with sidelobes inside a fixed multiple of $\sqrt{N}$; otherwise the core claim fails. Cyclotomic-aligned puncturing must beat random puncturing on sidelobes; otherwise the algebra adds nothing. The calibrated exponent must lower held-out misrouting relative to both $\alpha \to \infty$ and the small neural classifier; otherwise the aggregation layer is dropped and the contribution is the heterogeneous-mask construction alone.
References
- Low Ambiguity Zone: Theoretical Bounds and Doppler-Resilient Sequence Design in Integrated Sensing and Communication Systems
- The role of an ontology-based knowledge backbone in a circular factory
- The Non-linear Reflection Relationship Model of Remanufacturing Crankshaft Radial Runout Based on the Uncertainty Measure
- Linear Complexity of Two Classes of Binary Generalized Cyclotomic Sequences of Any Order
- Some new balanced and almost balanced quaternary sequences with low autocorrelation
- New Sequences Design From Weil Representation With Low Two-Dimensional Correlation in Both Time and Phase Shifts
- New Construction of Optimal Type-II Binary Z-Complementary Pairs
- Power mean pyramid scores for summarization evaluation
- Pythagorean Fuzzy Interaction Muirhead Means with Their Application to Multi-Attribute Group Decision-Making
- Some q-Rung Dual Hesitant Fuzzy Heronian Mean Operators with Their Application to Multiple Attribute Group Decision-Making
- FALCON: Feedback Adaptive Loop for Content-Based Retrieval
- Industrial Machine Tool Component Surface Defect Dataset
- Remanufacturing production planning and control: Conceptual framework for requirement definition