Machines that rotate, radios that transmit, and hearts that beat all produce signals with hidden rhythm. Their statistics repeat at a characteristic rate even when the raw waveform is buried in noise. Engineers already exploit this rhythm, known as cyclostationarity, to pull weak fault signatures out of heavy noise. They then hand the cleaned feature to a classifier or a hand-set threshold. When such a system watches a machine continuously for weeks, nobody can say what fraction of its alarms are false. This proposal builds a monitor that can make that guarantee. Its key move is that a periodic signal supplies its own null reference. In each time window, we compare the strength of the rhythm at the fault's predicted repetition rate against the same quantity measured at many nearby rates where no fault rhythm should exist. That comparison yields a valid p-value without any healthy-machine calibration data. We then combine evidence across harmonics and sensors, and feed the per-window p-values to an online procedure that bounds the expected fraction of false alarms over the whole stream.
In two sentences: the null distribution for "is there a fault rhythm at rate α?" can be read off the same window at rates where no rhythm is expected, so p-values recalibrate themselves as the machine's noise drifts. Stacking these p-values across harmonics and sensors, then applying online false discovery rate (FDR) control, turns cyclostationary diagnostics from an uncalibrated classifier into an alarm stream with a stated error budget.
Core Claim
The research file shows a clean gap. Cyclostationary feature work couples its features to uncalibrated decision rules. Liu et al. pair blind deconvolution with a BiLSTM classifier (An Optimized Maximum Second-Order Cyclostationary Blind Deconvolution and BiLSTM Model), and HiNoVa thresholds Kendall's τ between device histograms (HiNoVa). The p-value stacking and FDR literature has worked only on static image benchmarks. MLOD and Combine and Conquer fuse layer or detector p-values. eg-BH applies Benjamini–Hochberg thresholding, a rule that rejects tests whose ranked p-values fall under a rising line. No paper in the file links cyclostationary features to FDR control at all.
Simply chaining these pieces would inherit the weakness Ma et al. proved, which they call the calibration crunch. When p-values come from a held-out calibration set of size $m$, detection power is Beta-distributed and collapses when $m$ is small. On a real machine the problem is worse, because the healthy reference drifts with load, speed and temperature. Our claim is that cyclostationary structure dissolves this bottleneck. Off-cycle frequencies in the current window provide a fresh, local null sample at every step. The result is a three-tier monitor: within-window off-cycle p-values, then dependence-robust stacking across harmonics and sensors, then online FDR across windows. It should hold its stated false-alarm proportion under drift where healthy-calibrated pipelines do not.
Research Program
The building block is the cyclic spectral coherence estimate $S_t(\alpha)$. It measures how strongly the signal in window $t$ correlates with a copy of itself modulated at cyclic frequency $\alpha$, the rate at which the signal's statistics repeat. A bearing fault produces impulses at a kinematically predicted rate, such as the outer-race defect frequency, so $S_t$ peaks there. Stationary noise gives no peak, following the cross-term cancellation argument in the file's section 1.1.1.
Speed wanders and fault rates slip. We therefore score a candidate as the maximum of $S_t$ over a small band $B_k$ around the $k$-th predicted fault harmonic. We compute the identical band-maximum over a reference set $\mathcal{R}_t$ of equally wide bands. These bands are placed away from all known kinematic rates, meaning shaft harmonics and other fault families. The within-window p-value is
$$p_{t,k} = \frac{1 + \#\{B \in \mathcal{R}_t : \max_{\alpha \in B} S_t(\alpha) \ge \max_{\alpha \in B_k} S_t(\alpha)\}}{1 + |\mathcal{R}_t|}.$$
This is exactly valid when the band statistics are exchangeable under "no fault." That holds approximately for stationary or slowly varying background noise. Using the same band-max operation on both sides keeps the frequency search from inflating significance.
Any leaked structure in the reference bands makes p-values larger, not smaller. The expected failure mode is therefore conservativeness, which is the regime ADDIS was designed for. The real risk is unmodeled periodic interference landing near a fault band, which would make p-values anti-conservative. Characterizing when that happens is a first-class deliverable.
The second tier stacks the harmonics $k$ and the sensor channels $c$ into one p-value per window. These tests are strongly dependent because they share one impulse train. Fisher's method assumes independence, so we use the Cauchy combination statistic, which Li et al. evaluated in MLOD as valid under arbitrary dependence. The statistic is $T_t = \sum_{k,c} w_{k,c} \tan\big((0.5 - p_{t,k,c})\pi\big)$, with weights $w_{k,c}$ summing to one, converted back to a p-value through the standard Cauchy tail. Brown's correction, as used in Combine and Conquer, is a comparison arm.
The third tier treats each window as one hypothesis arriving online, "the machine is faulty in this window." It allocates an error budget with ADDIS, setting the target as the expected fraction of false alarms among all alarms raised. Successive windows overlap in their physical cause but not in their samples. We use non-overlapping windows separated by a short gap, which matches the local-dependence setting under which Tian and Ramdas prove FDR control. We also study the sensitivity to ADDIS's candidate and discard thresholds, which they flag as a known weakness.
The program also has a theory track. We will state conditions on the background process under which off-cycle band statistics are asymptotically exchangeable. Candidate conditions are short-memory noise and bounded cyclic leakage between bands separated by more than a spectral-resolution width. From these we derive the finite-$|\mathcal{R}_t|$ analogue of Ma et al.'s Beta result. Because $|\mathcal{R}_t|$ is fixed by bandwidth rather than by how much healthy data was recorded, power should stop depending on calibration budget. That prediction is directly testable.
Baselines
The naive reference point is the classical envelope-spectrum alarm. It computes $S_t$ at the fault band and raises an alarm when it exceeds a fixed multiple of its median on an initial healthy segment, one alarm rule per window, with no multiplicity control. This is roughly how condition monitoring is done in practice. It defines the false-alarm proportion that an uncalibrated system actually produces.
The strongest statistical competitor applies the eg-BH recipe (A Closer Look at Generalized BH) to the same features. Per-window statistics are converted to p-values against a healthy calibration archive, with the archive split into $L$ parts and the p-values pooled by a generalized mean. Windows then pass through the same ADDIS stage, so this arm differs from ours only in where the null comes from. It is the strongest competitor because Ma et al. built it specifically to fix small-calibration instability. On images, they report F1 rising from 59.34% to 73.25% with Energy scores on TinyImageNet. We will give it the same features, the same combination rule, the same online procedure, and calibration archives ranging from minutes to hours of healthy data. We beat it if, at matched realized FDR, our power is higher under drift and small archives. We must also stay within the nominal FDR level where eg-BH exceeds it after a speed or load change.
The strongest diagnostic competitor is CYCBD-BiLSTM from Liu et al. On the Case Western Reserve University (CWRU) bearing benchmark, they report 98.20% diagnostic accuracy at −9 dB SNR, about 9.89 points above their baselines. We will reproduce it on our streams, convert its per-window fault probability into a calibrated score, and pass it through both a fixed threshold and eg-BH plus ADDIS. That gives the learned feature a fair chance at error control. Accuracy on balanced segments is not FDR on a long, mostly healthy stream. Beating it means a lower realized false-alarm proportion at comparable detection delay, without fault-labeled training data. No reported result was found on Lacuna for streaming FDR or detection delay on CWRU.
Three ablations isolate the ingredients. The first replaces off-cycle p-values with healthy-calibrated ones, which is the eg-BH arm above. The second replaces Cauchy stacking with Fisher's method to expose inflation from dependent harmonics. The third replaces ADDIS with LORD++, SAFFRON and a per-window Bonferroni-style threshold to measure the value of handling conservative nulls. A feature ablation swaps cyclic coherence for broadband kurtosis inside the full statistical pipeline. This shows whether the cyclostationary statistic, rather than the error-control machinery, carries the power.
Why not just…?
Why not just calibrate on healthy data and run Benjamini–Hochberg or ADDIS? This is the real threat. On a stationary test rig it may work as well as our method, since CWRU recordings have fixed load conditions. The case for off-cycle nulls rests entirely on drift and scarce healthy data. Ma et al.'s Beta-distribution result predicts power instability when the calibration set is small. Nothing in the file predicts how healthy calibration behaves when the background noise changes after calibration. Our experiments therefore include controlled drift: resampled speed changes, changes in the level of added noise, and switches between load conditions. If healthy calibration matches us on these too, the core claim fails, and we will report it as such.
Why not just train a strong classifier like CYCBD-BiLSTM and threshold its output? Liu et al. show high accuracy at −9 dB. However, they need labeled fault data, they tune cyclic frequency and filter length with a metaheuristic, and they note that the multi-stage preprocessing can propagate errors. A softmax threshold offers no statement about what fraction of a month's alarms are false. Our comparison passes the classifier through the same error-control stage, so if learned features are simply better, that will show up as higher power at equal FDR.
Why not just fuse everything with Fisher's method, as MLOD and JTLA do? Harmonics of one impulse train are not independent. Fisher's method then double-counts evidence and inflates false alarms. MLOD and JTLA operate on network layers, where this matters less for ranking metrics like AUROC than it does for a stated error rate. The Fisher ablation measures the inflation directly on healthy streams.
Decisive Experiments
The primary testbed is a set of long monitoring streams assembled from CWRU recordings, the benchmark used by Liu et al. Each stream is mostly healthy segments, with fault segments of each defect type and size inserted at random onsets. White and colored noise is added down to −9 dB to match Liu et al.'s hardest setting. Speed drift is produced by time-resampling within the range CWRU motor speeds span. Ground truth comes from the insertion schedule.
A synthetic suite generates cyclostationary impulse trains in controlled noise. It also injects deliberate interference tones near fault bands to map where off-cycle p-values become anti-conservative. As a transfer check, we use the LoRa and WiFi IQ datasets of HiNoVa. The question there is "this window comes from an unenrolled transmitter," with cyclic features at the symbol and clock rates; HiNoVa's reported LoRa AUPRC of 0.95 is the reference.
The metrics are realized FDR averaged over many stream replicates, against nominal levels of 5% and 10%. We also measure power as the fraction of faulty windows flagged, and detection delay after onset. Healthy-only streams measure the false alarms per hour of a purely null process, the cleanest validity check.
Three results decide the project. The first is a calibration-budget sweep: our power should stay flat as healthy data shrinks, while eg-BH's power should fall, as Ma et al.'s theory predicts. The second is the drift test: eg-BH's realized FDR should exceed nominal after a load or speed change, while ours stays at or below nominal. The third is the interference map: it bounds how close an unmodeled periodic tone can sit to a fault band before validity breaks. That bound determines whether the method is safe to deploy, or needs a reference-band screening step that we would then add and re-evaluate.
References
- An Optimized Maximum Second-Order Cyclostationary Blind Deconvolution and Bidirectional Long Short-Term Memory Network Model for Rolling Bearing Fault Diagnosis
- HiNoVa: A Novel Open-Set Detection Method for Automating RF Device Authentication
- Enhancing Out-of-Distribution Detection with Multitesting-based Layer-wise Feature Fusion
- Combine and Conquer: A Meta-Analysis on Data Shift and Out-of-Distribution Detection
- A Closer Look at Generalized BH Algorithm for Out-of-Distribution Detection
- ADDIS: an adaptive discarding algorithm for online FDR control with conservative nulls.
- A General Framework For Detecting Anomalous Inputs to DNN Classifiers