Source-linked AI summary
Real-time and adaptive anomaly detection algorithm for cyclostationary models
Justyna Witulska, Tomasz Barszcz, Ireneusz Jabłoński, Agnieszka Wyłomańska
TL;DR
Cyclostationary signals challenge real-time anomaly detection because phase-dependent variability can be mistaken for abnormality. PeriodicCALM extends CALM with cycle-aware envelopes and filtering, improving detection reliability while retaining real-time efficiency and supporting compressor-vibration monitoring.
Problem
Existing real-time anomaly detection methods may misclassify natural cyclic variations in cyclostationary signals because CALM assumes independent observations.
Method
PeriodicCALM extends CALM with cyclic envelope construction and filtering that model cycle-dependent variability and distinguish normal cyclic behavior from genuine anomalies.
Results
Precision increased from 0.094 to 0.974 and F1-score from 0.171 to 0.986, while computational overhead remained practically negligible relative to CALM.
Takeaways & Limitations
PeriodicCALM provides an effective, computationally efficient approach for real-time anomaly detection in cyclostationary data, including compressor vibration signals.
Takeaways & Limitations
The current implementation requires the cycle length T to be specified by the user, and further work is needed for data drift, outages, missing transmissions, and evolving distributions.
Abstract
from arXiv · showhide
This article introduces PeriodicCALM, an effective real-time anomaly detection framework designed for cyclostationary data streams. While classical cyclostationary processes feature periodically time-varying statistical properties, real-world signals often contain recurring impulsive components that conceal abnormal behavior. Existing real-time methods for struggle with these dynamics, frequently misinterpreting phase-dependent variability as non-cyclic anomalies and causing excessive false alarms. To address this, PeriodicCALM incorporates cycle-dependent variability to systematically ignore regular cyclic impulses while accurately isolating genuine anomalies. Operating in real time with continuous retraining capabilities, the method adapts dynamically to evolving signal characteristics. Comparative evaluations against the baseline CALM framework using simulated data demonstrate significant improvements in detection accuracy and training efficiency, alongside a reduction in prediction latency. Furthermore, the practical utility of PeriodicCALM is validated on real-world vibration signals collected from a compressor monitoring system.
1. Introduction
PeriodicCALM targets real-time, unsupervised anomaly detection for cyclostationary signals whose periodic dependencies and disturbances challenge conventional methods. It extends CALM with cycle-aware modeling to reduce false alarms while preserving computational efficiency.
- Cyclostationary processes have statistical characteristics, including the mean or autocovariance function, that vary periodically over time and arise in monitoring, telecommunications, and physiological applications.
- CALM’s independence assumption can misclassify natural oscillations in cyclostationary signals as anomalies, increasing false positive rates.
- PeriodicCALM addresses the need for computationally efficient, unsupervised real-time anomaly detection that accounts for cyclostationary signals’ temporal dependencies and periodic structure.
- PeriodicCALM extends CALM with cyclic envelope construction and filtering to distinguish normal cyclic behavior from genuine anomalies while retaining the baseline’s simplicity and computational efficiency.
- The study evaluates PeriodicCALM on simulated cyclostationary data and demonstrates its application to real compressor vibration data.
2. Cyclostationary model with additive outliers
The paper models cyclostationary signals with additive impulses using PAR and compressor-vibration formulations, distinguishing their periodic mean and autocovariance structures. Simulations illustrate how impulsive contamination affects these periodic behaviors.
- Additive disturbances are modeled through i.i.d. impulse amplitudes whose amplitude range is controlled by a and b, while p controls occurrence frequency.
- Cyclostationary models have mean and autocovariance functions that vary periodically, with period T defined by the smallest integer satisfying these conditions.
- The PAR model has constant zero mean and periodic autocovariance, whereas the compressor-vibration model has periodic mean and time-independent autocovariance.
- In the PAR model, impulse signs follow sign(X_t), allowing positive or negative anomalies; in the compressor model, disturbances are restricted to positive additions.
- The PAR simulation uses periodic coefficients and additive outliers, while the compressor simulation compares clean and contaminated trajectories and their Monte Carlo variance or mean functions.
- For compressor vibrations, periodic oscillations arise from rotating components, with the modeled cycle period set to T_p = 1190.
3. Methodology for anomaly detection in cyclostationary signals
PeriodicCALM extends real-time CALM anomaly detection with a cyclic envelope that models phase-dependent variability in cyclostationary signals. It filters candidate detections through this envelope while retaining the baseline’s real-time workflow and supporting univariate-to-multivariate extension.
- Enhanced methodology for cyclostationary signals: PeriodicCALM retains CALM’s real-time anomaly-scoring and thresholding procedure while adding periodic-envelope construction and utilization during training and prediction.Figure 3 distinguishes baseline CALM steps from the PeriodicCALM extensions.
- Enhanced methodology for cyclostationary signals: The method applies baseline CALM to generate candidate anomalies, then validates each candidate against the cyclic envelope to retain only points outside normal cyclic bounds.Points inside the envelope are reclassified as false positives representing normal cyclic behavior.
- Enhanced methodology for cyclostationary signals: The procedure is presented for univariate signals, with multivariate generalization obtained by applying it independently to components and combining per-feature labels.The univariate formulation isolates the cyclic-envelope filtering step from multivariate label-combination complexity.
- Enhanced methodology for cyclostationary signals: PeriodicCALM adds a cyclic envelope to CALM, modeling phase-dependent variability so normal cyclic oscillations can be separated from genuine anomalies.The envelope is designed without explicit seasonal decomposition or preprocessing.
- Enhanced methodology for cyclostationary signals: The envelope is constructed from sliding-window local quantiles, training-data variance adjustment, and averaging bounds across complete cycles.Cycle averaging aligns the bounds by phase position within a known cycle.
- Enhanced methodology for cyclostationary signals: PeriodicCALM is intended to reduce false alarms in periodic data while preserving computational simplicity and real-time capability for industrial monitoring.The paper identifies rotating machinery, compressors, and vibration-based systems as relevant application settings.
4. Simulation study
The simulation study compares PeriodicCALM with CALM across two periodic-signal models and varied impulse configurations, showing that cyclic-envelope filtering reduces false positives while preserving detection of genuine anomalies. PeriodicCALM is especially beneficial for strongly dependent signals and remains effective across contamination levels and amplitudes.
- Performance assessment: CALM often labels regular cyclic peaks as anomalies, whereas PeriodicCALM filters these false positives by modeling cycle-dependent variability.Illustrative examples for both models show the baseline producing numerous false positives corresponding to normal cyclic peaks.
- Performance assessment: PeriodicCALM substantially reduces false positives compared with CALM while preserving high recall for genuine non-cyclic impulses.Mean Δ for true positives was 0 for almost all impulse configurations.
- Performance assessment: PeriodicCALM’s improvement is less pronounced for Model 1, suggesting that CALM can handle signals with weaker dependencies.The stronger gains occur for Model 2, where dependence is more pronounced.
- Performance assessment: Similar performance across impulse probabilities and amplitude ranges indicates robustness to varying anomaly occurrence and magnitude.The simulations varied impulse probabilities by model and covered six amplitude ranges defined relative to D.
- Runtime computational complexity and memory: Training dominates runtime and peak memory, while envelope construction and rule-based filtering each contribute less than one second.Computational cost depends mainly on training data spanning several complete cycles rather than total sample length.
5. Application to real data
PeriodicCALM was evaluated on vibration records from a reciprocating compressor, distinguishing normal periodic impulses from non-periodic anomalous events. Compared with CALM, it reduced false detections while retaining regime-dependent anomalies.
- The compressor data came from a condition-monitoring system on a 2-MW reciprocating compressor used to raise natural-gas pressure for pipeline transport.The records were collected by the same sensor at different time intervals.
- The nine 1-s vibration records ranged from normal periodic impulses to progressively increasing numbers of non-periodic impulses, creating increasingly difficult evaluation cases.Signals #1 and #2 contained only periodic impulses, while Signals #3–#9 included anomalous non-periodic impulses.
- PeriodicCALM isolated genuine non-cyclic disturbances while classifying normal periodic oscillations as normal compressor behavior.Figure 6 shows the PeriodicCALM detections, whereas Figure 7 presents the corresponding baseline CALM result.
- Unlike a global threshold, PeriodicCALM detected a lower-amplitude anomaly in Signal #4 by accounting for local signal characteristics.The threshold example used a value of 2.5 and missed the anomaly near 0.75 s because its magnitude stayed below the global threshold.
- The real-data results support PeriodicCALM’s suitability for continuous monitoring systems involving cyclic data, including industrial predictive maintenance.
6. Summary and conclusions
PeriodicCALM extends CALM for real-time anomaly detection in cyclostationary signals by modeling cyclic variability. Experiments show improved detection reliability and low computational overhead, while several adaptation and periodicity limitations remain.
- PeriodicCALM extends CALM for dependent cyclostationary data by exploiting repetitive structure to separate normal cyclic variation from genuine non-cyclic anomalies.
- Precision increased from 0.094 to 0.974 and F1-score from 0.171 to 0.986 with PeriodicCALM’s cyclic envelope filtering.The filtering suppressed false alarms from normal periodic oscillations without compromising detection of genuine non-cyclic anomalies.
- Additional cyclic-envelope construction and filtering caused only a slight, practically negligible increase in runtime and memory cost relative to CALM.This computational profile supports real-time monitoring use.
- The current implementation requires the user to specify the cycle length T, and future work targets automatic estimation from data.Proposed estimation routes include spectral analysis and autocorrelation.
- Future development should address data drift, sensor outages, missing transmissions, evolving distributions, and multiple or non-integer periodicities.
- The methodology was validated for cyclostationary signal processing but may extend to other real-time anomaly-detection problems involving dependent, temporally structured data.