⚡ electrical engineering

Unsupervised Vibroacoustic Anomaly Detection for End-of-Line Testing of High-Variation Hydraulic Machinery

This paper proposes ASD-BOD, an unsupervised framework combining Robust Principal Component Analysis and weighted cophenetic correlation-based clustering to effectively detect vibroacoustic anomalies in high-variation hydraulic machinery during End-of-Line Testing, significantly improving detection performance without requiring product-specific re-tuning.

Original authors: Maximilian Romeser, Reinhold von Schwerin

Published 2026-08-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Maximilian Romeser, Reinhold von Schwerin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern factory, the goal is no longer just to make millions of identical parts, but to build thousands of slightly different machines tailored to specific needs. This shift toward high-mix, low-volume production means that a single factory might churn out dozens of unique variations of a hydraulic pump, each with its own size, pressure settings, and control mechanisms. For the engineers responsible for quality control, this creates a difficult puzzle. They must ensure that every single unit works perfectly before it leaves the factory, but the sheer variety of products means there is no single "standard" sound or vibration to compare against. If a machine is built differently, its normal operating noise changes, making it nearly impossible to tell if a strange vibration is a sign of a broken part or simply the result of a different design. This challenge is particularly acute for axial piston units, powerful hydraulic pumps used in everything from construction excavators to agricultural tractors, where a hidden defect in a bearing can lead to catastrophic failure in the field.

To solve this, researchers Maximilian Romeser and Reinhold von Schwerin developed a new way to listen to these machines that does not rely on knowing what a "perfect" machine sounds like in advance. Instead of trying to build a universal model that accounts for every possible variation, they treated each group of machines being tested at the same time as its own unique family. Their method, called ASD-BOD, operates on the simple but powerful idea that within a single batch of identical units, the healthy machines will all sound very similar to one another, while the few faulty ones will stand out as rare outliers. By focusing on these groups rather than the entire history of production, the system can ignore the background noise that changes from one product design to the next and focus only on the strange, localized sounds that indicate a problem.

The core of their approach involves a mathematical technique that separates the common background hum from the unique, suspicious signals. Imagine a room full of people talking; the technique can isolate the general murmur of the crowd and the specific, sharp sound of a single person shouting, even if the background noise is loud and chaotic. In the context of the hydraulic pumps, this "background hum" includes the natural roar of fluid flowing through pipes and the vibrations caused by the test equipment itself. The researchers used a method called robust principal component analysis to strip away these shared, process-induced noises, leaving behind a clean signal that highlights only the specific, irregular vibrations caused by defects. This allows the system to detect subtle faults, such as damage to a rolling bearing, that would otherwise be completely masked by the machine's normal operation.

The team tested this framework using real-world data from axial piston units, deliberately introducing faults into some of the bearings to see if the system could find them. They simulated a production environment where the machines were tested under different speeds and pressures, creating a scenario where the data shifted constantly. The results were striking. Without their new method, the system struggled to distinguish between a healthy machine and a faulty one, often getting confused by the natural variations in the data. However, once they applied their noise-filtering technique, the system's ability to spot the bad units improved dramatically. The accuracy of the detection, measured by how well it could separate the good units from the bad ones, jumped from a score of roughly 0.4 to 0.8. More importantly, the number of false alarms—where a perfectly good machine was wrongly flagged as broken—dropped from about 6 percent to just 0.23 percent.

A key innovation in their work was how they handled the size of the groups being tested. In a real factory, the number of units in a batch can vary wildly, from a handful to dozens. Many existing methods fail when the group size changes because the rules for what counts as "strange" shift along with the numbers. The researchers created a new scoring system that remains consistent regardless of how many units are in the batch. This means that a factory can switch from testing small batches to large ones without needing to recalibrate the software or rewrite the rules. The system automatically adjusts, ensuring that the threshold for flagging a defect stays reliable whether the batch contains five units or thirty. This adaptability is crucial for the high-mix, low-volume manufacturing environment, where flexibility is the most valuable asset.

The researchers also explored different ways to measure how different the machines sounded from one another. They found that simply counting the total energy of the strange vibrations was more effective than trying to analyze the complex shape of the sound waves in detail. By focusing on the total amount of unusual energy after the background noise was removed, the system could identify the faulty units with greater precision. Visualizing the data showed that the method successfully isolated the faulty units, separating them from the healthy ones in a clear, logical pattern. Even in difficult cases where a healthy machine happened to be noisier than usual, the system correctly identified it as part of the healthy group, whereas older methods would have mistakenly flagged it as a defect.

This work demonstrates that it is possible to achieve high-quality fault detection in complex, variable production environments without needing vast amounts of historical data or labeled examples of broken machines. By treating each batch as a self-contained group and using advanced signal processing to filter out the inevitable noise of the manufacturing process, the researchers have created a tool that is both robust and adaptable. The method does not require engineers to know the specific frequencies of every possible fault or to tune the system for every new product variant. Instead, it learns the normal behavior of the current group of machines and instantly spots the anomalies. This approach offers a practical path forward for industries moving toward mass customization, ensuring that even in a world of endless variety, the quality of every single product can be guaranteed.

Technical Summary: Unsupervised Vibroacoustic Anomaly Detection for End-of-Line Testing of High-Variation Hydraulic Machinery

Problem Statement
The shift toward Industry 4.0 has driven manufacturing toward High-Mix, Low-Volume (HMLV) regimes, characterized by extensive product customization and small batch sizes. This trend complicates End-of-Line Testing (EoLT) for rotating machinery, specifically Axial Piston Units (APUs) used in mobile hydraulics. Traditional rule-based or supervised machine learning approaches struggle in this environment due to four primary challenges:

Weak Fault-to-Noise Ratio: Strong, operation-dependent background noise (e.g., turbulent flow, pressure pulsations) masks subtle fault signatures like bearing defects.
Variant-Specific Characteristics: High product variance leads to distinct structural dynamics per variant, preventing the use of universal diagnostic models.
Limited Fault Data: The rarity of actual faults in production restricts the availability of labeled data required for supervised learning.
Domain Shift Sensitivity: Systematic changes in statistical properties (domain shifts) caused by product variants, test rig differences, or procedural variations degrade models trained on specific batches, leading to high rates of "pseudo-anomalies" (false positives).

Existing unsupervised methods often lack robust preprocessing to handle hydraulic noise or rely on fixed thresholds that fail to transfer across varying batch sizes.

Methodology: ASD-BOD Framework
The authors propose ASD-BOD (Adaptive Spectral Decomposition-based Batch-wise Outlier Detection), an unsupervised framework that treats each production batch as an independent, localized domain. The framework operates on the assumption that while significant distribution shifts occur between batches, conditions within a single batch are relatively stable. The process consists of four sequential stages:

Data Aggregation and Preprocessing: Raw vibroacoustic time-series signals are normalized, filtered, and converted into spectral representations using Fast Fourier Transform (FFT) and computed order tracking. These are aggregated into a batch-level spectral matrix $M \in \mathbb{R}^{n \times m}$ , where rows represent units and columns represent frequency bins.
Adaptive Spectral Decomposition (ASD) via RPCA: The core innovation applies Robust Principal Component Analysis (RPCA) to decompose the spectral matrix $M$ $M$ into a low-rank matrix $L$ $L$ and a sparse matrix $S$ $S$ ( $M = L + S$ $M = L + S$ ).
- $L$ (Low-Rank): Captures shared background noise and batch-common structures (e.g., pressure pulsations, structural resonances).
- $S$ (Sparse): Isolates unit-specific anomalies (e.g., bearing faults).
- This data-driven approach requires no prior knowledge of fault frequencies or system dynamics, automatically suppressing process-induced variations while preserving deterministic operational features.
Dissimilarity Indication: The framework computes dissimilarity between units based on the sparse matrix $S$ . Two indicators are evaluated: a Sparsity Indicator (SI) (sum of absolute values in $S$ ) and the Wasserstein distance (preserving spectral morphology).
Clustering-Based Outlier Detection (OD): A hierarchical clustering algorithm is applied to the dissimilarity matrix. To address HMLV challenges, the authors introduce two key modifications to standard clustering:
- Weighted Cophenetic Correlation (wCC): A modified recursive partitioning algorithm uses a weighted cophenetic correlation coefficient. This ensures that the clustering decision is invariant to batch size by balancing the influence of intra-cluster and inter-cluster pairwise distances.
- Threshold-Variant Scoring: Instead of a fixed threshold, outlier scores are derived by tracking the highest clustering threshold at which an instance remains in an isolated cluster. This generates a continuous, bounded score in $[0, 1]$ that is transferable across different batch sizes.

Key Contributions

Unsupervised HMLV Framework: ASD-BOD extends fleet-based condition monitoring to EoLT, enabling fault detection from the first production batch without historical training data or labeled faults.
RPCA for ASD: This is the first application of RPCA to adaptively decompose vibroacoustic spectra ensembles in EoLT, effectively separating background noise from anomalies without manual parameter tuning.
Threshold-Transferable Scoring: The introduction of a weighted cophenetic hierarchical procedure generates interpretable anomaly scores that remain robust across varying batch sizes, addressing the scalability limitations of previous methods.
Real-World Validation: The framework is validated on a real-world APU dataset featuring authentic bearing faults under severe disturbances and distribution shifts.

Experimental Results
The framework was evaluated on a dataset comprising 40 APUs (30 healthy, 10 with induced bearing faults) tested under 12 different load points.

Performance Improvement: The RPCA-based preprocessing significantly enhanced detection capabilities. The Area Under the Precision-Recall Curve (AUPRC) increased from 0.421 (using simple sparsity indicators without RPCA) to 0.887 (using RPCA-Sparsity).
False Positive Reduction: The False Positive Rate (FPR) dropped from 6.00% to 0.23% at the optimal operating point.
Robustness: The RPCA-Sparsity pipeline outperformed all other configurations, including those using mean/median subtraction or Wasserstein distances. It demonstrated superior ability to distinguish faults from high-noise healthy units (e.g., units 34 and 37) that were misclassified as anomalies by non-RPCA methods.
Batch Size Transfer: While some methods (like Isolation Forest) showed high local discrimination, their decision thresholds were not portable across different batch sizes. The proposed weighted cophenetic (wCC) score maintained useful transferability across batch sizes ranging from 5 to 30, though RobustZ showed slightly better transfer in specific scalar regimes.

Significance and Claims
The paper claims that ASD-BOD provides a practical, interpretable, and robust solution for vibroacoustic fault detection in high-variance manufacturing environments. By shifting from global models to batch-wise unsupervised outlier detection, the framework overcomes the data scarcity and domain shift issues inherent in HMLV production. The integration of RPCA allows for the suppression of complex hydraulic background noise without requiring detailed physical models or historical baselines. The authors emphasize that the framework is designed to complement, rather than replace, existing rule-based or supervised workflows, offering a continuous, threshold-transferable metric for risk-based prioritization in quality assurance. The study concludes that while the method is highly effective for rolling element bearing faults (which manifest as sparse spectral peaks), its application to broadband faults or multimodal populations requires further investigation.

More like this