Datasets and Testbeds for Data-Driven Fault Diagnosis

34 concepts31 papers1 proposal

Data-driven fault diagnosis leverages machine learning and deep learning to identify, classify, and predict failures in complex engineering systems before they cause catastrophic damage. From autonomous underwater vehicles (AUVs) and liquid rocket engines to wind turbines and enterprise servers, modern industrial systems generate massive streams of sensor data (such as vibration, current, and temperature) that contain subtle signatures of degradation.

However, training deep learning models requires high-quality, labeled datasets that capture these failure modes. In the real world, machines rarely fail, and when they do, the resulting data is seldom captured cleanly or safely. To overcome this data scarcity bottleneck, the research community relies on three core paradigms: standardized public benchmarks, custom physical test-rigs, and high-fidelity simulation environments. Together, these approaches provide the foundation for developing and validating robust diagnostic algorithms.


Standardized Public Benchmarks

To evaluate new neural network architectures objectively, researchers need common baselines. Standardized public datasets act as the "ImageNet" of the fault diagnosis field, allowing different algorithmic approaches to be compared under identical conditions.

In the domain of rotating machinery—which forms the backbone of most industrial manufacturing—the Case Western Reserve University (CWRU) rolling bearing dataset is a widely recognized benchmark. As highlighted in a comprehensive two-decade review by Gawde et al., public repositories like the CWRU and FEMTO datasets provide the essential baseline data required to train and evaluate multi-fault diagnostic models. For example, Chengwei Li utilized the CWRU dataset to validate a multimodal information fusion framework that combines vibration, weak magnetic, and temperature signals (Rolling Bearing Fault Diagnosis Based on Deep Learning and Autoencoder Information Fusion). Similarly, Han et al. used the CWRU dataset to demonstrate the noise-robustness of a hybrid CNN-LSTM-GRU model designed to capture both spatial and long-term temporal dependencies in vibration data.

Beyond rotating parts, benchmarks exist for complex chemical and robotic systems. For large-scale industrial processes, Wang et al. leveraged the simulated Tennessee Eastman (TE) process dataset—a standard testbed modeling complex, interconnected chemical plant dynamics—to evaluate out-of-distribution fault detection.

In robotics, Ji et al. introduced a dedicated Autonomous underwater vehicle fault diagnosis dataset. By collecting 1,225 samples across 17 state variables (including motor signals, depth, and sensor readings) from a quadrotor AUV, the authors established an objective benchmark for model-free diagnostic methods across five distinct conditions, such as propeller damage and sensor failures. This public dataset has enabled subsequent algorithmic innovations; for instance, Wu et al. utilized it to validate a multi-channel convolutional neural network capable of diagnosing AUV faults even when up to 60% of the sensor data is missing.

+-----------------------------------------------------------------------+
|                       Fault Diagnosis Data Spectrum                   |
+-----------------------------------------------------------------------+
|  Public Benchmarks       Physical Test-Rigs       High-Fidelity Sim   |
|  (CWRU, TE, AUV)         (Custom Benches, Rigs)   (Simulink, Digital) |
|                                                                       |
|  High Standardization    True Physical Noise      Infinite Scalability|
|  Algorithmic Baselines   Safe Fault Induction     Zero Physical Risk  |
+-----------------------------------------------------------------------+

Custom Physical Test-Rigs

While public benchmarks are invaluable for algorithmic iteration, they do not cover every specialized industrial application. When a specific system lacks a public dataset, researchers build custom physical testbeds or instrument real-world machinery to capture true physical noise and environmental complexities safely.

To diagnose faults in specialized hydraulics and power generation, researchers build scaled down or emulated environments:

  • Wind Turbines: Stetco et al. developed a custom laboratory test-rig that emulates a Type IV wind turbine generator. This physical setup allowed the authors to safely induce rotor demagnetization faults at specific severity levels and capture high-frequency raw current signals for real-time condition monitoring.
  • Centrifugal Pumps: Zhu et al. constructed a custom experimental bench to generate vibration data for vertical Essential Service Water centrifugal pumps, mapping normal operations against six distinct fault types to train support vector machine classifiers.
  • Micro-Turbines: To capture the unique dynamics of micro-turbine environments, Yang et al. utilized an experimental rolling bearing-rotor system bench to collect vibration data across varying speeds for healthy bearings and bearings with slight-to-severe race and cage faults.

Physical testbeds can also be used to reproduce environmental stressors captured from the field. In the context of enterprise computing, Chan et al. used tri-axial accelerometers to capture real vibration profiles from a live datacenter. The authors then reproduced these exact profiles on an electrodynamic programmable vibrational table to study hard drive performance degradation and validate mitigation strategies under controlled laboratory conditions.

When building a physical rig is impossible, researchers must rely on long-term field collection. Zhou et al. gathered a massive dataset comprising approximately 3 million records from 107 excavators operating over 40 days. Because these were real-world operations, the authors relied on domain experts to manually label anomalies across 68 working condition metrics, creating a high-fidelity dataset rooted in actual operational wear and tear.


High-Fidelity Simulation Environments

For many critical systems, inducing physical faults—even on a test-rig—is too dangerous, expensive, or logistically unfeasible. In these scenarios, researchers turn to high-fidelity software simulations, such as MATLAB/Simulink or finite element models, to synthetically generate diverse, multi-severity fault datasets.

This approach is vital for aerospace and high-voltage electrical infrastructure:

  • Liquid Rocket Engines: Longfei Zhang et al. generated a diagnostic dataset using a component-wise Simulink model of a liquid rocket engine. The simulation allowed them to safely model six critical fault modes at multiple severity levels, sampling data at 1 kHz to extract interpretable causal features without destroying hardware.
  • MEMS Gyroscopes: To address the lack of physical fault data for micro-electromechanical systems, Cui et al. built a self-feedback simulation platform in MATLAB/Simulink. This platform synthetically generates seven distinct gyroscope signal types (including blocking, drift, and periodic faults) to train diagnostic models.
  • Circuit Breakers: Shao et al. developed a novel simulation method to model core jamming faults within circuit breaker operating mechanisms, generating a 600-sample dataset of time-frequency spectrograms across twelve distinct fault conditions and severities.
  • Structural Infrastructure: Xulong Zhang et al. engineered a custom simulation platform to generate acoustic data representing normal and loose connection states in bridge expansion joints, producing 1,500 data fragments to train edge-computing classifiers.

Simulation and curation principles also extend to text-based and operational data. For instance, Li et al. constructed a specialized dataset for virtual power plants by compiling and structuring technical documentation from various devices. This dataset includes non-standardized questions and key reference facts, allowing the authors to evaluate Retrieval-Augmented Generation (RAG) models designed to assist maintenance personnel with root-cause troubleshooting.


Why This Multi-Faceted Approach Works

Data-driven fault diagnosis works because engineering failures leave distinct physical footprints—whether as a frequency shift in a vibration spectrum, a transient drop in an electrical current, or an acoustic anomaly. However, a deep learning model can only recognize these footprints if its training data spans the boundary between normal operation and failure.

By combining public benchmarks, physical testbeds, and high-fidelity simulations, the field circumvents the impossibility of collecting real-world catastrophic failure data. Public benchmarks provide a playground for optimizing neural architectures; physical test-rigs inject the messy, unmodeled noise of the real world; and simulations allow researchers to safely map out dangerous, multi-fault scenarios. Together, these data generation strategies ensure that when a real-world system begins to degrade, diagnostic models have already seen the pattern before.

Key Foundational Work

Modelling and diagnosis of feedback-controlled processes using dynamic PCA and neural networks (2003) — contributed 1 concept elements to this direction.
Correcting vibration-induced performance degradation in enterprise servers (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_auv_fault_diagnosis_dataset_683a24a4,
  title = {AUV fault diagnosis dataset},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/auv-fault-diagnosis-dataset/rdr_683a24a4f0154d63a014f70158926738},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_auv_fault_diagnosis_dataset_683a24a4_v1,
  title = {AUV fault diagnosis dataset},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/auv-fault-diagnosis-dataset/rdr_683a24a4f0154d63a014f70158926738/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake