PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-driven Machinery Fault Diagnosis: A Comprehensive Review

Dhiraj Neupane, Mohamed Reda Bouadjenek, Richard Dazeley, Sunil Aryal

arXiv:2405.18843v2cs.AIcs.LG

TL;DR

Existing machinery fault-diagnosis reviews are often narrow in equipment or analysis method and give limited attention to implementation challenges. This survey synthesizes machine-learning approaches, datasets, condition-monitoring methods, and emerging techniques, concluding with a broad review of methodologies, datasets, and challenges for future research.

  • Problem

    Existing MFD reviews have limited applicability across manufacturing settings and often treat noisy data, feature selection, and unforeseen faults superficially.

  • Method

    The survey reviews traditional and advanced machine-learning approaches, datasets, condition-monitoring techniques, and domain-adaptation, federated-learning, and transformer methods.

  • Results

    The review provides a comprehensive overview of MFD approaches, datasets, predictive methods, and challenges spanning traditional, deep, federated, reinforcement, transfer, and physics-informed learning.

  • Takeaways & Limitations

    The survey identifies challenges and provides recommendations intended to guide future research and improve machinery-fault datasets and machine reliability.

  • Takeaways & Limitations

    MFD faces scarce faulty and labeled data, computationally demanding model training, and limited interpretability of deep-learning models.

Abstract

from arXiv · show

In this era of advanced manufacturing, it's now more crucial than ever to diagnose machine faults as early as possible to guarantee their safe and efficient operation. With the massive surge in industrial big data and advancement in sensing and computational technologies, data-driven Machinery Fault Diagnosis (MFD) solutions based on machine/deep learning approaches have been used ubiquitously in manufacturing. Timely and accurately identifying faulty machine signals is vital in industrial applications for which many relevant solutions have been proposed and are reviewed in many articles. Despite the availability of numerous solutions and reviews on MFD, existing works often lack several aspects. Most of the available literature has limited applicability in a wide range of manufacturing settings due to their concentration on a particular type of equipment or method of analysis. Additionally, discussions regarding the challenges associated with implementing data-driven approaches, such as dealing with noisy data, selecting appropriate features, and adapting models to accommodate new or unforeseen faults, are often superficial or completely overlooked. Thus, this survey provides a comprehensive review of the articles using different types of machine learning approaches for the detection and diagnosis of various types of machinery faults, highlights their strengths and limitations, provides a review of the methods used for condition-based analyses, comprehensively discusses the available machinery fault datasets, introduces future researchers to the possible challenges they have to encounter while using these approaches for MFD and recommends the probable solutions to mitigate those problems. The future research prospects are also pointed out for a better understanding of the field. We believe this article will help researchers and contribute to the further development of the field.

1. Introduction

Machinery fault diagnosis is important for maintaining safe, reliable, and efficient industrial operations, yet existing reviews remain fragmented and leave important methodological and coverage gaps. This review addresses those gaps by organizing machine-learning research across algorithms, datasets, monitoring techniques, and emerging approaches.

  • Machinery fault diagnosis supports the safety, reliability, and efficiency of critical machinery in modern industrial processes.
  • Fault diagnosis spans detection, anomaly detection, isolation, identification, and reconstruction, while maintenance may be preventive, predictive, or reactive.
  • Existing ML/DL reviews provide fragmented summaries and lack a broad account of how fault information is learned and represented.
  • The review covers ML algorithms, including under-explored reinforcement learning, and examines their advantages, disadvantages, and intelligent MFD applications.
  • It surveys more than 30 datasets and discusses fault detection, classification, federated learning, and condition-monitoring techniques beyond vibration signals.

2. Methodology Implemented and Taxonomy Developed

The review uses a narrative overview of machine-learning research in machinery fault diagnosis, drawing primarily on Google Scholar and Scopus searches. A taxonomy provides the structural foundation for organizing methods, datasets, applications, and related topics.

  • The authors used a narrative overview approach to synthesize the current state of machine-learning research in machinery fault diagnosis.
  • Table 1 compares this article with other reviews on machinery fault diagnosis studies.
  • Literature searches used Google Scholar and Scopus with combinations of machinery, equipment, fault, machine-learning, and deep-learning keywords.
  • A taxonomy was constructed to systematically organize fundamental MFD principles, machine learning, advanced learning techniques, applications, and data-related domains.

3. Machinery Fault Data (and Analysis)

Machinery fault diagnosis depends on sufficient, diverse data and a systematic path from sensing to feature extraction and fault identification. The review covers major machinery data types, available datasets, and augmentation strategies used to address data scarcity.

  • Data importance: Data quantity, quality, and diversity strongly affect the performance of data-driven machinery fault diagnosis models.Selecting appropriate target data is important for reliable and accurate predictive models.
  • Available datasets: The reviewed dataset resources span chemical processes, gears, bearings, wind turbines, and related vibration, acoustic, speed, and power measurements.Table 2 summarizes dataset characteristics and the corresponding MFD studies.
  • Data types and analyses: The general diagnosis process collects sensor data, extracts features, and identifies or classifies the equipment fault state.Common inputs include vibration, acoustic emission, temperature, and current signals.
  • Data types and analyses: Vibration is the most widely acquired predictive-maintenance data type, accounting for nearly three-quarters of methods because it detects many machinery faults.Other approaches include wear-debris, oil, acoustic-emission, thermography, and motor-current-signature analyses.
  • Data types and analyses: Vibration analysis is non-intrusive and sensitive to early faults, supporting diagnosis of unbalance, misalignment, and bearing problems.Accelerometers capture condition-dependent vibration signals for analysis and remaining-useful-life prediction.
  • Data types and analyses: Wear-debris, oil, acoustic-emission, thermographic, and motor-current analyses provide complementary indicators of machinery condition.These methods examine particles, lubricant properties, high-frequency sound, temperature distributions, or electrical-current spectra to identify faults.
  • Data augmentation: Rare failures and high collection costs produce limited, biased datasets, motivating data augmentation and transfer learning to improve diversity and generalization.Augmentation transforms existing samples in time or frequency domains and can help prevent overfitting.

4. Data-driven Approaches

The review organizes MFD methods into traditional and advanced data-driven approaches, spanning physical, expert-system, signal-processing, machine-learning, deep-learning, reinforcement-learning, transfer-learning, and federated-learning methods. It compares their applications, strengths, limitations, and reported diagnostic results across machinery and operating conditions.

  • Method taxonomy: MFD methods are organized into traditional approaches—physical-model, expert-system, signal-processing, and traditional machine-learning—and advanced approaches including deep, reinforcement, transfer, and federated learning.This taxonomy structures the review of data-driven machinery fault diagnosis.
  • Traditional Data-driven Approaches: Physical-model methods use mathematical or analytical machine models, whereas expert systems automate diagnosis using expert knowledge and reasoning strategies.Expert-system reasoning includes rule-based, fuzzy-logic, neural-network, and case-based approaches.
  • Traditional Data-driven Approaches: Signal-processing methods extract fault-relevant information through filtering and denoising, using techniques such as FFT, wavelet transforms, WPT, EMD, HHT, cepstrum, and envelope analysis.These methods emphasize characteristics associated with the machine’s fault status.
  • Traditional Data-driven Approaches: Traditional machine-learning methods include ANN, SVM, decision trees, random forests, KNN, Naive Bayes, HMM, clustering, PCA, and regression, but manual feature extraction can hinder early diagnosis in complex industrial environments.Dynamic, nonlinear, and multimodal processes make feature analysis and selection difficult for large datasets.
  • Advanced Data-driven Methods: Advanced approaches use raw or transformed sensor data for automated fault recognition, with CNNs applied to one-dimensional time series, two-dimensional representations, and spatial-temporal correlations.Reported examples include 97.1% accuracy on the IMS dataset, 99.77% accuracy on the CWRU dataset at a signal-to-noise ratio of 10, and 98.8% diagnostic accuracy for a 1-D CNN.
  • Advanced Data-driven Methods: The reviewed literature also covers anomaly detection for limited labeled-fault settings, transformer-based diagnosis, and federated learning for decentralized training across distributed devices.Examples report nearly 99.99% accuracy in noise-free conditions with robustness under added noise, while federated learning supports privacy-preserving collaborative training.

5. Challenges

Data-driven MFD faces practical challenges spanning datasets, sensors, learning paradigms, signal conditions, and model design. These constraints affect data quality, fault coverage, evaluation, generalization, and deployment.

  • Sensor challenges: Sensor networks and multi-sensor fusion must address integration complexity, uncertainty, synchronization, security, network traffic, energy efficiency, and acquisition costs.These issues become especially consequential when combining heterogeneous sources across edge or cloud systems.
  • Data challenges: Real-world machinery data is noisy, inconsistent, heterogeneous, non-stationary, and incomplete, while industrial datasets are often difficult to obtain.Sensor failures, intermittent faults, and communication issues can degrade collected data, and industries may hesitate to share datasets.
  • Data challenges: Faulty and labeled data are scarce, producing class imbalance and limiting generalization to unseen fault types.Faults vary in type and severity, while labeling is costly, time-consuming, and often lacks sufficient diversity.
  • Signal challenges: Complex movement, interacting faults, external noise, non-stationary behavior, and operating-condition sensitivity can obscure informative fault features.Improper sensor placement can further reduce signal quality and adversely affect deep-learning fault detection.
  • Learning challenges: Supervised, semi-supervised, and unsupervised methods face distinct problems involving labeling, class imbalance, threshold selection, evaluation, false positives, and unseen faults.Unsupervised methods may detect anomalies without labels but can be difficult to evaluate and explain.
  • Model challenges: ML/DL systems require careful feature selection and substantial computation while facing overfitting, noise sensitivity, limited interpretability, and weak generalization across engineering settings.Deep models can behave as black boxes, and many techniques remain tailored to specific situations rather than integrated environments.

6. Recommendations for Future researchers

The review recommends grounding MFD development in careful dataset analysis, realistic and well-documented data collection, suitable model selection, and stronger validation. It also highlights adaptation, interpretability, broader system coverage, and open research practices as future priorities.

  • Data and dataset design: Researchers should study datasets before modeling and design collections with high-quality sensors, secure placement, realistic conditions, varied loads and speeds, accurate labels, preprocessing, and systematic organization.Publicly shared datasets should include reliable hosting and comprehensive documentation.
  • Data and dataset design: Sensor counts and modalities should match the method, with multi-sensor fusion combining vibration, temperature, and acoustic signals when appropriate.The review distinguishes fewer sensors for classical methods from multiple locations for deep learning approaches.
  • Model selection: Model choice should reflect operating conditions: simpler or frequency-based methods suit simpler settings, whereas noisy or complex environments may require denoising and advanced deep learning.Architecture selection should balance complexity, interpretability, and computational requirements.
  • Learning strategies: Unsupervised or semi-supervised learning, cost-sensitive methods, resampling, and tailored loss functions are proposed to address scarce labels and imbalanced fault data.Anomaly detection is presented as a way to detect faults without labels, including potentially new or unseen faults.
  • Validation and robustness: Regularization, denoising, cross-validation, repeated experiments, and hybrid combinations of deep learning, traditional ML, signal processing, and feature engineering can support robust evaluation and generalization.The review also recommends exploring domain adaptation for differing operating conditions and machine configurations.
  • Future directions: Future work should improve reinforcement-learning formulations, explainable AI, continuous adaptation, digital-twin use, broader multi-component diagnosis, and open-source code and documentation.The review calls for methods covering interactions among bearings, gears, motors, and other components operating together.

7. Future Prospects

Future MFD research focuses on improving diagnostic robustness, adaptability, interpretability, and real-time efficiency through advanced learning, data fusion, and hybrid modeling approaches.

  • Robustness: Noise-resilient algorithms are needed to improve fault detection and classification under mixed-speed and noisy industrial conditions.Future work also targets crack-damage detection under mixed-speed operation.
  • Advanced models and data fusion: Transformer architectures, physics-informed neural networks, and sensor, feature, and decision fusion are proposed to improve learning across heterogeneous machinery data.Fusion may combine vibration, current, torque, acoustic, and visual measurements.
  • Learning with limited labels: Semi-supervised and reinforcement learning aim to reduce labeled-data dependence, support sequence-based fault prediction, and generalize across new environments.Offline reinforcement learning is emphasized for practical applications and proactive fault detection.
  • Adaptability and interpretability: Domain generalization, domain confusion, domain adaptation, and explainable AI are intended to address unforeseen faults, data imbalance, limited labels, and interpretability.These approaches are presented as ways to make MFD systems more transparent and adaptable.
  • Efficiency and hybrid modeling: Pre-training, neural architecture search, model compression, mechanism knowledge, digital twins, and hybrid data-driven models are proposed for efficient real-time diagnostics and improved maintenance strategies.These methods target computational challenges, interpretability, generalization, prediction error, and maintenance refinement.

8. Conclusion

The review synthesizes machinery fault diagnosis research across data sources, maintenance, datasets, and traditional and advanced learning methods. It identifies challenges and recommends improvements in dataset quality, algorithm selection, and practical application.

  • Review scope: The review covers machinery fault diagnosis from data sources and maintenance techniques to predictive maintenance strategies.Its scope is presented as a comprehensive overview of the field.
  • Methods: It discusses traditional approaches alongside deep, federated, reinforcement, and transfer learning, plus physics-informed neural networks.The review spans multiple machine-learning paradigms and physics-informed methods.
  • Data and maintenance: The review emphasizes data collection and maintenance analysis while surveying available machinery fault datasets.It links predictive methods with fault detection and machine reliability within its stated scope.
  • Challenges and recommendations: Identified challenges and recommendations target dataset quality, algorithm selection, and practical application considerations.These topics are presented as areas for future research and implementation.

Appendix B: Machine Fault Detection Datasets

The appendix describes two machinery datasets: an Airbus helicopter vibration dataset for healthy-versus-faulty sequence classification and a synthetic predictive-maintenance dataset with labeled machine failures.

  • Airbus helicopter dataset: The Airbus helicopter dataset contains 1,677 training and 594 validation vibration sequences recorded at 1,024 Hz, with half labeled abnormal.It supports automated flight-test data validation and was used in Airbus’s 2019 AI challenge.
  • AI4I 2020 dataset: The AI4I 2020 dataset contains 10,000 synthetic records with 14 features covering temperatures, speed, torque, tool wear, and machine failure.Its failure label indicates whether failure occurred but does not specify the failure mode.

C-MAPSS Dataset

The reviewed datasets span simulated turbofan degradation, bearing faults, high-speed aeronautical bearing tests, and gearbox gear-fault conditions, with C-MAPSS explicitly bounded by its simulated nature.

  • C-MAPSS: C-MAPSS contains 21 sensor variables across five subsets representing turbofan wear levels and degradation processes.It is derived from a dynamic simulator for large commercial turbofan engines and supports predictive-model development.
  • C-MAPSS: C-MAPSS is useful for deep-learning training because of extensive samples and diverse operating conditions, but simulation may not fully replicate real-world scenarios.This limits direct representativeness of real-world operating conditions.
  • Bearing datasets: The CWRU dataset provides experimental bearing-fault vibration signals covering inner-race, outer-race, and ball faults at different severities, loads, sensor positions, and sampling frequencies.Its single-point faults were introduced by electro-discharge machining.
  • Bearing datasets: The DIRG dataset examines high-speed aeronautical roller bearings above 6,000 rpm using two accelerometers across varied damages, speeds, and loads.It includes two experimental sessions involving localized bearing damage and prolonged testing.
  • Gear datasets: EDGFD contains gearbox vibration signals for healthy gears, one chipped tooth, and three worn teeth under specified helical-gear operating conditions.The tested gear-mesh frequency was 365 Hz, compared with a typical 355 Hz value.

FEMTO Bearing Dataset

The FEMTO bearing dataset contains multimodal sensor data collected from bearings under normal and accelerated degradation conditions on the PRONOSTIA platform.

  • The dataset includes time-domain vibration signals, temperature measurements, and speed data from bearings.
  • Data were collected under both normal and accelerated degradation conditions for real-time monitoring.

Gearbox Fault Diagnosis Data

These datasets cover gearbox, bearing, motor, and machinery fault diagnosis using vibration, temperature, speed, thermal-image, and multivariate measurements across varied operating conditions.

  • The gearbox dataset contains four-direction vibration measurements under loads from 0% to 90% for healthy and broken-tooth conditions.
  • The DSTG HUMS datasets capture vibration from planet-gear fatigue-crack tests in a Bell Kiowa helicopter gearbox.
  • The IMS Bearing dataset provides one-second vibration snapshots sampled at 20 kHz alongside temperature data from run-to-failure tests.
  • The IMTI dataset contains 369 thermal images of three-phase induction motors across 11 fault conditions at 320×240 resolution.
  • The JNU bearing datasets represent one health state and three fault modes across rotating speeds, yielding 12 classes by working condition.
  • MaFaulDa comprises 1,951 multivariate time series covering six simulated machinery states measured with accelerometers, a tachometer, and a microphone.

MFPT Bearing dataset

The MFPT bearing dataset provides fault and baseline measurements, while related datasets support steel-defect, wind-turbine gearbox, and gearbox fault-detection research.

  • The MFPT dataset includes three baseline conditions and ten outer- and inner-race fault conditions under varying loads.
  • The North East University dataset contains 1,800 balanced steel-surface images spanning six defect types.
  • The NREL wind-turbine benchmark includes healthy and damaged gearboxes operating at 1,800 rpm and 1,200 rpm with over 125 sensors.
  • The PHM-2009 gearbox dataset contains 20 test cases with vibration, temperature, and torque measurements and requires preprocessing and feature extraction before machine learning.

PHM2010 Milling Cutter Dataset

The PHM-2010 dataset supports remaining-useful-life estimation for CNC milling cutters using force, vibration, and acoustic-emission measurements under specified cutting conditions.

  • The dataset records cutting forces and vibrations along the X, Y, and Z axes using a three-component dynamometer and three accelerometers.
  • An acoustic-emission sensor monitors high-frequency stress waves during cutting.
  • The operating conditions include a spindle speed of 10400 rpm and a feed rate of 1555 mm/min in the x-direction.

PHME Datasets

The review describes PHME’s 2021 challenge as a real-world manufacturing production-line benchmark covering fault detection, classification, and root-cause identification. It also catalogs diverse machinery datasets spanning bearings, gears, chemical processes, and wind turbines.

  • The 2021 PHME challenge used real-world industrial production-line datasets for fault detection, classification, and root-cause identification.
  • The PU bearing dataset contains high-resolution vibration, torque, and temperature measurements from six healthy bearings and 26 damaged sets.Faults were induced artificially and through accelerated life tests under varied conditions.
  • The SEU dataset includes motor and gearbox vibrations, torque, five gear fault types, and five bearing fault types under two speed-load conditions.
  • The Tennessee Eastman Process benchmark contains 22 simulation runs, 52 variables, and 20 fault types for continuous-process diagnosis.
  • Additional datasets cover gearbox damage, bearings, and wind-turbine abnormalities using vibration, speed, acceleration, acoustic, and SCADA measurements.Examples include the THU, UA, UO, UoC, UoARED-VAFCLS, and wind-turbine SCADA datasets.

XJTU-SY Bearing Dataset

The XJTU-SY bearing dataset provides run-to-failure data from 15 rolling-element bearings under three speed-load conditions. It uses high-frequency measurements from two accelerometers and is intended for bearing-fault analysis.

  • The XJTU-SY dataset contains run-to-failure data from 15 rolling-element bearings under three speed-load conditions.The conditions are 2100 rpm with 12 kN, 2250 rpm with 11 kN, and 2400 rpm with 10 kN.
  • Two accelerometers collected data at 25.6 kHz, with 32768 points recorded per minute and stored in CSV files.
  • The review places the XJTU-SY dataset among broader machinery-fault resources that include simulators and datasets for varied equipment and operating conditions.Machine fault simulators support controlled studies using multichannel acquisition and analyses such as time waveforms, amplitude spectra, and frequency responses.
Loading 2405.18843v2…