PaperlayerView exact revision on arXiv

Source-linked AI summary

An Explainable Artificial Intelligence Approach for Unsupervised Fault Detection and Diagnosis in Rotating Machinery

Lucas Costa Brito, Gian Antonio Susto, Jorge Nei Brito, Marcus Antonio Viana Duarte

arXiv:2102.11848v1cs.AIcs.LG

TL;DR

Rotating-machinery monitoring needs explainable fault diagnosis despite the frequent absence of labeled historical data. The paper proposes a three-stage unsupervised approach using vibration features, anomaly detection, and explainability-based diagnosis, and reports successful fault detection with classification or root-cause information across studied datasets.

  • Problem

    Rotating-machinery AI often lacks explainability and requires labeled data that are unavailable in many industrial settings.

  • Method

    The method extracts time- and frequency-domain vibration features, detects faults with configurable unsupervised anomaly-detection algorithms, and uses SHAP-based feature rankings for classification or root cause analysis.

  • Results

    The methodology enables unsupervised fault detection while providing explainability and information for root cause analysis or unsupervised fault classification.

  • Takeaways & Limitations

    The approach can be adapted to different datasets and fault types by changing extracted features and does not require previously labeled data.

  • Takeaways & Limitations

    Fast-ABOD’s approximation quality depends on the number of nearest neighbors k, despite its improved computational cost.

Abstract

from arXiv · show

The monitoring of rotating machinery is an essential task in today's production processes. Currently, several machine learning and deep learning-based modules have achieved excellent results in fault detection and diagnosis. Nevertheless, to further increase user adoption and diffusion of such technologies, users and human experts must be provided with explanations and insights by the modules. Another issue is related, in most cases, with the unavailability of labeled historical data that makes the use of supervised models unfeasible. Therefore, a new approach for fault detection and diagnosis in rotating machinery is here proposed. The methodology consists of three parts: feature extraction, fault detection and fault diagnosis. In the first part, the vibration features in the time and frequency domains are extracted. Secondly, in the fault detection, the presence of fault is verified in an unsupervised manner based on anomaly detection algorithms. The modularity of the methodology allows different algorithms to be implemented. Finally, in fault diagnosis, Shapley Additive Explanations (SHAP), a technique to interpret black-box models, is used. Through the feature importance ranking obtained by the model explainability, the fault diagnosis is performed. Two tools for diagnosis are proposed, namely: unsupervised classification and root cause analysis. The effectiveness of the proposed approach is shown on three datasets containing different mechanical faults in rotating machinery. The study also presents a comparison between models used in machine learning explainability: SHAP and Local Depth-based Feature Importance for the Isolation Forest (Local- DIFFI). Lastly, an analysis of several state-of-art anomaly detection algorithms in rotating machinery is included.

1. Introduction and Related Work

The paper addresses limited explainability and scarce labeled data in rotating-machinery monitoring by proposing an unsupervised, explainable fault-detection and diagnosis approach.

  • Industrial rotating-machinery monitoring increasingly uses AI to improve machine reliability and availability.
  • Supervised fault-classification models are often impractical because industrial datasets lack labels covering all normal and fault conditions.
  • Unsupervised anomaly detection identifies deviations from normal behavior without requiring labeled historical data.
  • Explainability remains underexplored, although interpretable results are considered important for real-world adoption of machinery-monitoring AI.
  • The proposed approach extracts time- and frequency-domain vibration features, detects anomalies, and diagnoses faults using feature-importance rankings.
  • Diagnosis uses unsupervised classification or root cause analysis, while the modular framework supports different anomaly-detection models and fault types.
  • The authors describe the study as the first comparison of unsupervised state-of-the-art anomaly-detection algorithms for rotating-machinery monitoring.

2.1. Anomaly Detection Algorithms

This section introduces unsupervised anomaly detection as a way to identify unusual observations and reviews several algorithms used in the study, including kNN and MCD.

  • Anomaly detection identifies rare observations that differ from the normal data distribution.It can summarize the status of a multivariate system with an anomaly score.
  • The reviewed algorithms are arranged by increasing year of presentation.
  • 2.1.1. k-Nearest Neighbors (kNN): In kNN-based anomaly detection, the distance from an observation to its kth-nearest neighbor serves as the anomaly score.The distance may use the largest, mean, or median distance among the k neighbors.
  • MCD fits a robust multivariate location and covariance model by selecting n of N instances whose covariance matrix has the lowest determinant.Mahalanobis distance computed from this model is then used as the anomaly score.
  • For centered, non-normalized data, robust location and covariance are computed directly with FastMCD; otherwise, the estimates are recomputed from the robust support.

2.1.3. Local Outlier Factor (LOF) and Cluster-based Local Outlier Factor (CBLOF)

LOF detects anomalies through local density, while CBLOF first clusters the data to address LOF’s difficulty with clustered observations.

  • Local Outlier Factor (LOF): LOF treats observations in dense local regions as inliers and observations in low-density regions as anomalies.
  • Local Outlier Factor (LOF): LOF computes local reachability density and then uses it to evaluate the anomaly score sLOF.Reachability distance replaces ordinary distance to reduce statistical fluctuations and noise.
  • Local Outlier Factor (LOF): An sLOF value around 1 or below indicates similarity to neighboring points, whereas values above 1 indicate an outlier.
  • Cluster-based Local Outlier Factor (CBLOF): CBLOF extends LOF by applying clustering before density-based scoring to better handle clustered data.
  • Cluster-based Local Outlier Factor (CBLOF): CBLOF partitions the dataset into disjoint clusters and assigns scores according to whether an observation belongs to a small or large cluster.The parameters α and β determine the small- and large-cluster categories.

2.1.4. One-class Support Vector Machines (OCSVM)

This section describes OCSVM as a boundary-based anomaly detector and also outlines feature bagging as an ensemble alternative using subsets of features.

  • One-class Support Vector Machines (OCSVM): OCSVM projects training data into a high-dimensional space and separates them from the origin with a hyperplane.New samples inside the frontier-defined subspace are treated as inliers; others are anomalies.
  • One-class Support Vector Machines (OCSVM): OCSVM can use linear, polynomial, sigmoid, or Gaussian kernels to construct nonlinear decision boundaries.The parameters gamma and nu control the kernel coefficient and specified training or support-vector bounds.
  • Feature Bagging: Feature Bagging combines multiple outlier detectors trained on randomly selected subsets of the original features.Any anomaly detection approach can serve as the base estimator.
  • Feature Bagging: The final Feature Bagging anomaly score sums scores from all T detector iterations.Scores may alternatively be combined by averaging models or taking the maximum score.
  • Feature Bagging: The number of base estimators and the number of features drawn for each estimator are adjustable.

2.1.6. Angle-based Outlier Detector (ABOD) and Fast-ABOD

Fast-ABOD detects outliers through angle-based relationships, approximating ABOD with angles formed using k nearest neighbors. Isolation Forest instead isolates observations through random feature-based partitioning and interprets shorter paths as stronger anomaly evidence.

  • Angle-based Outlier Detector (ABOD) and Fast-ABOD: ABOD identifies outliers from the angles formed between a data point and pairs of other observations.Outliers typically form acute angles, whereas inliers form more varied angles.
  • Angle-based Outlier Detector (ABOD) and Fast-ABOD: Fast-ABOD reduces computation by evaluating only angles involving each observation and its k nearest neighbors.Its approximation quality depends on the selected k.
  • Isolation Forest: Isolation Forest recursively partitions randomly selected feature values until observations are isolated in binary trees.Each tree uses a randomly selected feature and threshold to split the data.
  • Isolation Forest: Isolation Forest assigns stronger anomaly scores to observations with shorter average paths from tree roots to leaves.Scores close to 1 indicate anomalies, values much smaller than 0.5 indicate normal instances, and values near 0.5 are ambiguous.
  • Isolation Forest: Isolation Forest performs well with high-dimensional data, irrelevant attributes, and training sets without anomalies, while supporting parallel computation.Its ensemble structure motivates using a dedicated interpretability method.

2.1.8. Histogram-based outlier score (HBOS)

HBOS detects anomalies using independently constructed feature histograms. It estimates each sample’s anomaly score from the inverse densities of its feature values, with several parameters requiring tuning.

  • Histogram-based outlier score (HBOS): HBOS uses normalized univariate histograms to estimate the density of each feature.For numerical data, each histogram uses k equal-sized bins, and k is tuned as a hyper-parameter.
  • Histogram-based outlier score (HBOS): HBOS computes each instance’s anomaly score as a product of the inverse estimated densities across its features.The formulation uses p features and hist_i(x) as the density estimate for feature i.
  • Histogram-based outlier score (HBOS): High HBOS scores indicate outliers, whereas low scores indicate inliers.The algorithm also tunes α to reduce overfitting and tolerance to handle samples outside histogram bins.

2.1.9. Lightweight on-line detector of anomalies (LODA)

LODA estimates anomaly scores from sparse random projections and one-dimensional histograms. Its lightweight design supports real-time processing and online detector updates, including settings with concept drift.

  • Lightweight on-line detector of anomalies (LODA): LODA combines k one-dimensional histograms, each estimating projected input-data density over n bins.Each histogram corresponds to a projection onto a single projection vector.
  • Lightweight on-line detector of anomalies (LODA): LODA outputs the average logarithm of probabilities estimated across individual projection vectors.The probability estimated by the ith histogram corresponds to the projection associated with vector w_i.
  • Lightweight on-line detector of anomalies (LODA): LODA is useful when many samples must be processed in real time or when the detector must be updated online.The method can also be applied in domains subject to concept drift.
  • Ensemble method: The ensemble method combines base-model outputs into a final result and obtains the final decision through majority voting.The binary outputs use 1 for anomalies and 0 for normal samples.

2.2. Explainable Artificial Intelligence (XAI)

The paper uses model-agnostic SHAP and model-specific Local-DIFFI to explain anomaly-detection predictions through feature-importance rankings. Local-DIFFI preserves Isolation Forest performance while substantially reducing explanation computation time relative to SHAP.

  • SHAP: SHAP is a model-agnostic method for interpreting machine-learning predictions in supervised and unsupervised tasks.It uses Shapley values from coalitional game theory.
  • SHAP: SHAP ranks features by their contribution to an individual model prediction, supporting anomaly troubleshooting by domain experts.The ranking explains the model at the individual data-point level.
  • SHAP: Because exact Shapley values are expensive to obtain, SHAP approximates them using a conditional expectation function of the original model.This approximation is part of the method’s computational design.
  • Local-DIFFI: Local-DIFFI provides model-specific local feature importance for Isolation Forest, with DIFFI offering global feature importance as well.It was introduced to improve interpretability of Isolation Forest predictions.
  • Local-DIFFI: Local-DIFFI is post-hoc, preserves the performance of Isolation Forest, and has computing time orders of magnitude smaller than SHAP.The method aims to increase users’ confidence in results by explaining trained Isolation Forest instances.

3. Proposed Approach

The proposed methodology combines vibration-feature extraction, unsupervised anomaly detection, and explainability-based fault diagnosis. It supports unsupervised classification for component-specific faults and root cause analysis when features relate to multiple faults.

  • Framework: The methodology extracts vibration features, detects anomalies with configurable algorithms, and diagnoses faults using explainability.The pipeline comprises feature extraction, fault detection, and fault diagnosis; its anomaly-detection model can be modified without changing the overall structure.
  • Feature extraction: Classic machine-learning techniques exploit established filtering approaches and ad-hoc features developed for rotating-machinery faults.The approach uses time-, frequency-, and time-frequency-domain vibration features rather than relying on deep learning to learn features implicitly.
  • Feature extraction: General features indicate system fault or degradation, whereas specific features support identifying or locating the fault.General features provide global detection; component-specific features enable more targeted diagnosis.
  • Fault detection: Anomaly detection is performed without fault labels by dividing extracted features into training and testing groups and tuning model hyperparameters.Samples are classified as normal or anomalous before explainability is applied to anomalous samples.
  • Fault diagnosis: SHAP or Local-DIFFI ranks feature importance after anomaly detection, enabling unsupervised classification or root cause analysis.General features are removed from the ranking; the most important specific feature can classify a fault, while broader rankings assist specialist-led analysis.
  • Explainability comparison: SHAP is model-agnostic but computationally costly, motivating comparison with the Isolation Forest-specific Local-DIFFI method.The study compares the resulting rankings using Kendall-Tau rank distance.

4. Experimental procedure

The experiments evaluate the methodology across three rotating-machinery datasets and dynamic or static monitoring scenarios. Detection, diagnosis, stability, and explainability rankings are assessed with task-specific metrics and repeated tests.

  • Datasets: Three datasets cover bearing and gearbox defects, misalignment, unbalance, mechanical looseness, and combined faults.Using different datasets and monitoring approaches is intended to validate the methodology in different scenarios.
  • Datasets: Bearing Dataset Case 1 contains 984 observations: 531 labeled normal and 453 labeled anomalies.The labels were used only to evaluate efficiency and were not used by the anomaly-detection model.
  • Datasets: The bearing analysis uses kurtosis and rms as general features, plus BPFI, BPFO, and BSF as fault-specific features.The specific frequencies indicate defect type, while kurtosis and rms indicate defect presence and global degradation.
  • Analysis approaches: The experiments distinguish dynamic sequential monitoring from static data without temporal correlation.Case 1 uses an updating sliding-window training group, while Cases 2 and 3 divide historical samples into training and test groups.
  • Analysis approaches: Static training groups contain 80% normal samples and 20% randomly selected anomalies.This setup represents mostly normal machine operation and evaluates implementation with anomalies present in the training set.
  • Evaluation: Detection results use F1-Score, PR-AUC, average confusion matrices, and standard deviations, emphasizing correct fault identification.Diagnosis uses accuracy for unsupervised classification and Kendall Tau distance to compare SHAP with Local-DIFFI.

5. Results and discussion

Across three cases, unsupervised anomaly detection identified rotating-machinery faults, including incipient faults, while explainability supported fault classification and root-cause analysis. Performance varied by dataset and task, with some models offering lower computational cost.

  • Fault detection: IF, HBOS, and MCD generally achieved the strongest fault-detection performance, although kNN performed better for some datasets.The best F1-Score models differed by case, and PR-AUC generally favored the same leading models.
  • Fault detection: Anomaly detection models successfully identified faults across the three rotating-machinery cases without supervised labels.The authors conclude that the studied models can detect faults in an unsupervised way, with faults behaving as anomalous observations.
  • Computational cost: HBOS combined strong results with the lowest computational time among the best-performing detection models, supporting potential industrial implementation.Most models had low average training and single-sample testing times, while FastABOD, MCD, and IF required more computation.
  • Fault detection: Isolation Forest classified most anomalies correctly, including 694 of 709 anomalies in Case 3, while normal-sample accuracy was lower in that case.Case 1 classified all normal samples correctly, and Case 2 misclassified two anomalies and two normal samples on average.
  • Fault detection: Anomaly scores separated normal and faulty samples and could reveal incipient faults, but static scores did not show fault evolution consistently.Case 2 showed higher scores for anomalies, whereas Case 3 remained visually difficult to distinguish by fault type; Case 1 showed a gradual score increase near incipient failure.

6. Conclusions

The proposed three-stage methodology enables unsupervised fault detection while providing explainability for root cause analysis and unsupervised fault classification. Its modular anomaly-detection design supports different algorithms and fault types without previously labeled data, while domain adaptation and transfer learning remain future work.

  • 6. Conclusions: The methodology combines feature extraction, unsupervised anomaly detection, and fault diagnosis through unsupervised classification or root cause analysis.Vibration features are extracted in the time and frequency domains using existing human knowledge.
  • 6. Conclusions: The proposed approach enables unsupervised fault detection and provides explainability, root cause information, and unsupervised fault classification.These outcomes were demonstrated across three datasets containing different mechanical faults.
  • 6. Conclusions: Different state-of-the-art anomaly-detection algorithms can be selected according to the dataset.The approach can also be adapted to different fault types by modifying features associated with potential faults.
  • 6. Conclusions: Because the approach does not require previously labeled data and uses vibration-analysis knowledge, it has possible industrial applications.The conclusion links this data requirement and knowledge basis to the methodology’s application scope.
  • 6. Conclusions: Future work will investigate domain adaptation and transfer learning with model-interpretability methods for different industrial scenarios.This is identified as a direction for improving applicability.
Loading 2102.11848v1…