A Stacked Autoencoder Neural Network based Automated Feature Extraction Method for Anomaly detection in On-line Condition Monitoring

Mohendra Roy, Sumon Kumar Bose, Bapi Kar, Pradeep Kumar Gopalakrishnan, Arindam Basu

Abstract

This paper presents a hybrid neural network architecture designed for real-time machine condition monitoring. By combining a Traditional Autoencoder (TAE) for automated feature extraction with an On-line Sequential Extreme Learning Machine (OSELM) for anomaly detection, the system eliminates the need for domain-specific handcrafted features. Evaluated on the NASA bearing dataset, the method achieves 100% detection accuracy with a lightweight structure suitable for Edge computing and IoT-based predictive maintenance.

Introduction

In industrial environments, predicting mechanical failures—particularly in critical components like motors and bearings—is essential to preventing costly unplanned downtime. Traditional condition monitoring typically relies on analyzing raw sensor data, which is computationally intensive and noise-prone, or on manually extracting "handcrafted" features such as root mean square, kurtosis, or skewness.

Handcrafting features requires significant domain expertise and extensive trial-and-error to identify metrics that reliably indicate faults for specific machinery. As the industry shifts toward IoT and edge computing, there is an increasing demand for automated, lightweight feature extraction methods that can operate in real-time on low-power devices without requiring prior domain knowledge.

Proposed Methodology

The authors propose a stacked neural network architecture that operates in two distinct phases: offline feature learning and online anomaly detection.

1. Automated Feature Extraction (Offline Training)

Instead of manual feature engineering, a Traditional Autoencoder (TAE) is employed to learn a compressed representation of raw vibration signals. Raw data is first smoothed by averaging every five samples to mitigate high-frequency noise, reducing dimensionality from 20,480 to 4,096 points. A single-hidden-layer TAE is then trained offline to compress these inputs into just five latent features.

2. Anomaly Detection (Online Learning & Inference)

The five extracted features are passed to an On-line Sequential Extreme Learning Machine (OSELM), which functions as a one-class classifier. The OSELM is trained online using data from the machine's initial healthy state. Because OSELMs update weights sequentially and only at the output layer, they offer exceptionally fast convergence.

Figure 1: The proposed hybrid architecture showing the transfer of the trained encoder from the TAE to the OSELM for online inference.

During the inference phase, the system monitors the OSELM's output. If the output deviates from the learned "healthy" baseline beyond an adaptive threshold—calculated using the mean and standard deviation of training errors—an anomaly is flagged.

Experimental Setup and Datasets

The model was validated using the NASA Bearing Dataset from the Center for Intelligent Maintenance Systems. This dataset includes vibration signals from 12 bearings across three datasets, recorded at a 20kHz sampling rate. It contains ground-truth labels for various health states, including inner race defects, roller element defects, and outer race failures.

Figure 5: Summary of the NASA Bearing Datasets used for training and validation, detailing sample counts and fault conditions.

The authors utilized a leave-one-out cross-validation strategy, training the autoencoder and optimizing thresholds on 11 bearings while testing on the 12th to ensure robust performance across different mechanical conditions.

Key Results and Discussion

  • Detection Accuracy: The automated feature extraction method achieved 100% detection accuracy in distinguishing healthy from faulty bearings, perfectly aligning with the NASA dataset's ground truth.
  • Performance Comparability: The data-driven features performed on par with a baseline model using five traditional handcrafted features. This demonstrates that automated feature learning can effectively replace manual engineering without sacrificing diagnostic precision.
  • Computational Efficiency: The OSELM classifier demonstrated rapid convergence, requiring an average of only 582 samples to define a healthy baseline. This allows the system to become operational within the first 20% of a typical data stream, making it ideal for rapid deployment on edge devices.

Conclusion and Future Work

The study proves that a TAE + OSELM stack can successfully automate feature extraction and anomaly detection for industrial condition monitoring. By learning machine-specific "healthy" baselines directly from data, the system bypasses the limitations of universal classifiers and domain-specific feature engineering.

While the current adaptive thresholding relies on a sensitivity hyperparameter ($K$) that requires initial calibration, the overall simplicity of the architecture is highly promising for hardware implementation. Future research will focus on deploying this model onto power-efficient VLSI hardware to create autonomous, edge-computing sensors for industrial IoT applications.

Summary updated

BibTeX
@inproceedings{roy2018stacked,
  author = {Mohendra Roy and Sumon Kumar Bose and Bapi Kar and Pradeep Kumar Gopalakrishnan and Arindam Basu},
  title = {{A Stacked Autoencoder Neural Network based Automated Feature Extraction Method for Anomaly detection in On-line Condition Monitoring}},
  booktitle = {SSCI},
  year = {2018},
  url = {https://openreview.net/forum?id=WGUsKw5zJF},
}
Figure