Deep learning-based anomaly-onset aware remaining useful life estimation of bearings

View article
Just published in @PeerJCompSci - Deep learning-based anomaly-onset aware remaining useful life estimation of bearings Read the full article https://t.co/JuFBKuvOkS Research by Kamat et al. @symbiosistweets #AI #DataMining #MachineLearning #DataScience https://t.co/5s5HOjMU5c
Just published in @PeerJCompSci - Deep learning-based anomaly-onset aware remaining useful life estimation of bearings Read the full article https://t.co/eNYRispZdS Research by Kamat et al. @symbiosistweets #AI #DataMining #MachineLearning #DataScience https://t.co/9RcRwwnkWG
PeerJ Computer Science

Main article text

 

Introduction

  • Sequential Dependencies: The numerous machinery components interact in a myriad fashion, resulting in intricate temporal correlations between different sensors. Exploiting sensor readings to capture complicated machine operating characteristics with respect to time dependencies is crucial and necessary (Xu et al., 2020).

  • Noisy Sensor Measurements: Sensor measurements are interfered with background environmental noise. Each sensor may have different noise bands, and capturing them in the model might be difficult (Luo et al., 2020).

  • Health Degradation curve: In model-based systems, developing a health degradation curve based on various components requires high domain expertise and an enormous amount of historical data (Lin et al., 2020).

  1. The application of unsupervised clustering approach using the k-means technique with Silhouette Coefficient for anomaly trend analysis.

  2. The development of a novel deep-learning based Anomaly-Onset Aware RUL (AOA-RUL) prediction framework for bearing machinery using selected time-domain extracted features based on probability ranking. The anomaly detector indicates the time stamp at which the first anomaly occurred. The anomaly detection is developed using the hybrid Autoencoder-LSTM technique.

  3. Further, the RUL regressor is trained and evaluated on degradation onset data alerted by the anomaly detector. The Anomaly-Onset Aware RUL prediction’s primary goal is to prevent RUL estimation during healthy operations when the RUL estimation is not necessary. Various LSTM variants are used to implement the RUL regressor, and a comparative analysis is presented.

Materials & Methods

Dataset description

Anomaly-onset aware RUL prediction framework

Stage 1: Data preprocessing and feature extraction

Stage 2: Feature ranking and feature selection

  • Linear Regressor: The primary objective of regression analysis is to develop an equation that can predict values for some predictor variables for all members of a population given explanatory variables. Linear regressor makes use of recursive elimination function for feature selection (Chen et al., 2017). RFE is widely used because it is simple to set up. It successfully identifies which features (columns) in a training dataset are more significant in predicting the target variable. The following is an example of a linear regression line equation:

    y=a+bX

    Here X is the independent variable, which is displayed on the x-axis, and y is the dependent variable, which is shown on the y-axis. The intercept (the value of y when x = 0) is a, while the slope of the line is b.

  • Random Forest Regressor: Many decision trees make up a random forest. Every node in the decision trees is a feature qualifier to divide the dataset into two groups with similar response values. Impurity is the criterion by which the (locally) best condition is selected. As a result, it is possible to calculate how much each feature reduces the weighted impurity in a tree while training it. The impurity decrease from each feature in a forest may be averaged, and the features can be ranked based on this metric. High-ranking features can then be selected for optimal model performance (Ubing et al., 2019). On a decision tree, the significance of each feature is computed as follows:

    fii=j:nodejsplitsonfeatureinijk  allnodesnik

    where fisub(i)=feature i’s importance

    nisub(j)= node j’s importance

    By dividing by the total of all feature importance values, they may be normalized to a number between 0 and 1:

    normfii=fiij  allfeaturesfij

    At the Random Forest stage, the average of all the trees is the most important attribute. The total number of trees is divided by the sum of the feature’s importance rating on each tree:

    RFfii=j  alltreesnormfijT

    where RFfisub(i)= feature i ’s importance calculated from all the trees in the Random Forest.

  • Mutual Info Regressor: Mutual information is calculated for fixed categories in a classification task or a continuous target variable in regression situations. The entropy of the variables is used to calculate mutual information. Mutual information (MI) between two random variables is a non-negative number that indicates how dependent the variables are on each other. By rating their mutual information with the target variable, we may choose the features from the feature space. The advantage of mutual information over the different statistical techniques is that it can handle non-linear relationships between features and target variables (Hashemi, Dowlatshahi & Nezamabadi-pour, 2021). Formally, the mutual information between two random variables X and Y is as follows:

    I(X;Y)=H(X)H(X|Y)

    where I(X;Y) represents mutual information between X and Y, H(X) represents entropy for X and H(X|Y) represents conditional entropy for X given Y.

Stage 3: Unsupervised clustering for generation of fault diagnosis data

where it consists of the following two elements:

  • The average distance between a sample and the rest of the samples in the same cluster a

  • The average distance between a sample and the neighboring cluster b’s samples

Stage 4: Semi-supervised anomaly detection for fault diagnosis

Stage 5: Anomaly-triggered remaining useful life prediction for fault prognosis

  • Vanilla LSTM: The Vanilla LSTM is a basic LSTM configuration. Vanilla LSTM is the name given to it in this to distinguish it from deeper LSTMs and the suite of more complex configurations. An LSTM model with a single hidden layer of LSTM units and a prediction layer is called a Vanilla LSTM (Xu et al., 2020). If we add multiple LSTM hidden layers, the architecture can be converted into a stacked-LSTM model.

  • Bi-Directional LSTM: By walking through input time steps in both the forward and backward directions, bidirectional LSTMs address the challenge of getting the most out of the input sequence. This model is implemented by cloning the first recurrent layer of the network such that two layers are now side-by-side, then feeding the input sequence to the first layer and a reversed duplicate of the input sequence to the second layer (Wang et al., 2019).

  • Conv-LSTM: In the Conv-LSTM architecture, Convolutional Neural Network (CNN) layers for feature extraction on input data are combined with LSTMs to give sequence prediction. The ConvLSTM is a kind of LSTM similar to the CNN-LSTM in that each LSTM unit contains a convolutional reading of the input. The ConvLSTM was created to read two-dimensional spatial-temporal data, but it may also be used to forecast multivariate time series (Li et al., 2021).

  • Encoder-Decoder LSTM: Sequence-to-sequence prediction issues (seq2seq) are a more difficult form of problem in which a sequence is given as input and a sequence prediction is required as output. The Encoder-Decoder is made up of two sub-models: one that reads the input sequence and encodes it into a fixed-length vector, and the other that decodes the fixed-length vector and predicts the sequence (Liu et al., 2021). The encoder is often a Vanilla LSTM model, although alternative encoder models such as Stacked, Bidirectional, and CNN models can also be utilized. This model may be used to forecast multi-step time series.

Results

Feature extraction

Feature ranking and feature selection

Unsupervised clustering for generation of fault diagnosis data

Semi-supervised anomaly detection for fault diagnosis

Anomaly-triggered RUL prediction for fault prognosis

Discussion

Comparative analysis of anomaly detection stage

Comparative analysis of anomaly triggered RUL estimation stage

Conclusions

  1. Development of clustering technique for anomaly pattern analysis on unlabeled data.

  2. Combined deep-learning-based anomaly triggered RUL estimation technique using hybrid semi-supervised Autoencoder-LSTM for anomaly detection having accuracy in the range of 60 to 90% across various bearings.

  3. LSTM based RUL estimators having prediction accuracy of over 90% for variants of Vanilla LSTM, Bidirectional LSTM, ConvLSTM, and Encoder-Decoder LSTM.

Additional Information and Declarations

Competing Interests

The authors declare that they have no competing interests.

Author Contributions

Pooja Vinayak Kamat conceived and designed the experiments, performed the experiments, analyzed the data, performed the computation work, prepared figures and/or tables, authored or reviewed drafts of the paper, and approved the final draft.

Rekha Sugandhi conceived and designed the experiments, analyzed the data, authored or reviewed drafts of the paper, and approved the final draft.

Satish Kumar conceived and designed the experiments, analyzed the data, performed the computation work, authored or reviewed drafts of the paper, and approved the final draft.

Data Availability

The following information was supplied regarding data availability:

The code is available at GitHub: https://github.com/PoojaPaiKamat/AnomalyOnsetAwareRUL.

The Pronostia IEEE 2012 bearing dataset (Femto dataset) is available at NASA: “FEMTO Bearing Data Set”, NASA Ames Prognostics Data Repository (http://ti.arc.nasa.gov/project/prognostic-data-repository), NASA Ames Research Center, Moffett Field, CA. https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/.

Funding

This work is funded by the Research Support Fund of Symbiosis International (Deemed) University. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

30 Citations 2,717 Views 600 Downloads

Your institution may have Open Access funds available for qualifying authors. See if you qualify

Publish for free

Comment on Articles or Preprints and we'll waive your author fee
Learn more

Five new journals in Chemistry

Free to publish • Peer-reviewed • From PeerJ
Find out more