Synthesizing Rolling Bearing Fault Samples in New
Conditions: A Framework Based on a Modified CGAN (N2FGAN).
​
Accessing fault data is crucial for developing data-driven fault diagnosis tools that can improve both the performance and safety of operations. Industrial machines mainly operate in normal conditions, so, there are more normal data than fault data, resulting in data imbalance. To this end, a novel algorithm based on conditional generative adversarial networks (CGANs) was introduced in this work.
​
This work draws inspiration from the pix2pix framework; an image-to-image translation algorithm introduced by Phillip Ishola et al., (See paper) which involves the translation of images from one domain to another, e.g., a day time image to a night image using a modified CGAN architecture.
​
In our work, a modification of Pix2Pix was adopted on vibration signals because of its generic nature and ability to work well on problems framed as an image-to image translation.
PROJECT INFO
PRODUCT
A journal paper and
Software Solution
TIMELINE
6 Months
ROLE
Research, Writing, Review & Editing
FIRST AUTHOUR
Maryam Ahang
PhD Student
Advanced Control and Intelligent Systems Lab
University of Victoria.
maryamahang@uvic.ca
Other contributors
M. Jalayer - masoudjalayer@uvic.ca
A. Shojaeinasab - ardeshir@uvic.ca
T. Charter - toddch@uvic.ca
H.Najjaran - najjaran@uvic.ca
PROJECT OBJECTIVES
Background / Problem.
-
Condition monitoring is essential for reducing operational costs and asset downtown; machine learning-based condition-monitoring algorithms require data when the machine is in faulty and normal states to make inferences.
-
Normal data are ample as systems usually work in desired conditions, however, fault data are rare.
-
To solve the problem of imbalance dataset, generative algorithms can be used to generate fault data. Generative algorithms are unsupervised learning paradigms that automatically discover patterns in input data so the model can produce new examples.
-
In this paper, a novel method inspired by image-to-image translation by Phillip Ishola et al. [see paper] was introduced and tested on vibration signals to generate fault data from normal data.
IMPLEMENTATION DETAILS
-
​​A literature review of generative algorithms and their applications in condtion monitoring algorithms in the literature was carried out.
-
Implementation of the proposed model, N2FGAN - (normal to fault GAN) on vibration data was done. Pairs of normal and fault data were fed as inputs to the network at given conditions. After the training phase, the network was able to generate new fault data under different conditions.
-
Neural network-based classifiers were used to validate the generated data alongside a statistical comparison with the t-distributed stochastic neighbor embedding (t-SNE) to represent the statistical distributions.
-
Four different deep learning classifiers were used to evaluate the quality of the faulty data generated, a binary LSTM classifier, three multiclass classifiers; convolutional LSTM (ConvLSTM), convolutional neural network(CNN) and convolutional autoencoder (ConvAE) were used.
-
In order to evaluate the performance of N2FGAN compared to other similar algorithms, a comparison panel was developed. Classical augmentation and two state-of-the-art generative algorithms; the Wasserstein GAN with gradient penalty (WGAN-GP) and CGAN, were chosen and implemented for comparison with the N2FGAN. The convolutional LSTM classifier was used to evaluate the data generated from the algorithms in the comparison panel and the N2FGAN performed better than all.
TOOLS AND ALGORITHMS
Programming Languages
Applied ML Algorithms
Data Visualization
-
Python NumPy Stack (NumPy, SciPy, Pandas, Matplotlib).
-
Generative Algorithms: GAN, CGAN and WGAN.
-
Classifiers: Binary LSTM, Convolutional LSTM, Convolutional Autoencoders (ConvAE), Convolutional Neural Network (CNN).
-
T-distributed stochastic neighbor embedding (t-SNE)