Supervised Health Stage Prediction Using Convolutional Neural Networks for Bearing Wear

Abstract

Early detection of faults in rotating machinery systems is crucial in preventing system failure, increasing safety, and reducing maintenance costs. Current methods of fault detection suffer from the lack of efficient feature extraction method, the need for designating a threshold producing minimal false alarm rates, and the need for expert domain knowledge, which is costly. In this paper, we propose a novel data-driven health division method based on convolutional neural networks using a graphical representation of time series data, called a nested scatter plot. The proposed method trains the model with a small amount of labeled data and does not require a threshold value to predict the health state of rotary machines. Notwithstanding the lack of datasets that show the ground truth of health stages, our experiments with two open datasets of run-to-failure bearing demonstrated that our method is able to detect the early symptoms of bearing wear earlier and more efficiently than other threshold-based health indicator methods.

Generalized multiscale feature extraction for remaining useful life prediction of bearings with generative adversarial networks

“…The assumption is labeling a small portion of the run-to-failure dataset where the signals are clearly distinguishable into healthy and unhealthy when no labels for health stages are given. In the previous study [26], we found that the FPT is not sensitive to p ranging from 0.05 to 0.5. p = 0.05 is used in the experiments in Section 4. By the CNN-based binary regression model, the HS and the FPT of the jth bearing, T F P T j , can be determined.…”

“…Although unsupervised methods [25], such as auto-encoder, promise the extraction of efficient health indicators(HI); however, they possess a challenge assigning an appropriate threshold for determining FPT. Recently, the CNN-based HS division methods have been proposed, but they require an appropriate threshold to determine the FPT [24] or need to determine bandwidths, where there are significant differences between normal and faulty conditions [21,26]. Li et al [24] proposed a GAN-based RUL prediction approach that consists of two stages.…”

“…However, it still requires an appropriate threshold to determine the FPT and needs the target domain data to train the DNN. Our previous works [21,26] proposed a supervised HS division method using CNN and an NSP, which transforms the time-series data into an image for feature extraction, rids the necessity of designating a threshold of health index. However, it still needs to determine bandwidths of bandpass filters manually in the data preprocessing procedure.…”

“…NSP is a data wrangling method that transforms correlated time series data into an image for multi-variate correlation analysis [21,26]. It is a matrix representation similar to the heat map of the quantized value of time series data.…”

“…In the previous works [21,26], the combination of NSP, which allowed us to integrate a frequency analysis that helped to extract faulty wear features, and a CNN-based binary regression model, rid the necessity of designating a threshold, one of the most crucial problems in early fault detection. The CNN based model effectively distinguished healthy and unhealthy states unambiguously by changing the anomaly detection problem into a binary regression problem, utilizing a simple trigger mechanism.…”