rolling element bearing

Fault detection of rolling element bearing (REB), has a very effective role in increasing the reliability of machinery and improving future decisions for rotating machinery operation. In this study, a new method based on a convolutional neural network (CNN) is developed for fault detection of REB. Its performance will be compared with other artificial intelligence (AI) techniques, 2-layer, and deep feedforward neural network (FFNN). In this regard, a set of accelerated-life tests has been implemented on an experimental platform. The models are aimed to recognize the impact pattern in the raw signals generated by faulty REBs. The innovation of the present study is to convert the high-dimensional input as a raw temporal signal to low-dimensional output. The developed method does not need preprocessing of data. Using several types of accelerated tests prevents overfitting. The result shows that the accuracy of the developed CNN-based method is 98.6% for all data sets and 94.6% for the validation dataset. The accuracy of the 2-layer FFNN is 85% for all datasets and 74.2% for the validation dataset and the accuracy of the deep FFNN is 82% for all datasets and 67% for the validation dataset. Therefore, the developed CNN-based method has better performance than the FFNN-based models.

Estimation of remaining useful life (RUL) of rolling element bearings (REBs) has a major effect on improving the reliability in the industrial plants. However, due to the complex nature of the fault propagation in these components, their prognosis is affected by various uncertainties. This effect is intensified when the recorded data is offline, which is very common for many industrial machines due to the lower cost rather than the online monitoring strategy. In the present paper, in order to overcome the shortcoming of the feed-forward neural network (FFNN) in REBs prognostics, a new method for considering two main uncertainties (caused by the measurement and process noises) is proposed, in the presence of offline data acquisition. Inthe proposed method, the primary RUL probability distribution corresponded to each offline measured data is predicted, utilizing the outputs of trained FFNNs. Then, the predicted RUL distribution will become more robust in confronting the temporal changes, by taking into account the approval of pervious stage predictions to the present prediction. As a result, the overall probability distribution of REBs RUL and also its confidence levels (CLs) areobtained. Finally, the evaluation of the proposed method is performed byutilizing bearing experimental datasets. The results show that the proposed method has the capability to express the estimated RUL CLs in the offline data acquisition method, effectively. By providing a probabilistic perspective, the proposed method can improve the reliability of the asset and also the decision-making about the future of the industrial plants.

In this paper, the temperature feature was employed to track down the degradation trend of rolling element bearings. The remaining useful life(RUL) of the rolling element bearing was predicted byassuming root mean square growth (RMS) of the acceleration signal to exponential function form and extraction of two other features. Then, the performance of these features was investigated in the prediction using a recurrent neural network(RNN). The experimental data of the accelerated life test on the rolling element bearing have been extracted from the prognostic. Contrary to the previous works, this paper considers the temperature feature instead of the time feature and also assuming the RMS of the acceleration signal to the exponential function form and using a RNN which causes a newmodel more applicable than previous models.