Search

Distributed bearing faults are the most common ones in industry and create random vibration patterns, which make their detection difficult. They are caused by lubrication issues, contamination issues, electrical erosion, bearing roughness, or the spread of a local fault. This research mainly focuses on the distrusted bearing faults diagnosis using a multi-sensory kit. For this purpose, a novel deep-learning framework is proposed to detect these faults using 3 axis vibrations and one stray magnetic flux signal. The data is collected at 50 operating points, i.e., 10 speed and 5 torque levels. The proposed architecture benefits from a multi-input pipeline consisting of time-frame signals and extracted features. A feature-rich architecture is proposed combining convolutional and high-level information. Although a deep learning structure coherently learns from the features through convolutional and LSTM layers, 20 predefined features sampled from each instance are also fed into the network to improve accuracy. The robustness of the overall system is validated with train/test split data. Deep learning results are compared with two more classification algorithms, SVM and XGBoost. The high accuracy of the proposed model demonstrates the superiority of the deep learning architecture for distributed bearing fault detection.

Developing a simple real-time bearing fault detection algorithm is essential for small cooling fan motors, ensuring power electronics systems and data center stability. Our study introduces a real-time, low-resource algorithm for microcontrollers, utilizing motor current data to detect the most common bearing faults in these applications. Notably, small motors exhibit significant current changes due to lubrication and contamination issues, unlike larger motors, where such changes are minimal. We comprehensively assess distributed bearing faults stemming from lubrication and contamination across seven motors under various conditions. These motors are tested under both no-load and fan-load scenarios at ten different speeds, with controlled aging of motor bearings. Key time-domain features, such as Root Mean Square (RMS), peak values, and crest factors of motor current, are scrutinized to create our proposed algorithm. We rigorously evaluate sensitivity, false detection scenarios, and compare our algorithm to a machine learning model. In practical experiments using the TI F280049 microcontroller, our algorithm outperforms, demanding minimal instruction cycles and memory resources. Achieving an accuracy rate exceeding 92% and consistently demonstrating an F1 score of over 90%, the algorithm is proven to be a robust and practical solution for precisely and rapidly detecting distributed bearing faults in small cooling fan motors.

Distributed bearing faults are highly common in industrial applications and display unpredictable vibration patterns impeding their detection. These faults stem from issues such as lubrication deficiencies, contamination, electrical erosion, roughness of the bearing surface, or the propagation of localized faults. This study aims to detect distributed bearing faults by utilizing a multisensory approach consisting of current, accelerometer, and fluxgate sensors. A novel 2-D deep learning framework is proposed, leveraging signals from six distinct sources, including three-axis vibration signals, stray magnetic flux signal, and two-phase current signals. Data are collected from 3- and 10-hp induction motors at 50 operational points, spanning ten speed levels and five torque levels. These six signals are transformed into matrices and combined to create a comprehensive matrix that provides an overall depiction of the bearing condition. The proposed deep learning architecture employs a 2-D convolutional model, which takes 2-D images as input and determines the bearing status. To evaluate the system's robustness, the data are divided into training and testing sets. The proposed model demonstrates remarkable effectiveness in detecting distributed bearing faults, achieving an impressive accuracy rate of 99.92%. Furthermore, a comprehensive comparison is provided, highlighting the impact of using various sets of inputs as sources for the deep learning model on the accuracy rate for each set. Through the analysis of the obtained results, a clear conclusion can be drawn: the model performs at its best when all six input sources are utilized.