Development of deep learning-based methods for rotating machinery fault diagnosis under varying speed conditions

  • Author / Creator
    Rao, Meng
  • Rotating machines are widely used in industrial applications, such as driving motors in elevators and gearboxes in wind turbines. Machines in these applications often operate under varying speed conditions due to variable operation demand, ever-changing environment conditions and so on. As time goes on, machines in service would be inevitably deteriorated. When the deterioration is accumulated to a certain level, faults may occur. Faults if not detected timely and maintained properly would result in the shutdown of a machine, then economic loss and even catastrophic disasters. To avoid these unexpected consequences, fault diagnosis, whose goal is to detect the occurrence and then classify the type and the severity of a fault, is of vital importance.
    Deep learning is widely used for fault diagnosis in the current big data era thanks to its automation nature and capability of processing massive data. However, performances of deep learning-based fault diagnosis are highly influenced by speed variation. Models perform well under constant speed conditions may fail under varying speed conditions. Specifically, when deep learning is adopted for fault diagnosis, we input condition monitoring data which are usually vibration signals to a deep learning model. The model performs fault feature learning and pattern recognition, and ultimately outputs diagnosis results. The problem is that speed variation induces additional features to vibration signals, and unfortunately, features induced by speed variation are often overlapped with features of faults. This increases the difficulty of learning sensitive fault features and thus impedes the fault diagnosis performances. Therefore, how to address the effects of speed variation is a key concern to facilitate deep learning-based fault diagnosis under varying speed conditions.
    The objective of this research is to develop deep learning models that can address the effects of speed variation, and ultimately achieve effective fault diagnosis for rotating machinery that operated under varying speed conditions. This research includes three topics. First, considering that rotating speed is frequently required for effective fault diagnosis but sometimes is not feasible to be measured, a new deep learning model named many-to-many-to-one bi-directional long short-term memory (MMO-BiLSTM) is proposed to extract rotating speed from vibration signals. The proposed model can work like a virtual speed meter, that is, automatically output synchronized rotating speed corresponding to given vibration signals. Second, a new deep learning model named speed normalized autoencoder (SN-AE) is proposed for fault detection under varying speed conditions. The proposed model automatically removes speed variation induced amplitude modulation in vibration signals and thus achieves better fault detection performances. Given that a fault being detected, the last topic aims to classify the type and severity of this fault. An auxiliary branch named speed adaptive gate (SAG) is proposed for existing deep learning models to improve their fault classification accuracy under varying speed conditions. The proposed SAG addresses speed induced fault information imbalance and therefore yields higher fault classification accuracies. Both the second topic and the third topic require the rotating speed as an auxiliary input. The rotating speed can be measured or extracted from the first topic.
    This research would promote the frontier of deep learning-based fault diagnosis, especially for varying speed conditions. The outcome of this research could serve as a good reference for engineering practitioners in industrial applications for a better maintenance. This research only considers the varying speed condition. The load is assumed constant. In the future, we will investigate the fault diagnosis of rotating machinery under varying load conditions.

  • Subjects / Keywords
  • Graduation date
    Spring 2023
  • Type of Item
  • Degree
    Doctor of Philosophy
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.