Automatic Data Processing Framework Based Deep Neural Network for Condition Monitoring

With the rapid growth of data volumes and complexity in the field of condition monitoring (CM) of machinery, the need to automate tasks such as information extraction and classification has become more important than ever. Artificial intelligence (AI) remains a promising solution to such challenging tasks. From a learning perspective, the majority of AI based shallow learning methods for CM have been applied for classification, whereas feature extraction task is still manually processed, which requires hand-crafted features based on expertise knowledge. Hence, the classification accuracy of the shallow learning method relies entirely on the quality of the extracted features. Contrariwise, AI based deep learning methods, and in particular the convolutional neural network (CNN) for CM have the capability to address the shortcomings of shallow learning methods by integrating both feature extraction and classification tasks into a single model. With deep CNN architecture, severe problems can appear caused by the activation function layer, which significantly affect the network performance, these include the vanishing gradient and dying ReLU problems. To automate the task of data processing and achieve high classification accuracy for CM, an improved AI framework has been developed in this research: Firstly, a CNN architecture has been designed for automatic feature extraction and classification. CNN was chosen as it has been found to offer several benefits; it can be trained in a supervised learning manner, representative features are extracted directly from the raw data, data dimensionality can be reduced, and it can automatically identify different classes for a given data set. Secondly, to addresses the shortcoming of the existing activation functions, and enhance the learning ability of the network, a hybrid activation function has been developed called the Improved Rectified Linear Unit and Hyperbolic Tangent function (IReLU-Tanh). The developed framework has been implemented and evaluated using both simulated and experimental vibration data. The results shown that the developed CNN architecture with the proposed IReLU-Tanh yields robust classification with high diagnostic accuracy and outperforms the commonly used activation functions Tanh, ReLU, LReLU, and ELU