Session: VIB-15-01: Machine Learning Applications in Vibrations and Dynamics

Paper Number: 88881

88881 - A Multi-Scale Convolutional Network With Attention Mechanism for Fault Diagnosis of Rotating Machines 

Rotating machines consist of a large part of industrial systems. Developing effective fault detection and diagnosis techniques for rotating machines have been studied to ensure safe operation and prevent unexpected economic losses. In general, fault diagnosis is performed in two steps: feature extraction and fault classification. In the feature extraction step, features that are distinctive for different fault modes are extracted. Then in fault classification steps, these features are comprehensively considered to classify each fault mode. Fault extraction is an important step that can greatly affect diagnosis performance, and many feature extraction techniques such as statistical analysis and frequency analysis are popularly used. Recently Convolutional Neural Networks(CNN) shows competitive performance in fault diagnosis tasks, due to their ability for automatic feature extraction and classification. However, designing an effective CNN network for certain tasks required the selection of proper kernel size since kernel size in CNN layers determines temporal scales where the feature is extracted. Fault-related information can be located in different scales depending on the fault, which means that certain scale information may or may not be very important for fault diagnosis tasks. To solve these issues, this paper proposes a multi-scale CNN structure with an attention algorithm. The algorithm consists of multiple convolution paths with different kernel sizes. These paths were connected in parallel to extract features from different scales. Features extracted from different scales are multiplied by weights calculated from the attention mechanism. The attention mechanism measures the scale importance between features and is expressed as a value between 0 and 1. Then, the feature mixing layer extract features from these multiple scales. This process forms a single feature extraction block, and these blocks are stacked to extract complicated fault features. A residual connection is also employed inside each block to prevent saturation problems of the attention mechanism. An experimental study of the PHM data challenge 2009 gearbox testbed is conducted to verify the effectiveness of the proposed algorithm. The accelerometer signal of a sampling frequency of 66.67kHz was used. There were 6 kinds of fault modes, which contain both single fault and combined faults. The input shaft was operated with a rotating speed of 50Hz. Signals of the same operating conditions were obtained twice, and each was used as training data and test data. Raw sensor signals were transformed into temporal segments by the sliding window algorithm and were used as deep learning input. The algorithm contains three additional components compared to a general convolutional neural network: multi-scale path convolution, attention mechanism, and residual connection. The models with and without some of these components were designed and tested on the same dataset to compare model performance. Accuracy, recall, and precision was used as the evaluation metrics of the models. The results show that each of these components contributes to enhancing the feature learning capability of the model and demonstrate that the proposed structure has effective diagnosis performance for rotating machines. Multi-scale convolution structure has contributed most to model performance, and path attention and residual connection followed. Path attention showed an increase in the model accuracy with few parameters. Also, t-distributed stochastic neighbor embedding is conducted to visualize the feature clustering performance of the comparative models. Features from the global average pooling layer which is the final layer of the feature extractor are selected to analyze between models. The results coincide with the performance of the algorithms and test data clustered much better in the proposed method than general CNN. Finally, mean attention values for each fault mode were visualized to analyze how attention is activated inside the algorithm. The results show that attention that is close to the end of the network activates more largely than the one that is close to the start of the network.

Presenting Author: Hyeongmin Kim Seoul National University

Presenting Author Biography: Hyeongmin Kim received a B.S. degree in Mechanical Engineering from Seoul National University, Seoul, Republic of Korea, in 2019. He is currently pursuing a Ph.D. degree at the Department of Mechanical and Aerospace Engineering at Seoul National University, Seoul, Republic of Korea. His current research topics are prognostics and health management for electric drive systems, rotating machines, and thermal power plants.

Authors:

Hyeongmin Kim Seoul National University
Chan Hee Park Seoul National university
Chaehyun Suh Seoul National University
Minseok Chae Seoul National university
Heonjun Yoon Soongsil University
Byeong D. Youn Seoul National University