Publication

Hyperautomation Artificial Intelligence

Window Size and Sampling Rate Selection for Cost-optimal Deep Learning-based Fault Diagnosis

본문

Conference
Asian Congress of Structural and Multidisciplinary Optimization (ACSMO) 2024
Author
Hyeongmin Kim and Byeong D. Youn
Date
2024-05-21
Presentation Type
Oral

Abstract 


Diagnosing faults in rotating equipment is essential to prevent economic losses due to equipment downtime. Recently, due to the development of IoT sensors and advancements in computing and data storage technology, many deep learning-based rotating machine fault diagnosis methods have been developed. Deep learning-based fault diagnosis methods can automatically extract fault features and have high diagnosis performance when enough data is trained. However, to effectively apply a deep learning-based fault diagnosis method in industry, it is necessary to minimize computing and data storage costs. For cost-effective diagnosis, 1) the sampling rate to acquire the signal and 2) the window length of the deep learning(DL) input[3] should be carefully set. If the sampling rate and window size become too large, the amount of data storage and the time required for model learning and analysis increases. Conversely, if the sampling rate and window size are too small, fault characteristics cannot be extracted, causing diagnostic performance to deteriorate. The problem is that the sampling rate and window size of most of the deep learning-based fault diagnosis methods are set heuristically. To solve this problem, this paper proposes a novel method to automatically determine the optimal sampling rate and window size without degrading diagnostic performance. First, the proposed method measures the performance and diagnosis uncertainty of a DL model of the maximum sampling rate and window size. Next, the proposed method maintains the performance of data samples while reducing the window length and sampling rate and selects a minimum point where uncertainty does not increase significantly. The proposed method was verified on an actual rotating machine dataset. The result shows that the proposed method can set the sampling rate and window minimal without deteriorating diagnostic performance.