Image representation of vibration signals and its application in intelligent compound fault diagnosis in railway vehicle wheelset-axlebox assemblies

Axlebox vibration signals contain critical status information regarding the operating state of a wheelset-axlebox assembly in railway vehicles, and an increasing number of studies have demonstrated the superiority of operating state diagnostic methods based on machine learning that apply image representations of vibration signals as input data. In this regard, the Markov transition field algorithm has been demonstrated to represent the state features of a time series efficiently. However, the frequency spectrum of axlebox vibrations has been shown to provide superior performance in these diagnostic methods. Unfortunately, the conventional Markov transition field algorithm is not applicable to the frequency domain. The present study addresses this problem by proposing a novel spectral Markov transition field (SMTF) algorithm to represent the frequency spectrum characteristics of vibration signals in image form through the construction of the firstorder Markov transition matrix of the frequency-domain signal. The proposed algorithm employs no parameters requiring prior knowledge. Moreover, the machine learning network is efficiently trained to recognize specific features related to compound faults using the transfer learning technique, which enables the network to obtain highly separable and intelligent image classification results under small sample conditions. The performance of the SMTF algorithm is evaluated experimentally using a scaled test rig to collect vibration data for three types of artificial wheelset-axlebox faults under three different velocities. In addition, three other state-of-the-art image representation algorithms are employed for comparison. The results demonstrate that the proposed SMTF algorithm effectively represents compound fault features comprehensively, and the developed fault diagnosis framework based on the SMTF representations obtains superior classification results with much greater separability under all three velocity conditions. The present study represents a foundation for developing intelligent condition monitoring for other rotating machinery. Axlebox vibration signals contain critical status information regarding the operating state of a wheelset-axlebox assembly in railway vehicles, and an increasing number of studies have demonstrated the superiority of operating state diagnostic methods based on machine learning that apply image representations of vibration signals as input data. In this regard, the Markov transition field algorithm has been demonstrated to represent the state features of a time series efficiently. However, the frequency spectrum of axlebox vibrations has been shown to provide superior performance in these diagnostic methods. Unfortunately, the conventional Markov transition field algorithm is not applicable to the frequency domain. The present study addresses this problem by proposing a novel spectral Markov transition field (SMTF) algorithm to represent the frequency spectrum characteristics of vibration signals in image form through the construction of the first order Markov transition matrix of the frequency-domain signal. The proposed algorithm employs no parameters requiring prior knowledge. Moreover, the machine learning network is efficiently trained to recognize specific features related to compound faults using the transfer learning technique, which enables the network to obtain highly separable and intelligent image classification results under small sample conditions. The performance of the SMTF algorithm is evaluated experimentally using a scaled test rig to collect vibration data for three types of artificial wheelset-axlebox faults under three different velocities. In addition, three other state-of-the-art image representation algorithms are employed for comparison. The results demonstrate that the proposed SMTF algorithm effectively represents compound fault features comprehensively, and the developed fault diagnosis framework based on the SMTF representations obtains superior classification results with much greater separability under all three velocity conditions. The present study represents a foundation for developing intelligent condition monitoring for other rotating machinery. (c) 2020 Elsevier Ltd. All rights reserved.