**1. Introduction**

Elevator is a large-scale complex equipment integrating mechanical, electrical and control. If it breaks down, it will directly affect the safe and efficient operation of the elevator. Nowadays, elevators are used more and more frequently in daily life and production. Therefore, the number of elevators shows a trend of continuous growth [1], and the accompanying elevator failure and maintenance problems are becoming more and more prominent [2,3]. As the power device of the elevator, the traction machine determines whether the elevator can operate normally. With the vigorous development of science and technology, the state recognition technology of elevator traction machine is also constantly improving [4,5].

In recent years, the fault diagnosis technology of elevator traction machine with the artificial intelligence [6] or image processing [7] has been widely applied and developed. However, these methods face problems, such as non-universality of diagnostic model, high cost of model training, and requirement for massive fault samples. In addition, the selection of fault features is also of great significance to the optimization of diagnosis model.

Traction machine is a complex mechanical structure, which is closely connected by various parts. Therefore, the state identification of the traction machine can be diagnosed by various signals, such as vibration, noise, current, temperature, braking torque, speed, and power. Many useful information is hidden in the vibration signal of the traction machine [8]. These signal characteristics can reflect the working condition of the equipment. By analyzing the vibration characteristics of the equipment, the safety operation, accident prevention and maintenance cost reduction can all be accomplished.

Based on the vibration signals, a lot of research have been done in which signal feature extraction methods are the most important section of fault diagnosis [9,10]. The

**Citation:** Li, D.; Yang, J.; Liu, Y. Research on State Recognition Technology of Elevator Traction Machine Based on Modulation Feature Extraction. *Sensors* **2022**, *22*, 9247. https://doi.org/10.3390/ s22239247

Academic Editors: Yongbo Li and Ruqiang Yan

Received: 6 October 2022 Accepted: 24 November 2022 Published: 28 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

time domain method and frequency domain method have been commonly used in early fault diagnosis engineering [11]. Filtering, amplification, statistical feature calculation, correlation analysis, and other time-domain signal processing are referred to as timedomain signal analysis. However, it only reflects the change of amplitude with time and lack frequency bands information. The frequency domain analysis method is to describe the raw signal in the frequency domain, which is more intuitive than the time domain analysis method. However, traditional frequency domain analysis might fail to extract the characteristics information of traction machines due to the heavy background noise and complicated excitation sources [12]. Therefore, only relying on the frequency domain analysis method is far from meeting the current requirements of traction machine fault diagnosis. This brings huge challenges to the status identification and fault diagnosis of traction machines. Therefore, a time-frequency combination processing method has been proposed. Short-time Fourier transform (STFT) [13,14] and wavelet transform (WT) [15] are the common used processing tools with fine time localization and frequency resolution. These methods are realized by superposition of Fourier transform in different fixed window length. However, due to lacking self-adaptability, the quality of feature extraction might be affected by the selection of window function or wavelet basis function.

In addition, the above methods can extract the characteristics of vibration signals, but the collected vibration signals usually contain background noise and unknown frequency interference. To eliminate noise component and extract the fault feature information of raw vibration signals, several demodulation techniques have been applied to past research, such as Hilbert transform (HT) [16], empirical mode composition (EMD) [17], spectral kurtosis (SK) [18], nonstationary analysis [19–21], and cyclostationary analysis [22–24]. These methods have been applied to modulation frequency extraction already, which noted the modulation mechanism in a rotating machine.

Feng et al. [25,26] proposed an adaptive iterative generalized demodulation method to extract the modulation features in nonstationary analysis. The vibration characteristics of hydraulic turbine and planetary gearbox have been successfully found in the joint timefrequency domain. Most vibration signals of traction machine are non-stationary signals, but they are cyclostationary signals, namely, the correlation function of traction machine signals is periodic function of time. In view of the cyclostationary analysis theory, a variety of methodologies have been proposed, in which cyclic modulation spectrum (CMS) and fast spectral correlation (Fast-SC) are two typical cyclostationary tools [22]. However, they did not gain its deserved attention because of high computational cost.

Wang et al. [27] improved the cyclostationary methods with an application of Teager Kaiser energy operator (TKEO), which can enhance fault feature recognition with low computational burden. Song et al. [28,29] proposed a demodulation method based on time-frequency analysis (TSA) and principal component analysis (PCA) and applied it to the modulation frequency extraction of pump and permanent magnet synchronous motor (PMSM). Moreover, due to dimensionality reduction of time-frequency distribution matrix, the burden of high computational cost was greatly relieved. The main process of the algorithm is as follows: Firstly, the raw vibration signal is transformed into time-frequency domain by STFT. Then, the PCA method is used to reduce the dimensionality of the timefrequency spectrum in order to extract the eigenvalues of the principal components. Finally, the principal components are reconstructed to obtain the modulation signals.

Among the above demodulation methods, it could be found that the demodulation method base on PCA (DPCA) has great potential for applications in traction machine. In addition, although the fault diagnosis technology of elevator traction machine based on artificial intelligence or image processing has been widely applied and developed. However, few investigations have been done to extract and analyze the modulation features of traction. The modulation mechanism of traction machine has also rarely been involved. These above issues have greatly hindered the development of elevator fault diagnosis technology.

In this paper, the modulation characteristics of the traction machine vibration signal were extracted through a demodulation method based on time-frequency analysis and principal component analysis (DPCA). The characteristics extracted by DPCA is more prominent under the interference of background noise and unknown frequency, which is helpful to the state identification of the traction machine. The principle of signal demodulation method and experiential setting are introduced respectively in Section 2. In Section 3, the vibration signal of the traction machine is processed by FFT, STFT and DPCA methods. The influence of different working conditions on the vibration of traction machine is discussed, which shows the superiority of the demodulation technology. Finally, the conclusions are drawn in Section 4.
