Wire rope is used for lifting, pulling, tensioning, and bearing in material handling machinery. It has many advantages such as high strength, light deadweight, and stable operation. However, the working environment of the wire rope is harsh, and it is prone to wear, breakage and other damages in use. Once the wire rope is damaged, it will affect its bearing capacity, which is easy to break and causes economic losses and casualties. Therefore, it is significant to implement damage detection for wire rope.
There are many non-destructive testing methods for wire ropes, including ultrasonic testing [1], radiographic testing [2], electromagnetic testing [3], etc. MFL testing is non-destructive testing with high efficiency and reliability, which comprehensively detects the internal and external defects of ferromagnetic materials [4, 5]. This testing method is suitable for wire rope inspection [6, 7]. The detection principle is to conduct axial magnetization of the wire rope through DC coils or permanent magnet instruments, detect the MFL at the damaged part through the Hall sensor and output the MFL signal. According to the MFL signal, the damage type of the wire rope can be determined.
The MFL signal under dreadful working conditions is interfered with by noise, and it is difficult to recognize the damage. To deal with the damage identification of the MFL signal under noise background, many scholars have done relevant research work. Among them, Shan et al proposed an adaptive shift average algorithm method to denoise the MFL signal [8]. They found the optimal window width matrix through iterative optimization and obtained a high signal-to-noise ratio. Yao et al. conducted denoising processing based on wavelet multi resolution analysis methods, and effectively eliminated abnormal points, power frequency interference, and background noise through feature decomposition and reconstruction of broken wire damage signals [9]. The method suppresses the strand-waveform noise and highlights the signal. The above methods can denoise the MFL signal, but they cannot intelligently find the damage.
With the development of machine learning, quick identification and classification are deeply concerned by scholars in fault diagnosis [10–14]. At the same time, the combination of the denoising method and machine learning provides a new direction for wire rope damage identification. Zhang et al conducted wavelet denoising for the MFL signal of the wire rope [15]. Then they conducted damage classification and feature extraction in space, and finally input the features into a wavelet neural network for damage identification. Kim and Park used Hilbert transform to denoise and quantified the MFL signal by using the damage indexes [16]. Then they used an artificial neural network to automatically estimate the severity of the damage. These methods have high accuracy for the identification of the wire rope damage under noise background. However, signal denoising and eigenvalue extraction are complex and tedious. It is very necessary to find a method that combines machine learning with damage identification without denoising and feature extraction.
CNN in machine learning has fast speed, high accuracy, and good robustness for image classification [17–19], so it is widely used in fault identification and classification [20–22]. Some scholars use the characteristics of CNN to convert the damage signals in the fault field into images and input them into CNN for damage identification. Chen et al fused the horizontal and vertical vibration data of the gearbox into a two-dimensional matrix and used DCNN to identify the health of the gearbox [23]. Sun et al transformed the 1D fault signals of the gas sensor into 2D gray images and used the CNN to improve the accuracy of fault diagnosis of hydrogen sensors [24]. Yang et al converted the multi-source vibration signal into a two-dimensional matrix and used CNN for fault diagnosis of reciprocating compressors [25]. At the same time, CNN also has a certain noise resistance. Hoang et al proposed a rolling bearing fault diagnosis method based on CNN. They converted vibration signals into images. Without denoising, it effectively classified damage in noisy background, and had certain robustness and fault tolerance [26]. EfficientnetV2, as a newly proposed classification network, increases the network width, depth, and resolution to improve the performance [27, 28]. It has been used in plant disease detection [29, 30], mechanical fault diagnosis [31], and some other fields [32, 33].
Compared with the signal, the image will be more intuitive. Many scholars are committed to finding the conversion relationship between the MFL signal and the MFL image and observing the damage more intuitively through the MFL image. For example, Li et al and Zheng et al both obtain multiple MFL signals through multiple Hall sensor arrays and the obtained numerical matrix forms the MFL image [34, 35]. However, multiple MFL signals to MFL images require a large amount of computation. It is necessary to transfer the single MFL signal to the MFL image. Inspired by the transformation of the bearing vibration signal into the image signal [36–38], the MFL signal is preprocessed by wavelet transform and array segmentation to form a numerical matrix, and the numerical matrix is transformed into the image. Compared with the bearing vibration signal, the MFL signal is mostly the pulse signal or the abnormal amplitude signal, which makes it easier to see the damage of the wire rope in the MFL image.
Previous studies of denoising and machine learning have been able to identify the damage of the wire rope. However, the denoising and feature extraction for MFL signals are complicated, and it is also unable to combine CNN with higher classification accuracy. To deal with the problems, we propose a method of wire rope damage identification via Light-EfficientNetV2 and MFL image. The overall flow chart is shown in Fig. 1. Firstly, the MFL signal of the wire rope under noise background is segmented to form a numerical matrix and then turn into an MFL image. Then, the Light-EfficientnetV2 is selected in the CNN, which reduces the number of layers and accelerates the training efficiency. Finally, the obtained images are divided into the training set and the verification set and then brought into the network for damage identification.
This paper is arranged as follows: Section 2 introduces the principle of MFL testing, the characteristics of the noise, and the test rig. Section 3 transforms the MFL signal to the MFL image. Section 4 presents the Light-EfficientNetV2 network and the training results for MFL images. The last section gives the conclusion of this paper.