A Deep Learning Based Method for Detecting of Wear on the Current Collector Strips’ Surfaces of the Pantograph in Railways

In pantographs, current collector strips transmit the electrical energy they receive from the catenary to the locomotive and provide the necessary power for the locomotive’s movement. In order for the current collector strips to transmit electricity to the locomotive in a healthy way, their surface must be smooth. Wear on the surface of the current collector strips reduces conductivity and can create arcs, endangering the health and safety of the pantograph and catenary system. In this paper, a Convolutional Neural Network (CNN) architecture is developed to detect wear on the current collector strips. Images obtained from pantographs used on railways were created with a clean and improved data set using Hough Transform and Power Law Transform. The created dataset contains 909 pantograph images. This dataset was trained and tested with both the developed CNN architecture and classic deep learning architectures (ResNet50, VGG16). The experimental results show that the developed CNN architecture’s training results and test results are more successful than classical architectures.


I. INTRODUCTION
Rail systems constitute the backbone of the economy throughout the world, especially considering its freight transportation potential. With the developing technology, the increasing speed in rail systems has exponentially increased the importance of safety [1]. For this reason, condition monitoring techniques are becoming widespread in the railway industry to increase safety and reduce maintenance costs [2]. In a constantly growing network of countless objects and dangers, such as railways, manual condition monitoring and periodic maintenance are becoming increasingly difficult [3]. In recent years, studies on automatic condition monitoring and predictive maintenance methods have been carried out considering these situations. The power required for the operation of electric trains, whose usage prevalence has increased gradually in recent years and is expected to continue increasing, is obtained from pantograph-catenary systems. Pantographs, one of the basic components of the pantograph-catenary system, provides electrical power to the locomotive by contacting the catenary while the train is The associate editor coordinating the review of this manuscript and approving it for publication was Tallha Akram . running [4]. Pantographs draw electrical power from the catenary using plates called current collector strips, which are made of wear-resistant materials such as carbon or graphite [5]. Their surface must be sufficiently smooth so that the catenary can move steadily from side to side on the pantograph. These plates are in constant contact with the catenary wire, causing wear on the collector strips over time. On the surface of the collector strips, different types of wear occur, such as groove-shaped wear, partial wear, excessive wear and cracks, etc. When this wear is horizontal or vertical, it causes arcs by increasing the friction between the pantograph and the catenary while the train is in motion. The wear form, called a groove or notch, which may occur on the sides of the plates, can hold the catenary wire laterally over time, pulling the catenary wire and causing deformation or even breakage in the catenary wire. On the other hand, symmetrical wear on the plates, impairs the transmission of electricity over time, as it decreases the plate's vertical length. For this reason, it is very important to detect these abrasions before they reach certain levels and to carry out the necessary maintenance. Predictive maintenance based on the system 's condition can thus reduce maintenance costs and increase the lifetime of the pantograph's plates. To ensure stable and safe operation VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ of the pantograph catenary system, the contact force between the pantograph and the catenary must be kept to a certain extent. Small contact force can lead to arcs caused by wear on the pantograph's collector strips due to the unexpectedly high temperature from the stationary power supply, while massive contact force increases the wear on the strips due to friction [6]. Some studies in the literature to detect faults in the pantograph and catenary are described in this section. Luo et al. [7] proposed a method to detect abnormal behavior in the pantograph-catenary system. In their study, a modified region-based convolutional neural network (RCNN) was proposed to detect pantograph faults. In their study, they performed an arc detection after determining the head area of the pantograph. Yu and Su [8] proposed a method based on the photoelectric transformation mechanism for arc detection in pantographs. In addition, Qu et al. [9] proposed a method based on the Adadelta deep neural network and used a genetic optimization method to predict the state of pantograph and catenary. Liu et al. [10] proposed a deeply separable convolution-based method for detecting faults of the droppers in the pantograph. They divided the failures that occurred in the dropper into three classes. Similarly, Yi et al. [11] used ultrasonic sensors to detect cracks and wear on the surface of the pantograph collector strips. In their proposed method, wear status can be estimated to some extent. It is also sensitive to the electrical environment and temperature. Situation monitoring applications have turned to the more recently developed technologies, such as high resolution cameras and image processing technology. Landi et al. [12] analyzed the anomalies in the pantograph-catenary system by interpreting the hot spots on the catenary line and pantograph with thermographic images obtained with infrared cameras. Zhu et al. [13] performed a study to detect wear on pantograph strips using edge extraction algorithms. However, the method developed in their study needs to be improved. Furthermore, Karakose et al. [14] proposed an approach based on image processing to monitor whether a pantograph catenary system contact point is safe. They focused on the contact point of the pantograph and catenary, and therefore did not do a study to detect current collection strips of the pantograph. Furthermore, Hamey et al. [15] developed an online monitoring system for detecting wear on the pantograph. A method for wear detection in collector strips was proposed by Ostlund et al. [16]. In their work, they used voltage and current to detect wear and arc in current collector strips but did not work on the image. Image detection is a more preferred method since it is a non-contact method. Vazquez et al. [17] provided a sensor system for measuring changes in the height of the contact wire as a pantograph catenary passes through the line. The purpose of this sensor system was to prevent faults due to the interaction between the changing height and the pantograph-catenary. With an infrared camera, which prevented the resulting image from being affected by light and vibration, they monitored the contact point of the catenary with the pantograph. Barmada et al. [18] proposed the use of arc detection in pantograph-catenary systems with a support vector-based classification. In their study, they used current and voltage data instead of images. Tang et al. [19] proposed a probabilistic Bayesian approach to observe pantograph wear, combining the image, scale, and location of the pantograph and determine the visual anomaly. They detected the pantograph headline in their work. Our study focused on wear in the current collection strips of the pantograph. The regulation of the contact between the pantograph and catenary was investigated by Mokrani et al. [20]. Sacchi et al. [21], Capece et al. [22] developed an automatic imaging system to detect locomotive pantograph failures. In addition, Wei et al. [23] developed a high-speed photography and arc detection lab simulation system for pantographs. They performed their work in a laboratory simulation environment and made an arc detection process; however, they did not address the wear in the current collection strips. The newest method of measuring radiation due to pantograph contact with the catenary on high-speed trains was developed by Ma et al. [24]. Song et al. [25] carried out a study that took into account the contact surface and the dynamic performance of pantograph catenary systems. Koyama et al. [26] proposed a method in which they placed sensors on the contact line to detect wear in the pantograph and developed a sensor system with this method. In their work, they obtained data by placing a sensor system on the pantograph head. In the method they proposed, the system is in contact with the pantograph, but there is the possibility that contact methods can damage the pantograph. In our study, we suggest a completely non-contact method. Massat et al. [27] simulated the interaction between the pantograph and catenary. Using these simulation results, they offered methods for detecting errors in this system. Judek et al. [28] proposed a wavelet transform based approach to identify wear in the pantographs' carbon collector strips. In addition, they performed a simulation and tested their proposed method on the simulation. Zhou et al. [29] proposed a pantograph catenary monitoring system based on the state-based recognition of the pantograph. Their system included data collection modules, positioning modules and data analysis and recognition modules consisting of sensors. Kang et al. [30] proposed a three-stage wear detection system for contact wire support. First, a faster R-CNN network was adopted to limit key catenary components, and image areas containing contact wire support components were obtained. The contact wire supported components were then segmented with a conventional catenary component, and partitioning network, using Bayesian segmentation, and combining different level features of the backbone network. As a result, they determined wear in accordance with the criteria defined by the geometry of the components. Siyang et al. [31] proposed an improved rapid R-CNN architecture to detect deformation of the pantograph's contact plates. They used 257 pantograph images and divided them into four classes. Wei [4] divided the failures in the pantograph collector strips into four classes and detected them with a deep learning network. They also used image processing techniques. However, the number of images they used while training the deep learning network was 158, and the number of test images was 80, which are not sufficient for deep learning. More images should be used to increase accuracy and precision values. In this study, a special Convolutional Neural Network (CNN) architecture was developed to increase the detection success of the wear in the pantograph's current collecting strips. The images used in this study were obtained from the electric railway line at The Republic of Turkey State Railways. The success of the model was tested using these images and was compared with the results of RESNET50 and VGG16 deep learning models frequently used in the literature.
The contributions of this study are as follows: • Although the pantograph slide plate was first diagnosed and described by Wei et al. in 2019, it is known that the pantograph has a fragmented structure, so it is possible to determine in which part of the pantograph (right front, left front, right rear, left rear) the failure occurred, thereby allowing only the defective part to be replaced. This is much more economical than previous process, and it is extremely important in terms of perceiving many other systemic problems (pantograph curvature, one-sided high pressure problem). One of the greatest contributions of this study from a methodological perspective is the fault diagnosis and determination of the class of the fault as well as its location.
• The second contribution of this paper is to propose a novel approach for fault detection, recognition and region detection and to present a CNN network suitable for this subject. The pantograph data used in the training and testing phase of the proposed method is considerably higher than its counterparts in the literature. This increases system reliability.
• Third, four basic criteria are introduced for assessing the wear condition. In particular, the problem of wear edge prediction is investigated using image processing techniques such as image enhancement and Hough transform applied to wear depth estimation and other wear condition assessments. Effectiveness of the proposed method is demonstrated using real fault images and video taken from the Republic of Turkey State Railways.
• Fourth, method in the literature used 158 images for training, 80 images for the test, and recognized five classes. When the total number of images was divided by the number of classes, an average of 47 images were evaluated for each class. In our study, a total of 909 images and images belonging to four classes were recognized. This means that an average of 228 images belonging to each class were evaluated. Although the success rate seems high in the study in the literature, it is not an effective result because the amount of data is low and the number of classes is high. As the number of images increases, the success rate of the architecture is likely to decrease. One of the basic logics of deep learning is to optimize weights by working with lots of data. Optimizing the weight is to be discussed in the literature. In our study, weights are optimized because the amount of data is higher and the number of classes is less. The remainder of the paper is organized as follows: In Section II gives a detailed introduction of the image processing techniques and deep neural networks used in this paper. Section III presents the image processing and classification results and discussions. In Chapter 4, conclusions and future works are mentioned.

II. THE PROPOSED MODEL
In this study, a method is proposed for detecting wear in the current collector strips of the pantograph. A CNN architecture is proposed using pantograph images obtained from railways. The images obtained were first processed, and then a data set suitable for deep learning was created. The data set created was tested in classical CNN models and compared with the proposed CNN architecture. The flow chart of the studies carried out in this study is shown in Figure 1.

A. IMAGE ACQUISITION
To obtain images of the pantograph's current collector strips, image data was collected from electric trains operating in the Republic of Turkey State Railways. The images were recorded using an experimental setup consisting of a camera VOLUME 8, 2020 and a light source. The schematic drawing of the image collection device is shown in Figure 2. Examples of the images obtained are presented in Figure 3.

B. LABELLING IMAGES
The degree of wear occurring in the pantographs were measured by dismantling the pantograph in the maintenance unit of the Republic of Turkey State Railways. Some criteria for evaluating the surface condition of the pantograph were discussed in this unit. If we express the first wear depth in the collector strips of the pantograph as A d , then the maximum wear depth can be expressed as in Equation 1 [4].
where, A(x) represents the amount of wear at position x. L is the length of the collector strip. The second criterion is surface roughness, denoted by R. This can be calculated by Equation 2 [4].
where A m refers to the average wear on the collector strip. From the pixels obtained by image processing methods, the thickness of the collector strips can be found. If the maximum thickness of the strip is considered to be 40 mm, then the thickness can be calculated from the image with Equation 3 [4].
where d represents the value of the estimated thickness of the pantograph in the image in pixels. D is the value of the maximum thickness in pixels. The degree of wear in the collector strips on the railways is measured by removing the collector strips with various equipment working with the above principle. In accordance with these measurements, the images taken from the pantograph are classified and labeled. The 909 images obtained from the Republic of Turkey State Railways were separated and labeled by wear class with the assistance of the Republic of Turkey State Railways personnel serving in the maintenance unit. As a result, four wear classes were created: solid, low wear, excess wear and groove-shaped wear. Figure 4 shows some examples of wear patterns. Tagged image classes are named as shown in Table 1.

C. PRE-PROCESSING AND RESIZING OF THE IMAGES
Lines in the images were detected by applying the Canny Edge Detection Algorithm and Hough Transform. By using the start and end points of these lines and the region between two lines, the study can focus only on the areas where current collecting strips are located in the pantograph head. The obtained image is divided into four regions by Hough lines to evaluate each strip separately and to detect partial wear. The graphical representation of the transactions performed is shown in Figure 5. The pseudo code of the Hough transform is given in Table 2. A Power Law Transformation was applied to the images to improve them. The Power Law Transformation is a transformation used in general purpose contrast enhancement applications [32]. The general formula of the power law transformation is as in Equation 4 [32]: where O represents the output image, 'I' represents the input image, and c and are positive constants. According to the  different values of , the change of O according to I is as in Figure 6. The improved images are resized so that they are all the same size. All images are set to 150 × 430 size.

D. PROPOSED CONVOLUTIONAL NEURAL NETWORK (CNN) ARCHITECTURE
CNNs are one of the most frequently used image classification models of neural networks. In the CNN the corresponding filters can capture the spatial and time dependencies of an image [35]. CNN reduces the image properties to an easier-to-edit system without dropping features essential to a good classification. CNN's architecture comprises of a layer array that uses a differential function for each layer to convert one layer to another. There are normally three layers for creating a CNN model: the convolution layer, the pooling layer, and the fully connected layer. The pooling layers have two types of pooling. The first type is maximum pooling, which returns the maximum value from the kernel part of the image and eliminates the noisy activation and reduces both noise and dimension. The second type is an average pooling that reduces the size of the matrices and uses this reduction as a noise control mechanism. Consequently, maximum pooling is better than average pooling. In general, CNN uses the input 3rd row tensor which takes n rows, m columns and three color channels in an image matrix (R, G and B) into account, and the image 's spatial structural structure. The input passes through the convolution layer, the pooling layer, and the fully connected layer, where each layer 's output is used as the input for the layer that follows. The network begins with input neurons of 3 * m * n used for encoding pixel densities for input properties for a 2-D image. This is followed by a convolution layer in the f X f local receptor area. The result is a layer of 3 × (m-f + 1) × (n-f + 1) hidden feature neurons when using a 1 epoch step. The pooling layer will be applied to 2×2 regions on each of the 3 feature maps to obtain 3 × (m-3 × (m-r + 1) / 2 × (n-r + 1) / 2 hidden feature neurons. The feature map is generated usually with the process of convolution by multiplying the filter or kernel elements by the input matrix element. The result is then collected to obtain a pixel from the feature map. The filter is sliding through the input matrix and creates other features. The convolution process can be written mathematically, as shown in Equation 5 below [35].
VOLUME 8, 2020  where, i is from 1 to m-f + 1, and j is from 1 to n-f + 1. The pseudo code of the operations performed in the convolution layer is given in Table 3. Activation functions are applied to the data after the convolution process. The selection of the activation function affects the success rate. The most commonly used activation functions are ReLu and Sigmoid functions. The main logic of the ReLu activation function is that it generates zero output value for negative values that come as input, and for positive values it returns the input value exactly. ReLu was preferred as the activation function in the CNN architecture developed in this paper. The pseudo code of the process performed in the fully connected layer is given in Table 4. Where w is the neuron weights, and X represents the inputs of the layer. A CNN architecture was created using Python Jupiter. Labelled images were used after this architecture was created. For this model, 771 training data and 138 test data were created. After training, the success of the model was determined by performing the test process with the test data that was not entered during training. This paper focuses on the current collector strips in the pantograph images to identify and classify the wear in these strips and aims to eliminate the possibility of damaging the possible arcs and the pantograph and catenary system by providing partial wear.   In this paper, a deep CNN network is created to detect wear on the pantograph collector strips. The block diagram of the created CNN architecture is shown in Figure 7. The parameters of the developed CNN architecture are as in Table 5.
As can be seen from Table 5, the proposed CNN has 10 layers in total, including 1 input layer, 2 convolutional layers, 2 max pooling layers, 2 dropout layers, 1 fully connected layer, 1 Dense layer, and 1 softmax calculation layer. There are only 5 layers (2 convolutional layers and 1 fully connected layers) whose weight parameters need to be trained. In comparison with ResNet, VGG16Net has better feature extraction ability despite its simple structure. With this architecture, a feature map was created from the training data. In the developed architecture, 2 multiple convolution layers, max pooling and dropout layers followed, as shown Figure 7. In the convolution layers, the input values were multiplied by the weight values of the neurons in the layers, and the output values were produced.
In addition, filters were used for learning in these layers. Initially, filters were produced randomly, and the optimum VOLUME 8, 2020 filters were attempted to be identified during the training. Feature maps of the relevant regions of the image were subject to the convolution process using these filters. In this way, activation maps or feature maps of the images were obtained. A max pooling process was performed in the pooling layers. The flattened layer contained layers with 1×253376, 19765× 128, and 128 × 4 neurons, respectively.

E. VGG16 TRANSFER MODEL
VGG16 is an architecture consisting of 16 convolution layers, 5 pooling layers, and 3 fully connected layers. The convolution process was carried out with 3 × 3 kernels in the convolution layers of this architecture. Although it has very successful performance values, having this many parameters makes it difficult to use. The structure of the VGG16 architecture is shown in Figure 8. Transfer learning was performed using the labeled pantograph images and the VGG16 model.

F. RESNET50 TRANSFER MODEL
ResNet [33], which trains very deep networks with its permanent connections, has a deeper, bigger, and slower structure than GoogleNet. With this model, which is 50 layers deep, higher performance is achieved in the ILSVRC data set compared to GoogleNet. The architecture of the ResNet50 model [33] is as shown in Figure 9. The dataset created was trained by giving entries to the last layer using the ResNET50 Transfer Model and tested with test data.

A. EXPERIMENTAL SETUP
In this study, a data set was created by taking images of the current collection strips of pantographs from the railways. A camera and a light source were used to take the images, which were transferred to the computer and prepared and labeled for use in the deep learning network by preprocessing. The tools and technologies used in this study are given in Table 6.
The algorithms mentioned in this paper were developed using the Python Jupiter software platform.

B. EXPERIMENTAL RESULTS
Lines were obtained by applying Canny Edge Detection and Hough Transform to the current collector strips images obtained from pantographs in the Turkey Republic State Railways. Using these lines, each pantograph head image was divided into four separate images. By finding 50 Hough lines, some of these lines were eliminated and four meaningful lines were obtained. The recorded images are horizontal pantograph images. For this reason, the angle value range was determined to obtain horizontal lines. The Hough chart is shown in Figure 10.
The data set with Hough lines was created with a 6% error rate. A clean data set was created by selecting the wrong images and eliminating them from the data set. Of the 50 lines obtained after applying the Hough Transform,  the unnecessary lines were eliminated by adjusting the angle value range. As a result of the elimination process, four lines were obtained. An example of the display of these four lines in the original image is shown in Figure 11.
A Power Law Transformation was applied to improve the obtained current collection strip images. The preprocessed images were resized. All images were set to a 150 × 430 size. A data set consisting of 909 images was formed with images obtained the Republic of Turkey State Railways. In this data set, 771 image data were used for training, and 138 other image data were used as test data. The 909 images were labeled by dividing them into four wear classes, with support from maintenance personnel the Republic of Turkey State   Railways. These classes were determined as solid, low wear, excess wear, and groove-shaped wear. A sample image of each wear class is given in Figure 12.
The distribution of the images used in the network education by class as test and training data are given in Table 7 and Figure 13.
The developed CNN network architecture, VGG16, and Resnet50 were trained with 771 images of four classes labeled and tested with 138 images. The network training was carried out with 500 epochs and 64 batch size. The developed CNN architecture and ResNet50 and VGG16 transfer models were tested by using the untrained images in the data set. The confusion matrices created with the results obtained with the test made from untrained data are shown in Figure 14. Table 8 shows the classification results obtained from Precision, Recall, and Fscore values for each class of the three deep learning models used in the study. As can be seen from the confusion matrices, the suggested achitecture gave error-free results in three architectures in Class_0 images, namely ''robust'' images. The suggested CNN architecture and ResNet50's prediction ratio was higher in images belonging to Class_1, in other words, called ''low wear''. The accuracy of the predicted rate of the proposed CNN architecture was higher in Class_2 images, that is, ''excess wear'' images, and in Class_3 images, that is, ''groove-shaped wear''. A performance analysis was done by calculating the class based Precision, Recall, and Fscore values of each model by using the confusion matrices of the models. Considering these results, the Precision, Recall, and Fscore values were obtained as 1 with three models for Class0.This means that they are very successful in recognizing Class_0 in 3 models. The proposed CNN and ResNet50 models for Class1's Precision value resulted in a value of 0.697, and the Precision value of these two models is higher than the VGG16 model. For Class 1 Recall values, the proposed CNN model gave the highest value with 0.7667. Likewise, according to Class_1, the Fscore value was obtained as 0.7302 with the proposed CNN architecture. This value was higher than the results of other models, showing that the proposed CNN model in the recognition of Class_1 is more successful than the other two models.
When the Precision values of Class_2 were examined, a higher value of 0.5 was obtained with the proposed CNN model compared to other models. When the Recall value of Class_2 is taken into consideration, a value of 0.4516 was obtained with the proposed CNN model. This value is higher than the Recall values obtained with the other two models.   Similarly, with the proposed CNN model, the Fscore values of Class_2 had a higher value than other models obtained with 0.4746.
These results show that the proposed CNN architecture was better than the other two models for recognizing Class_2 images. Finally, when the Precision, Recall, and Fscore values of Class_3 are evaluated, the values are 0.64 with the proposed CNN architecture. This value was higher than the values obtained from other models. As can be seen from these results, the average performance accuracy of the proposed CNN architecture was higher than the other two deep learning architectures as seen in Table 9.
As a result, in this paper, a study was carried out on the detection of wears in the current collection strips of pantographs powered by electric trains on railways. In this study, a CNN model is proposed for wear detection. A dataset was created with images from pantographs used on railways. The proposed CNN architecture and ResNet50 and VGG16 architectures have been trained and tested with this dataset. Test results of these three deep learning models were obtained. Confusion matrices have been created containing these results. Precision, Recall, Fscore values of the confusion matrix were obtained for each model based on class. The obtained results were compared. In comparison results, the proposed CNN model was found to be more successful in recognizing the dataset classes used compared to the other two models.

C. DISCUSSIONS
In this study, a CNN architecture was proposed to detect and classify the wear that occurs in the current collector strips of the pantographs. In addition, the data set created was also  used in ResNET50 and VGG16 architectures. The success rate of the proposed architecture was compared with the success rate of these two classical models, and the proposed architecture was found to be 70.14% more successful than these two classical deep learning architectures. Studies that detect faults in the pantograph and catenary in the literature were examined. In Table 10, these studies and the method we proposed are compared in terms of detection type, used signals, and detection method. As can be seen from Table 10, some studies aimed at detecting faults in the catenary (such as [19], [30]). Some other studies, unlike ours, are studies aimed at detecting arcs that occur when contacting the pantograph and catenary (such as [18], [23]). Another study in the literature used voltage and current signals (such as [16]) to detect wear in the current collector strips of pantographs. In this study, the mechanisms contacting the pantograph were used to obtain current and voltage data from the pantograph's current collector strips. This contact can, in time, damage the pantograph.
In the method we propose, it is not possible to damage the pantograph, as the wear on the current collector strips of the pantographs are detected without contact using images.
The closest method to the method we propose is the study mentioned in the reference number [31] in the literature. The comparison of this study and the method we propose is given in the Table 11. Although the accuracy rate of the study in the literature appears to be higher in Table 11, our study is more successful considering the amount of data used in training and testing and the number of detected classes. The study in the literature used 158 images for training, 80 images for testing, and recognized 5 classes. When the total number of images was divided by the number of classes, an average of 47 images were evaluated for each class. In our study, 771 pantograph images were used for training and 138 pantograph images for testing, and 4 classes were detected. Therefore, a total of 909 images were used. When the total number of images is divided by the number of classes, it is seen that an average of 228 images belonging to each class were evaluated. Therefore, success rate alone is not a sufficient criterion.

IV. CONCLUSION AND FUTURE WORK
Wear and failures in pantograph current collector strips can cause costly pantograph and catenary accidents. Therefore, it is very important to detect wear in current collector strips at an early stage. This study was carried out to determine the wear on the current collection strips of the pantograph using images of pantographs containing wear obtained from the railways and belonging to four wear classes. These images were enhanced using Power Law Transform to create a clean dataset. The pantograph images were then divided into four parts using lines obtained with Hough Transform to determine the locations of the wear. In this way, the wear class for each part of the current collector strips was determined. Training and testing phases were carried out using a CNN architecture that was developed with a dataset containing 909 images, 771 of which were used in training and 138 in the test. The accuracy rate of the proposed architecture was found to be 70.14%. The same dataset was tested on ResNET50 and VGG16 architectures. The proposed architecture was compared with these two classical architectures and was found to be more successful. The proposed study was compared with the studies in the literature in detail, and it was revealed that the proposed architecture is more effective and improvable.
This study can be improved for use in autonomous systems that can detect faults in pantograph-catenary systems. By obtaining more data, weights of the CNN architecture can be optimized, thus increasing the success rate. By detecting the worn area of the pantograph and increasing the speed, it will be possible to detect in real time. In this way, periodic maintenance on railways will not be required. With the developing technology, autonomous systems desired in every field can be used for detecting pantograph faults.