Pantograph-catenary arc detection method based on machine learning

Abstract

The invention provides a pantograph-catenary arc detection and identification method based on machine learning. According to the method, a pantograph catenary arc with overall offset of a pantograph head area is extracted by adopting a method based on pantograph head area guidance, so that the problem of large background interference in arc detection is solved; a new machine learning model is established; a heuristic V3 network is adopted as a backbone network to extract arc features more quickly and more accurately; a region-of-interest alignment algorithm is adopted in the region generation network, and feature mapping operation of candidate bounding boxes is carried out, so that the problem of region mismatching in a region-of-interest pooling algorithm is solved, and the accuracy of target positioning is improved; a soft non-maximum suppression algorithm is adopted to solve the problems of missing detection and false detection of a non-maximum suppression algorithm, so that a candidate frame is more accurately selected. According to the invention, the arc phenomenon can be rapidly and accurately detected.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer digital image processing and pattern recognition, and in particular relates to a bow-catenary arc detection method based on machine learning. Background Art

[0002] In electrified railways, the pantograph-contact wire system is a key component for trains to obtain electrical energy. The train obtains electrical energy through the contact between the pantograph on the top and the contact wire, thereby ensuring the smooth operation of the train. During the operation of the train, due to various factors, the pantograph and the contact wire are prone to separation, resulting in pantograph arcing. The arc phenomenon will affect the current collection quality of the pantograph and the contact wire, increase the wear of the pantograph slide and the contact wire, damage the electrical equipment inside the train, interfere with the signal equipment along the line, and pose a huge threat to the safety of train operations.

[0003] At present, the common method of detecting arcs is to record and monitor the operation of the pantograph-catenary system through cameras, and to analyze the monitoring video by reviewing it after the arc occurs. The key to ensuring the smooth operation of trains is how to quickly extract accurate and effective arc information from the monitoring video. The traditional model has the following three problems when it comes to arc detection:

[0004] 1. Complex video background greatly interferes with arc recognition;

[0005] 2. Low accuracy and high computational complexity;

[0006] 3. The false detection rate or missed detection rate is high. Summary of the invention

[0007] The present invention aims to overcome the above-mentioned defects and proposes a method for detecting and identifying arcs in a bow-net based on machine learning. That is, a method based on the guidance of the pantograph bow head area is adopted to extract the bow-net arc with the overall offset of the bow head area to solve the problem of large background interference in arc detection; a new machine learning model is established: the heuristic V3 network is used as the backbone network to extract arc features faster and more accurately; the region of interest alignment algorithm is used in the region generation network to perform feature mapping operations on the candidate bounding box to solve the problem of region mismatch in the region of interest pooling algorithm, thereby improving the accuracy of target positioning; the soft non-maximum suppression algorithm is used to solve the problems of missed detection and false detection of the non-maximum suppression algorithm, thereby more accurately selecting the candidate border. It can realize fast and accurate detection of arc phenomena.

[0008] The present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is characterized by comprising the following steps:

[0009] S1. Extract the pantograph head area and obtain the pantograph-catenary arc area guided by the pantograph head area;

[0010] S2. Use the region of interest alignment algorithm to perform feature mapping of candidate bounding boxes;

[0011] S3. Progressively reduce the score of overlapping detection boxes based on the soft non-maximum suppression algorithm;

[0012] S4. The heuristic V3 network is used as the backbone network and combined with the region generation network to extract arc features.

[0013] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0014] In S1, the following steps are included:

[0015] S1.2, apply rotation and scaling to the original captured images to enlarge the dataset;

[0016] S1.3. From the perspective of image feature extraction, a method for extracting the arc area of ​​the pantograph-catenary based on the guidance of the pantograph head area is designed;

[0017] S1.4, all processed images constitute the experimental dataset;

[0018] Or follow these steps:

[0019] S1.1, apply rotation and scaling to the original captured images to enlarge the dataset;

[0020] S1.2, after intercepting the arc video of the bow-net into continuous frame images, grayscale processing is performed to convert the images into grayscale images;

[0021] S1.3. From the perspective of image feature extraction, a method for extracting the arc area of ​​the pantograph-catenary based on the guidance of the pantograph bow head area is designed; S1.4. All processed images constitute the experimental data set.

[0022] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0023] In S1.3, the method of coordinate transformation of the bow head positioning frame is used to guide the bow head area, and the area where the arc phenomenon may occur is accurately obtained.

[0024] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0025] In S1.3, the specific steps are as follows:

[0026] S1.3a: Obtain the four values ​​x of the bow head positioning frame through the coordinates of the upper left corner and the lower right corner of the positioning frame 左 、x右 ,y 上 and 下 , the coordinate transformation is performed based on these four values;

[0027] S1.3b: The width w and height h of the positioning box are directly obtained through the coordinates of the two corner points of the positioning box:

[0028] w=x 右 -x 左

[0029] h=y 上 -y 下

[0030] S1.3c: Divide the positioning frame into M equal parts in height direction and N equal parts in width direction, and take one of them as the height direction offset factor h' and width direction offset factor w' of the positioning frame. The offset factor is used as the basic amount to better control the precision of the offset:

[0031] h'=h / M

[0032] w'=w / N

[0033] S1.3d: Generate 4 random numbers n1, n2, n3 and n4 within a certain data range (0-1). The offset of each side is determined by the offset factor and the random number. The upward offset of the top edge of the positioning box is Δy. 上 , the upward offset of the lower side is Δy 下 , the left side is offset to the right by Δx 左 , the right side is offset to the left by Δx 右 , their values ​​are determined by the following formulas respectively;

[0034] Δy 上 =n1×h'

[0035] Δy 下 =n2×h'

[0036] Δx 左 =n3×w'

[0037] Δx 右 =n4×w'.

[0038] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0039] In S2, the following specific steps are included:

[0040] S2.1: traverse each candidate region without quantizing the floating point coordinates after mapping;

[0041] S2.2: Divide the candidate region into k×k units without quantizing the boundaries of the divided units;

[0042] S2.3: Divide each unit into 4 small units, take the center position as the sampling point, and determine 4 sampling points in each unit;

[0043] The present invention selects 4 small units because when the training samples (data sets) are not complicated, selecting 4 small units can reduce the model training time while ensuring a certain accuracy and improve the speed of bow-catenary arc detection.

[0044] S2.4: Calculate the image value at the sampling point coordinate pixel using bilinear interpolation method;

[0045] S2.5: Perform maximum pooling on the calculated four sampling points to obtain a fixed-dimensional region of interest output.

[0046] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0047] In S2.4, the bilinear interpolation method comprises the following steps:

[0048] Assume that function f is in Q 11 =(x1,y1),Q 12 =(x1,y2),Q 21 =(x2,y1) and Q 22 =(x2,y2) The values ​​of these four points are known. Find the value of function f at P = (x,y);

[0049] S2.4.1. Perform linear interpolation in the x direction and obtain:

[0050]

[0051]

[0052] S2.4.2. Perform linear interpolation in the y direction and obtain:

[0053]

[0054] S2.4.3. Get the required f(x,y)

[0055]

[0056] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0057] In S3, the following steps are involved:

[0058] S3.1: All detection boxes are sorted according to their scores, and the detection box with the highest score is selected as the reference box. For each remaining detection box, the score is updated according to the intersection-and-union ratio between the detection box and the reference box to determine the optimal target.

[0059] Regarding the above scores: In the machine learning model proposed in the present invention, image recognition will be performed on the training samples. During the image recognition process, the network will propose candidate boxes (rectangular areas). The differences between different candidate boxes and the actual targets are different. The more the match is consistent with the actual target contour, the higher the value detected by the model (i.e., the score).

[0060] The intersection-to-union ratio is defined as the ratio of the intersection area of ​​the predicted box P and the true annotation box GT to the union area, as shown in the following formula:

[0061]

[0062] S3.2: When the current candidate box and the candidate bounding box with the highest score overlap, if their intersection-over-union ratio exceeds a given threshold, the soft non-maximum suppression algorithm uses a linear weighted method to update the confidence of the candidate box to S i (1-IoU), as shown below:

[0063]

[0064] In the formula, S i is the confidence level, N t is the preset threshold, B is the candidate bounding box with the highest score at present, and b i is the candidate box to be processed;

[0065] S3.3: Select the next detection box with the highest score as the reference box, and repeat the score update process in the above steps until all detection boxes are processed;

[0066] S3.4: Re-rank the detection boxes based on the updated scores and select the box based on the final score.

[0067] Furthermore, the present invention provides a method for detecting a pantograph-catenary arc based on machine learning, which is also characterized by:

[0068] In S4, the model training method is as follows:

[0069] S4.2.1: Take the training set image as input, use the heuristic V3 network as the backbone network of the model, extract the input image features to obtain the feature map, and apply the feature map to the subsequent network;

[0070] S4.2.2: The network uses these feature maps to generate candidate regions, and obtains feature maps of a specified number of feature candidate regions by generating anchor boxes of different sizes;

[0071] S4.2.3: Unify them into the same size and input them all into the fully connected layer for target recognition and positioning, and use the normalized exponential function to classify specific categories, and complete the regression operation to obtain the precise location of the object.

[0072] The method for detecting arc between a pantograph and a catenary based on machine learning as claimed in claim 8, characterized in that:

[0073] The training method of the heuristic V3 network is as follows: the feature extraction module pre-trains the heuristic V3 network on the training set;

[0074] S4.1.1: The input image is passed through three convolutional layers with a kernel size of 3 pixels × 3 pixels, a batch normalization layer, a nonlinear activation layer, and a 3 pixels × 3 pixels maximum pooling layer;

[0075] S4.1.2: After two convolutional layers with a kernel size of 3 pixels × 3 pixels, a batch normalization layer, a nonlinear activation layer, and a 3 pixels × 3 pixels maximum pooling layer;

[0076] S4.1.3: Feature extraction through three heuristic modules;

[0077] S4.1.4: After a 8-pixel × 8-pixel maximum pooling layer;

[0078] S4.1.5: Pass through a dropout layer and a 3×3 pixel convolution layer, a batch normalization layer, and a nonlinear activation layer.

[0079] In addition, a bow-catenary arc detection method based on machine learning of the present invention is applied to the detection and identification of bow-catenary arc phenomena. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 The overall flow chart of the present invention;

[0081] Figure 2 Structural diagram of the overall model;

[0082] Figure 3 Test result graph;

[0083] Figure 4 Comparison diagram of the system of the present invention and the traditional system. DETAILED DESCRIPTION

[0084] The method of the present invention can be used for bow-catcher arc detection, and can perform accurate and rapid arc detection on various bow-catcher monitoring video images.

[0085] like Figure 1 As shown, the present invention is realized by the following technical solutions: a method based on pantograph bow head area guidance is adopted to extract the bow-net arc with overall offset in the bow head area, so as to solve the problem of large background interference in arc detection; a new machine learning model is established: a heuristic V3 network is adopted as the backbone network to extract arc features faster and more accurately; a region of interest alignment algorithm is adopted in the region generation network to perform feature mapping operations on candidate bounding boxes, so as to solve the problem of region mismatch in the region of interest pooling algorithm, thereby improving the accuracy of target positioning; a soft non-maximum suppression algorithm is adopted to solve the problems of missed detection and false detection of the non-maximum suppression algorithm, thereby more accurately selecting candidate borders.

[0086] Specifically, it includes the following technical solutions:

[0087] The first part is the algorithm for extracting the bow head area. The camera installed on the top of the train collects the arc video of the bow and catenary at the same angle and position. The video includes different operating environments such as day, night, and tunnels. The steps are as follows:

[0088] Step 1: Cut the arc video of the bow-net into continuous frame images and then perform grayscale processing to convert the image into a grayscale image.

[0089] Step 2: Apply rotation and scaling to the original captured images to enlarge the dataset.

[0090] Step 3: From the perspective of image feature extraction, a method for extracting the arc area of ​​the pantograph network based on the guidance of the pantograph head area is designed. The pantograph head positioning frame is offset simultaneously from the top, bottom, left and right directions to accurately obtain the area where the arc phenomenon may occur.

[0091] Step 4: All processed images constitute the experimental dataset.

[0092] The second part is the region of interest alignment algorithm:

[0093] The region of interest alignment algorithm is selected instead of the original region of interest pooling to unify the feature map size to solve the problem of region mismatch caused by region of interest pooling, thereby improving the accuracy of target positioning.

[0094] The candidate bounding box coordinates generated by the region generation network correspond to the original image space. According to the mapping relationship between the feature map and the original image, the candidate bounding box can be mapped to the feature map. The data type of the candidate bounding box coordinate value is usually a floating point number. During the pooling process, the region of interest pooling needs to be quantized into an integer type first, and then evenly divided into several units. After two quantization operations, there will be a certain deviation between the candidate bounding box and its position in the original image, that is, there will be a problem of regional mismatch, which affects the accuracy of target positioning. The present invention selects the region of interest alignment algorithm to perform feature mapping operations on the candidate bounding box.

[0095] The third part is the soft non-maximum suppression algorithm:

[0096] At the same target location, the region candidate network will generate a large number of candidate bounding boxes, which overlap with each other. The non-maximum suppression algorithm suppresses the maximum by searching for the local maximum. The non-maximum suppression algorithm deletes all bounding boxes greater than a given threshold. When the target is occluded, it is easy to cause false detection and missed detection. The present invention adopts an improved version of the non-maximum suppression algorithm, the soft non-maximum suppression algorithm.

[0097] The fourth part is the design based on machine learning model.

[0098] The heuristic V3 network is used as the backbone network to extract arc features faster and more accurately.

[0099] Traditional networks have problems such as long training time, difficulty in adjusting parameters, large storage capacity required, and gradient vanishing and overfitting during training as the network width and depth increase. The heuristic V3 network can reduce the dimension of larger matrices while aggregating visual information at different sizes, making it convenient to extract features at different scales and speeding up network training time. Therefore, the present invention chooses to use the heuristic V3 network as the backbone network, combined with the region generation network, so that the machine learning-based model proposed in the present invention can be better optimized.

[0100] Here are the steps:

[0101] Step 1: Take the image of the experimental dataset as input and use the heuristic V3 network as the backbone network of the model, which acts as a feature extractor. Extract features from the input image to obtain a feature map, which is applied to the subsequent region generation network and fully connected layer.

[0102] Step 2: The region generation network uses these feature maps to generate candidate regions, that is, by generating anchor boxes of different sizes to obtain feature maps of a specified number of feature candidate regions.

[0103] Step 3: Unify them into the same size and input them all into the fully connected layer for target recognition and positioning, and use the normalized exponential function to classify specific categories, and complete the regression operation to obtain the precise location of the object.

[0104] Finally, the model structure of the present invention is as follows Figure 2 As shown, the final output of the entire model includes the width and height of the prediction box, the coordinates of the center point, the confidence level, and the probability of each category.

[0105] The fifth part is model training

[0106] After the network is constructed, the machine learning model proposed in the present invention is trained using the experimental data set in step 1.

[0107] The sixth part is the arc detection of the bow-net

[0108] Here are the steps:

[0109] Step 1: Selection of performance indicators

[0110] In order to verify the detection effect of pantograph arc, the mean average precision is used to evaluate the trained model.

[0111] Step 2: Model testing

[0112] The experimental results show that the pantograph arc can be accurately detected and located through this technical solution.

[0113] The present invention gives full play to the advantages of machine learning models, maintains high precision and good robustness, and at the same time improves the accuracy through a bow-catenary arc extraction method based on the overall offset of the pantograph bow head area, so that the bow-catenary arc phenomenon can be detected and identified quickly and accurately.

[0114] The following is a description of a specific optional embodiment.

[0115] The first part is the algorithm for extracting the bow area. The camera installed on the top of the train collects videos at the same angle and position. The video includes different operating environments such as day, night, and tunnels. The steps are as follows:

[0116] Step 1: Intercept the arc video of the bow net into continuous frame images, grayscale the obtained images, and convert the images into grayscale images. The mean of the RGB value of each pixel is used as the grayscale value of the pixel, so as to obtain the grayscale image set G = {g1, g2, ...g k-1 ,g k ,...g 9625}.

[0117] Step 2: Apply rotation and scaling to the grayscale image set G to expand the data set, and finally collect 9625 experimental data images with a resolution of 1920 pixels × 720 pixels. Then the image set I = {i1, i2, ...i k-1 ,i k ,...i 9625}, where each element in I is a 1920*720 matrix.

[0118] Step 3: From the perspective of image feature extraction, a method for extracting the arc area between the pantograph and the catenary is designed based on the guidance of the pantograph head area. The pantograph head positioning frame is simultaneously offset from the top, bottom, left, and right directions to accurately obtain the area where the arc phenomenon may occur.

[0119] Step 4: All processed images constitute the final experimental dataset. In the entire dataset, 80% of the samples are selected as training sets and validation sets, and the remaining 20% ​​are used as test sets.

[0120] In the above steps, the present invention proposes a method for extracting the arc from the pantograph-catenary circuit based on the guidance of the pantograph head area in step 3.

[0121] The method of coordinate transformation of the bow head positioning frame is used to guide the bow head area, and a more accurate area where arcing may occur is obtained, which can effectively eliminate irrelevant interference in the image. Although the contact position of the bow and the net continues to change during the relative movement of the bow and the net, resulting in the spatial position of the bow and net arc always moving randomly, the angle and position of the camera recording the bow and net arc video are fixed, and the area where the arc phenomenon occurs is always fixed relative to the pantograph bow head area. It can be determined that the area where the arc phenomenon may occur should be a rectangular area above the pantograph bow head area, containing the pantograph slide bar and with a width approximately equal to the working length of the slide bar.

[0122] The specific pantograph bow head area guidance has the following 4 steps:

[0123] Step 1.3a: Get the four values ​​x of the bow head positioning frame given by the model in the previous section 左 、x 右 ,y 上 and 下 , namely the coordinates of the upper left corner and the lower right corner of the positioning box. The coordinate transformation is based on these four values.

[0124] Step 1.3b: Directly obtain the width w and height h of the positioning box through the coordinates of the two corner points of the positioning box:

[0125] w=x 右 -x 左

[0126] h=y 上 -y下

[0127] Step 1.3c: Divide the positioning frame into two equal parts in height and four equal parts in width, and take one of them as the height offset factor h' and width offset factor w' of the positioning frame. Use the offset factor as the basic amount to better control the precision of the offset:

[0128] h'=h / 2

[0129] w'=w / 4

[0130] Step 1.3d: Generate 4 random numbers n1, n2, n3 and n4 within a certain data range. The offset of each side is determined by the offset factor and the random number. The upward offset of the top edge of the positioning box is Δy 上 , the upward offset of the lower side is Δy 下 , the left side is offset to the right by Δx 左 , the right side is offset to the left by Δx 右 , and their values ​​are determined by the following formulas respectively.

[0131] Δy 上 =n1×h'

[0132] Δy 下 =n2×h'

[0133] Δx 左 =n3×w'

[0134] Δx 右 =n4×w'

[0135] The second part is the region of interest alignment algorithm:

[0136] Select ROI alignment to solve the problem of region mismatch caused by ROI pooling and improve the accuracy of target positioning.

[0137] Here are the steps:

[0138] Step 1: Traverse each candidate region without quantizing the mapped floating point coordinates.

[0139] Step 2: Divide the candidate region into k×k units ( Figure 2 2×2 in the figure), and the boundaries of the equally divided cells are not quantized.

[0140] Step 3: Divide each unit into 4 small units, take the center position as the sampling point, and determine 4 sampling points in each unit.

[0141] Step 4: Use bilinear interpolation to calculate the image value at the sampling point coordinate pixel.

[0142] Step 5: Perform maximum pooling on the four calculated sampling points to obtain a fixed-dimensional region of interest output.

[0143] The implementation principle and steps of the bilinear interpolation method in step 4 are as follows:

[0144] Assume that function f is in Q 11 =(x1,y1),Q 12 =(x1,y2),Q 21 =(x2,y1) and Q 22 The values ​​of the four points P=(x2,y2) are known. Find the value of the function f at P=(x,y).

[0145] First, perform linear interpolation in the x direction and obtain:

[0146]

[0147]

[0148] Secondly, linear interpolation is performed in the y direction to obtain:

[0149]

[0150] Finally, we get the desired f(x,y)

[0151]

[0152] The third part is the soft non-maximum suppression algorithm:

[0153] The soft non-maximum suppression algorithm proposes a method to gradually reduce the scores of overlapping detection boxes. That is, instead of directly removing these detection boxes, detection boxes with high overlap but still likely to contain useful information participate in the final result with a lower priority.

[0154] Here are the steps:

[0155] Step 1: All detection boxes are sorted by score, and the detection box with the highest score is selected as the reference box. For each remaining detection box, the score is updated according to its intersection-with-union ratio with the reference box, rather than directly removed. The optimal target is then determined by comparing the intersection-with-union ratio.

[0156] The intersection over union (IoU) is defined as the ratio of the intersection area of ​​the predicted box (P) and the true annotation box (GT) to the union area, as shown in the following formula:

[0157]

[0158] In the present invention, the candidate area with the highest intersection-and-union ratio with the real target area and the candidate area with an intersection-and-union ratio greater than 0.8 with any real marked box area are marked as positive samples, and the candidate area with an intersection-and-union ratio less than 0.2 with any real marked box area is marked as negative samples.

[0159] Step 2: When the current candidate box and the candidate bounding box with the highest score overlap, if their intersection-over-union ratio exceeds a given threshold, the soft non-maximum suppression algorithm uses a linear weighted method to update the confidence of the candidate box to S i (1-IoU), so that the candidate box also has a chance to become the final candidate box, avoiding missing the optimal target. As shown in the following formula:

[0160]

[0161] In the formula, S i is the confidence level, N t is the preset threshold, B is the candidate bounding box with the highest score at present, and b i is the candidate box to be processed.

[0162] Step 3: Select the next detection box with the highest score as the reference box, and repeat the score update process in the above steps until all detection boxes are processed.

[0163] Step 4: Re-rank the detection boxes based on the updated scores and select the box based on the final score.

[0164] The fourth part is the design based on machine learning model.

[0165] The heuristic V3 network is used as the backbone network to extract arc features faster and more accurately. The heuristic V3 network can reduce the dimension of large matrices while aggregating visual information at different sizes, making it convenient to extract features at different scales and speeding up network training time. The present invention selects the heuristic V3 network as the backbone network to better optimize the machine learning-based model proposed in the present invention.

[0166] Here are the steps:

[0167] Step 1: The feature extraction module is constructed by pre-training a heuristic V3 network on the training set.

[0168] Step 2: The input image first passes through three convolutional layers with a kernel size of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a 3 pixels × 3 pixels maximum pooling layer.

[0169] Step 3: Pass through 2 convolutional layers with a kernel size of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a 3 pixels × 3 pixels maximum pooling layer.

[0170] Step 4: Feature extraction through three heuristic modules.

[0171] Step 5: Pass through a 8-pixel × 8-pixel maximum pooling layer.

[0172] Step 6: Pass through a dropout layer and a 3×3 convolutional layer, a batch normalization layer, and a nonlinear activation layer.

[0173] The trained heuristic V3 network is obtained. The trained heuristic V3 network is used as the backbone network of the present invention to continue the training of the overall model.

[0174] Here are the steps:

[0175] Step 1: Take the training set image as input and use the heuristic V3 network as the backbone network of the model, which acts as a feature extractor. Extract features from the input image to obtain a feature map, which is applied to the subsequent network.

[0176] Step 2: The network uses these feature maps to generate candidate regions, that is, by generating anchor boxes of different sizes to obtain feature maps of a specified number of feature candidate regions.

[0177] Step 3: Unify them into the same size and input them all into the fully connected layer for target recognition and positioning, and use the normalized exponential function to classify specific categories, and complete the regression operation to obtain the precise location of the object.

[0178] Finally, the model structure of the present invention is as follows Figure 2 As shown, the final output of the entire model includes confidence and the probability of each category.

[0179] The fifth part is model training

[0180] After the network is built, the machine learning model proposed in the present invention is trained using the experimental data set in step 1. The number of model iterations is set to 400 times, and the batch size is set to 128. Stochastic gradient descent is selected as the model parameter optimizer, the initial value of the learning rate is 0.002, the momentum is set to 0.9, and the weight decay coefficient is 0.0005. Starting from the initial value, the learning rate is gradually reduced by following the half cosine learning rate. The test results are as follows Figure 3 shown.

[0181] The sixth part is the arc detection of the bow-net

[0182] Here are the steps:

[0183] Step 1: Selection of performance indicators

[0184] In order to verify the detection effect of pantograph arc, the trained model is evaluated using mean average precision (mAP). The mean average precision is the average of the average precision of each category, which is equal to the sum of the average precision of all categories divided by the number of all categories.

[0185]

[0186] In the formula, AP i represents the average precision value of each type, and N refers to the number of types detected by the model.

[0187] Step 2: Model testing

[0188] In order to verify the reliability of the bow-catenary arc detection method based on machine learning proposed in the present invention, the method proposed in the present invention was experimentally verified.

[0189] The experimental results show that the machine learning-based method proposed in the present invention has better performance indicators, that is, it can better identify arc phenomena and improve recognition accuracy and speed. The machine learning-based model can accurately detect arcs at a processing speed of 25 frames per second, with an average accuracy of 96.1%. Through this technical solution, the pantograph arc can be accurately detected and located.

[0190] Comparative test 1:

[0191] The proposed method is trained and compared with the traditional Faster R-CNN, Mask R-CNN, RefineDet and YOLOv3 on the test set to compare and evaluate their performance. Figure 4 shown.

[0192] Comparative test 2:

[0193] The proposed method is tested on the test set with the traditional Faster R-CNN, Mask R-CNN, RefineDet and YOLOv3. Since the key to the detection of arc in the bow-net is how to identify it quickly and accurately, the mean average precision (mAP) and running time of the detection model are selected as evaluation indicators to more intuitively reflect the performance of different models. The network performance comparison is shown in the following table:

[0194]

[0195] It can be found that compared with traditional algorithms, the present invention gives full play to the advantages of machine learning models, maintains high precision and good robustness, and at the same time improves the accuracy through the bow-catenary arc extraction method based on the overall offset of the pantograph bow head area, so that the bow-catenary arc phenomenon can be detected and identified quickly and accurately.

Claims

1. A method for detecting arc in a pantograph-catenary system based on machine learning, characterized in that: The following steps are included: S1. Extract the pantograph head area and obtain the pantograph-catenary arc area guided by the pantograph head area; S2. Use the region of interest alignment algorithm to perform feature mapping of candidate bounding boxes; S3. Progressively reduce the score of overlapping detection boxes based on the soft non-maximum suppression algorithm; S4. The heuristic V3 network is used as the backbone network and combined with the region generation network to extract arc features.

2. A method for detecting pantograph-catenary arc based on machine learning as claimed in claim 1, characterized in that: In S1, the following steps are included: S1.

1. After the arc video of the bow-net is intercepted into continuous frame images, grayscale processing is performed to convert the images into grayscale images; S1.2, apply rotation and scaling to the original captured images to expand the dataset; S1.

3. From the perspective of image feature extraction, a method for extracting the arc area of ​​the pantograph-catenary based on the guidance of the pantograph head area is designed; S1.4, all processed images constitute the experimental dataset; Or follow these steps: S1.1, apply rotation and scaling to the original captured images to enlarge the dataset; S1.2, after intercepting the arc video of the bow-net into continuous frame images, grayscale processing is performed to convert the images into grayscale images; S1.

3. From the perspective of image feature extraction, a method for extracting the arc area of ​​the pantograph-catenary based on the guidance of the pantograph head area is designed; S1.

4. All processed images constitute the experimental dataset.

3. A method for detecting pantograph-catenary arc based on machine learning as claimed in claim 2, characterized in that: In S1.3, the method of coordinate transformation of the bow head positioning frame is used to guide the bow head area, and the area where the arc phenomenon may occur is accurately obtained.

4. The method for detecting pantograph-catenary arc based on machine learning as claimed in claim 2, characterized in that: In S1.3, the specific steps are as follows: S1.3a: Obtain the four values ​​x of the bow head positioning frame through the coordinates of the upper left corner and the lower right corner of the positioning frame 左 、x 右 ,y 上 and 下 , the coordinate transformation is performed based on these four values; S1.3b: The width w and height h of the positioning box are directly obtained through the coordinates of the two corner points of the positioning box: w=x 右 -x 左 h=y 上 -y 下 S1.3c: Divide the positioning frame into M equal parts in height direction and N equal parts in width direction, and take one of them as the height direction offset factor h' and width direction offset factor w' of the positioning frame. The offset factor is used as the basic amount to better control the precision of the offset: h'=h / M w'=w / N S1.3d: Generate four random numbers n1, n2, n3 and n4 within a certain data range. The offset of each side is determined by the offset factor and the random number. The upward offset of the top edge of the positioning box is Δy 上 , the upward offset of the lower side is Δy 下 , the left side is offset to the right by Δx 左 , the right side is offset to the left by Δx 右 , their values ​​are determined by the following formulas respectively; Δy 上 =n1×h' Δy 下 =n2×h' Δx 左 =n3×w' Δx 右 =n4×w'。 5. The method for detecting pantograph-catenary arc based on machine learning as claimed in claim 1, characterized in that: In S2, the following specific steps are included: S2.1: traverse each candidate region without quantizing the floating point coordinates after mapping; S2.2: Divide the candidate region into k×k units without quantizing the boundaries of the divided units; S2.3: Divide each unit into 4 small units, take the center position as the sampling point, and determine 4 sampling points in each unit; S2.4: Calculate the image value at the sampling point coordinate pixel using bilinear interpolation method; S2.5: Perform maximum pooling on the calculated four sampling points to obtain a fixed-dimensional region of interest output.

6. A method for detecting pantograph-catenary arc based on machine learning as claimed in claim 5, characterized in that: In S2.4, the bilinear interpolation method comprises the following steps: Assume that function f is in Q 11 =(x1,y1),Q 12 =(x1,y2),Q 21 =(x2,y1) and Q 22 =(x2,y2) The values ​​of these four points are known. Find the value of function f at P = (x,y); S2.4.

1. Perform linear interpolation in the x direction and obtain: S2.4.

2. Perform linear interpolation in the y direction and obtain: S2.4.

3. Get the required f(x,y) 7. The method for detecting pantograph-catenary arc based on machine learning as claimed in claim 1, characterized in that: In S3, the following steps are involved: S3.1: All detection boxes are sorted according to their scores, and the detection box with the highest score is selected as the reference box. For each remaining detection box, the score is updated according to the intersection-and-union ratio between the detection box and the reference box to determine the optimal target. The intersection-to-union ratio is defined as the ratio of the intersection area of ​​the predicted box P and the true annotation box GT to the union area, as shown in the following formula: S3.2: When the current candidate box and the candidate bounding box with the highest score overlap, if their intersection-over-union ratio exceeds a given threshold, the soft non-maximum suppression algorithm uses a linear weighted method to update the confidence of the candidate box to S i (1-IoU), as shown below: In the formula, S i is the confidence level, N t is the preset threshold, B is the candidate bounding box with the highest score at present, and b i is the candidate box to be processed; S3.3: Select the next detection box with the highest score as the reference box, and repeat the score update process in the above steps until all detection boxes are processed; S3.4: Re-rank the detection boxes based on the updated scores and select the box based on the final score.

8. The method for detecting pantograph-catenary arc based on machine learning as claimed in claim 1, characterized in that: In S4, the model training method is as follows: S4.2.1: Take the training set image as input, use the heuristic V3 network as the backbone network of the model, extract the input image features to obtain the feature map, and apply the feature map to the subsequent network; S4.2.2: The network uses these feature maps to generate candidate regions, and obtains feature maps of a specified number of feature candidate regions by generating anchor boxes of different sizes; S4.2.3: Unify them into the same size and input them all into the fully connected layer for target recognition and positioning, and use the normalized exponential function to classify specific categories, and complete the regression operation to obtain the precise location of the object.

9. A method for detecting pantograph-catenary arc based on machine learning as claimed in claim 8, characterized in that: The training method of the heuristic V3 network is as follows: the feature extraction module pre-trains the heuristic V3 network on the training set; S4.1.1: The input image is passed through three convolutional layers with a kernel size of 3 pixels × 3 pixels, a batch normalization layer, a nonlinear activation layer, and a 3 pixels × 3 pixels maximum pooling layer; S4.1.2: After two convolutional layers with a kernel size of 3 pixels × 3 pixels, a batch normalization layer, a nonlinear activation layer, and a 3 pixels × 3 pixels maximum pooling layer; S4.1.3: Feature extraction through three heuristic modules; S4.1.4: After a 8-pixel × 8-pixel maximum pooling layer; S4.1.5: Pass through a dropout layer and a 3×3 pixel convolution layer, a batch normalization layer, and a nonlinear activation layer.

10. Application of a bow-catenary arc detection method based on machine learning as described in any one of claims 1 to 9 in detecting and identifying bow-catenary arc phenomena.

Images