A machine learning-based pantograph-catenary arc detection method

Abstract

The application provides a method for pantograph-catenary arc detection and recognition based on machine learning. That is, a method based on the pantograph head region guidance is adopted to extract the pantograph head region overall offset pantograph-catenary arc to solve the problem of large background interference in arc detection; a new machine learning model is established; a heuristic V3 network is used as a backbone network to extract arc features more quickly and accurately; a region of interest alignment algorithm is used in the region generation network to perform feature mapping operation on the candidate boundary box to solve the problem of region mismatch in the region of interest pooling algorithm, thereby improving the accuracy of target positioning; and a soft non-maximum suppression algorithm is used to solve the problem of missed detection and false detection of the non-maximum suppression algorithm, thereby more accurately selecting the candidate boundary box. The method can realize rapid and accurate detection of the arc phenomenon.

Description

Technical Field

[0001] This invention belongs to the field of computer digital image processing and pattern recognition technology, and specifically relates to a method for detecting arcing in a pantograph-catenary circuit based on machine learning. Background Technology

[0002] In electrified railways, the pantograph-contact system is a crucial component for trains to obtain electrical energy. Trains receive power through the contact between the pantograph and the contact wire on the overhead wire, ensuring smooth train operation. During operation, due to various factors, the pantograph and contact wire are prone to separation, generating pantograph arcing. Arcing affects the quality of current collection, accelerates wear on the pantograph contactor and contact wire, damages internal electrical equipment, interferes with signaling equipment along the line, and poses a significant threat to train operational safety.

[0003] Currently, the common method for detecting electric arcs is to record and monitor the operation of the pantograph-catenary system using cameras, and then analyze the monitoring video after an arc occurs. The key to ensuring smooth train operation is how to quickly extract accurate and effective arc information from the monitoring video. Traditional models for arc detection have the following three problems:

[0004] 1. Complex video backgrounds cause significant interference to arc recognition;

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

[0006] 3. The false positive rate or false negative rate is relatively high. Summary of the Invention

[0007] This invention aims to overcome the aforementioned shortcomings and proposes a machine learning-based method for pantograph-catenary arc detection and recognition. Specifically, it employs a method guided by the pantograph head region to extract the pantograph-catenary arc due to the overall offset of the head region, thus addressing the problem of significant background interference in arc detection. A novel machine learning model is established: a heuristic V3 network is used as the backbone network for faster and more accurate arc feature extraction; a region of interest alignment algorithm is employed in the region generation network for feature mapping of candidate bounding boxes, addressing the region mismatch problem in the region of interest pooling algorithm and improving target localization accuracy; and a soft non-maximum suppression algorithm is used to solve the problems of missed and false detections in non-maximum suppression algorithms, thereby selecting candidate bounding boxes more accurately. This method enables rapid and accurate detection of arc phenomena.

[0008] This invention provides a machine learning-based method for detecting arcing in a pantograph-catenary system, characterized by the following steps:

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

[0010] S2. Feature mapping of candidate bounding boxes is performed using a region of interest alignment algorithm;

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

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

[0013] Furthermore, the present invention provides a machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

[0014] S1 includes the following steps:

[0015] S1.2 Apply rotation and scaling to the original captured images to expand the dataset;

[0016] S1.3 From the perspective of image feature extraction, a method for extracting the arc region of the pantograph and catenary based on pantograph head region guidance was designed.

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

[0018] Alternatively, the following steps can be taken:

[0019] S1.1 Apply rotation and scaling to the original captured images to expand the dataset;

[0020] S1.2 After extracting the pantograph-catenary arc video into continuous frame images, perform grayscale processing to convert the images into grayscale images;

[0021] S1.3 From the perspective of image feature extraction, a method for extracting the arc region of the pantograph based on the pantograph head region was designed; S1.4 All processed images constitute the experimental dataset.

[0022] Furthermore, the present invention provides a machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

[0023] In S1.3, the bow head region is guided by the coordinate transformation method of the bow head positioning frame to accurately obtain the area where electric arcing may occur.

[0024] Furthermore, the present invention provides a machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

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

[0026] S1.3a: Obtain the four values ​​x of the bow head positioning frame using the coordinates of the top left and bottom right corners of the positioning frame. 左 x右 y 上 and y 下 The coordinate transformation is based on these four values;

[0027] S1.3b: The width w and height h of the positioning frame can be obtained directly from the coordinates of the two corner points of the positioning frame.

[0028] w = x 右 -x 左

[0029] h = y 上 -y 下

[0030] S1.3c: Divide the positioning frame into equal parts in the height direction M and the width direction N, and take one part of each as the height direction offset factor h' and the width direction offset factor w' of the positioning frame. Using the offset factor as the basic quantity, the fineness of the offset can be better controlled.

[0031] h' = h / M

[0032] w'=w / N

[0033] S1.3d: Generate four 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 numbers. The top edge of the positioning box is offset upwards by Δy. 上 The upward offset of the bottom is Δy 下 The offset from left to right is Δx 左 The offset from right to left is Δx 右 Their values ​​are determined by the following formulas;

[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 machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

[0039] S2 includes the following specific steps:

[0040] S2.1: Traverse each candidate region without quantizing the mapped floating-point coordinates;

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

[0042] S2.3: Divide each unit into 4 smaller units, and take the center position of each unit as the sampling point. Four sampling points can be determined in each unit.

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

[0044] S2.4: The image value at the pixel coordinates of the sampling point is calculated using bilinear interpolation.

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

[0046] Furthermore, the present invention provides a machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

[0047] In S2.4, the use of bilinear interpolation includes the following steps:

[0048] Assume function f in Q 11 = (x1, y1), Q 12 = (x1, y2), Q 21 = (x2, y1) and Q 22 Given the values ​​of the four points (x2, y2) in the interval P, find the value of the function f at (x, y).

[0049] S2.4.1. Performing linear interpolation in the x-direction yields:

[0050]

[0051]

[0052] S2.4.2. Performing linear interpolation in the y-direction yields:

[0053]

[0054] S2.4.3. Obtain the desired f(x,y)

[0055]

[0056] Furthermore, the present invention provides a machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

[0057] S3 includes the following steps:

[0058] S3.1: All detection boxes are sorted according to their scores. 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-union ratio with the reference box to determine the optimal target.

[0059] Regarding the above scores: In the machine learning model proposed in this invention, image recognition is performed on the training samples. During the image recognition process, the network proposes candidate boxes (rectangular regions). Different candidate boxes have different differences from the actual target. The more closely the candidate boxes match the actual target contour, the higher the detected value (i.e., the score).

[0060] The intersection-union ratio (IUGR) is defined as the ratio of the area of ​​the intersection region of the predicted bounding box P and the area of ​​the union region of the ground truth bounding box GT, as shown in the following formula:

[0061]

[0062] S3.2: When the current candidate box and the highest-scoring candidate bounding box overlap, if their intersection-union ratio exceeds a given threshold, the soft non-maximum suppression algorithm updates the confidence of the candidate box to S using a linear weighting method. i (1-IoU), as shown in the following formula:

[0063]

[0064] In the formula, S i For confidence level, N t Let B be the preset threshold, and B be the candidate bounding box with the highest score. i Candidate boxes to be processed;

[0065] S3.3: Select the next highest-scoring detection box as the baseline box, and repeat the score update process in the above steps until all detection boxes have been processed;

[0066] S3.4: Reorder the detection boxes based on the updated scores and select boxes according to the final scores.

[0067] Furthermore, the present invention provides a machine learning-based method for detecting pantograph-catenary arcs, characterized in that:

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

[0069] S4.2.1: Using the images in the training set as input, the heuristic V3 network is used as the backbone network of the model to extract features from the input images to obtain feature maps, which are then applied to subsequent networks.

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

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

[0072] The pantograph-catenary arc detection method based on machine learning as described in claim 8 is characterized in that:

[0073] The training method for 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 passes through three convolutional layers with kernel sizes of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a 3-pixel × 3-pixel max pooling layer;

[0075] S4.1.2: After two convolutional layers with kernel size of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a 3-pixel × 3-pixel max pooling layer;

[0076] S4.1.3: Feature extraction is performed using three heuristic modules;

[0077] S4.1.4: After one 8-pixel × 8-pixel max pooling layer;

[0078] S4.1.5: It uses a random deactivation layer, a 3-pixel × 3-pixel convolutional layer, a batch normalization layer, and a non-linear activation layer.

[0079] Furthermore, the machine learning-based pantograph-catenary arc detection method of the present invention has been applied to the detection and identification of pantograph-catenary arc phenomena. Attached Figure Description

[0080] Figure 1 Overall flowchart of the present invention;

[0081] Figure 2 Overall model structure diagram;

[0082] Figure 3 Test result image;

[0083] Figure 4 The comparison results between the system of this invention and the traditional system are shown in the figure. Detailed Implementation

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

[0085] like Figure 1 As shown, this invention is achieved through the following technical solutions: A method based on the pantograph head region guidance is used to extract the pantograph-catenary arc with overall offset of the pantograph head region, thus solving the problem of high background interference in arc detection; a new machine learning model is established: a heuristic V3 network is used as the backbone network to extract arc features faster and more accurately; an interest region alignment algorithm is used in the region generation network to perform feature mapping operations on candidate bounding boxes, solving the problem of region mismatch in the interest region pooling algorithm, thereby improving the accuracy of target localization; a soft non-maximum suppression algorithm is used to solve the problems of missed detection and false detection in the non-maximum suppression algorithm, thus selecting candidate bounding boxes more accurately.

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

[0087] The first part is the algorithm for extracting the pantograph head region. A camera mounted on the top of the train captures video of the pantograph-catenary arc at the same angle and position. This video includes different operating environments such as daytime, nighttime, and tunnels. The steps are as follows:

[0088] Step 1: Extract continuous frame images from the pantograph-catenary arc video and then perform grayscale processing to convert the images into grayscale images.

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

[0090] Step 3: From the perspective of image feature extraction, a pantograph-catenary arc region extraction method based on pantograph head region guidance was designed. The pantograph head positioning frame is simultaneously offset from four directions: top, bottom, left, and right to accurately obtain the region where arcing 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 chosen to replace the original region of interest pooling to unify the feature map size, thereby solving the problem of region mismatch that occurs in region of interest pooling and improving the accuracy of target localization.

[0094] The candidate bounding box coordinates generated by the region generation network correspond to the original image space. Based on the mapping relationship between the feature map and the original image, the candidate bounding boxes can be mapped onto the feature map. The data type of the candidate bounding box coordinate values ​​is usually floating-point numbers. During pooling, region of interest pooling requires first quantizing these values ​​into integers and then dividing them into several units. These two quantization operations cause a certain deviation between the candidate bounding box and its position in the original image, resulting in a region mismatch problem that affects the accuracy of target localization. This invention selects a region of interest alignment algorithm for the feature mapping operation of the candidate bounding boxes.

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

[0096] At the same target location, the region candidate network generates a large number of candidate bounding boxes, which overlap with each other. Non-maximum suppression algorithms (NMS) suppress maxima by searching for local maxima. However, NMS deletes all bounding boxes larger than a given threshold, which can easily lead to false positives and false negatives when the target is occluded. This invention employs an improved version of the NMS algorithm: a soft non-maximum suppression algorithm.

[0097] Part Four is the design based on machine learning models.

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

[0099] Traditional networks suffer from problems such as long training times, difficult parameter tuning, large storage requirements, and the tendency to exhibit gradient vanishing and overfitting during training as network width and depth increase. The heuristic V3 network, however, can reduce the dimensionality of large matrices while simultaneously aggregating visual information at different sizes, facilitating feature extraction at various scales and accelerating network training. Therefore, this invention chooses the heuristic V3 network as the backbone network, combined with a region generation network, to better optimize the proposed machine learning-based model.

[0100] The steps are as follows:

[0101] Step 1: Using images from the experimental dataset as input, a heuristic V3 network is used as the backbone of the model, functioning as a feature extractor. Feature maps are extracted from the input images, and these feature maps are then applied to the subsequent region generation network and fully connected layers.

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

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

[0104] Finally, the model structure of this 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 score, and the probability of each class.

[0105] The fifth part is model training.

[0106] After the network is constructed, the machine learning-based model proposed in this invention is trained using the experimental dataset from step 1.

[0107] Part 6 is pantograph-catenary arc detection.

[0108] The steps are as follows:

[0109] Step 1: Selection of performance indicators

[0110] To verify the detection effectiveness of the pantograph arc, the trained model was evaluated using the mean accuracy.

[0111] Step 2: Model Detection

[0112] Experimental results show that this technical solution can accurately detect and locate the pantograph arc.

[0113] This invention fully leverages the advantages of machine learning models, maintaining high accuracy and robustness, while further enhancing accuracy through a pantograph-catenary arc extraction method based on the overall offset of the pantograph head region. This enables rapid and precise detection and identification of pantograph-catenary arc phenomena.

[0114] The following description uses a specific optional embodiment as an example.

[0115] The first part is the algorithm for extracting the bow-head region. Cameras mounted on the top of the train capture video at the same angle and position, including different operating environments such as daytime, nighttime, and tunnels. The steps are as follows:

[0116] Step 1: Extract consecutive frames from the pantograph-catenary arc video, and convert the resulting images to grayscale. Use the average RGB value of each pixel as its grayscale value, thus obtaining a 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 dataset, ultimately collecting 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 pantograph-catenary arc region extraction method based on pantograph head region guidance was designed. The pantograph head positioning frame was simultaneously offset from four directions: top, bottom, left, and right to accurately obtain the region where arcing may occur.

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

[0120] In the above steps, step 3 of this invention proposes a method for extracting the pantograph-catenary arc based on the guidance of the pantograph head region.

[0121] By utilizing the coordinate transformation method of the pantograph head positioning frame for pantograph head region guidance, a more accurate area where arcing may occur can be obtained, effectively eliminating irrelevant interference in the image. Although the contact position of the pantograph and catenary continuously changes during relative movement, causing the spatial position of the pantograph-catenary arc to always be in random motion, the angle and position of the camera recording the pantograph-catenary arc video are fixed, and the area where the arcing phenomenon occurs is always fixed relative to the pantograph head region. It can be determined that the area where arcing may occur should be a rectangular area above the pantograph head region, including the pantograph sliding strip, and with a width approximately equal to the working length of the sliding strip.

[0122] The specific guidance of the pantograph head area involves the following four steps:

[0123] Step 1.3a: Obtain the four values ​​x of the bow head positioning box given in the previous section. 左 x 右 y 上 and y 下 These are the coordinates of the top left and bottom right corners of the positioning frame, and the coordinate transformation is based on these four values.

[0124] Step 1.3b: Obtain the width w and height h of the positioning frame directly using the coordinates of its two corner points:

[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 the height direction and four equal parts in the width direction, and take one part of each as the height direction offset factor h' and width direction offset factor w' of the positioning frame. Using the offset factors as basic quantities, the precision of the offset can be better controlled.

[0128] h' = h / 2

[0129] w' = w / 4

[0130] Step 1.3d: 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 numbers. The upward offset of the top edge of the positioning box is Δy. 上 The upward offset of the bottom is Δy 下 The offset from left to right is Δx 左 The offset from right to left is Δx 右 Their values ​​are determined by the following formulas.

[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] Selecting region of interest alignment can solve the problem of region mismatch that occurs in region of interest pooling, thereby improving the accuracy of target localization.

[0137] The steps are as follows:

[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), the boundary of the evenly divided unit is not quantized.

[0140] Step 3: Divide each unit into 4 smaller units and take the center position of each unit as the sampling point. Four sampling points can be determined in each unit.

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

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

[0143] The implementation principle of step 4, which uses bilinear interpolation, is as follows:

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

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

[0146]

[0147]

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

[0149]

[0150] Finally, we obtain 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 progressively reduce the scores of overlapping detection boxes. That is, instead of directly removing these detection boxes, detection boxes with high overlap but still potentially containing useful information participate in the final result with a lower priority.

[0154] The steps are as follows:

[0155] Step 1: Sort all bounding boxes by score, select the bounding box with the highest score as the baseline box. For each remaining bounding box, update its score based on its intersection-union ratio (IU) with the baseline box, instead of removing it directly. Then, determine the optimal target by comparing the IU values.

[0156] The Intersection over Union (IoU) ratio is defined as the ratio of the area of ​​the intersection of the predicted bounding box (P) and the area of ​​the union of the ground truth bounding boxes (GT), as shown in the following formula:

[0157]

[0158] In this invention, candidate regions with the highest intersection-union ratio (IU) with the real target region and candidate regions with an IU greater than 0.8 with any real bounding box region are labeled as positive samples, while candidate regions with an IU less than 0.2 with any real bounding box region are labeled as negative samples.

[0159] Step 2: When the current candidate box and the highest-scoring candidate bounding box overlap, if their intersection-union ratio exceeds a given threshold, the soft non-maximum suppression algorithm updates the confidence of the candidate box to S using a linear weighting method. i (1-IoU) is used to give this candidate box a chance to become a final candidate box, thus avoiding missing the optimal target. This is illustrated in the following formula:

[0160]

[0161] In the formula, S i For confidence level, N t Let B be the preset threshold, and B be the candidate bounding box with the highest score. i These are the candidate boxes to be processed.

[0162] Step 3: Select the next highest-scoring detection box as the baseline box, and repeat the score update process in the previous steps until all detection boxes have been processed.

[0163] Step 4: Reorder the detection boxes based on the updated scores and select boxes according to the final scores.

[0164] Part Four is the design based on machine learning models.

[0165] A heuristic V3 network is employed as the backbone network to extract arc features more quickly and accurately. Using the heuristic V3 network allows for dimensionality reduction of large matrices while simultaneously aggregating visual information at different sizes, facilitating feature extraction at various scales and accelerating network training time. This invention chooses the heuristic V3 network as the backbone network to better optimize the proposed machine learning-based model.

[0166] The steps are as follows:

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

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

[0169] Step 3: The process involves two convolutional layers with kernel sizes of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a max pooling layer of 3 pixels × 3 pixels.

[0170] Step 4: Feature extraction is performed using three heuristic modules.

[0171] Step 5: Pass through an 8-pixel × 8-pixel max pooling layer.

[0172] Step 6: Use a random deactivation layer, a 3-pixel × 3-pixel convolutional layer, a batch normalization layer, and a non-linear activation layer.

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

[0174] The steps are as follows:

[0175] Step 1: Using the images from the training set as input, a heuristic V3 network is used as the backbone network of the model, which functions as a feature extractor. Feature maps are extracted from the input images, and these feature maps are applied to subsequent networks.

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

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

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

[0179] The fifth part is model training.

[0180] After the network was constructed, the machine learning-based model proposed in this invention was trained using the experimental dataset from step 1. The number of model iterations was set to 400, and the batch size to 128. Stochastic gradient descent was chosen as the model parameter optimizer, with an initial learning rate of 0.002, momentum of 0.9, and weight decay coefficient of 0.0005. Starting from the initial values, a half-cosine learning rate was gradually applied to progressively decrease the learning rate. The detection results are as follows... Figure 3 As shown.

[0181] Part 6 is pantograph-catenary arc detection.

[0182] The steps are as follows:

[0183] Step 1: Selection of performance indicators

[0184] To verify the effectiveness of pantograph arc detection, the trained model was evaluated using mean average precision (mAP). Mean average precision is the average of the average precision for each class, and is equal to the sum of the average precisions for all classes divided by the number of classes.

[0185]

[0186] In the formula, AP i This represents the average precision value for each type, where N is the number of types detected by the model.

[0187] Step 2: Model Detection

[0188] To verify the reliability of the machine learning-based pantograph-catenary arc detection method proposed in this invention, the method was experimentally verified.

[0189] Experimental results show that the machine learning-based method proposed in this invention has better performance indicators, namely, it can better identify electric arc phenomena and improve recognition accuracy and speed. The machine learning-based model achieves accurate detection of electric arcs at a processing speed of 25 frames per second, with an average accuracy of 96.1%. This technical solution enables precise detection and location of pantograph arcs.

[0190] Comparative Experiment 1:

[0191] The proposed method was compared with traditional Faster R-CNN, Mask R-CNN, RefineDet, and YOLOv3 on the test set to evaluate its performance. The detection results are as follows: Figure 4 As shown.

[0192] Comparative Experiment 2:

[0193] The proposed method was tested against traditional Faster R-CNN, Mask R-CNN, RefineDet, and YOLOv3 on a test set. Since the key to pantograph-catenary arc detection is rapid and accurate identification, the mean average precision (mAP) and runtime of the detection models were chosen as evaluation metrics to more intuitively reflect the performance of different models. The network performance comparison is shown in the table below:

[0194]

[0195] It can be seen that, compared with traditional algorithms, this invention fully leverages the advantages of machine learning models, maintaining high accuracy and good robustness, while improving accuracy through a pantograph-catenary arc extraction method based on the overall offset of the pantograph head region. Thus, it can quickly and accurately detect and identify pantograph-catenary arc phenomena.

Claims

1. A machine learning-based method for detecting pantograph-catenary arcs, characterized in that, It includes the following steps: S1. Extract the pantograph head region and obtain the pantograph-catenary arc region guided by the pantograph head region; S2. Feature mapping of candidate bounding boxes is performed using a region of interest alignment algorithm; S3. Progressively reduce the score of overlapping detection boxes based on a soft non-maximum suppression algorithm; S4. Using a heuristic V3 network as the backbone network, combined with a region generation network, arc features are extracted; Specifically, S2 includes the following steps: S2.1: Traverse each candidate region without quantizing the mapped floating-point coordinates; S2.2: Divide the candidate region into k×k units without quantizing the boundaries of the units; S2.3: Divide each unit into 4 smaller units, and take the center position of each unit as the sampling point. Four sampling points can be determined in each unit. S2.4: The image value at the pixel coordinates of the sampling point is calculated using bilinear interpolation. S2.5: Max pooling is performed on the four calculated sampling points to obtain a fixed-dimensional region of interest output; S3 includes the following steps: S3.1: All detection boxes are sorted according to their scores. 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-union ratio with the reference box to determine the optimal target. Wherein, the intersection-union ratio is defined as the predicted box. P True annotation box GT The ratio of the area of ​​the intersection region to the area of ​​the union region is shown in the following formula: S3.2: When the current candidate box and the highest-scoring candidate bounding box overlap, if their intersection-union ratio (IU) exceeds a given threshold, the soft non-maximum suppression algorithm updates the confidence of the candidate box using a linear weighting method. As shown in the following formula: In the formula, For confidence level, The preset threshold, B This is the candidate bounding box with the highest score so far. Candidate boxes to be processed; S3.3: Select the next highest-scoring detection box as the baseline box, and repeat the score update process in the above steps until all detection boxes have been processed; S3.4: Reorder the detection boxes based on the updated scores and select the boxes according to the final scores; In S4, the model is trained as follows: S4.2.1: Using the images in the training set as input, the heuristic V3 network is used as the backbone network of the model to extract features from the input images to obtain feature maps, which are then applied to subsequent networks. S4.2.2: The network uses these feature maps to generate candidate regions, and obtains feature maps of a specified number of candidate regions by generating anchor boxes of different sizes; S4.2.3: Unify them to the same size and input them all into the fully connected layer for target recognition and localization, and use the normalized exponential function to classify specific categories, while completing the regression operation to obtain the precise location of the object; The training method for 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 passes through three convolutional layers with kernel sizes of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a 3-pixel × 3-pixel max pooling layer; S4.1.2: After two convolutional layers with kernel size of 3 pixels × 3 pixels, a batch normalization layer, a non-linear activation layer, and a 3-pixel × 3-pixel max pooling layer; S4.1.3: Feature extraction is performed using three heuristic modules; S4.1.4: After one 8-pixel × 8-pixel max pooling layer; S4.1.5: It uses a random deactivation layer, a 3-pixel × 3-pixel convolutional layer, a batch normalization layer, and a non-linear activation layer.

2. The machine learning-based arc detection method for pantograph-catenary circuits as described in claim 1, characterized in that: S1 includes the following steps: S1.1 After extracting the pantograph-catenary arc video into continuous frame images, perform grayscale processing 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 region of the pantograph and catenary based on pantograph head region guidance was designed. S1.

4. All processed images constitute the experimental dataset; Alternatively, the following steps can be taken: S1.1 Apply rotation and scaling to the original captured images to expand the dataset; S1.2 After extracting the pantograph-catenary arc video into continuous frame images, perform grayscale processing to convert the images into grayscale images; S1.3 From the perspective of image feature extraction, a method for extracting the arc region of the pantograph and catenary based on pantograph head region guidance was designed. S1.

4. All processed images constitute the experimental dataset.

3. The pantograph-catenary arc detection method based on machine learning as described in claim 2, characterized in that: In S1.3, the bow head region is guided by the coordinate transformation method of the bow head positioning frame to accurately obtain the area where electric arcing may occur.

4. The machine learning-based pantograph-catenary arc detection method as described 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 using the coordinates of the top left and bottom right corners of the positioning frame. 左 x 右 y 上 and y 下 The coordinate transformation is based on these four values; S1.3b: The width of the positioning frame can be obtained directly from the coordinates of its two corner points. w and high h : w=x 右 -x 左 h=y 上 -y 下 S1.3c: Divide the positioning frame into M equal parts in the height direction and N equal parts in the width direction, and take one part of each as the height direction offset factor of the positioning frame. and width direction offset factor Using the offset factor as the basic quantity allows for better control over 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 numbers. The top edge of the positioning box is offset upwards by... The upward offset of the bottom is Δy 下 The offset from left to right is Δx 左 The offset from right to left is Δx 右 Their values ​​are determined by the following formulas; Dy 上 = n1×h' Dy 下 = n2×h' Δx 左 =n3×w' Δx 右 =n4×w'。 5. The pantograph-catenary arc detection method based on machine learning as described in claim 1, characterized in that: In S2.4, the use of bilinear interpolation includes the following steps: Assume function f in Q 11 = (x1, y1), Q 12 = (x1, y2), Q 21 = (x2, y1) and Q 22 Given the values ​​of the four points (x2, y2), find the function f in... The value; S2.4.

1. Performing linear interpolation in the x-direction yields: S2.4.

2. Performing linear interpolation in the y-direction yields: S2.4.

3. Obtain the desired result 6. The application of the machine learning-based pantograph-catenary arc detection method as described in any one of claims 1-5 in the detection and identification of pantograph-catenary arc phenomena.

Images