A pantograph system electric spark real-time detection method and device and storage medium

Abstract

The application discloses a pantograph-catenary system electric spark real-time detection method and device and a storage medium, and belongs to the technical field of rail transit. The application is suitable for real-time spark detection of a pantograph monitoring system, adopts a low-power and high-efficiency processing method, can realize a real-time detection speed of 25 frames per second under extremely small computing power consumption, and can effectively adapt to complex environmental light conditions, thereby guaranteeing high precision and high robustness of detection. Through combination of a lightweight neural network and time series analysis, the accuracy and stability of spark detection are improved, and the application is suitable for pantograph monitoring and abnormal detection in the fields of high-speed rail and electric power.

Description

Technical Field

[0001] This application relates to the field of rail transit technology, and in particular to a method, device and storage medium for real-time detection of electric sparks in a pantograph-catenary system. Background Technology

[0002] like Figure 1 The diagram shown illustrates an implementation scenario of this application. The pantograph is a key component in electric trains used to acquire electrical energy. Installed on the train roof, it transmits current via a carbon sliding plate through contact with the contact wire. The pantograph and contact wire together form the pantograph-catenary system, ensuring a stable and continuous power supply for the electric train while minimizing wear on the contact surface.

[0003] However, during current transmission, factors such as wear at the contact points, snowfall, and freezing can increase the resistance of the electrical contact points, leading to electric arcs and sparks, which affect the stability of power transmission. The generation of electric arcs and sparks can not only damage electrical equipment but also cause trains to lose traction due to frequent power outages. Therefore, pantograph condition monitoring is crucial for ensuring the safe operation of trains, especially the real-time detection of electric arcs and sparks.

[0004] Current monitoring methods largely rely on manual video observation, which is inefficient and makes it difficult to capture instantaneous sparks or arcs in real time. Existing intelligent monitoring systems also face the problem of light interference, such as sunlight, tunnel and bridge light sources, which can affect the accurate detection of sparks or arcs. At the same time, pantograph status detection has high computational requirements and needs to be achieved in real time with limited computing power.

[0005] The following are the patent information that the applicant obtained from searching the technical solutions of this application, sorted in ascending order of relevance.

[0006] Patent No. CN112132789B proposes to judge the pantograph horn state by combining target detection model and segmentation model, but does not involve the solution of spark and arc problem of pantograph-catenary system; Patent No. CN112016544A proposes a pantograph intelligent monitoring and analysis method, device and system, which detects arc by arc sensor, does not involve specific image analysis method, and the technical path of spark monitoring is different from that of the present application; Patent No. CN115797699A proposes a pantograph electric spark fault detection method, which detects spark by improved full convolution single-stage target detection model, and the detection target and method are different from those of the present application; Patent No. CN110458126B focuses on the spark detection method, and detects the state of the pantograph by a pre-trained model, but the determination and screening method of the spark detection range is different from that of the present application, and is not suitable for night environment; Patent No. CN107633266B proposes a power locomotive catenary pantograph arc detection method, which detects arc by training Adaboost cascade classifier, and the method of screening suspected spark frame and the time continuity analysis method are different from those of the present application.

[0007] Therefore, through the retrieval of the technical scheme of the present application, the existing monitoring method has certain technical limitations, and the present application can effectively avoid external light source interference and improve the detection accuracy of spark and arc by combining the monotonicity increase and decrease law of time sequence spark frame, especially suitable for train operation monitoring in variable environment. SUMMARY

[0008] The embodiment of the present application provides a pantograph-catenary system electric spark real-time detection method, device and storage medium, and the technical scheme is as follows:

[0009] On the one hand, a pantograph-catenary system electric spark real-time detection method is provided, and the method comprises:

[0010] obtaining a pantograph monitoring image;

[0011] locating a pantograph head region image from the pantograph monitoring image;

[0012] calculating the horizontal movement rate of the highlight target in the pantograph head region image to determine whether to enter the spark detection step;

[0013] performing time sequence analysis on the highlight area in the pantograph head region image to determine whether a suspected spark appears;

[0014] extracting the maximum image frame of the suspected spark from the pantograph head region image;

[0015] inputting the maximum image frame of the suspected spark into a neural network for classification and screening, excluding non-spark images, and retaining image frames judged as sparks.

[0016] Optionally, the acquiring pantograph monitoring images comprises:

[0017] acquiring real-time video stream through RTSP protocol;

[0018] converting the video stream into an image sequence in RGB format, the image sequence being pantograph monitoring images, and storing the image sequence in a queue for subsequent processing.

[0019] Optionally, the locating pantograph head region images from the pantograph monitoring images comprises:

[0020] extracting features and locating the pantograph head region images based on the pantograph monitoring images through a target detection algorithm in a convolutional neural network.

[0021] Optionally, the calculating a lateral movement change rate of highlighted targets in the pantograph head region images to determine whether to enter a spark detection step comprises:

[0022] calculating a lateral movement change rate of highlighted targets in the pantograph head region images;

[0023] if the lateral movement change rate is less than a set threshold, it is considered that it is strong light interference such as the sun;

[0024] if the lateral movement change rate is greater than the set threshold, the spark detection step is continued.

[0025] Optionally, the performing time series analysis on highlighted areas in the pantograph head region images to determine whether a suspected spark appears comprises:

[0026] calculating an area of a highlighted target region in the pantograph head region images to generate a time series of highlighted areas;

[0027] determining whether it is a flickering feature of a spark according to a fluctuation amplitude of area changes in the time series of highlighted areas, and if the highlighted area change amplitude is large, it is determined that it is a suspected spark.

[0028] Optionally, the inputting the largest image frame of the suspected spark into a neural network for classification and screening comprises:

[0029] extracting the largest image frame of the suspected spark from the pantograph head region images, the largest image frame being an image frame that exhibits the largest highlighted area in the time series;

[0030] inputting the largest image frame of the suspected spark into a pre-trained neural network model, the neural network model identifying whether there is a spark feature in the image by extracting features and classifying the image;

[0031] According to the classification result output by the neural network, the suspected spark maximum image frame of non-spark is excluded, and only the image frame classified as spark is reserved.

[0032] In another aspect, an overhead line system electric spark real-time detection device is provided, the device comprising:

[0033] An image acquisition module is configured to acquire a pantograph monitoring image.

[0034] An image positioning module is configured to position a pantograph head region image from the pantograph monitoring image.

[0035] A detection judgment module is configured to calculate a horizontal movement rate of a highlight target in the pantograph head region image to determine whether to enter a spark detection step.

[0036] A spark judgment module is configured to perform time series analysis on a highlight area in the pantograph head region image to determine whether a suspected spark appears.

[0037] An image extraction module is configured to extract a suspected spark maximum image frame from the pantograph head region image.

[0038] An image screening module is configured to input the suspected spark maximum image frame into a neural network for classification screening, to exclude non-spark images and reserve image frames determined as sparks.

[0039] In another aspect, a computer readable storage medium is provided, the storage medium storing at least one instruction, the at least one instruction being used to be executed by a processor to implement the overhead line system electric spark real-time detection method of the above aspect.

[0040] In another aspect, a computer program product is also provided, the computer program product storing at least one instruction, the at least one instruction being loaded and executed by the processor to implement the overhead line system electric spark real-time detection method of the above aspect.

[0041] The present application discloses an overhead line system electric spark real-time detection method, which can effectively realize real-time spark detection with low algorithm consumption by acquiring and processing pantograph monitoring images, combining highlight target change rate detection and highlight area time series analysis. The method can exclude environmental light interference, especially strong light interference such as sunlight, and provides efficient and accurate spark detection capability. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 An implementation scenario schematic diagram of the present application is shown;

[0043] Figure 2A flow chart of the real-time detection method of the pantograph system electric spark provided by an example embodiment of the present application is shown.

[0044] Figure 3 A pantograph monitoring diagram is shown.

[0045] Figure 4 A stage structure diagram of ShuffleNetV2 is shown.

[0046] Figure 5 A pantograph head region positioning schematic diagram is shown.

[0047] Figure 6 A high-light target marking schematic diagram is shown.

[0048] Figure 7 A high-light target transverse position change rate in the pantograph head region schematic diagram is shown.

[0049] Figure 8 A high-light area sequence when the spark flashes schematic diagram is shown.

[0050] Figure 9 A high-light area sequence when the spark flashes schematic diagram is shown.

[0051] Figure 10 A network model structure schematic diagram is shown. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the present application more clear, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0053] In this paper, "a plurality of" refers to two or more. "And / or", which describes the association between the associated objects, means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. The character " / " generally represents that the associated objects before and after are a "or" relationship.

[0054] Before each embodiment, the specific technical problems to be overcome in the background art are further analyzed.

[0055] First of all, the problem of low efficiency of artificial watching of pantograph state video. The vehicle mechanic is a position that follows the motor vehicle safety fault to do processing, and the vehicle mechanic is responsible for monitoring the running state of the motor train set on line and the processing of emergency failure after the motor train set leaves the depot. In the past, the mechanic needs to pay attention to the running state of the pantograph in the video at all times, and can respond in time to the continuous abnormality such as pantograph failure, but for the flickering transient phenomenon such as spark and arc, the mechanic is difficult to pay continuous attention, and cannot accurately capture and timely alarm.

[0056] Secondly, the spark and arc detection is easily interfered by sunlight, bridge, tunnel, trackside light, and high-brightness light at night. Since 2019, the CR400AF / BF and other intelligent motor trains have gradually installed intelligent analysis hosts on the basis of the original video monitoring system for the purpose of real-time automatic detection of the pantograph operation state and reducing the manual warning pressure of the train mechanic. However, due to the fact that the motor train shuttles in various terrains and landscapes across the country, strong light interference is easily generated in some scenes such as bridge hollow suspension, tunnel entry and exit, and night light changes, so a stable and effective detection method is needed.

[0057] Furthermore, in order to save cabinet space on the motor train and save energy consumption, the intelligent analysis host uses an edge embedded device with limited computing power. Since the video monitoring system reaches 25 frames per second, the spark occurs only for a short time, so the intelligent analysis host needs to complete the spark detection task under the condition of extremely limited computing power while completing other identification tasks.

[0058] Therefore, the present patent proposes a pantograph-catenary system electric spark real-time detection method based on a lightweight neural network and time series analysis. The method detects sparks frame by frame through image processing and assists in screening and judgment by a lightweight neural network, realizes a detection speed of 25 frames per second under the condition of extremely small computing power consumption, and can effectively exclude strong light interference. The specific detection process is described through the following embodiments.

[0059] Please refer to Figure 2 which shows a flowchart of the pantograph-catenary system electric spark real-time detection method provided by an example embodiment of the present application, and the method comprises:

[0060] Step 201: acquiring pantograph monitoring images.

[0061] In a possible implementation, step 201 includes the following contents.

[0062] Content one: acquiring real-time video streams through the RTSP protocol;

[0063] Content two: converting the video streams into image sequences in RGB format, the image sequences being pantograph monitoring images, and storing the pantograph monitoring images in a queue for subsequent processing.

[0064] First, the rtsp address provided by the pantograph monitoring video server is read to acquire the pantograph monitoring video stream, the video stream is converted into an image matrix sequence in RGB format, and the images are stored in a first-in-first-out queue. The current frame of the pantograph monitoring image is acquired from the queue in sequence, as shown in Figure 3 .

[0065] Step 202, locating the pantograph head region image from the pantograph monitoring image.

[0066] In one possible implementation, feature extraction is performed based on the pantograph monitoring image through a target detection algorithm in a convolutional neural network, and the pantograph head region image is located.

[0067] In one example, a lightweight convolutional network target detection technology is used to accurately and efficiently locate the key region of the pantograph head. This method processes the input image through a series of steps, including scaling, feature extraction, feature fusion, prediction, and post-processing, and finally outputs the target box with the highest probability of the pantograph head and its specific coordinates. The detailed process is as follows:

[0068] 1. Data preparation

[0069] The training data mainly comes from the pantograph images collected in actual railway scenes, covering pantograph images under different lighting conditions (such as sunny, cloudy, and night), different angles (such as front, side, and oblique angles), and different environments (such as inside tunnels and on bridges). At the same time, in order to increase the diversity and robustness of the data, the original collected images are also subjected to data enhancement processing, including random cropping, rotation, flipping, brightness adjustment, etc.

[0070] Using professional image annotation tools, the pantograph head region in the collected pantograph image is manually annotated, and the annotation information includes the rectangular frame coordinates (x, y, w, h) of the pantograph head region and the corresponding class label (pantograph head).

[0071] 2. Image preprocessing

[0072] The pixel resolution of the current frame image is scaled to 352x352 to ensure uniform input image size for subsequent processing in a fixed network structure. The scaling operation uses a bilinear interpolation method to minimize the loss of image information.

[0073] 3. Feature extraction (Backbone)

[0074] ShuffleNetV2 structure is used as the backbone network for feature extraction. This structure uses channel shuffling and depth separable convolution operations to reduce computational complexity while maintaining high feature extraction capability. ShuffleNetV2 is divided into multiple stages (Stage), each containing multiple ShuffleV2Block. The output feature maps (C2, C3) of Stage3 and Stage4 are 22x22 and 11x11 in size, respectively, as the final feature input.

[0075] Specifically, as Figure 4As shown, ShuffleNetV2 is divided into several stages. The initial convolutional layer mixes the input from three channels and performs batch normalization. After passing through an activation function, the input is fed into a max-pooling layer, where downsampling forms the initial feature map. Stages 2, 3, and 4 each consist of multiple ShuffleV2Blocks, with different numbers of repetitions: 4, 8, and 4, respectively. Each ShuffleV2Block contains two branches: a main branch and a projection branch. The main branch extracts features using 1×1 convolutions, 3×3 depthwise separable convolutions, and 1×1 convolutions. The projection branch downsamples the input feature map with a stride of 2. The first ShuffleV2Block has a stride of 2 for downsampling; the remaining ShuffleV2Blocks have a stride of 1 for feature refinement. To simplify the model data, only the feature maps C2 and C3 output from Stages 3 and 4 are used, with dimensions of 22×22 and 11×11, respectively.

[0076] 4. Feature Fusion (LightFPN)

[0077] Efficient feature fusion and extraction are achieved using a Lightweight Feature Pyramid Network (LightFPN) combined with Depthically Separable Convolutions (DWConv). The main function of LightFPN is to fuse feature maps of different scales to fully utilize feature information at different scales, thereby improving feature extraction efficiency. Specifically, LightFPN contains two 1×1 convolutional layers that perform channel compression on 11×11 and 22×22 feature maps respectively. The 11×11 feature map is then upsampled to a 22×22 size and concatenated with the 22×22 feature map. Finally, a series of depthwise separable convolutional layers further extract and fuse features from the concatenated feature map.

[0078] Depthwise Separable Convolution (DWConv): DWConv is an efficient convolution operation that breaks down traditional convolution operations into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs independent convolution operations on each channel, while pointwise convolution fuses features from different channels. By using DWConv, the computational cost and number of parameters in convolution operations can be significantly reduced, improving the model's operational efficiency.

[0079] 5. Activation function processing

[0080] The Sigmoid activation function is used to process the coordinates and confidence probabilities of the predicted bounding boxes, while the Softmax activation function is used to process the target category. The Sigmoid function maps values ​​to the (0,1) interval, making it suitable for handling probability values; the Softmax activation function converts the scores of multiple categories into a probability distribution, ensuring that the sum of the probabilities of all categories is 1. This activation function processing enhances the model's stability and facilitates subsequent prediction layer operations.

[0081] 6. Target Prediction (Prediction Layer)

[0082] Anchor Boxes are used to predict the target's class and coordinates. Anchor Boxes are a predefined set of rectangles with different scales and aspect ratios used to initialize the target's position and size. During training, the model predicts the target's offset and class probability based on the input feature map and Anchor Boxes.

[0083] In one example, for each grid cell in the 11×11 and 22×22 feature maps, three bounding boxes are predicted. Each bounding box contains five values: x (x-coordinate of the top-left corner), y (y-coordinate of the top-left corner), w (width), h (height), and confidence (target confidence probability). Each box corresponds to only one target class. AnchorBoxes are a predefined set of bounding boxes with different scales and aspect ratios used to initialize the target's position and size. During training, the model predicts the target's offset and class probability based on the input feature maps and AnchorBoxes.

[0084] 7. Post-processing

[0085] The target boxes output by the prediction layer are filtered and subjected to maxima suppression. A threshold is set to filter out target boxes with low confidence probabilities, retaining only those with confidence probabilities higher than the set value. Maximal suppression is used to eliminate overlapping target boxes, improving the final detection accuracy, and selecting the box with the highest confidence as the final prediction.

[0086] In one possible implementation, after the R-case, maxima suppression is applied to the retained target boxes to remove overlapping target boxes. The main idea of ​​maxima suppression is to sort the target boxes according to their confidence probabilities, then select the target box with the highest confidence probability as the final predicted box, and delete other target boxes whose overlap with the predicted box is higher than a set threshold. Through maxima suppression, multiple overlapping target boxes can be avoided from being detected simultaneously, improving the accuracy of the detection results. Finally, the target box with the highest probability of the pantograph head is selected, and its specific coordinates (x, y, w, h) are output.

[0087] 8. Model Training Evaluation

[0088] The dataset is divided into training, validation, and test sets. The training set is used to train the model, the validation set is used to tune the model's hyperparameters, and the test set is used to evaluate the model's final performance. To train the prediction layer, a multi-task loss function (including bounding box regression loss, object confidence loss, and class classification loss) is used. The bounding box regression loss measures the difference between the coordinates of the predicted box and the ground truth box, using the CIoU loss function; the object confidence loss measures the difference between the confidence probabilities of the predicted box and the ground truth box, using the Binary Cross Entropy loss function; and the class classification loss measures the difference between the class probabilities of the predicted box and the ground truth box, using the Softmax Cross Entropy loss function. By minimizing the multi-task loss function, the model can learn a more accurate object prediction capability.

[0089] During training, the model's performance is periodically evaluated on the validation set, and hyperparameters such as learning rate and batch size are adjusted based on the evaluation results. Finally, the model's performance is evaluated on the test set using metrics such as precision, recall, and mean average precision (mAP). Precision is the proportion of samples predicted as positive that were actually positive; recall is the proportion of samples that were actually positive that were predicted as positive; and mean average precision is the average of the average precision across different classes.

[0090] 9. Model Testing Application

[0091] Image data is obtained from the image queue in step 201, and the images are input into the pre-trained lightweight convolutional network object detection model. The model outputs the coordinate information of the pantograph head region, thereby obtaining the pantograph head region image. The final execution result is as follows: Figure 5As shown. Since the vertical vibration amplitude of the pantograph is relatively small during operation, and considering that the pantograph position will not change significantly within 5 frames at a frame rate of 25 frames / second, the pantograph position is detected every 5 frames to save computing resources and ensure real-time performance and accuracy.

[0092] Step 203: Calculate the lateral movement rate of the bright target in the pantograph head area image to determine whether to proceed to the spark detection step.

[0093] In one possible implementation, step 203 includes the following.

[0094] Content 1: Calculate the rate of change of lateral movement of the highlighted target in the pantograph head region image;

[0095] Content 2: If the rate of change of lateral movement is less than a set threshold, it is considered to be interference from strong light such as sunlight;

[0096] Content 3: If the rate of change of the lateral movement is greater than the set threshold, then the spark detection step continues.

[0097] The above content will be explained in detail below.

[0098] During high-speed train operation, spark video detection needs to cope with various extreme weather conditions and sudden changes in light. When dealing with scenarios such as rain, snow, fog, frost, or sudden changes in light caused by entering or exiting tunnels or bridges, appropriate scene recognition algorithms are used to avoid false alarms. This patent focuses on discussing false spark detection alarms caused by bright sunlight appearing in the bow area during the daytime, which is a relatively frequent occurrence among various types of interference.

[0099] When sunlight shines on the pantograph head area, the sunlight may be blocked by the pantograph's vibrating contact plate, producing a spark-like flickering phenomenon that can easily be misidentified as a spark. Because the lateral movement speed of the sun is relatively slow, its lateral change rate is small, while the relative movement of the pantograph contact plate and the contact wire is faster, resulting in a larger lateral change rate when sparks occur. To effectively eliminate sunlight interference, this method distinguishes between bright targets by calculating the lateral movement change rate of the bright targets. Specifically, a lightweight convolutional network target detection method is used to detect bright targets such as the sun in the image. If a bright target is detected, the coordinate information of the bounding box (e.g., ...) is output. Figure 6 As shown, the red rectangle represents the sun. By calculating the intersection-over-union (IoU) ratio between the highlighted target rectangle and the bow-head region rectangle, it can be determined whether sunlight affects the bow-head region. If the IoU ratio is greater than 0, it is considered that there is interference from a highlighted target in the bow-head region; if it is less than 0, it is considered that there is no interference.

[0100] like Figure 7The rate of change of the lateral position of the highlighted target in the bow-head region is shown. Using the upper left corner of the bow-head region as the origin of the X-axis coordinate, a first-in-first-out queue with a capacity of 100 is created to cyclically record the X-axis position of the highlighted target in 100 frames. If there is no highlighted target interference in the bow-head region, the X-axis position is recorded as 0. The coefficient of variation of this queue is calculated in real time. First, the mean of the queue is calculated. Given the standard deviation s, the coefficient of variation is... The coefficient of variation reflects the rate of change of the x-axis position of the bright target. If the rate of change is small, it is determined to be solar interference. If the rate of change is large, it is further determined in step 4 whether it is a spark.

[0101] This paper further elaborates on the refinement of the judgment criteria, especially its adaptability in real-world complex environments (such as weather and light changes).

[0102] (1) During high-speed train operation, spark video detection needs to cope with various extreme weather and light changes, especially under conditions such as rain, snow, fog, frost, or sudden changes in light caused by environmental changes such as tunnels and bridges. Through scene recognition algorithms, the system can effectively avoid false alarms caused by these interference sources. By introducing scene recognition algorithms, the system can accurately judge the difference between light interference and spark signals, reducing the false alarm rate and ensuring stability and high-precision detection capabilities under various operating environments. This patent mainly focuses on the problem of false spark recognition alarms that may occur in daytime scenes due to strong sunlight shining on the bow area, especially among various interference sources, where this problem is quite common.

[0103] (2) However, in practical applications, how to effectively distinguish between interference signals and real sparks using this algorithm, especially in complex environments, still requires further clarification and optimization of the judgment criteria. Therefore, the embodiments of this application provide...

[0104] The judgment criteria are as mentioned in step 204. When a spark appears, the bow head area will be bright and flashing, and the bright area will first increase and then decrease, that is, it flashes by. The rate of change of the X-axis position of sunlight is relatively small.

[0105] Step 204: Perform time series analysis on the bright area in the pantograph head region image to determine whether there is a suspected spark.

[0106] In one possible implementation, step 204 includes the following.

[0107] Content 1: Calculate the area of ​​the highlighted target region in the pantograph head region image and generate a time series of the highlighted area;

[0108] Content 2: Based on the fluctuation range of the area change in the time series of the bright area, determine whether it is a flashing feature of a spark. If the fluctuation range of the bright area is large, it is determined to be a suspected spark.

[0109] Specifically, when a spark occurs, the pantograph head area will exhibit a bright and flashing characteristic, and the bright area will initially increase and then decrease. This continuous change in the bright area can be used to detect the spark. The specific steps are as follows: Considering that general video surveillance operates at 25 frames per second, and according to the "Interim Technical Conditions for Video Monitoring Systems of EMU Pantographs TJ / CL409-2023," a spark lasting more than 700ms is considered a large spark, i.e., lasting 18 frames. Therefore, a first-in-first-out queue with a capacity of 50 is created to cyclically record the number of pixels in the bright area of ​​the pantograph head region in 50 frames of images. The bright area can be segmented using a fixed threshold; the number of pixels with a brightness greater than 250 is the bright area.

[0110] Define the sequence as S = {s1, s2, ..., s} n}, where n = 50, s i It is the i-th element in the sequence, representing the area in pixels. Let A be 1 / 20 of the area of ​​the bow-head region. thresh This serves as the amplitude threshold for the spark area; only sparks exceeding this threshold are considered sparks. For example... Figure 8 As shown, first identify the area wave peak value, iterate through i (1 < i < n), if s i ≥s i-1 And s i ≥s i+1 And s i >A thresh , then s i It is a spark wave peak. Simply put, within a continuous interval, the zero point of monotonically increasing and monotonically decreasing is the peak point.

[0111] Then, starting from i-1, search leftwards in the sequence until a position j (1≤j≤i-1) is found such that s j ≤A thresh or s j ≤s j-1 Position j is the monotonic starting point of the rising edge. Therefore, p s =max{j|1≤j≤i-1,s j >s j-1 If such a j does not exist, then p s =0.

[0112] Similarly, starting from i+1, search to the right of the sequence until a position k (i+1≤k≤n) is found such that s k ≤A thresh or s k ≤sk+1 The position k is the monotonic endpoint of the falling edge. Therefore, p e =min{k|i+1≤k≤n,s k >s k+1 If such a k does not exist, then p e =n, that is, p e =50.

[0113] In summary, based on the bright area sequence, the peak interval (p) during spark generation can be obtained. s p e When a spark flickers multiple times within a 50-frame (2-second) frame, multiple peak intervals will exist within the same bright area sequence, such as... Figure 8 As shown, if the end point of the current peak and the start point of the subsequent peak are adjacent within 10 frames, then the two peak intervals are merged into one interval, which is called the suspected spark interval.

[0114] Step 205: Extract the largest image frame suspected of being a spark from the pantograph head region image.

[0115] Next, the sequence index of the largest bright area in the spark interval extracted from the 50-frame sequence is the brightest moment of the spark in that interval. The image of the brightest moment is extracted, which is the largest image frame suspected to be the spark.

[0116] Step 206: Input the largest image frame of the suspected spark into the neural network for classification and filtering, exclude non-spark images, and retain the image frames that are determined to be sparks.

[0117] In one possible implementation, step 206 includes the following.

[0118] Content 1: Extract the largest image frame suspected of being a spark from the pantograph head area image, wherein the largest image frame is the image frame that exhibits the largest bright area in the time series;

[0119] Content 2: The image frame with the largest suspected spark is input into a pre-trained neural network model. The neural network model extracts and classifies features from the image to identify whether spark features exist in the image.

[0120] Content 3: Based on the classification results output by the neural network, the largest image frame that is not a spark but a suspected spark is excluded, and only the image frames that are classified as sparks are retained.

[0121] In practical applications, some targets that exhibit bright flashing and a high rate of position change may not actually be sparks, for example... Figure 9As shown, raindrops shimmer under the illumination of the supplementary lighting, which can easily be misidentified as sparks. Therefore, further filtering is required. Thus, this step uses a lightweight backbone model to classify the frames with the largest suspected sparks, excluding non-spark images. The specific steps are described below:

[0122] 1. Data Preparation

[0123] We collected continuous pantograph videos from actual railway scenarios, covering pantograph video scenes under different lighting conditions (e.g., sunny, cloudy, night), different angles (e.g., front, side, oblique angle), and different environments (e.g., inside tunnels, on bridges). The videos were segmented into continuous frames, and the frame image with the largest suspected spark was extracted as the dataset. Data augmentation was performed (e.g., random cropping, rotation, flipping, brightness adjustment) to increase data diversity and robustness.

[0124] 2. Image preprocessing

[0125] The image resolution was scaled down to 352×352 to unify the input size and reduce the loss of image information.

[0126] 3. Model Structure

[0127] like Figure 10 As shown, the lightweight backbone model ShuffleNetV2 is used as the feature extraction layer to output the C3 feature map. The feature map is then pooled through a global average pooling layer and finally connected to a fully connected layer for classification. In the output layer, the Softmax activation function is used to process the classification results and output the predicted probabilities (confidence scores) of spark and non-spark binary classification.

[0128] 4. Model Training Evaluation

[0129] The dataset is divided into training, validation, and test sets. The training set is used to train the model, the validation set is used to tune the model's hyperparameters, and the test set is used to evaluate the model's final performance. A class classification loss function, Softmax Cross Entropy, is used to measure the difference between the probability of the input image's class and the true class. By minimizing the multi-task loss function, the model can learn more accurate classification prediction capabilities. During training, the model's performance is periodically evaluated on the validation set, and hyperparameters such as learning rate and batch size are adjusted based on the evaluation results. Finally, the model's performance is evaluated on the test set using metrics such as precision, recall, and mean average precision (mAP). Precision is the proportion of samples predicted as positive that were actually positive; recall is the proportion of samples that were actually positive that were predicted as positive; and mean average precision is the average of the average precision across different classes.

[0130] 5. Model Testing Application

[0131] During the formal detection process, images of suspected maximum spark frames are extracted and input into the trained lightweight convolutional network classification model. The model outputs the binary classification probability of the suspected maximum spark frame (confidence score between spark frame and non-spark frame), and the one with the highest probability is the classification prediction result. To improve the model's compatibility in various scenarios, the model is initially applied to the device. Images that the system classifies as spark frames are further collected, and images that are misclassified as spark frames are then fed into a new round of model training to optimize the classification model and improve classification accuracy.

[0132] In summary, this embodiment achieves automated detection of the pantograph status, particularly accurate identification of transient phenomena such as sparks and arcing, improving detection efficiency and accuracy. Specifically, it can capture and warn of abnormal situations such as sparks and arcing in real time, effectively shortening fault response time and ensuring the safe operation of the EMU. To address strong light interference in complex environments such as sunlight, bridges, tunnels, trackside lighting, and high-brightness light at night, a stable and effective detection algorithm was designed, enabling the system to adapt to various complex and changing operating environments and ensuring detection accuracy and reliability. Through algorithm optimization, low-power operation is achieved, allowing application to edge-embedded devices, reducing energy consumption and saving cabinet space on the EMU. Therefore, the use of advanced image processing and recognition algorithms effectively solves the problems of low efficiency in manually viewing pantograph status videos and susceptibility to interference in spark and arcing detection, significantly improving the safety level and overall operating efficiency of the EMU.

[0133] In addition, this application also provides a real-time electric spark detection device for a pantograph-catenary system, the device comprising:

[0134] Image acquisition module, used to acquire pantograph monitoring images;

[0135] The image positioning module is used to locate the pantograph head area image from the pantograph monitoring image;

[0136] The detection and judgment module is used to calculate the lateral movement rate of the bright target in the pantograph head area image and determine whether to enter the spark detection step.

[0137] The spark detection module is used to perform time series analysis on the bright area in the pantograph head region image to determine whether a suspected spark has occurred.

[0138] The image extraction module is used to extract the largest image frame suspected of being a spark from the image of the pantograph head area;

[0139] The image filtering module is used to input the largest image frame of the suspected spark into the neural network for classification and filtering, excluding non-spark images and retaining the image frames that are determined to be sparks.

[0140] This application also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the real-time detection method for electric sparks in a pantograph-catenary system as provided in the above embodiments.

[0141] Optionally, the computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. The random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0142] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0143] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0144] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for real-time detection of electric sparks in a pantograph-catenary system, characterized in that, The method includes: Acquire pantograph monitoring images; Locating the pantograph head region image from the pantograph monitoring image; the process of locating the pantograph head region image from the pantograph monitoring image includes: extracting features from the pantograph monitoring image and locating the pantograph head region image using a target detection algorithm in a convolutional neural network, wherein the ShuffleNetV2 structure is used as the backbone network and a Lightweight Feature Pyramid Network (LightFPN) is used for the feature extraction. The lateral movement rate of the bright targets in the pantograph head area image is calculated to determine whether to proceed to the spark detection step. Time series analysis is performed on the highlighted area in the pantograph head region image to determine whether any suspected sparks have appeared; Extract the largest image frame suspected of sparking from the pantograph head region image; The largest image frame suspected of being a spark is input into a neural network for classification and filtering, excluding non-spark images and retaining the image frames identified as sparks.

2. The method according to claim 1, characterized in that, The acquisition of pantograph monitoring images includes: Acquire real-time video streams via the RTSP protocol; The video stream is converted into an RGB format image sequence, which is a pantograph monitoring image and stored in a queue for subsequent processing.

3. The method according to claim 1, characterized in that, The step of calculating the lateral movement rate of the bright target in the pantograph head region image and determining whether to proceed to the spark detection step includes: Calculate the rate of change of lateral movement of the highlighted target in the pantograph head region image; If the rate of change of lateral movement is less than a set threshold, it is considered to be interference from strong light such as sunlight; If the rate of change of lateral movement is greater than the set threshold, the spark detection step continues.

4. The method according to claim 1, characterized in that, The step of performing time-series analysis on the highlighted area in the pantograph head region image to determine whether a suspected spark has occurred includes: Calculate the area of ​​the highlighted target region in the pantograph head region image and generate a time series of the highlighted area; The fluctuation range of the area change in the time series of the bright area is used to determine whether it is a flashing feature of a spark. If the fluctuation range of the bright area is large, it is determined to be a suspected spark.

5. The method according to claim 1, characterized in that, The step of inputting the largest image frame of the suspected spark into a neural network for classification and filtering includes: Extract the largest image frame suspected of sparks from the pantograph head area image, wherein the largest image frame is the image frame that exhibits the largest bright area in the time series; The image frame with the largest suspected spark is input into a pre-trained neural network model, which extracts and classifies features from the image to identify whether spark features exist in the image. Based on the classification results output by the neural network, the largest image frame that is not a spark but is suspected to be a spark is excluded, and only the image frames that are classified as sparks are retained.

6. A real-time electric spark detection device for a pantograph-catenary system, characterized in that, The device includes: Image acquisition module, used to acquire pantograph monitoring images; The image positioning module is used to locate the pantograph head area image from the pantograph monitoring image; The detection and judgment module is used to calculate the lateral movement rate of the bright target in the pantograph head area image and determine whether to enter the spark detection step. The spark detection module is used to perform time series analysis on the bright area in the pantograph head region image to determine whether a suspected spark has occurred. The image extraction module is used to extract the largest image frame suspected of being a spark from the image of the pantograph head area; The image filtering module is used to input the largest image frame of the suspected spark into the neural network for classification and filtering, excluding non-spark images and retaining the image frames that are determined to be sparks.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the real-time electric spark detection method for a pantograph-catenary system as described in any one of claims 1 to 5.

Images