A method for identifying defects of a carbon slide plate of a contact rail current collector shoe
By using deep learning algorithms to locate and identify defects in the carbon slide plate of the current collector shoe, the problem of human visual observation being affected by light was solved, and efficient and accurate defect identification was achieved.
Patent Information
- Application Number
- CN202210902425.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-29
- Publication Date
- 2026-02-17
- Estimated Expiration
- 2042-07-29
AI Technical Summary
In existing technologies, the human eye's observation of current collector shoe defects is greatly affected by changes in external lighting, leading to inaccurate identification results.
Deep learning algorithms, especially the FCOS deep learning network model, are used to locate the contact rail current collector shoe and carbon slide plate area and identify defects. The model includes a backbone network module, a feature regression module, and a detection module. The LabelImg tool is used to annotate the images to train the current collector shoe localization, carbon slide plate localization, and defect identification models.
It is unaffected by changes in external lighting, improving the accuracy and efficiency of defect identification, and can efficiently identify defects in the current collector shoe carbon slide plate.
Smart Images
Figure CN115439406B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of defect identification, in particular to a contact rail current collector shoe carbon slide plate defect identification method. BACKGROUND
[0002] The current collector shoe is an important part of the electric locomotive, which is a key component for introducing the current on the third rail into the train body to provide power, has a great influence on the safety and operation cost of the train, and the physical defects of the current collector shoe itself, such as missing nuts, detecting current collector shoe carbon slide plate cracks and current collector shoe carbon slide plate damage, have a direct negative impact on the introduction of electric current into the train and the safety of train operation. In the prior art, the defects of the current collector shoe are mainly observed by the naked eye, but this method is greatly disturbed by changes in external light, and the results obtained are not accurate. SUMMARY
[0003] In view of the above shortcomings in the prior art, the present application provides a contact rail current collector shoe carbon slide plate defect identification method to solve the problem of inaccurate results of observing the defects of the current collector shoe by the naked eye.
[0004] In order to achieve the above-mentioned application purpose, the technical scheme adopted by the present application is as follows: a contact rail current collector shoe carbon slide plate defect identification method, comprising the following steps:
[0005] S1, collecting an image of a contact rail current collector shoe to be identified;
[0006] S2, positioning the contact rail current collector shoe using a deep learning algorithm on the contact rail current collector shoe image;
[0007] S3, positioning the carbon slide plate region of the contact rail current collector shoe using a deep learning algorithm;
[0008] S4, identifying defects in the carbon slide plate region of the contact rail current collector shoe using a deep learning algorithm;
[0009] S5, outputting the defect identification result of the carbon slide plate region of the contact rail current collector shoe.
[0010] Further, the deep learning algorithm is completed through an FCOS deep learning network model, and the FCOS deep learning network model comprises a backbone network module, a feature regression module and a detection module.
[0011] Further, the backbone network module is composed of a BottleneckCSP, and a feature map is extracted from the contact rail current collector shoe image as the input of the feature regression module.
[0012] Further, the feature regression module adopts a PFN structure, and each pixel in the input feature map is calculated by regression to generate a feature pyramid.
[0013] Further, the detection module comprises a Center-ness and NMS algorithm, which is used to suppress the generation of low-quality Bounding Box.
[0014] Further, the LabelImg image labeling tool is used to sequentially label the collector shoe region, the carbon slide plate in the collector shoe region, and the defects in the carbon slide plate.
[0015] The collector shoe positioning model training set, the collector shoe carbon slide plate positioning model training set, and the carbon slide plate defect detection model training set are generated respectively.
[0016] The FCOS deep learning network model is trained using the collector shoe positioning model training set, the collector shoe carbon slide plate positioning model training set, and the carbon slide plate defect detection model training set respectively, to obtain the collector shoe positioning model, the collector shoe carbon slide plate positioning model, and the carbon slide plate defect recognition model.
[0017] Further, the LabelImg image labeling tool is used to sequentially label the collector shoe region, the carbon slide plate in the collector shoe region, and the defects in the carbon slide plate.
[0018] The LabelImg image labeling tool is used to label the collector shoe region in the collected collector shoe image.
[0019] The collector shoe positioning model is used to extract the collector shoe region from the collected collector shoe image, and the LabelImg image labeling tool is used to label the carbon slide plate in the collector shoe region.
[0020] The collector shoe carbon slide plate positioning model is used to extract the carbon slide plate region from the collected collector shoe image, and the LabelImg image labeling tool is used to label the defects in the carbon slide plate region.
[0021] Further, the positioning step of the contact rail collector shoe is:
[0022] S21, the trained collector shoe positioning model is used to perform convolution operation on the to-be-recognized contact rail collector shoe image to extract semantic information of the image.
[0023] S22, when the similarity of the extracted semantic information and the collector shoe semantic information saved in the collector shoe positioning model is not less than 95%, it indicates that the contact rail collector shoe image contains a collector shoe, and enters step S23, otherwise it indicates that the contact rail collector shoe image does not contain a collector shoe, and enters step S24.
[0024] S23, the collector shoe positioning model is used to perform deconvolution operation on the contact rail collector shoe image to obtain pixel position information of the collector shoe, and the pixel position information is output as the collector shoe positioning result of the contact rail collector shoe image.
[0025] S24, performing a collector shoe positioning operation on the next contact rail collector shoe image, and returning to step S21.
[0026] Further, the collector shoe carbon slide area positioning step is:
[0027] S31, performing a convolution operation on the collector shoe positioning result image using the trained collector shoe carbon slide positioning model to extract semantic information of the image;
[0028] S32, when the similarity between the extracted semantic information and the collector shoe carbon slide semantic information saved in the collector shoe carbon slide positioning model is not less than 95%, it is indicated that the collector shoe positioning result image contains the collector shoe carbon slide, and step S33 is entered, otherwise it is indicated that the collector shoe positioning result image does not contain the collector shoe carbon slide, and step S34 is entered;
[0029] S33, performing an inverse convolution operation on the collector shoe positioning result image using the collector shoe carbon slide positioning model to obtain pixel position information of the collector shoe carbon slide, and outputting the pixel position information as the collector shoe carbon slide positioning result of the collector shoe positioning result image;
[0030] S34, performing a collector shoe carbon slide positioning operation on the next collector shoe positioning result image, and returning to step S31.
[0031] Further, the collector shoe carbon slide area defect recognition step is:
[0032] S41, dividing the collector shoe carbon slide positioning result image into several blocks, and performing a convolution operation on each block area using the trained collector shoe carbon slide defect recognition model to extract semantic information of the image;
[0033] S42, comparing the extracted semantic information with the semantic information of various defects saved in the collector shoe carbon slide defect recognition model, when there is semantic information with the highest similarity and not less than 70%, it is indicated that a defect is detected and the semantic information of the defect is output, step S43 is entered, otherwise the semantic information is not the semantic information of the collector shoe defect, step S44 is entered;
[0034] S43, performing an inverse convolution operation on each block area using the collector shoe carbon slide defect recognition model to extract pixel position information of the collector shoe defect in the collector shoe carbon slide positioning result image, and outputting the pixel position information and the defect semantic information as the result;
[0035] S44, performing a collector shoe defect recognition operation on the next collector shoe carbon slide positioning result image, and returning to step S41.
[0036] The application has the beneficial effects that the application is the application of the deep learning network model in the contact rail trolley shoe slide plate defect identification, the application is not disturbed by external light changes, is only related to the contact rail trolley shoe slide plate area features in the current image, and has high recognition rate. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 The flowchart of the application is shown in the figure.
[0038] Figure 2 The overall framework diagram of the FCOS deep learning network model in the application is shown in the figure. DETAILED DESCRIPTION
[0039] The specific embodiments of the application are described below to facilitate the understanding of the application by those skilled in the art, but it should be clear that the application is not limited to the scope of the specific embodiments, and for those skilled in the art, as long as various changes are within the spirit and scope of the application defined and determined by the appended claims, all the application and creation utilizing the concept of the application are included in the protection.
[0040] As shown in the figure, a contact rail trolley shoe carbon slide plate defect identification method comprises the following steps: Figure 1
[0041] S1, collecting a contact rail trolley shoe image to be identified;
[0042] S2, using a deep learning algorithm to locate the contact rail trolley shoe on the contact rail trolley shoe image;
[0043] The positioning step of the contact rail trolley shoe is:
[0044] S21, using a trained trolley shoe positioning model to perform convolution operation on the contact rail trolley shoe image to be identified to extract semantic information of the image;
[0045] S22, when the similarity of the extracted semantic information and the semantic information of the trolley shoe saved in the trolley shoe positioning model is not less than 95%, it is indicated that the contact rail trolley shoe image contains a trolley shoe, and step S23 is entered, otherwise it is indicated that the contact rail trolley shoe image does not contain a trolley shoe, and step S24 is entered;
[0046] S23, using the trolley shoe positioning model to perform deconvolution operation on the contact rail trolley shoe image to obtain pixel position information of the trolley shoe, and outputting the pixel position information as the trolley shoe positioning result of the contact rail trolley shoe image;
[0047] S24, performing trolley shoe positioning operation on the next contact rail trolley shoe image, and returning to step S21.
[0048] The training step of the current collector shoe positioning model is:
[0049] The current collector shoe region in the current collector shoe image is labeled using the LabelImg image labeling tool to generate a current collector shoe positioning model training set, a total of 1000 images, and the training set is composed as follows:
[0050] Table 1 current collector shoe positioning model training set
[0051] With shoe Without shoe Proportion 90％ 10％
[0052] The current collector shoe positioning model is obtained by training the FCOS deep learning network model using the current collector shoe positioning model training set.
[0053] S3, using a deep learning algorithm to locate the carbon slide plate region of the contact rail current collector shoe;
[0054] The positioning step of the carbon slide plate region of the contact rail current collector shoe is:
[0055] S31, using the trained current collector shoe carbon slide plate positioning model to perform convolution operation on the current collector shoe positioning result image to extract the semantic information of the image;
[0056] S32, when the similarity of the extracted semantic information and the semantic information of the current collector shoe carbon slide plate saved in the current collector shoe carbon slide plate positioning model is not less than 95%, it is indicated that the current collector shoe positioning result image contains the current collector shoe carbon slide plate, and step S33 is entered, otherwise it is indicated that the current collector shoe positioning result image does not contain the current collector shoe carbon slide plate, and step S34 is entered;
[0057] S33, using the current collector shoe carbon slide plate positioning model to perform deconvolution operation on the current collector shoe positioning result image to obtain the pixel position information of the current collector shoe carbon slide plate, and outputting the pixel position information as the current collector shoe carbon slide plate positioning result of the current collector shoe positioning result image;
[0058] S34, performing current collector shoe carbon slide plate positioning operation on the next current collector shoe positioning result image, and returning to step S31.
[0059] The training step of the current collector shoe carbon slide plate positioning model is:
[0060] The current collector shoe region is extracted using the current collector shoe positioning model, and the carbon slide plate in the extracted current collector shoe region is labeled using the LabelImg image labeling tool to generate a current collector shoe carbon slide plate positioning model training set, a total of 2000 images, and the training set is composed as follows:
[0061] Table 2 carbon slide plate positioning model training set
[0062] With carbon slide plate Without carbon slide plate Proportion 85％ 15％
[0063] The FCOS deep learning network model is trained using the pantograph carbon slide plate positioning model training set to obtain a pantograph carbon slide plate positioning model.
[0064] S4, using a deep learning algorithm to identify defects in the pantograph carbon slide plate area of the contact rail;
[0065] The pantograph carbon slide plate area defect identification step is:
[0066] S41, the pantograph carbon slide plate positioning result image is segmented into several blocks, and a trained pantograph carbon slide plate defect identification model is used to perform convolution operation on each block area to extract semantic information of the image;
[0067] S42, compare the extracted semantic information with the semantic information of various defects saved in the pantograph carbon slide plate defect identification model, when there is semantic information with the highest similarity and not less than 70%, it means that a defect is detected and the semantic information of this type of defect is output, step S43 is entered, otherwise the semantic information represented is not the semantic information of the pantograph defect, step S44 is entered;
[0068] S43, using the pantograph carbon slide plate defect identification model to perform deconvolution operation on each block area to extract pixel position information of the pantograph defect in the pantograph carbon slide plate positioning result image, and the pixel position information and the defect semantic information are output as the result;
[0069] S44, pantograph defect identification operation is performed on the next pantograph carbon slide plate positioning result image, and step S41 is returned.
[0070] The pantograph carbon slide plate defect identification model training step is:
[0071] The pantograph carbon slide plate area is extracted using the pantograph carbon slide plate positioning model, and the LabelImg image labeling tool is used to label the defects in the extracted pantograph carbon slide plate area to generate a pantograph carbon slide plate defect identification model training set, a total of 2600 images, the training set is composed of the following table:
[0072] Table 3 pantograph carbon slide plate defect identification model training set
[0073] Nut loosening Nut falling off Carbon slide plate crack Carbon slide plate defect Proportion 20％ 10％ 35％ 35％
[0074] The FCOS deep learning network model is trained using the pantograph carbon slide plate defect identification model training set to obtain a pantograph carbon slide plate defect identification model.
[0075] S5, output the contact rail pantograph slide plate area defect identification result.
[0076] The FCOS is adopted in the deep learning network model in the application, and the FCOS has the advantages of faster detection rate and higher monitoring accuracy. Figure 2 The network model of the FCOS mainly comprises three parts, i.e., a backbone network module, a feature regression module and a detection module.
[0077] The backbone network module mainly comprises a BottleneckCSP, and the purpose is to extract a feature map from an input image as an input of a subsequent module. The BottleneckCSP module greatly reduces the calculation amount while enhancing the learning performance of the entire convolutional neural network. The BottleneckCSP is a relatively deep network, and the gradient value of back propagation between layers is enhanced by increasing the structure of the residual network, so that the gradient disappearance caused by deepening the network is avoided, and finer-grained features are extracted without worrying about network degradation.
[0078] The feature regression module adopts a PFN structure, and each pixel in the input feature map is calculated by regression to generate a feature pyramid as the basis for subsequent detection and classification. The FCOS algorithm added with the PFN has two advantages: (1) ensuring a high Best Possible Recall (BPR). In the feature map regression calculation, a larger step usually leads to a lower BPR, and the BPR is crucial for the detection task. The FCOS algorithm added with the PFN can still obtain a high BPR even if a larger step is used in the regression calculation, and the main reason is that the PFN detects different sizes of targets on different feature maps. (2) solving the ambiguity problem caused by the overlapping of Bounding Box (BB) during training. Different feature_map levels in the FPN need to regress BBs of different size ranges. For a feature level Pi, exp(six) with a trainable scalar si is used instead of the standard exp(x), where si is a trainable parameter for self-adapting the adjustment of different levels.
[0079] The detection module mainly comprises a Center-ness and a NMS algorithm. The FCOS introduces some low-quality BBs, i.e., BBs far away from the center of the target, and the Center-ness is used to well suppress the generation of these low-quality BBs. The Center-ness has the advantage of being relatively simple and does not need to introduce other hyperparameters. The position of the Center-ness is below the classification network branch of the Head network. For a given regression target of a position l (left upper horizontal coordinate), t (left upper vertical coordinate), r (right lower horizontal coordinate) and b (right lower vertical coordinate), the definition of the center-ness target is as follows:
[0080]
[0081] where the meaning of * is the quality of the current target corresponding position, that is, the coordinate point confidence.
[0082] The range is between [0, 1], and the BCE loss function is used for training. It can be used to reduce the weight of the bounding box far from the object center. It can be regarded as a soft threshold, which can be learned in the process of network training and does not need to be adjusted. In addition, due to the existence of NMS, these low-quality BBs can be filtered to a certain extent.
Claims
1. A method for identifying defects of a carbon slide of a contact rail current collector shoe, characterized in that, The method comprises the following steps: S1, collecting an image of a contact rail shoe to be identified; S2, positioning the contact rail shoe using a deep learning algorithm on the image of the contact rail shoe; S3, positioning the carbon slide area of the contact rail shoe using the deep learning algorithm; S4, identifying defects in the carbon slide area of the contact rail shoe using the deep learning algorithm; S5, outputting the carbon slide area defect identification result of the contact rail shoe; The defect identification step of the carbon slide area of the contact rail shoe comprises: S41, segmenting the carbon slide positioning result image of the contact rail shoe into several blocks, and performing convolution operation on each block area using the trained shoe carbon slide defect identification model to extract semantic information of the image; S42, comparing the extracted semantic information with the semantic information of various defects saved in the shoe carbon slide defect identification model, when there is semantic information with the highest similarity and not less than 70%, it indicates that a defect is detected and the semantic information of the defect is output, and step S43 is entered, otherwise the semantic information is not the semantic information of the shoe defect, and step S44 is entered; S43, performing deconvolution operation on each block area using the shoe carbon slide defect identification model to extract pixel position information of the shoe defect in the shoe carbon slide positioning result image, and outputting the pixel position information and the defect semantic information as the result; S44, performing shoe defect identification operation on the next shoe carbon slide positioning result image, and returning to step S41; The deep learning algorithm is completed through an FCOS deep learning network model, and the FCOS deep learning network model comprises a backbone network module, a feature regression module and a detection module.
2. The method according to claim 1, wherein The backbone network module is composed of a BottleneckCSP, and a feature map is extracted from the contact rail shoe image as the input of the feature regression module.
3. The method according to claim 1, wherein The feature regression module adopts a PFN structure, performs regression calculation on each pixel in the input feature map, and generates a feature pyramid.
4. The method according to claim 1, wherein The detection module comprises a Center-ness and a NMS algorithm, which are used to suppress the generation of low-quality BoundingBox.
5. The contact rail shoe carbon slide defect identification method according to claim 1, wherein a LabelImg image labeling tool is used to sequentially label the contact rail shoe region, the carbon slide region of the contact rail shoe, and the defects in the carbon slide region of the contact rail shoe; The shoe positioning model training set, the shoe carbon slide positioning model training set and the carbon slide defect detection model training set are generated respectively; The FCOS deep learning network model is trained using the shoe positioning model training set, the shoe carbon slide positioning model training set and the carbon slide defect detection model training set respectively, and the shoe positioning model, the shoe carbon slide positioning model and the carbon slide defect identification model are obtained.
6. The contact rail shoe carbon slide defect identification method according to claim 5, wherein The contact rail shoe region, the carbon slide plate region of the contact rail shoe, and defects in the carbon slide plate region of the contact rail shoe are sequentially labeled by using a LabelImg image labeling tool, specifically: The contact rail shoe region in the collected contact rail shoe image is labeled by using a LabelImg image labeling tool; The contact rail shoe region is cropped from the collected contact rail shoe image by using the contact rail shoe positioning model, and the carbon slide plate in the contact rail shoe region is labeled by using a LabelImg image labeling tool; The carbon slide plate region is cropped from the collected contact rail shoe image by using the contact rail shoe carbon slide plate positioning model, and defects in the carbon slide plate region are labeled by using a LabelImg image labeling tool.
7. The method according to claim 1, wherein The positioning step of the contact rail shoe is: S21, performing convolution operation on the contact rail shoe image to be recognized by using the trained contact rail shoe positioning model to extract semantic information of the image; S22, when the similarity of the extracted semantic information and the semantic information of the contact rail shoe saved in the contact rail shoe positioning model is not less than 95%, it is indicated that the contact rail shoe image contains a contact rail shoe, and step S23 is entered, otherwise it is indicated that the contact rail shoe image does not contain a contact rail shoe, and step S24 is entered; S23, performing deconvolution operation on the contact rail shoe image by using the contact rail shoe positioning model to obtain pixel position information of the contact rail shoe, and outputting the pixel position information as the contact rail shoe positioning result of the contact rail shoe image; S24, performing contact rail shoe positioning operation on the next contact rail shoe image, and returning to step S21.
8. The method according to claim 1, wherein The positioning step of the carbon slide plate region of the contact rail shoe is: S31, performing convolution operation on the contact rail shoe positioning result image by using the trained contact rail shoe carbon slide plate positioning model to extract semantic information of the image; S32, when the similarity of the extracted semantic information and the semantic information of the carbon slide plate saved in the contact rail shoe carbon slide plate positioning model is not less than 95%, it is indicated that the contact rail shoe positioning result image contains a carbon slide plate, and step S33 is entered, otherwise it is indicated that the contact rail shoe positioning result image does not contain a carbon slide plate, and step S34 is entered; S33, performing deconvolution operation on the contact rail shoe positioning result image by using the contact rail shoe carbon slide plate positioning model to obtain pixel position information of the carbon slide plate, and outputting the pixel position information as the carbon slide plate positioning result of the contact rail shoe positioning result image; S34, performing carbon slide plate positioning operation on the next contact rail shoe positioning result image, and returning to step S31.