Defect identification method for catenary dropper structure abnormity

Abstract

The invention discloses a defect identification method for structural abnormity of a catenary dropper, and relates to the field of dropper defect identification. According to the method, the structural change of the dropper in the standard state is judged by using the spatial transformation network STN and the unsupervised feature discrimination network SA-AnnoNet. The method has the characteristics of high efficiency and stability, is suitable for dropper images obtained in different acquisition states, and can complete the establishment of the defect identification model without any abnormal dropper sample.

Description

Technical Field

[0001] The invention relates to the field of hanger string defect identification, and in particular to a method for identifying abnormal defects in a contact network hanger string structure. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] In electrified railways, droppers serve as important load-bearing structures, suspending contact wires and load-bearing cables. This allows the contact network to limit its elasticity and relaxation within a certain reasonable range without adding supports, thereby ensuring that the train pantograph can smoothly receive current from the contact network.

[0004] According to the standard specifications for electrified railway construction, the hanging string should ensure:

[0005] 1. The body shall not have loose strands or broken strands, the wire clamp shall not have cracks or deformation, and the fastening bolts shall not be loose;

[0006] 2. The suspension string should be free of loose or broken strands and in good stress state. It should be vertical to the line at any temperature and the torsion deviation should not exceed 20mm;

[0007] 3. The crimping surfaces of the two crimping tubes should be in the same plane, one on this side and two on the current-carrying ring side;

[0008] 4. The current-carrying ring of the contact wire suspension string should be in the same direction as the train's forward direction, while the load-bearing wire should be in the opposite direction. The tail of the current-carrying ring should be fixed on the tail side of the suspension string clamp bolt, with the head raised at 45°;

[0009] 5. The distance between the heart-shaped ring and the suspension string and the crimping tube should not exceed 5mm;

[0010] 6. The threaded clip has no cracks or burns, and the installation direction is opposite to the current-carrying ring, with 5mm exposed;

[0011] 7. The maximum height difference between two adjacent suspension strings shall not exceed 10mm. The height difference between the first suspension string in the span and the adjacent elastic suspension string must be less than 5mm. The height difference between the elastic suspension string and the positioning wire clamp shall be 0.

[0012] 8. The long limb of the stop washer locks the body, and the short limb locks the nut;

[0013] 9. When the hanging string is less than 60mm away from other charged objects, it should be adjusted;

[0014] It can be seen that the suspension string has strict construction standards and daily maintenance requirements. In actual working conditions, under the influence of various complex external conditions, the structure of the suspension string may change, which will have a negative impact on the stable and safe operation of the contact network and electric locomotives. In order to promptly check the occurrence of the above situation, manual inspection or manual analysis of the contact network suspension state image acquisition data is often used in reality. Summary of the invention

[0015] The purpose of the present invention is: due to the large number of suspension strings and the long mileage of the line, in order to improve the effectiveness of the analysis, the present invention uses deep learning image processing knowledge and theory to design a set of computer algorithms for automatically analyzing the appearance of suspension strings, which can achieve rapid, stable and automatic identification of significant structural defects such as suspension strings not being stressed, suspension strings current-carrying rings being missing, and suspension strings current-carrying rings being out of position, greatly reducing labor costs and improving the timeliness of defect elimination, thereby solving the effectiveness and stability problems caused by the need to manually perform structural analysis on a large number of contact network suspension strings in the process of defect analysis of contact network images collected by the contact network suspension status detection and monitoring system (referred to as 4C system).

[0016] The technical solution of the present invention is as follows:

[0017] A method for identifying defects of abnormal catenary suspension string structure, comprising:

[0018] Step S1: cropping the hanging string area from the 4C image acquired by the 4C system to obtain a hanging string image;

[0019] Step S2: transforming the hanging string image based on the space transformation network STN so that the transformed hanging string image has a consistent standard state;

[0020] Step S3: Using the unsupervised feature discrimination network SA-AnoNet, the structure of the standardized hanging string image is extracted, and the features are compared with those of the normal hanging string image without structural problems, so as to realize defect identification.

[0021] Furthermore, a deep learning target detection model is used to crop the hanging string area from the 4C image acquired by the 4C system to obtain a hanging string image.

[0022] Furthermore, the step S1 comprises:

[0023] Step S11: collecting the original 4C image containing the suspension string, and using annotation software to respectively annotate the suspension string load-bearing cable clamp and the suspension string contact wire clamp in the image;

[0024] Step S12: using the annotated 4C images to train a deep learning target detection model, the trained deep learning target detection model can detect the suspension string structure in various types of 4C images, and crop the corresponding suspension string area to obtain a suspension string image;

[0025] Step S13: Detect the 4C image acquired by the 4C system through a deep learning target detection model, and crop the hanging string area according to the detection result.

[0026] Furthermore, the step S13 includes:

[0027] If no load-bearing cable clamp or contact wire clamp is detected, the detection process is exited;

[0028] If only one load-bearing cable clamp and one contact wire clamp are detected, the minimum circumscribed rectangle method is used to obtain the complete suspension string area;

[0029] If there is more than one load-bearing cable clamp or contact wire clamp, a weighted bipartite graph minimum matching algorithm is used to match pairs of dropper string load-bearing cable clamps and contact wire clamps, and cut them into separate dropper string areas.

[0030] Furthermore, the consistent standard state means that the hanging string images processed by the space transformation network STN have similar apparent features, including: the hanging string inclination direction, inclination degree, and the orientation of the current-carrying ring are similar.

[0031] Furthermore, the space transformation network STN comprises:

[0032] Positioning networks, mesh generators, and samplers;

[0033] The localization network is a convolutional neural network. The output of the localization network is an affine transformation matrix that describes the spatial transformation of the input image.

[0034] The grid generator applies an affine transformation matrix to the input image, generating a grid that can be used to sample the input image;

[0035] The sampler samples the input image using the generated grid, producing the output image.

[0036] Furthermore, the unsupervised feature discrimination network SA-AnoNet comprises:

[0037] Encoder G E (x) and decoder G D (z) constitutes the generator G, the discriminator D and the reconstructed image feature extractor E(x ′ ); the encoder G E (x) and decoder G D(z) are connected by a Skip-Connection structure; the encoder G E The spatial attention mechanism is introduced in (x).

[0038] Furthermore, after the standardized hanging string image is input into the unsupervised feature discrimination network SA-AnoNet, the positioning difference value is calculated, and based on the positioning difference value, it is judged whether the input hanging string image is an abnormal image.

[0039] Furthermore, the positioning difference value is calculated by the following formula:

[0040] Diff(x)=‖G E (x)-E(G(x))‖1

[0041] in:

[0042] Diff(x) represents the positioning difference value;

[0043] Setting Thresholds when , the input image can be considered as an abnormal image.

[0044] Furthermore, the formula of the spatial attention mechanism is as follows:

[0045]

[0046] In encoder G E In (x), for each feature map F of size C×H×W ′ , the channel dimension uses maximum pooling and average pooling to obtain two two-dimensional features, and splices them together in the channel dimension, that is, 2×H×W features, and obtains the weight coefficient M through a 7×7 convolution and Sigmoid activation function s , and finally through M s ×F ′ The scaled new features can be obtained.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention uses the space transformation network STN and the unsupervised feature discrimination network SA-AnoNet to determine the structural changes of the suspension string under the standard state. The method is efficient and stable, and is applicable to the suspension string images obtained under different acquisition states, and does not require any abnormal suspension string samples to complete the establishment of the defect recognition model.

[0049] 2. The present invention utilizes the space transformation network STN to unify the apparent features of the suspension strings to be detected (relatively uniform orientation, size, etc.), thereby improving the accuracy of feature extraction in the defect identification stage.

[0050] 3. In constructing the feature discrimination model, the present invention adopts a mode that does not require defective samples. The high-dimensional feature space distribution is constructed only through the features of samples with normal structures. The features of samples with abnormal structures are outliers in the high-dimensional space and do not conform to the above spatial distribution, so abnormal distinction can be made.

[0051] 4. In the process of feature extraction, in order to further focus its high-dimensional features on the suspension string structure and reduce the impact of noise caused by changes in non-structural factors such as illumination and background on the features, the present invention innovatively introduces a spatial attention mechanism into the feature extraction network to improve the robustness and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a defect identification method of abnormal structure of overhead contact network suspension string;

[0053] Figure 2 is the original 4C image;

[0054] Figure 3 It is the annotated 4C image;

[0055] Figure 4 is the final hanging string area;

[0056] Figure 5 The structure of the spatial transformer network STN;

[0057] Figure 6 It is the hanging string image before the spatial transformation network STN transformation;

[0058] Figure 7 It is the hanging string image after the transformation by the space transformation network STN;

[0059] Figure 8 This is the structural diagram of Ganomaly;

[0060] Fig. 9 This is the structure diagram of the AnoNet network;

[0061] Fig.10 This is the structure of the spatial attention mechanism. DETAILED DESCRIPTION

[0062] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0063] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0064] Embodiment 1

[0065] See also Figure 1 A method for identifying abnormal defects in a catenary suspension string structure specifically comprises the following steps:

[0066] Step S1: cropping the hanging string area from the 4C image acquired by the 4C system to obtain a hanging string image;

[0067] Step S2: Based on the space transformation network STN, the hanging string image is transformed so that the transformed hanging string image has a consistent standard state; that is, the hanging string image processed by the space transformation network STN has similar appearance characteristics, including: the hanging string tilt direction, tilt degree, and current-carrying ring orientation are similar;

[0068] Step S3: Using the unsupervised feature discrimination network SA-AnoNet, the structure of the standardized hanging string image is extracted, and the features are compared with those of the normal hanging string image without structural problems, so as to realize defect identification.

[0069] In this embodiment, specifically, in order to obtain the precise position of the hanging string and the local sub-image in the 4C image, a deep learning target detection model is used to crop the hanging string area from the 4C image acquired by the 4C system to obtain a hanging string image.

[0070] In this embodiment, specifically, step S1 includes:

[0071] Step S11: Collect a large number of original 4C images containing hanging strings (such as Figure 2 Use annotation software to annotate the suspension string load-bearing cable clamp and the suspension string contact wire clamp in the image (as shown in Figure 3 shown);

[0072] Step S12: using the annotated 4C images to train a deep learning target detection model, the trained deep learning target detection model can detect the suspension string structure in various types of 4C images, and crop the corresponding suspension string area to obtain a suspension string image;

[0073] Step S13: The 4C image acquired by the 4C system is detected by the deep learning target detection model, and the hanging string area is cut out according to the detection result (the final hanging string area is as shown in FIG. Figure 4 shown).

[0074] In this embodiment, specifically, step S13 includes:

[0075] If no load-bearing cable clamp or contact wire clamp is detected, the detection process is exited;

[0076] If only one load-bearing cable clamp and one contact wire clamp are detected, the minimum circumscribed rectangle method is used to obtain the complete suspension string area;

[0077] If there is more than one load-bearing cable clamp or contact wire clamp, a weighted bipartite graph minimum matching algorithm (KM) is used to match pairs of dropper string load-bearing cable clamps and contact wire clamps and cut them into separate dropper string regions.

[0078] It should be noted that since the 4C image data is obtained by using a high-speed inspection vehicle to take pictures of the contact network, in this process, the inspection vehicle itself (speed, vibration, tilt, etc.) and the difference in the installation of the suspension string itself caused by the change in the height of the contact network will cause the shape of the suspension string to change in the 4C image, mainly manifested in different degrees of inclination of the suspension string and inconsistent directions of the current-carrying rings of the suspension string. In order to ensure the stability of the subsequent feature analysis process, the suspension string image needs to be preprocessed to improve the uniformity of the suspension string image.

[0079] In this embodiment, a spatial transformation network STN is used for implementation; it should be noted that STN is a neural network structure that can adaptively perform spatial transformation on the input image, thereby improving the accuracy and robustness of the model. The STN network consists of three main components: a localization network, a grid generator, and a sampler.

[0080] The localization network is a convolutional neural network, whose output is an affine transformation matrix that describes the spatial transformation of the input image; the grid generator applies the affine transformation matrix to the input image to generate a grid that can be used to sample the input image; the sampler uses the generated grid to sample the input image to produce the output image. The structure of the spatial transformer network STN is as follows Figure 5 shown.

[0081] The main function of the localization network is to learn an affine transformation matrix Used to describe the spatial transformation of the input image. θ is initialized to the identity transformation matrix at the beginning of network training. θ is continuously modified through network iteration and loss function to finally achieve transformations such as rotation, scaling, translation and distortion.

[0082] The grid generator applies the affine transformation matrix to the input image to generate a grid that can be used to sample the input image. The output of the grid generator is a two-dimensional grid where each point represents a pixel position in the output image. The grid generator can generate a corresponding grid based on the spatial transformation of the input image, thereby realizing the spatial transformation of the input image. The formula of the grid generator is as follows:

[0083]

[0084] here represents the coordinate point of the input image s, Represents the most punctuated point of the output image t, where the coordinates of s and t are normalized to -1 to 1.

[0085] After the spatial transformation is completed, a sampling method (i.e., a sampler) is needed to calculate the pixel values ​​corresponding to the transformed coordinate points. The calculation based on bilinear interpolation is used here:

[0086]

[0087] is the gray value of a point on the cth channel of the output feature map, is the grayscale of the cth channel point (n,m) on the input feature map. or When it is greater than 1, the corresponding max() value is 0, that is, only The grayscale values ​​of the four surrounding points determine the grayscale of the target pixel. and The smaller it is, the greater the impact and the greater the weight.

[0088] The image of the hanging string before the spatial transformation network STN transformation is as follows Figure 6 As shown, the image of the hanging string after the spatial transformation network STN transformation is as follows Figure 7 shown.

[0089] It can be seen that the images of the hanging strings processed by the space transformation network STN have similar appearance characteristics (the inclination direction, inclination degree, and direction of the current-carrying ring of the hanging strings are similar, etc.).

[0090] After standardizing the image of the suspension string, it is necessary to determine whether there is a structural problem with the suspension string. The traditional method is to use a supervised classification model for identification, which has the following two disadvantages:

[0091] 1. A large number of defect samples are required to build a classification model with generalization availability, which is often not feasible in practice.

[0092] 2. A model needs to be established for each specific defect type, and unknown defect types cannot be identified.

[0093] In order to solve the above drawbacks, this embodiment first introduces an anomaly detection network with a Skip-Connection structure, referred to as AnoNet.

[0094] AnoNet is modified from Ganomaly network. The structure of Ganomaly is as follows Figure 8 As shown:

[0095] The Ganomaly network structure consists of three parts: encoder G E (x) and decoder G D (z) constitutes the generator G, the discriminator D and the reconstructed image feature extractor E(x ′ ), E(x ′ ) and G E (x) has the same structure. The input image x passes through G E (x) obtains the high-dimensional space feature vector z, z passes through G D (z) Get the reconstructed image E(x ′ ) Extract the feature z of the reconstructed image ′ The discriminator D judges the original input x as true and reconstructs The judgment is false. Thus, the codec in the generator is continuously optimized so that the reconstructed image is indistinguishable from the original image.

[0096] Then there is an obvious problem in directly using the Ganomaly network: due to the complex and diverse collection environment of the hanging string image, a more complex network is required. E (x) and decoder G D(z)) can improve the accuracy of network feature extraction, which usually causes the network gradient to diffuse and fail to converge, resulting in network unavailability. To solve this problem, this embodiment uses the Skip-Connection structure to connect the encoder G E (x) and decoder G D (z), the obtained network is called AnoNet, and its network structure is as follows Fig. 9 shown.

[0097] For image segmentation tasks or tasks that focus on spatial domain information, the pooling layer in the encoder of the traditional network has greatly reduced the size of the feature map, which is not conducive to image reconstruction tasks. Skip-Connection allows the network to fuse the feature maps of the corresponding positions of the encoder on the channel during each level of upsampling. By fusing the underlying features with the high-level features, the network can retain more high-resolution detail information contained in the high-level feature maps, thereby improving the reconstruction accuracy. While the reconstruction of adjacent layer features and cross-layer feature reconstruction enhances the transmission of feature information between layers, it further utilizes the rich detail information in the high-level convolutional feature layer, thereby maximizing the utilization of feature information in each layer of the network.

[0098] The AnoNet network is trained with standardized hanging string samples without structural anomalies. When a standardized hanging string image with structural anomalies is input, the encoder and decoder in the model will not be applicable to abnormal samples. The decoder will add features that only normal samples have to the reconstructed image of abnormal samples. This results in the high-dimensional space feature vector z after encoding being different from There is a huge difference. The positioning difference value is:

[0099] Diff(x)=‖G E (x)-E(G(x))‖1

[0100] Setting Thresholds when , the input image can be considered as an abnormal image.

[0101] Secondly, in order to further enable the AnoNet network to focus on the feature extraction of the suspension string structure information and reduce the impact of various objective conditions such as imaging background, illumination changes, vehicle body movement, lens blur, etc. on the feature extraction of the suspension string image, this embodiment further introduces a spatial attention mechanism into the network structure to improve the robustness and stability of AnoNet feature extraction.

[0102] The structure of the spatial attention mechanism is as follows Fig.10 As shown, the formula is as follows:

[0103]

[0104] In AnoNet encoder G E In (x), for each feature map F of size C×H×W ′ , the channel dimension uses maximum pooling and average pooling to obtain two two-dimensional features, and splices them together in the channel dimension, that is, 2×H×W features, and obtains the weight coefficient M through a 7×7 convolution and Sigmoid activation function s , and finally through M s ×F ′ The scaled new features can be obtained. The spatial attention mechanism can effectively change the response of AnoNet network features to spatial positions and increase the feature weights of important areas.

[0105] That is, the unsupervised feature discrimination network SA-AnoNet is the above-mentioned unsupervised anomaly detection model that integrates the spatial attention mechanism.

[0106] It should be noted that the unsupervised feature discrimination network SA-AnoNet can identify structural anomalies of suspension strings, including but not limited to the following types: missing current-carrying rings of suspension strings, suspension strings not under stress, missing suspension strings, and wrong installation direction of the current-carrying rings of suspension strings.

[0107] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

[0108] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A method for identifying abnormal defects in the structure of a contact network suspension string, characterized in that: include: Step S1: cropping the hanging string area from the 4C image acquired by the 4C system to obtain a hanging string image; Step S2: transforming the hanging string image based on the space transformation network STN so that the transformed hanging string image has a consistent standard state; Step S3: Using the unsupervised feature discrimination network SA-AnoNet, the structure of the standardized hanging string image is extracted, and the features are compared with those of the normal hanging string image without structural problems, so as to realize defect identification.

2. A method for identifying defects of abnormal structure of overhead wire suspension string according to claim 1, characterized in that: The deep learning target detection model is used to crop the hanging string area from the 4C image collected by the 4C system to obtain the hanging string image.

3. The method for identifying abnormal defects of a contact network suspension string structure according to claim 2, characterized in that: The step S1 comprises: Step S11: collecting the original 4C image containing the suspension string, and using annotation software to respectively annotate the suspension string load-bearing cable clamp and the suspension string contact wire clamp in the image; Step S12: using the annotated 4C images to train a deep learning target detection model, the trained deep learning target detection model can detect the suspension string structure in various types of 4C images, and crop the corresponding suspension string area to obtain a suspension string image; Step S13: Detect the 4C image acquired by the 4C system through a deep learning target detection model, and crop the hanging string area according to the detection result.

4. A method for identifying defects of abnormal structure of overhead wire suspension string according to claim 3, characterized in that: The step S13 comprises: If no load-bearing cable clamp or contact wire clamp is detected, the detection process is exited; If only one load-bearing cable clamp and one contact wire clamp are detected, the minimum circumscribed rectangle method is used to obtain the complete suspension string area; If there is more than one load-bearing cable clamp or contact wire clamp, a weighted bipartite graph minimum matching algorithm is used to match pairs of dropper string load-bearing cable clamps and contact wire clamps, and cut them into separate dropper string areas.

5. The method for identifying abnormal defects of overhead wire suspension structure according to claim 1, characterized in that: The consistent standard state means that the images of the suspension strings processed by the space transformation network (STN) have similar apparent features, including the inclination direction, degree of inclination, and orientation of the current-carrying ring of the suspension strings.

6. The method for identifying abnormal defects of overhead wire suspension structure according to claim 1, characterized in that: The space transformation network STN comprises: Positioning networks, mesh generators, and samplers; The localization network is a convolutional neural network. The output of the localization network is an affine transformation matrix that describes the spatial transformation of the input image. The grid generator applies an affine transformation matrix to the input image, generating a grid that can be used to sample the input image; The sampler samples the input image using the generated grid, producing the output image.

7. The method for identifying abnormal defects of overhead wire suspension structure according to claim 1, characterized in that: The unsupervised feature discrimination network SA-AnoNet comprises: Encoder G E (x) and decoder G D (z) constitutes the generator G, the discriminator D and the reconstructed image feature extractor E(x ′ ); the encoder G E (x) and decoder G D (z) are connected by a Skip-Connection structure; the encoder G E The spatial attention mechanism is introduced in (x).

8. A method for identifying defects of abnormal structure of overhead contact network suspension string according to claim 7, characterized in that: After the standardized hanging string image is input into the unsupervised feature discrimination network SA-AnoNet, the positioning difference value is calculated, and based on the positioning difference value, it is judged whether the input hanging string image is an abnormal image.

9. A method for identifying defects of abnormal structure of overhead wire suspension string according to claim 8, characterized in that: The positioning difference value is calculated by the following formula: Diff(x)=‖G E (x)-E(G(x))‖1 in: Diff(x) represents the positioning difference value; Setting Thresholds when , the input image can be considered as an abnormal image.

10. The method for identifying abnormal defects of overhead wire suspension string structure according to claim 7, characterized in that: The formula of the spatial attention mechanism is as follows: In encoder G E In (x), for each feature map F of size C×H×W ′ , the channel dimension uses maximum pooling and average pooling to obtain two two-dimensional features, and splices them together in the channel dimension, that is, 2×H×W features, and obtains the weight coefficient M through a 7×7 convolution and Sigmoid activation function s , and finally through M s ×F ′ The scaled new features can be obtained.

Images