Catenary dropper unsupervised defect detection method based on convolution automatic encoder

Abstract

The invention discloses a catenary dropper unsupervised defect detection method based on a convolution automatic encoder, and aims to solve the problems of low efficiency, high cost, high false detection rate and high omission ratio in a traditional dropper detection method based on manual inspection or simple image processing. The method comprises the following steps that firstly, image data of the catenary dropper is collected and preprocessed, and the preprocessing step comprises image graying, denoising and normalization processing so as to eliminate the influence of external factors and enhance the image quality; then, the preprocessed image is input into a convolution automatic encoder model, a convolution automatic encoder is composed of an encoder and a decoder, the encoder extracts local features of the image through convolution and pooling operation and carries out dimension reduction, potential space representation is generated, and the potential space representation is extracted; the decoder restores the potential spatial representation to an output image similar to the input image by a deconvolution or up-sampling operation. By calculating a reconstruction error between an input image and a reconstruction image, an abnormally large reconstruction error is identified so as to judge whether the dropper has defects or not.

Description

Technical Field

[0001] The invention belongs to the technical field of high-speed railway contact network, and in particular relates to an unsupervised defect detection method for contact network suspension strings based on convolutional autoencoder. Background Art

[0002] The catenary dropper is an important part of the railway catenary system. Its function is to connect the contact wire to the supporting structure to ensure the normal power supply of the train. With the rapid development of railway transportation, the operation safety of the catenary is particularly important. However, in the long-term use of the dropper, due to natural aging, mechanical stress, environmental conditions (such as wind, rain, etc.) and frequent passage of trains, it is easy to have defects such as breakage, looseness, and deformation. If these defects are not discovered and handled in time, they may cause serious operational failures and affect the safety and stable operation of the train. Traditional methods for detecting dropper defects usually rely on manual inspections or monitoring systems based on simple image processing. These methods have problems such as low efficiency, high cost, and insufficient reliability. Especially in large-scale railway networks, manual inspection is particularly difficult. In addition, the detection system based on simple image processing methods is difficult to adapt to complex environmental conditions (such as light changes, noise interference, etc.), resulting in low detection accuracy and high missed detection rate. Therefore, an intelligent automatic detection method is urgently needed to improve the detection efficiency and accuracy of the dropper defects of the catenary.

[0003] In recent years, with the development of deep learning technology, especially the widespread application of convolutional neural networks (CNN), image-based defect detection technology has made significant progress. As an unsupervised learning method in deep learning, convolutional autoencoder (CAE) has gradually become a popular choice in the field of image anomaly detection due to its excellent performance in feature extraction and image reconstruction. By learning the low-dimensional potential representation of normal images, convolutional autoencoders can effectively identify abnormal features in input images. Therefore, the use of convolutional autoencoders for unsupervised defect detection of contact wire suspension strings has great application prospects.

[0004] By building a model based on convolutional autoencoders, unsupervised learning can be performed without a large amount of labeled defect data, and key features can be extracted from normal cable images. During detection, the reconstruction error between the input image and the reconstructed image is calculated, and the area with a larger error may represent the defective area of ​​the cable. This method can automatically and quickly detect potential defects in the cable, reduce manual intervention, and improve the operation and maintenance efficiency and safety of the railway contact network. Summary of the invention

[0005] The present invention relates to an unsupervised defect detection method for contact network droppers based on convolutional autoencoder, aiming to solve the problems of low efficiency, high cost and high false detection and missed detection rate in traditional detection methods, especially the difficulties in detecting contact network droppers in large-scale railway networks. The method realizes automatic detection of potential defects in contact network dropper images by adopting a convolutional autoencoder (CAE) model combined with unsupervised learning technology.

[0006] To achieve the above object, the technical solution adopted by the present invention is: an unsupervised defect detection method for contact network suspension strings based on convolutional autoencoder, comprising the steps of:

[0007] The image data of the overhead wire suspension string is collected and preprocessed; the preprocessed image data is input into a convolutional autoencoder model, which includes an encoder part and a decoder part; the encoder part performs a convolution operation on the input image to extract local features of the image, and performs dimensionality reduction through a pooling operation to generate a latent space representation; the decoder part restores the latent space representation to an output image similar to the input image through a deconvolution or upsampling operation; the reconstruction error is calculated, and potential suspension string defects are identified according to the size of the reconstruction error, wherein an abnormally high reconstruction error indicates that the suspension string has defects; a defect detection result is generated based on the reconstruction error; the encoder performs feature extraction and dimensionality reduction on the input image through multiple convolutional layers and pooling layers to generate a low-dimensional latent space representation z=f(x), wherein x is the input image and z is the low-dimensional representation; the decoder gradually restores the latent space representation through a deconvolution or upsampling layer to generate a reconstructed image similar to the input image. in, is the reconstructed image.

[0008] The convolution operation formula of the encoder part is:

[0009] z l =σ(W l *x+b l )

[0010] Among them, z l is the output of the i-th layer, W l is the convolution kernel weight matrix, * represents the convolution operation, b l is the bias term and σ is the nonlinear activation function.

[0011] The deconvolution operation formula of the decoder part is:

[0012]

[0013] Among them, W l ′ and b l′ are the weight matrix and bias term of the deconvolution layer respectively.

[0014] Furthermore, the preprocessing includes image grayscale, denoising, and normalization processing to reduce the impact of external factors such as ambient light and noise on the image and enhance the feature information of the image.

[0015] Furthermore, the defect detection method uses the overhead wire hanging string image data in a normal state to perform unsupervised training on the convolutional autoencoder model. The training goal is to minimize the error between the input image and the reconstructed image, which can be defined by the mean square error (MSE):

[0016]

[0017] Among them, x i and They represent the i-th pixel value of the input image and the reconstructed image respectively, and n is the total number of image pixels.

[0018] Furthermore, after the model training is completed, the defect detection method inputs the catenary string image to be detected, and calculates the reconstruction error between the input image and the reconstructed image. Whether the string has defects is determined based on the size of the reconstruction error: when the reconstruction error is small (lower than a preset threshold), it indicates that the image is a normal string image; when the reconstruction error is large (exceeding the preset threshold), it indicates that there may be defects in the image, and further marking and analysis are performed.

[0019] Furthermore, for the detected defects, based on the spatial distribution of the image reconstruction error, the defective area is located and marked, and a visual detection report of the contact network hanger string defects is generated, providing image marking and analysis results of the defective area.

[0020] The beneficial effects of adopting this technical solution are:

[0021] The beneficial effects of adopting this technical solution are mainly reflected in the following aspects:

[0022] 1. Improve detection efficiency and accuracy: By adopting the convolutional autoencoder (CAE) model, this method can automatically extract complex local features from the catenary string image and accurately identify potential defects by calculating the image reconstruction error. Compared with the traditional method based on manual inspection or simple image processing, this method has faster detection speed and higher accuracy, and can reduce missed detection and false detection.

[0023] 2. Unsupervised learning reduces data requirements: This technology does not rely on a large amount of labeled training data, but uses unsupervised learning to train the model using only images of hanging strings in normal states. This greatly reduces the cost and complexity of data annotation and is suitable for actual scenarios where large-scale labeled data is lacking.

[0024] 3. Strong robustness and adaptability to complex environments: In complex railway environments, hanging string images are easily affected by external conditions (such as lighting, weather, etc.). Through image preprocessing (such as denoising and normalization) combined with the characteristics of the CAE model, this method has strong robustness to environmental changes and can maintain high detection performance under different conditions.

[0025] 4. Automated and visualized inspection results: This method automatically identifies and locates the defective area of ​​the suspension string by reconstructing the spatial distribution of the error, and generates a visual inspection report. The operation and maintenance personnel can intuitively understand the specific location and nature of the defect through the image, which reduces the analysis time and facilitates timely maintenance measures.

[0026] 5. Save labor costs: Traditional manual inspection is time-consuming and labor-intensive, especially in long-distance overhead line systems, where the workload is huge. The use of this technical solution can significantly reduce the need for manual inspection and achieve automated real-time detection, thereby reducing labor costs and improving detection efficiency.

[0027] In summary, the scheme of the present invention greatly improves the automation, accuracy and efficiency of contact network hanger string defect detection by introducing convolutional autoencoders and unsupervised learning. At the same time, it has good robustness and cost-effectiveness, and has broad application prospects in the field of railway contact network operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flow chart of an unsupervised defect detection method for contact network suspension strings based on a convolutional autoencoder. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings.

[0030] In this embodiment, see Figure 1 As shown, the present invention proposes an unsupervised defect detection method for contact network suspension strings based on convolutional autoencoder, comprising the steps of:

[0031] The image data of the contact network suspension string is collected and preprocessed; the preprocessed image data is input into a convolutional autoencoder model, wherein the convolutional autoencoder model comprises an encoder part and a decoder part; the encoder part performs a convolution operation on the input image to extract local features of the image, and performs dimensionality reduction through a pooling operation to generate a latent space representation; the decoder part restores the latent space representation to an output image similar to the input image through a deconvolution or upsampling operation; the reconstruction error is calculated, and potential suspension string defects are identified according to the size of the reconstruction error, wherein an abnormally high reconstruction error indicates that a suspension string defect exists; and a defect detection result is generated based on the reconstruction error.

[0032] The preprocessing includes image grayscale, denoising, and normalization processing to reduce the impact of external factors such as ambient light and noise on the image and enhance the feature information of the image;

[0033] The encoder extracts features and reduces the dimension of the input image through multiple convolutional layers and pooling layers to generate a low-dimensional latent space representation z = f(x), where f(x) is the input image and z is the low-dimensional representation; the decoder gradually restores the latent space representation through deconvolution or upsampling layers to generate a reconstructed image similar to the input image. in, is the reconstructed image.

[0034] The convolution operation formula of the encoder part is:

[0035] z l =σ(W l *x+b l )

[0036] Among them, z l is the output of the i-th layer, W l is the convolution kernel weight matrix, * represents the convolution operation, b l is the bias term and σ is the nonlinear activation function.

[0037] The deconvolution operation formula of the decoder part is:

[0038]

[0039] Among them, W l ′ and b l ′ are the weight matrix and bias term of the deconvolution layer respectively.

[0040] The defect detection method uses the normal catenary string image data to perform unsupervised training on the convolutional autoencoder model. The training goal is to minimize the error between the input image and the reconstructed image, which can be defined by the mean square error (MSE):

[0041]

[0042] Among them, x i and They represent the i-th pixel value of the input image and the reconstructed image respectively, and n is the total number of image pixels.

[0043] After the model training is completed, the defect detection method inputs the catenary string image to be detected, and calculates the reconstruction error between the input image and the reconstructed image. The magnitude of the reconstruction error is used to determine whether the string has defects: when the reconstruction error is small (lower than a preset threshold), it indicates that the image is a normal string image; when the reconstruction error is large (exceeding the preset threshold), it indicates that there may be defects in the image, and further marking and analysis are performed.

[0044] For the detected defects, the defective area is located and marked based on the spatial distribution of the image reconstruction error, and a visual detection report of the contact network hanger defects is generated, providing image marking and analysis results of the defective area.

[0045] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An unsupervised defect detection method for contact network suspension strings based on convolutional autoencoder, characterized in that: Includes steps: The image data of the overhead wire suspension string is collected and preprocessed; the preprocessed image data is input into a convolutional autoencoder model, wherein the convolutional autoencoder model includes an encoder part and a decoder part; the encoder part performs a convolution operation on the input image, extracts local features of the image, and performs dimensionality reduction through a pooling operation to generate a latent space representation; The decoder part restores the latent space representation to an output image similar to the input image through deconvolution or upsampling operations; calculates the reconstruction error, and identifies potential suspension string defects based on the size of the reconstruction error, wherein an abnormally high reconstruction error indicates that the suspension string has defects; generates defect detection results based on the reconstruction error; the encoder extracts features and reduces the dimension of the input image through multiple convolutional layers and pooling layers to generate a low-dimensional latent space representation z=f(x), where x is the input image and z is the low-dimensional representation; the decoder gradually restores the latent space representation through deconvolution or upsampling layers to generate a reconstructed image similar to the input image in, is the reconstructed image; The convolution operation formula of the encoder part is: z l =σ(W l *x+b l ) Among them, z l is the output of the i-th layer, W l is the convolution kernel weight matrix, * represents the convolution operation, b l is the bias term, σ is the nonlinear activation function; The deconvolution operation formula of the decoder part is: Among them, W l ′ and b l ′ are the weight matrix and bias term of the deconvolution layer respectively.

2. The method according to claim 1, characterized in that The preprocessing includes image grayscale, denoising, and normalization processing to reduce the impact of external factors such as ambient light and noise on the image and enhance the feature information of the image.

3. The method according to claim 1, characterized in that The defect detection method uses the normal catenary string image data to perform unsupervised training on the convolutional autoencoder model. The training goal is to minimize the error between the input image and the reconstructed image, which can be defined by the mean square error (MSE): Among them, x i and They represent the i-th pixel value of the input image and the reconstructed image respectively, and n is the total number of image pixels.

4. The method according to claim 1, characterized in that After the model training is completed, the defect detection method inputs the catenary string image to be detected, and calculates the reconstruction error between the input image and the reconstructed image. The magnitude of the reconstruction error is used to determine whether the string has defects: when the reconstruction error is small (lower than a preset threshold), it indicates that the image is a normal string image; when the reconstruction error is large (exceeding a preset threshold), it indicates that there may be defects in the image, and further marking and analysis are performed.

5. The method according to claim 1, characterized in that For the detected defects, the defective area is located and marked based on the spatial distribution of the image reconstruction error, and a visual detection report of the contact network hanger defects is generated, providing image marking and analysis results of the defective area.

Images