Contact network positioning point detection method based on lightweight CNN classification algorithm model

Abstract

The invention discloses a catenary positioning point detection method based on a lightweight CNN classification algorithm model, and relates to the technical field of catenary intelligent detection, and the method comprises the steps: obtaining each piece of distance information scanned and returned by a plurality of laser displacement sensors disposed at the top of a catenary detection engineering vehicle, each piece of distance information is transmitted to the embedded control unit; judging whether each piece of distance information is in a preset range or not by using the embedded control unit, and if at least one piece of distance information is in the preset range, obtaining a shooting instruction generated by the embedded control unit; obtaining a positioning point image obtained by the shooting device according to the shooting instruction, and inputting the positioning point image into a model which is trained in the GPU and deployed in the CPU for processing to obtain a detection result of the positioning point image; laser scanning and image recognition technologies are fused, and a camera is triggered based on laser scanning to shoot, so that the image recognition amount is reduced, and the current contact network positioning point recognition precision and the real-time performance of visual recognition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of contact network intelligent detection, and more specifically, to a contact network positioning point detection method based on a lightweight CNN classification algorithm model. Background Art

[0002] In electrified railway design, the catenary is a power supply system installed above the railway line. It consists of the following components: the contact wire: the conductor in the catenary that directly contacts the pantograph of the electric locomotive; the positioning device: used to fix and adjust the position of the contact wire to ensure proper contact with the pantograph; the support device: a structure that provides physical support for the contact suspension and positioning device; the pillar: a column that supports the entire catenary structure; and related infrastructure: including the power supply system, protective devices, and other necessary auxiliary facilities. The main function of the catenary is to transmit electrical energy from the transmission line to the pantograph of the electric locomotive, thereby providing power for the locomotive. There are two types of catenary conductors: flexible conductors: suitable for railway lines in outdoor environments and can adapt to different weather conditions; rigid conductors: due to their relatively strong structure, they are mainly used for lines in tunnel environments such as subways.

[0003] Accurate positioning point detection technology can serve as the basis for the storage and management of data on a per-pole basis, playing a vital role in the contact network defect detection system. Existing technologies primarily rely on visual recognition or radar scanning for contact network positioning point detection. Currently, other applicable technical solutions exist, such as electronic tag-based detection and GPS-based detection. During the operation of the detection system, the vehicle is moving at high speed, and neither visual recognition nor radar scanning can avoid positioning errors due to the delay in recognition calculations. Pure laser scanning methods are prone to misidentification of positioning points. Pure visual recognition methods based on continuous video, if ensuring good recognition results is required, will require significant time overhead, resulting in insufficient real-time performance. Otherwise, a large number of misidentifications will occur.

[0004] Most current open-source image classification algorithm models require RGB three-channel image input and are based on GPU inference. In a CPU environment, the real-time performance of model inference is poor, failing to meet the requirements for fast response and low resource consumption. These models typically require large computing resources, making real-time processing difficult to achieve in practical applications. Therefore, there is a demand for lightweight and efficient classification models to ensure real-time inference classification effects in a CPU environment. Summary of the Invention

[0005] The purpose of the present invention is to provide a contact network positioning point detection method based on a lightweight CNN classification algorithm model. This scheme integrates laser scanning and image recognition technology, and constructs a method based on a lightweight CNN classification algorithm model to reduce the amount of image recognition by triggering camera shooting based on laser scanning, so as to improve the current contact network positioning point recognition accuracy and the real-time performance of visual recognition.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions: In the first aspect, the present application provides a contact network positioning point detection method based on a lightweight CNN classification algorithm model, comprising the following specific steps: Acquire the distance information scanned and returned by multiple laser displacement sensors installed on the top of the catenary inspection vehicle. The distance information includes the distance data between the corresponding laser displacement sensors and the positioning points, and transmit the distance information to the embedded control unit. Using the embedded control unit to determine whether each distance information is within a preset range, if at least one distance information is within the preset range, obtaining a shooting instruction generated by the embedded control unit, and transmitting the shooting instruction to the shooting device; The image of the anchor point obtained by the shooting device according to the shooting instruction is obtained, and the image of the anchor point is input into the TorchScript model trained on the GPU and deployed on the CPU for processing to obtain the detection result of the anchor point image, which is whether the anchor point is included or not.

[0007] On the basis of the above technical solution, the present invention can also be improved as follows.

[0008] Furthermore, there are four laser displacement sensors, and all of them scan continuously at a frequency of 1000 Hz.

[0009] Furthermore, the above TorchScript model is trained on the GPU through a lightweight CNN classification algorithm model, which includes a convolutional layer, a relu layer, a pooling layer, a first affine layer, a second affine layer, and a Sofmax layer.

[0010] Furthermore, the convolutional layer includes 6 convolution kernels with kernel_size of 5×5; in the relu layer, the activation function is the relu function; in the first affine layer and the second affine layer, the number of neurons is 38400 and 64 respectively.

[0011] Furthermore, the above-mentioned positioning point image is input into the TorchScript model for processing after image processing, and the positioning point image after image processing is resized to 640×640.

[0012] Furthermore, the output of the Sofmax layer is specifically: ; Where, Indicates the k The output value of the item, Represents the output of the second Affine layer k item, express The exponential function value of Indicates from i =1 to n All The sum of the values.

[0013] Furthermore, the training end condition of the above-mentioned lightweight CNN classification algorithm model is that the loss function reaches the training end condition. The specific loss function is: ; Where, Represents the loss function value, which is the difference between the model prediction value and the true value. k represents the total number of categories, indicator variables representing true values, Represents the predicted value of the model output.

[0014] In a second aspect, the present application provides a contact network positioning point detection system based on a lightweight CNN classification algorithm model, comprising: Multiple laser displacement sensors are installed on the top of the catenary inspection vehicle, which are used to continuously scan at a frequency of 1000Hz and return the corresponding distance information; An embedded control unit, configured to determine whether each piece of distance information is within a preset range, and further configured to generate a shooting instruction when at least one piece of distance information is within the preset range; A shooting device, used to obtain and store the positioning point image according to the shooting instruction generated by the embedded control unit; The TorchScript model is trained on the GPU and deployed on the CPU. It is used to obtain the corresponding detection results based on the anchor point image, including or excluding the anchor points.

[0015] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any one of the methods in the first aspect when executing the computer program.

[0016] In a fourth aspect, the present application provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute any one of the methods in the first aspect.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: In this application, first, the distance information scanned and returned by multiple laser displacement sensors is judged by an embedded control unit. Secondly, when one or more of the distance information is within a preset range, a shooting instruction for shooting the positioning point image is generated, and the shooting device shoots the positioning point image according to the shooting instruction and then stores it. Finally, the positioning point image is input into the TorchScript model trained in the GPU and deployed in the CPU for processing, so as to obtain the detection result of the positioning point image, and the detection result includes two results: including the positioning point or not.

[0018] In this application, the TorchScript model is obtained by training a lightweight CNN classification algorithm model. The lightweight CNN classification algorithm model is trained on the GPU and deployed on the CPU. When the model is inferred on the CPU, it has excellent resource usage, high real-time performance and extremely low CPU usage. At the same time, the classification model reasoning does not rely on the GPU, which can significantly reduce computing power overhead while ensuring classification accuracy and meet the needs of real-time processing.

[0019] In this application, the convolution layer is composed of 6 convolution kernels with kernel_size of 5×5, which not only greatly improves the computational efficiency, but also significantly reduces the number of model parameters, thereby reducing memory usage and computational overhead; the nonlinear transformation is performed through the relu function to enhance the expressiveness of the model; and the downsampling processing of the maximum pooling layer further reduces the size of the feature map, while retaining important feature information and improving the robustness of the model to translation and deformation; finally, the feature map output by the pooling layer is sent to the first Affine layer with 38400 neurons, and then enters the second Affine layer with 64 neurons. The output result dimension of the second Affine layer is 2, thereby integrating the extracted features for final positioning detection.

[0020] In this application, the cross-entropy loss function is used to calculate the model loss, which can effectively measure the difference between the model prediction and the true label, providing important feedback for model training. The training process of the image classification model includes data preparation, model construction, selection of loss function and optimizer, iterative training of forward propagation and backpropagation, as well as verification and tuning. Through the above series of carefully designed hierarchical structures, the overall model can achieve good classification performance while maintaining high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings: Figure 1 A flow chart of a detection method according to an embodiment of the present invention; Figure 2 Schematic diagram of a lightweight CNN classification algorithm model in an embodiment of the present invention; Figure 3 is a schematic diagram of a positioning point image including positioning points in an embodiment of the present invention; Figure 4 is a schematic diagram of a positioning point image that does not include positioning points in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0024] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0025] In the description of the embodiments of the present invention, "a plurality of" means at least two.

[0026] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0027] Example 1: In order to solve the problems that most of the current open source image classification algorithm models require RGB three-channel image input and are based on GPU reasoning, and in the CPU environment, the real-time performance of model reasoning is poor, which cannot meet the requirements of fast response and low resource consumption; and the existing models usually require large computing resources, which makes it difficult to achieve real-time processing in practical applications, this embodiment provides a contact network positioning point detection method based on a lightweight CNN classification algorithm model, such as Figure 1 As shown, the following specific steps are included: S1, obtains the distance information scanned and returned by multiple laser displacement sensors installed on the top of the contact network inspection engineering vehicle, the distance information including the distance data between the corresponding laser displacement sensors and the positioning points, and transmits the distance information to the embedded control unit.

[0028] Among them, the laser displacement sensor continuously scans at high frequency and returns distance information to the embedded control unit. In this solution, four laser displacement sensors can be set up. In order to avoid the contact line above the vehicle as much as possible, the laser displacement sensors are installed at ±300mm and ±500mm relative to the vehicle's central axis; the four laser emitters scan continuously at a frequency of 1000Hz and return all distance information to the embedded control unit.

[0029] Optionally, there are four laser displacement sensors, and all of them scan continuously at a frequency of 1000 Hz.

[0030] S2, using the embedded control unit to determine whether each distance information is within a preset range. If at least one distance information is within the preset range, obtaining a shooting instruction generated by the embedded control unit and transmitting the shooting instruction to the shooting device.

[0031] Among them, the embedded control unit determines whether the distance information is within the set range and issues a shooting command; the embedded control unit receives the distance information of all laser displacement sensors. If there is a laser whose distance is within the set range, it is considered that there is a locator above it. At this time, the embedded control unit issues a command to the camera and LED fill light. The above-mentioned shooting equipment includes a camera and an LED fill light.

[0032] Furthermore, the camera receives the instruction to start exposure and take pictures for storage; after the camera receives the instruction sent by the embedded control unit, the camera and the fill light are turned on at the same time, and the camera starts exposure and takes and stores images.

[0033] S3, obtains the positioning point image obtained by the shooting device according to the shooting instruction, and inputs the positioning point image into the TorchScript model trained in the GPU and deployed in the CPU for processing, and obtains the detection result of the positioning point image. The detection result is whether it contains the positioning point or not. Among them, the positioning point can be a locator installed in the contact network, and the detection of the positioning point can be understood as the detection of the locator, such as Figure 3 and Figure 4 As shown, Figure 3 is the positioning point image containing the locator, and Figure 4 This is the image without the locator.

[0034] Among them, the positioning point image is input into the TorchScript model for processing after image processing, and the positioning point image after image processing is resized to 640×640.

[0035] Optionally, the above TorchScript model is trained on the GPU using a lightweight CNN classification algorithm model, such as Figure 2 As shown in the figure, the lightweight CNN classification algorithm model includes a convolution layer, a relu layer, a pooling layer, a first Affine layer, a second Affine layer, and a Sofmax layer.

[0036] The convolution layer includes 6 convolution kernels with kernel_size of 5×5. In the relu layer, the activation function is the relu function. In the first affine layer and the second affine layer, the number of neurons is 38400 and 64 respectively.

[0037] Specifically, a lightweight CNN classification algorithm model that can perform real-time inference on the CPU is constructed and trained. The model architecture starts from the input layer, such as Figure 2 As shown in the figure, taking the anchor point image containing the anchor points as an example, it first passes through the convolution layer, in which only 6 convolution kernels are used; this design not only greatly improves the computational efficiency, but also significantly reduces the number of model parameters, thereby reducing memory usage and computational overhead; then, the output of the convolution layer is nonlinearly transformed through the ReLU activation function, enhancing the expressiveness of the model; subsequently, after the downsampling processing of the maximum pooling layer, the size of the feature map is further reduced, while retaining important feature information, thereby improving the robustness of the model to translation and deformation; next, the feature map is sent to the first fully connected layer with 38400 neurons, and then enters the second fully connected layer with 64 neurons and an output dimension of 2 to integrate the extracted features for final classification.

[0038] Optionally, the output of the Sofmax layer is: ; Where, Indicates the k The output value of the item, Represents the output of the second Affine layer k item, express The exponential function value of Indicates from i =1 to n All The sum of the values.

[0039] Optionally, the training end condition of the above-mentioned lightweight CNN classification algorithm model is that the loss function reaches the training end condition. The specific loss function is: ; Where, Represents the loss function value, which is the difference between the model prediction value and the true value. k represents the total number of categories, indicator variables representing true values, Represents the predicted value of the model output.

[0040] The training process for an image classification model involves data preparation, model construction, selection of a loss function and optimizer, iterative training using forward and backpropagation, and validation and tuning. First, an image dataset is collected and preprocessed, which is then used to train the constructed model. Forward propagation is used to calculate output and evaluate loss, followed by backpropagation to update model parameters. This process is repeated for multiple training cycles to optimize performance. After each cycle, a validation set is used to monitor model performance and hyperparameters are adjusted based on the results. After training is complete, an independent test set is used to evaluate the model's generalization ability, and the trained model is saved.

[0041] Specifically, the training process of the above-mentioned lightweight CNN classification algorithm model can be: 1. Preparation of a data set, including multiple images in the data set, including images with positioning points and images without positioning points; 2. Inputting each image in the data set into the lightweight CNN classification algorithm model for processing, and calculating the loss function until the number of iterations is reached or the value of the loss function does not exceed the threshold, then the above-mentioned TorchScript model is obtained.

[0042] Among them, the trained model can be converted into a TorchScript model for inference in a C++ environment. Then, the running environment is prepared, the API is developed to handle requests, pre-processing is performed to adapt to the input format, inference is performed to obtain output results, and post-processing is performed to parse and optimize the results. Finally, the model and configuration are added to the system. The pre-processing, inference and post-processing are compiled into a dynamic library for system calls. After loading the pictures taken by laser triggering, our trained image classification model is used for inference to identify and classify the contact network locators in the image and filter out pictures of non-locators.

[0043] Furthermore, based on the above positioning point detection results, the system positioning information is updated in real time to improve the positioning accuracy of the detection system, providing a basis for the contact network detection system to realize "one pole one file" data storage and management.

[0044] Example 2: The embodiment of the present application provides a contact network positioning point detection system based on a lightweight CNN classification algorithm model, including: Multiple laser displacement sensors are installed on the top of the catenary inspection vehicle, which are used to continuously scan at a frequency of 1000Hz and return the corresponding distance information; An embedded control unit, configured to determine whether each piece of distance information is within a preset range, and further configured to generate a shooting instruction when at least one piece of distance information is within the preset range; A shooting device, used to obtain and store the positioning point image according to the shooting instruction generated by the embedded control unit; The TorchScript model is trained on the GPU and deployed on the CPU. It is used to obtain the corresponding detection results based on the anchor point image, including or excluding the anchor points.

[0045] Specifically, the above-mentioned positioning point detection system can be implemented in the following manner when in use: S11, obtaining each distance information scanned and returned by multiple laser displacement sensors, and transmitting each distance information to the embedded control unit.

[0046] S12, using the embedded control unit to determine whether each distance information is within a preset range. If at least one distance information is within the preset range, obtaining a shooting instruction generated by the embedded control unit and transmitting the shooting instruction to the shooting device.

[0047] S13, obtaining the positioning point image obtained by the shooting device according to the shooting instruction, and inputting the positioning point image into the TorchScript model trained in the GPU and deployed in the CPU for processing to obtain the detection result of the positioning point image.

[0048] Example 3: An embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, any one of the methods in Embodiment 1 is implemented.

[0049] Example 4: An embodiment of the present application provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute any one of the methods in Example 1.

[0050] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A contact network positioning point detection method based on a lightweight CNN classification algorithm model is characterized by: The specific steps include: Acquire various distance information scanned and returned by multiple laser displacement sensors installed on the top of the catenary inspection engineering vehicle, the distance information including the distance data between the corresponding laser displacement sensors and the positioning points, and transmit the various distance information to the embedded control unit; Determining, by means of the embedded control unit, whether each piece of distance information is within a preset range, and if at least one piece of distance information is within the preset range, obtaining a shooting instruction generated by the embedded control unit, and transmitting the shooting instruction to a shooting device; Obtain the positioning point image obtained by the shooting device according to the shooting instruction, and input the positioning point image into the TorchScript model trained in the GPU and deployed in the CPU for processing to obtain the detection result of the positioning point image, which is whether the positioning point is included or not.

2. The method for detecting contact network positioning points based on a lightweight CNN classification algorithm model according to claim 1, characterized in that: There are four laser displacement sensors, and all of them scan continuously at a frequency of 1000 Hz.

3. The method for detecting contact network positioning points based on a lightweight CNN classification algorithm model according to claim 1, characterized in that: The TorchScript model is trained in a GPU through a lightweight CNN classification algorithm model, and the lightweight CNN classification algorithm model includes a convolution layer, a relu layer, a pooling layer, a first Affine layer, a second Affine layer, and a Sofmax layer.

4. The method for detecting contact network positioning points based on a lightweight CNN classification algorithm model according to claim 3 is characterized in that: The convolution layer includes 6 convolution kernels with kernel_size of 5×5; in the relu layer, the activation function is the relu function; in the first affine layer and the second affine layer, the number of neurons is 38400 and 64 respectively.

5. The method for detecting contact network positioning points based on a lightweight CNN classification algorithm model according to claim 1, characterized in that: The positioning point image is input into the TorchScript model for processing after image processing, and the positioning point image after image processing is resized to 640×640.

6. The method for detecting contact network positioning points based on a lightweight CNN classification algorithm model according to claim 3, characterized in that: The output of the Sofmax layer is specifically: ； Where, Indicates the k The output value of the item, Represents the output of the second Affine layer k item, express The exponential function value of Indicates from i =1 to n All The sum of the values.

7. The method for detecting contact network positioning points based on a lightweight CNN classification algorithm model according to claim 6, characterized in that: The training end condition of the lightweight CNN classification algorithm model is that the loss function reaches the training end condition, and the loss function is specifically: ； Where, Represents the loss function value, which is the difference between the model prediction value and the true value. k represents the total number of categories, indicator variables representing true values, Represents the predicted value of the model output.

8. The contact network positioning point detection system based on the lightweight CNN classification algorithm model is characterized by: include: Multiple laser displacement sensors are installed on the top of the catenary inspection vehicle, which are used to continuously scan at a frequency of 1000Hz and return the corresponding distance information; an embedded control unit, configured to determine whether each piece of distance information is within a preset range, and further configured to generate a shooting instruction when at least one piece of distance information is within the preset range; A shooting device, configured to obtain and store an image of the positioning point according to the shooting instruction generated by the embedded control unit; A TorchScript model, which is trained on a GPU and deployed on a CPU, is used to obtain corresponding detection results based on the anchor point image, including or excluding the anchor points.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the method according to any one of claims 1 to 7 is implemented when the processor executes the computer program.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, which enable a computer to execute the method of any one of claims 1 to 7.

Images