Brake disc gluing defect detection method and system based on pulse neural network
By using a multi-angle image acquisition and processing method based on spiking neural networks, a lightweight detection model was built, which solved the problems of low accuracy, low efficiency and high cost in traditional detection methods, and achieved efficient and accurate detection of brake disc adhesive defects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-10
AI Technical Summary
Traditional manual visual inspection and existing computer vision inspection methods are inaccurate, inefficient and costly in detecting defects in brake disc adhesive coating. Furthermore, the models lack generalization and robustness, making it difficult to meet the needs of large-scale production.
A detection method based on spiking neural networks is adopted. Through multi-angle image acquisition, image registration and preprocessing, a lightweight pulse-driven Transformer structure and a pulse mask autoencoder are built. The training method of spiking neural networks is optimized to improve the performance of the detection model.
It enables automated, rapid, and accurate detection of adhesive defects on brake discs, reducing computational complexity and training costs, and improving the real-time performance and reliability of the detection.
Smart Images
Figure CN121639615A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of automobile parts quality detection technology, in particular to a brake disc glue coating defect detection method and system based on a pulse neural network. BACKGROUND
[0002] The brake disc is a core component in the automobile braking system, and the vehicle is decelerated and braked through the friction between the surface attached material and the brake pad. The surface glue coating process is usually used to bond the friction material in industrial production of brake discs. If the glue coating does not meet the technical index, it may cause the friction material to be not firmly attached, and problems such as slipping and shaking may occur during braking, reducing the braking effect and affecting the safety of driving. Therefore, it is crucial to check the unqualified brake disc in time, which can prevent unqualified products from entering the market, improve the overall quality and reliability of the products, and enhance the trust of consumers on the brand.
[0003] The traditional brake disc glue coating defect detection adopts the manual visual inspection method, relying on the worker's naked eye observation and experience judgment. However, this method has low detection accuracy and is unstable, and is easily affected by the worker's fatigue degree and subjective factors. The manual detection has low detection efficiency and is difficult to meet the needs of large-scale production, and is easy to cause secondary damage to the glued brake disc. The existing method realizes automatic visual detection based on a computer, uses a camera to image the brake disc, and uses image processing and deep learning methods to more comprehensively and accurately detect glue coating defects. Through learning and training of a large number of brake disc glue coating image data sets, the neural network model can automatically identify and locate various defects. However, the process of analyzing the image by the model has high computational complexity, requires high-cost training resources, has large energy consumption, and the generalization and robustness of the model are insufficient, which affects the accuracy of brake disc glue coating defect detection. SUMMARY
[0004] In view of the technical problems of the above-mentioned prior art, the present application provides a brake disc glue coating defect detection method and system based on a pulse neural network. Based on the multi-angle collected brake disc images, a light-weight pulse-driven Transformer structure and a pulse mask autoencoder are built using a pulse neural network to realize automatic detection of brake disc glue coating defects. The improved training method is used to optimize the pulse neuron firing pattern in the pulse neural network, improve the model performance, and reduce the training cost.
[0005] The present application provides a brake disc glue coating defect detection method based on a pulse neural network, comprising: (1) photographing the brake disc from different angles to collect multi-angle image data of the glued brake disc; (2) The multi-angle brake disc image is preprocessed, the multi-angle images are fused through an image registration method, comprehensive information of a glue coating area of the brake disc is obtained, a target area for glue coating defect detection is positioned according to inherent hole features on the brake disc, and the multi-angle brake disc fused image is cropped; (3) A detection model based on a pulse neural network is built, a convolution and a self-attention mechanism are used to encode the brake disc image, glue coating defect information of the brake disc is extracted from the image based on pulse driving, the detection model is trained through a pulse firing approximation method, and a pre-training detection model with optimal performance is obtained; (4) The cropped multi-angle brake disc fused image is input into the pre-training detection model, a glue coating defect type of the brake disc is recognized, a position of the defect is positioned, and corresponding processing measures are taken.
[0006] The application also provides a brake disc glue coating defect detection system based on a pulse neural network, comprising: A multi-angle image acquisition module is configured to capture multi-angle images of the brake disc from different angles. An image preprocessing module is configured to preprocess the multi-angle brake disc image, fuse the multi-angle images through an image registration method, obtain comprehensive information of a glue coating area of the brake disc, position a target area for glue coating defect detection according to inherent hole features on the brake disc, and crop the multi-angle brake disc fused image. A model building and training module is configured to build a detection model based on a pulse neural network, encode the brake disc image through a convolution and a self-attention mechanism, extract glue coating defect information of the brake disc from the image based on pulse driving, train the detection model through a pulse firing approximation method, and obtain a pre-training detection model with optimal performance. An automatic defect detection module is configured to input the cropped multi-angle brake disc fused image into the pre-training detection model, recognize a glue coating defect type of the brake disc, position a position of the defect, and take corresponding processing measures.
[0007] Further, the pulse neural network (SNN) used in the application is a neural network simulating a biological neural system, and the basic unit thereof is a pulse neuron. Unlike a traditional artificial neural network (ANN) using continuous decimal values to transmit information, the pulse neuron receives continuous values and converts them into a pulse sequence: The physical explanation for the pulse neuron generating the pulse sequence is that the pulse neuron receives generated current and accumulates membrane potential, and compares the membrane potential with a threshold value to determine whether to generate a pulse. A leaky integrate-and-fire pulse neuron is commonly used, and the dynamic model thereof is represented as:
[0008]
[0009]
[0010] wherein, denotes the membrane time constant, denotes the time step the input current at time t; when the membrane potential exceeds the discharge threshold , the spiking neuron will fire a spike . is the Heaviside step function, which is 1 when the input value is greater than or equal to 0, and 0 otherwise; denotes the membrane potential after a firing event, which is equal to if no spike is generated, and equal to the reset potential otherwise.
[0011] Further, the method for training the spiking neural network includes ANN-to-SNN conversion and direct training: ANN-to-SNN conversion replaces the ReLU activation layer with a spiking neuron, approximates the activation value of the ANN using the spike firing rate, and migrates the high-performance pre-trained ANN network parameters to the SNN; in the direct training method, the SNN is unfolded in the simulation time step, and is trained in the time backward propagation manner; due to the non-differentiable event-driven mechanism in the spiking neuron, the backward propagation uses a proxy gradient, and the time dynamic characteristics are retained, but the time step simulation requires a large amount of computing resources, increases the training difficulty, and limits the scalability and performance improvement of the spiking neural network; The quantization error is the main problem in the existing training method, and the binary spike firing is a fundamental mechanism defect, which weakens the ability of the spiking neuron to evaluate the importance of the input signal (spatial representation) and memory and forgetting (time dynamic characteristics); the existing method solves the spatial representation method through rate coding, but the training efficiency is low, and the limitation of time dynamic characteristics cannot be overcome, therefore, the present application adopts a spike firing approximation method, uses integer activation values during training, and converts them into spike sequences during inference, forms an asynchronous continuous spike firing mode, and improves the training efficiency, performance and network expansion capability.
[0012] The present application discloses the following technical effects: The application provides a brake disc glue coating defect detection method based on a pulse neural network, which is based on a visual detection principle, processes and analyzes multi-angle images of a glue coated brake disc, extracts glue coating defect features by using a pulse driven detection model, and realizes automatic detection of glue coating defects of the brake disc. BRIEF DESCRIPTION OF DRAWINGS
[0013] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings of the embodiments of the application will be briefly introduced below, and the flowcharts are used to illustrate the operations performed by the system according to the embodiments of the application in this application. It should be understood that the foregoing or the following operations are not necessarily performed in sequence. On the contrary, various steps can be processed in reverse order or simultaneously according to needs. At the same time, other operations can be added to these processes, or one or more steps of operations can be removed from these processes.
[0014] Figure 1 The flowchart of the brake disc glue coating defect detection method based on the pulse neural network provided by the embodiments of the application is shown.
[0015] Figure 2 The structural diagram of the detection model based on the pulse neural network provided by the embodiments of the application is shown.
[0016] Figure 3 The structural diagram of the brake disc glue coating defect detection system based on the pulse neural network provided by the embodiments of the application is shown. DETAILED DESCRIPTION
[0017] The above description is only a summary of the technical solutions of the application, in order to more clearly understand the technical means of the application, the embodiments of the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described.
[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0019] In the following description, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.
[0020] Example 1: This application provides a method for detecting defects in brake disc adhesive coating based on a spiking neural network, such as... Figure 1 As shown, the method includes: Step S10: Take pictures of the brake disc from different angles to collect multi-angle image data of the brake disc after applying adhesive.
[0021] In this embodiment, a high-resolution and high-frame-rate industrial camera is selected, and the resolution is set to 2000×2000 pixels. The brake disc is fixed on the shooting platform, and a ring light source is set to make the lighting on the surface of the brake disc uniform, reduce shadows and reflections, and dynamically adjust according to environmental conditions. Adjust the camera's exposure time, aperture, and focal length, and take pictures from the front, side, and oblique angles of the brake disc to obtain the overall image, adhesive thickness, and edge information of the adhesive coating area, as well as the three-dimensional information of the adhesive coating area, for analysis of the distribution, overflow, and depth of the adhesive coating. Record the shooting angle and camera parameter settings for each brake disc image, save the images in JPEG format, and name and archive them according to shooting angle, date, and batch to achieve classified storage.
[0022] Step S20: Preprocess the multi-angle brake disc images, fuse the multi-angle images using image registration method to obtain comprehensive information on the adhesive coating area of the brake disc, locate the target area for adhesive coating defect detection based on the inherent circular hole features on the brake disc, and crop the fused multi-angle brake disc images.
[0023] In this embodiment, the multi-angle brake disc images are preprocessed to improve image quality and prepare for image registration and fusion. Read the multi-angle color images of the brake disc captured by the camera, and calculate the value of each pixel according to the following formula. grayscale value : ,in , and Representing pixels The pixel values corresponding to the red, green, and blue channels are used to calculate grayscale values and assign them to each pixel to generate a grayscale image. Set a Gaussian filter with a filter window size of 3×3, calculate the weight matrix of the Gaussian filter, move the filter to the grayscale image of the brake disc at multiple angles, and perform a weighted average of the pixel values within the window to obtain new pixel values. This smooths out random noise in the image and protects the edge features of the coated area. The grayscale values of the denoised multi-angle brake disc grayscale images are normalized to the range [0, 255], and the minimum grayscale value of each image is calculated. and maximum grayscale value Normalization is achieved using the following formula:
[0024] in, Represents the number of pixels in each image. The normalized grayscale values are calculated; the cumulative histogram of the normalized multi-angle brake disc grayscale image is calculated, and the image contrast is enhanced by linear mapping to highlight the glued area.
[0025] Based on the preprocessed multi-angle brake disc images, feature points are extracted from the images, and a descriptor for each feature point is generated. The ratio test matching method is used to match the feature points in the images from different angles to generate matching point pairs. Reliable matching point pairs are further selected through geometric consistency rules. Based on the selected matching point pairs, the geometric transformation relationship between the multi-angle images is estimated using the least squares method to generate a geometric transformation matrix. Using the front-shot brake disc image as a reference image, the estimated transformation matrix is used to perform geometric transformation on the preprocessed multi-angle brake disc images to align the feature points between the images, generating a registered image. A multi-angle brake disc fusion image is then generated using a weighted average fusion method to integrate multi-angle information.
[0026] The Hough transform is used to detect circular hole features in a multi-angle brake disc fusion image, including the following steps: Using an edge detection algorithm, dual threshold detection and edge linking are employed. Based on the principle of non-maximum suppression, an edge image of the multi-angle brake disc fusion image is generated, and edge points are extracted. Initialize the Hough transform parameters, including the minimum radius of the detected circular hole, the maximum radius of the detected circular hole, the accumulator threshold, and the minimum distance between the centers of the circular holes; For each edge point, the point is mapped to the parameter space according to the set equation of the circle. In the parameter space, each edge point corresponds to a conical surface. The number of votes for each parameter combination is recorded by an accumulator. Calculate all possible combinations of circle centers and radii for each edge point in the parameter space, and add the vote count at the corresponding position in the accumulator. Find the local maximum in the accumulator, filter the circles based on the threshold and the circle radius range, and remove duplicate circle detection results. Extract the center coordinates and radius information from the selected circular holes, delineate the feature boundary of the circular holes, and use it as the target area for brake disc adhesive defect detection. Then, crop the multi-angle brake disc fusion image according to the boundary and remove image areas that are irrelevant to defect detection.
[0027] Step S30: Build a detection model based on a spiking neural network. Use convolution and self-attention mechanisms to encode the brake disc image. Extract the adhesive defect information of the brake disc from the image based on pulse drive. Train the detection model through the pulse firing approximation method to obtain the pre-trained detection model with the best performance.
[0028] In this embodiment, the detailed steps for obtaining the pre-trained detection model include: Collect image data of different types of defects in the adhesive coating area of the brake disc, and use the Labellmg annotation tool to mark the location and type of adhesive coating defects in the images. Defect types include uneven adhesive distribution, adhesive dripping, pore mixing, adhesive loss and adhesive overflow. The annotation information of each image is organized into an XML format annotation file, which records the bounding box position coordinates, the glue coating defect type code, and the corresponding image data to form the training dataset of the detection model, ensuring that the dataset contains brake disc glue coating defect images under different angles and lighting conditions. In PyCharm, create a workspace and write the model and training code: Import the PyTorch-based deep learning framework and the snntorch library related to spiking neural networks; Construct a convolution-based spiking neural network module to extract local features of the brake disc. It includes spiking separable convolutional layers and channel convolutional layers. Each convolutional layer is connected to a spiking neuron and a pooling layer. Set the convolutional kernel size, input and output dimensions, and spiking parameters. A self-attention-based spiking neural network module is constructed, consisting of a spiking separable convolutional layer, a spiking-driven self-attention layer, and a channel multilayer perceptron. It dynamically focuses on important regions in the brake disc image, and spiking neurons replace the activation functions of each convolutional and linear layer in the module, with relevant parameters set. Construct a target detection head, use a decoder as the core to recover the pulse code into an image, configure the parameters of each layer in the Transformer structure based on artificial neural network in the decoder, set the output dimensions of the linear layer to be the number of glue defect types and the defect location information dimension, and calculate the loss function. Load the training set data using DataLoader, select IoU loss function and cross-entropy loss for glue coating defect localization and classification tasks respectively, and set SGD optimizer; During model training, the impulse method approximation method is used for training, the model parameters are updated, the model performance after each round of training is evaluated using the test set, the accuracy, recall and F1 score are calculated, the model hyperparameter settings are optimized and adjusted based on the evaluation results, and the model parameters with the best performance are saved as the pre-trained detection model.
[0029] Step S40: Input the cropped multi-angle brake disc fusion image into the pre-trained detection model to identify the type of adhesive defect in the brake disc, locate the defect, and take corresponding processing measures.
[0030] Example 2: This embodiment of the invention provides a detailed structure of a detection model based on a spiking neural network for automatic detection of adhesive defects in brake discs, as follows: Figure 2 As shown: The detection model integrates convolutional and Transformer-based spiking neural network modules, each containing a disc mixer and a channel mixer. The input multi-angle brake disc fusion image is downsampled to reduce feature dimensionality and then fed into the convolutional spiking neural network module to extract local features. This module consists of pulse-separable convolutions and channel convolutions. Separable spiking convolution employs a bottleneck structure, sequentially expanding and then compressing the feature channel dimension through pointwise convolution, depthwise separable convolution, and pointwise convolution, refining the feature information of each channel and improving the comprehensiveness of the acquired fine-grained information. However, in spiking neural networks, the direct connection between depthwise separable convolution and pointwise convolution, coupled with subsequent parameter resetting operations, can lead to severe energy consumption issues. Therefore, batch normalization layers and spiking neurons are added between convolutional layers, using pulse-driven masks to achieve sparse convolution, reducing model parameters and computational complexity. After four consecutive downsampling layers and processing by a convolution-based spiking neural network module, the output feature size is reduced to 1 / 8 of its original size. Then, the local features of the output are fed into the downsampling layer and then into the Transformer-based spiking neural network module to extract global features. This module adds a convolutional layer before calculating self-attention to improve the integration of global features. This convolutional layer uses a lightweight sparse spiking convolution operation. The module is based on a residual structure, combining a spiking separable convolutional layer, a spiking-driven self-attention layer, and a channel multilayer perceptron. In the spiking-driven self-attention layer, linear operations and spiking neuron activation are used to generate the query matrix in the self-attention mechanism. Key matrix Sum matrix Instead of energy-intensive reparameterized convolutions, to enhance the representation effect of self-attention layers, Channel expansion The original feature dimensions are restored after the attention calculation is completed.
[0031] The output of the last Transformer-based spiking neural network module is fed into the target detection head. After being processed by the decoder and linear layer mapping, the detection result of the brake disc adhesive defect can be obtained.
[0032] The process of the pulse firing approximation method for training the detection model is as follows: First, it is assumed that the spiking neurons transmit the pulse firing rate in the spatial dimension; second, during the training phase, integer values are used to activate the approximate pulse firing rate to reduce quantization error and improve the model's learning ability; finally, during the inference phase, the integer activation is converted into binary pulses with multiple time steps to extend the time step size and achieve efficient inference.
[0033] Example 3: The brake disc adhesive coating defect detection system based on a spiking neural network provided in this embodiment of the invention can execute the brake disc adhesive coating defect detection method based on a spiking neural network provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method, such as... Figure 3 As shown, it includes the following modules: Multi-angle image acquisition module: used to capture images of the brake disc from different angles, acquiring multi-angle image data of the brake disc after adhesive application; Image preprocessing module: used to preprocess multi-angle brake disc images, fuse multi-angle images through image registration method, obtain comprehensive information of the adhesive coating area of brake disc, locate the target area for adhesive coating defect detection based on the inherent circular hole features on brake disc, and crop the multi-angle brake disc fused image. Model building and training module: used to build a detection model based on spiking neural network, using convolution and self-attention mechanism to encode brake disc images, extracting information on adhesive defects of brake discs from images based on pulse drive, and training the detection model through pulse firing approximation method to obtain the pre-trained detection model with the best performance; Automated Defect Detection Module: This module is used to input the fused image of the cut multi-angle brake disc into the pre-trained detection model, identify the type of adhesive defect in the brake disc, locate the defect, and take corresponding measures.
[0034] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.
[0035] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
Claims
1. A method for detecting glue coating defects of a brake disc based on a pulse neural network, characterized in that, The method comprises: (1) photographing the brake disc from different angles to collect multi-angle image data of the brake disc after gluing; (2) preprocessing the multi-angle brake disc image, fusing the multi-angle image through an image registration method to obtain comprehensive information of the gluing area of the brake disc, and positioning the target area of the gluing defect detection according to the inherent circular hole feature on the brake disc, and cutting the multi-angle brake disc fused image; (3) building a detection model based on a pulse neural network, encoding the brake disc image using convolution and self-attention mechanism, extracting the gluing defect information of the brake disc from the image based on pulse driving, and training the detection model through a pulse firing approximation method to obtain a pre-trained detection model with optimal performance; (4) inputting the cut multi-angle brake disc fused image into the pre-trained detection model to identify the gluing defect type of the brake disc, positioning the defect location, and taking corresponding processing measures. 2.The pulse neural network based brake disc glue coating defect detection method of claim 1, wherein, In step (1), an industrial camera is selected, the brake disc is fixed on a shooting platform, a ring-shaped light source is set to make the light on the surface of the brake disc uniform, reduce shadows and reflections, and dynamically adjust according to environmental conditions; Adjust the exposure time, aperture size and focal length of the camera, shoot from the front, side and oblique angles of the brake disc, respectively obtain the overall image of the gluing area of the brake disc, the gluing thickness and edge information, and the stereoscopic information of the gluing area, which are used to analyze the distribution, overflow and depth of the gluing; Record the shooting angle and camera parameter setting of each brake disc image, save the image as a JPEG format, and name and archive it according to the shooting angle, date and batch to realize classified saving. 3.The pulse neural network based brake disc glue coating defect detection method of claim 1, wherein, In step (2), the multi-angle brake disc color images collected by the camera are read, and the gray value of each pixel point is calculated according to the following formula : : , , and respectively represent the pixel values corresponding to the red, green and blue channels of the pixel point , and the calculated gray value is assigned to each pixel point to generate a gray image; Set a Gaussian filter, the filter window size is 3x3, calculate the weight matrix of the Gaussian filter, move the filter to the multi-angle brake disc gray image, and perform weighted average on the pixel values in the window to obtain new pixel values, smooth the random noise in the image and protect the edge features of the gluing area; The gray value of the denoised multi-angle brake disc gray scale image is normalized to the range of [0, 255], and the gray minimum value of each image is calculated and the gray maximum value Normalization is realized by the following formula: in, Represents the number of pixels in each image. The normalized grayscale values are calculated; the cumulative histogram of the normalized multi-angle brake disc grayscale image is calculated, and the image contrast is enhanced by linear mapping to highlight the glued area.
4. The pulse neural network-based brake disc glue coating defect detection method of claim 3, wherein, The multi-angle brake disc gray image extracts feature points through an image registration algorithm, generates a descriptor for each feature point, matches the feature points in different angle images using a ratio test matching method to generate matching point pairs, and further screens out reliable matching point pairs through geometric consistency rules; According to the screened matching point pairs, the least squares method is used to estimate the geometric transformation relationship between the multi-angle images to generate a geometric transformation matrix; Take the front shot brake disc image as the reference image, use the estimated transformation matrix to perform geometric transformation on the preprocessed multi-angle brake disc image to align the feature points between the images, generate a registered image, and generate a multi-angle brake disc fused image through a weighted average fusion method to integrate multi-angle information. 5.The brake disc glue coating defect detection method based on pulse neural network according to claim 4, wherein, The multi-angle brake disc fused image detects the circular hole feature through Hough transformation, including the following steps: Through an edge detection algorithm, using double threshold detection and edge linking, based on the non-maximum suppression principle, an edge image of the multi-angle brake disc fused image is generated, and edge points are extracted; Initialize the Hough transformation parameters, including the minimum radius of the detected circular hole, the maximum radius of the detected circular hole, the accumulator threshold, and the minimum distance between the centers of the circular holes; For each edge point, the point is mapped to the parameter space according to the set equation of the circle, in the parameter space, each edge point corresponds to a conical surface, and the number of votes of each parameter combination is recorded in the accumulator; Calculate all possible center and radius combinations of each edge point in the parameter space, and increase the number of votes in the corresponding position in the accumulator, find the local maximum value in the accumulator, filter the holes according to the threshold and the radius range of the hole, and remove the repeated hole detection results; The center coordinates and radius information of the filtered hole are extracted, and the hole feature boundary is drawn, which is used as the target area of the brake disc coating defect detection, and the multi-angle brake disc fusion image is cropped according to the boundary, and the image area irrelevant to the defect detection is cropped. 6.The pulse neural network based brake disc glue coating defect detection method of claim 1, wherein, In step (3), the step of obtaining the pre-trained detection model comprises: Collecting image data of different types of defect features in the coating area of the brake disc, labeling the position and type of the coating defect in the image, the defect type including uneven distribution of glue, glue drop, air hole mixing, glue missing and glue overflow, and the labeled information of each image and the corresponding image data are combined to form a training data set of the detection model, including brake disc coating defect images under different angles and different light conditions; A workspace is established in Pycharm software, model code and training code are written, a deep learning framework based on Pytorch and a pulse neural network related library snntorch are imported, a convolution-based pulse neural network module, a self-attention-based pulse neural network module and a target detection head are constructed, a data loader is used to load the training set data, an IoU loss function and a cross entropy loss are selected for coating defect positioning and classification tasks, and an SGD optimizer is set. During the model training process, the pulse method approximation method is used for training, the model parameters are updated, the model performance after each training is evaluated using the test set, the accuracy, recall rate and F1 score are calculated, the model hyperparameter settings are optimized and adjusted according to the evaluation results, and the model parameters with the best performance are saved as the pre-trained detection model.
7. The pulse neural network-based brake disc glue coating defect detection method of claim 6, wherein, The convolution-based pulse neural network module is used to extract local features and is composed of pulse separable convolution and channel convolution: The pulse separable convolution adopts a bottleneck structure, and the feature channel dimension is first expanded and then compressed through pointwise convolution, depth separable convolution and pointwise convolution, so as to refine the feature information of each channel and improve the comprehensiveness of the acquired fine-grained information; batch normalization layers and pulse neurons are added between the convolution layers, sparse convolution is realized through pulse-driven masks, and the model parameters and computational complexity are reduced. 8.The pulse neural network based brake disc glue coating defect detection method of claim 6, wherein, The pulse neural network module based on the Transformer is used to extract global features, and a convolution layer is added before calculating the self-attention to improve the integration effect of global features, and the convolution layer adopts a lightweight sparse pulse convolution operation; the module is based on a residual structure, and combines a pulse separable convolution layer, a pulse-driven self-attention layer and a channel multilayer perceptron: In the pulse-driven self-attention layer, linear operation and pulse neuron activation are used to generate the query matrix in the self-attention mechanism , the key matrix and the value matrix , instead of the energy-consuming re-parameterized convolution; in order to enhance the representation effect of the self-attention layer, the channel of is expanded times, and the original feature dimension is restored after the attention calculation is completed. 9.The pulse neural network based brake disc glue coating defect detection method of claim 6, wherein, The pulse firing approximation method trains the detection model in the following process: first, it is assumed that the pulse neuron transmits the pulse firing rate in the spatial dimension; second, the integer value activation is used to approximate the pulse firing rate in the training stage, which reduces the quantization error and improves the learning ability of the model; finally, the integer activation is converted into binary pulses in multiple time steps in the inference stage, and the time step length is expanded to achieve efficient inference.
10. A brake disc glue coating defect detection system based on an impulse neural network, characterized in that, The system is used to implement the brake disc coating defect detection method based on the pulse neural network according to any one of claims 1-9, and the system comprises: A multi-angle image acquisition module is used to capture the brake disc from different angles and acquire multi-angle image data of the coated brake disc. An image preprocessing module is used to preprocess the multi-angle brake disc image, fuse the multi-angle image through an image registration method, obtain comprehensive information of the coating area of the brake disc, and locate the target area of the coating defect detection according to the inherent circular hole feature on the brake disc, and crop the multi-angle brake disc fusion image. A model building and training module is used to build a detection model based on the pulse neural network, encode the brake disc image using convolution and self-attention mechanism, extract the coating defect information of the brake disc from the image based on pulse driving, train the detection model through the pulse firing approximation method, and obtain a pre-trained detection model with optimal performance. An automatic defect detection module is used to input the cropped multi-angle brake disc fusion image into the pre-trained detection model, identify the coating defect type of the brake disc, locate the defect position, and take corresponding processing measures.