A method for identifying abnormal high-speed moving filter rod blasting beads

Abstract

This invention discloses a method for identifying anomalies in high-speed moving filter beads. First, images of high-speed moving filter beads are acquired and preprocessed. Binarization is used to separate the background and foreground of the image, locating the target region for anomaly detection. The target location information and extracted texture features are fused and used as input to a target anomaly detection network. Then, normal samples are collected and labeled to generate a normal sample training set. A deep convolutional neural network integrating a CBAM attention mechanism module is built to automatically extract effective features of the target and provide anomaly detection results, achieving accurate detection of anomalies in high-speed moving targets. This method utilizes image preprocessing and feature extraction to ensure effective feature representation of the target region. The feature extraction method and network structure can be adjusted according to different sample types. Under training with normal samples, a high anomaly detection recognition rate can be guaranteed, achieving efficient detection of target anomalies.

Description

Technical Field

[0001] This invention belongs to the field of anomaly detection technology, specifically relating to a method for identifying anomalies in high-speed moving filter rods that cause beads to burst. Background Technology

[0002] Anomaly detection refers to the technique of training a specific model on a given normal dataset to determine whether test data is abnormal. If the data is similar to most data in the given dataset, it indicates that the data has no obvious anomalies. Conversely, if the data differs significantly from most data in the given dataset, it indicates that the data may be anomaly. With the improvement and development of high-speed camera performance and vision systems, high-speed object anomaly detection has become an application scenario in anomaly detection. Due to the blurry imaging and difficulty in locating high-speed moving objects, automatic detection of high-speed anomalous objects is of great significance in the field of industrial production.

[0003] Traditional methods for high-speed object anomaly detection include template matching, image decomposition, frequency domain analysis, and classification surface construction. However, their applicability is limited and they are not suitable for variable or complex scenes. In recent years, deep learning has made significant progress in various fields of computer vision. Compared with traditional methods, deep learning has been widely adopted in image anomaly detection tasks due to its advantages such as not requiring manual feature design and higher algorithm versatility. In particular, the emergence of various neural networks has provided effective algorithms and directions for high-speed object anomaly detection.

[0004] The literature "Ruff L, Vandermeulen RA, The deep support vector data description method proposed by N, Deepecke L, Siddiqui SA, Binder A, et al. Deep one-class classification. In: Proceedings of the 35th International Conference on Machine Learning. Stockholm, Sweden: PMLR, 2018. 4390-4399” is a commonly used method in deep learning. This method first manually designates a point in the feature space as the feature center, and then uses the sum of the distances from normal samples to this point as the main body of the loss function to train the feature extraction network. However, the feature center point needs to be manually specified and cannot be modified, which is prone to model degradation.

[0005] The paper "Schlegl, T., et al. Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery" (Springer, Cham (2017)) first applied Generative Adversarial Networks (GANs) to the field of anomaly detection. In the training phase, it uses given normal samples to train a standard GAN so that the generator learns the distribution of normal samples. In the subsequent testing phase, it selects a representation vector through an iterative method so that the normal sample generated by the vector is closest to the test sample. However, its model training is complex, and it has insufficient algorithm efficiency. In addition, it does not utilize other information in the image, which may lead to low quality of generated samples. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a method for identifying abnormal bead bursting in high-speed moving filter rods.

[0007] The technical solution adopted in this invention is: a method for identifying abnormal bursting of beads in a high-speed moving filter rod, the specific steps of which are as follows:

[0008] S1. Acquire images of high-speed moving filter rods bursting with beads;

[0009] S2. Preprocess the high-speed moving filter rod bursting bead image acquired in step S1.

[0010] S3. Perform binarization on the image obtained in step S2 to separate the foreground and background;

[0011] S4. The target location information and the extracted target texture features are fused together and used as the input to the target anomaly detection network;

[0012] S5. Collect and label normal samples to generate a normal sample training set;

[0013] S6. Combine the Convolutional Block Attention Module (CBAM) to construct a deep convolutional neural network;

[0014] S7. Training a deep neural network.

[0015] Furthermore, in step S1, a high-speed camera is used to acquire RGB images of the high-speed moving filter rod popping beads.

[0016] Furthermore, in step S2, the collected RGB images are converted to grayscale using an averaging method.

[0017] Further, in step S3, a threshold is set based on the background and foreground, and the pixel positions of the foreground are obtained using a binarization method. Returning the original image yields an RGB image containing only the target. The specific steps are as follows:

[0018] Set the grayscale value of background pixels in the image to 0 and the foreground (target) to 1. Then, set a threshold based on the grayscale values ​​of the foreground and background to separate them. The expression is as follows:

[0019]

[0020] Among them, piexl old and piexl new These represent the old and new values ​​of the pixel, respectively. Threshold represents the pre-set threshold, which can be set according to the features of the foreground and background. After binarizing the image, the pixel position of the foreground can be obtained. Then, the position is returned to the original image to obtain an RGB color image containing only the foreground. Then, feature extraction is performed on this image.

[0021] Furthermore, in step S4, different parameters are set to extract the texture features of the image target using a Gabor filter, and the target's location information is combined as input to the target anomaly detection network. The specific steps are as follows:

[0022] The general formula for the two-dimensional expression of the Gabor function is as follows:

[0023]

[0024] Where x and y represent two-dimensional coordinates, G(x,y) is the Gabor two-dimensional function expression, and λ represents the wavelength of the sine function. σ represents the phase shift, γ represents the aspect ratio (ellipticity of the Gabor function) in the x and y directions, and σ represents the phase shift. x σ y Let x' and y' represent the constants of the Gaussian envelope along the x and y directions, respectively. x' and y' must satisfy the following conditions:

[0025]

[0026]

[0027] Where b1 represents the bandwidth, θ represents the direction of the Gaussian function, and σ represents the standard deviation of the Gaussian function.

[0028] By processing the image using a series of Gabor functions with different parameters, different spatial feature outputs can be obtained. After Gaussian smoothing and normalization, the amplitude response of Gabor is extracted as a new classification feature, forming the final feature representation of the pixel, which is then input into the target anomaly detection network.

[0029] Furthermore, in step S5, all images obtained in steps S1-S4 are labeled to construct a training set of normal samples.

[0030] Furthermore, in step S6, the deep neural network mainly consists of convolutional layers and a CBAM attention mechanism module. After obtaining the target feature matrix in step S4, it passes through pooling layers and convolutional layers, with two shortcut connections designed in between. Then, it passes through the CBAM attention mechanism module, which assigns adaptive weights to each position and channel of the feature map. After passing through two convolutional layers, the final feature map is obtained. After iterative training in step S7, the features automatically extracted by the neural network are input into the fully connected layer to obtain the classification label of the image sample.

[0031] Further, in step S7, the normal sample training set obtained in step S5 is input into the deep convolutional neural network constructed in step S6 for forward propagation, the loss function value is calculated, and the parameters of the deep convolutional neural network are updated using a gradient descent-based backpropagation algorithm. The forward and backward propagation is performed iteratively until the loss function converges. The specific steps are as follows:

[0032] S71, Forward Propagation

[0033] by Let represent the t-th feature map of the l-th layer (l≥2). If the l-th layer is a convolutional layer, then:

[0034]

[0035] in, This represents the convolution kernel that connects the s-th input feature map and the t-th output feature map. φ represents the bias term, φ(·) represents the nonlinear activation function, and the symbol “*” represents the convolution operation.

[0036] If the l-th layer is a max-pooling layer, then:

[0037]

[0038] in, This represents the t-th feature map of the l-th layer (l≥2). The coordinates of the position, r1 and r2 represent the size of the pooling window, and u1 and u2 represent the variables pre-set within the pooling window.

[0039] If the l-th layer is an average pooling layer, then:

[0040]

[0041] Where r1 and r2 represent the size of the pooling window.

[0042] If the l-th layer is a fully connected layer, then:

[0043] ν (l) =φ(w (l) ν (l-1) +b (l) (8)

[0044] Where, ν (l) w represents the feature map of the l-th layer. (l) Let b represent the weights of the l-th layer. (l) This is the bias term for this layer.

[0045] If the Lth layer is the output layer, then the posterior probability p(i|z(i) of the current sample belonging to the i-th class is... L ))for:

[0046]

[0047] Among them, z (L) This represents the input to the classification output layer. This represents the input of the i-th category in the classification output layer, where c is a pre-defined variable. This represents the input of the c-th category in the classification output layer, where C represents the total number of categories.

[0048] S72. Calculate the loss function value.

[0049] The loss function is constructed based on the binary cross-entropy function, and its calculation formula is as follows:

[0050]

[0051] Where LF(w) represents the loss function, w represents the set of weights and biases in the network, n represents the total number of samples, and y i The sample categories are represented by 0 for normal samples and 1 for abnormal samples. A represents the output feature map matrix, and its 1-norm is defined as follows: λ1 and λ2 represent the size of the output feature map matrix.

[0052] S73. The backpropagation algorithm based on gradient descent updates the network parameters, as shown in the following formula:

[0053]

[0054]

[0055] Where α is the learning rate, LF represents the loss function, and w old w new Let b represent the set of old and new weights in the network. old b new This represents the set of old and new values ​​of the bias term in the network.

[0056] Furthermore, the anomaly identification method also includes step S8, which involves testing the detection performance of the network trained in step S7, inputting the test sample into the network trained in step S7 for forward propagation to obtain the final feature map of the test sample, determining whether the sample is a normal sample or an abnormal sample based on whether there is an abnormal score, and then locating the abnormal region by upsampling back to the original image size.

[0057] The beneficial effects of this invention are as follows: The method of this invention first acquires images of high-speed moving filter beads bursting, preprocesses the images, and uses binarization to separate the background and foreground, locating the target region requiring anomaly detection. The target location information and extracted target texture features are fused and used as input to the target anomaly detection network. Then, normal samples are collected and labeled to generate a normal sample training set. A deep convolutional neural network integrating a CBAM attention mechanism module is built, automatically extracting effective features of the target and providing anomaly detection results, thus achieving accurate detection of anomalies in high-speed moving targets. This invention utilizes image preprocessing and feature extraction to ensure effective feature representation of the target region. The feature extraction method and network structure can be adjusted according to different sample types. Under training with normal samples, it can guarantee a high anomaly detection recognition rate, achieving efficient detection of target anomalies. Attached Figure Description

[0058] Figure 1 This is a flowchart of a method for identifying abnormal bursting beads in a high-speed moving filter rod according to the present invention.

[0059] Figure 2 This is a schematic diagram of a binarized segmented image provided in an embodiment of the present invention.

[0060] Figure 3 This is a schematic diagram of a neural network mapping provided in an embodiment of the present invention.

[0061] Figure 4 This is a schematic diagram of a neural network structure provided in an embodiment of the present invention.

[0062] Figure 5 This is a schematic diagram of the CBAM self-attention mechanism module provided in an embodiment of the present invention.

[0063] Figure 6 This is a schematic diagram of the anomaly detection results provided in an embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0065] like Figure 1 The flowchart of the method for identifying abnormal burst beads in a high-speed moving filter rod according to the present invention is shown below. The specific steps are as follows:

[0066] S1. Acquire images of high-speed moving filter rods bursting with beads;

[0067] S2. Preprocess the high-speed moving filter rod bursting bead image acquired in step S1.

[0068] S3. Perform binarization on the image obtained in step S2 to separate the foreground and background;

[0069] S4. The target location information and the extracted target texture features are fused together and used as the input to the target anomaly detection network;

[0070] S5. Collect and label normal samples to generate a normal sample training set;

[0071] S6. Combine the CBAM self-attention mechanism module to construct a deep convolutional neural network;

[0072] S7. Training a deep neural network.

[0073] In this embodiment, in step S1, a high-speed camera is used to acquire RGB images of the high-speed moving filter rod popping beads.

[0074] In this embodiment, in step S2, the collected RGB image is converted to grayscale using the averaging method, transforming the three-channel image into a single channel to facilitate subsequent binarization processing.

[0075] In this embodiment, in step S3, the grayscale value of the background pixels in the image is set to 0, and the grayscale value of the foreground (target) is set to 1. A threshold is set based on the grayscale values ​​of the foreground and background to separate the foreground and background, as shown in the following expression:

[0076]

[0077] Among them, piexl old and piexl new These represent the old and new values ​​of the pixel, respectively. `threshold` represents a pre-set threshold, which can be set according to the features of the foreground and background. In this embodiment, it is set to 125. After binarizing the image, the pixel positions of the foreground can be obtained. Figure 2This is a schematic diagram of a binarized segmented image provided in an embodiment of the present invention. It can be seen that the image after binarization can accurately provide the location information of the target. Then, the location is returned to the original image to obtain an RGB color image containing only the foreground. Then, feature extraction is performed on the image.

[0078] In this embodiment, in step S4, different parameters are set to extract the texture features of the image target using a Gabor filter, and the target's location information is used as the input to the target anomaly detection network.

[0079] Pixel values ​​in the three color dimensions are the most direct numerical features of an RGB image, but the information they contain is far from sufficient for anomaly detection. Therefore, further feature extraction is needed to enhance the information representation ability of the samples. Fusing the target's location information with the extracted image spatial features can provide more effective features for classifier training. Furthermore, there are many spatial features that can be used to describe images, including Gabor features, pixel gradient features, and morphological profile features. Among these, Gabor features are widely used as spatial features.

[0080] Gabor features are a type of feature used to describe image texture information. They utilize the Gabor function to window the signal in the frequency domain, thereby describing the signal's local frequency information. Different Gabor functions are obtained by modulating different Gaussian functions with a sine function. Their two-dimensional general expression is as follows:

[0081]

[0082] Where x and y represent two-dimensional coordinates, G(x,y) is the Gabor two-dimensional function expression, and λ represents the wavelength of the sine function. σ represents the phase shift, γ represents the aspect ratio in the x and y directions, and σ represents the phase offset. x σ y Let x' and y' represent the constants of the Gaussian envelope along the x and y directions, respectively. x' and y' must satisfy the following conditions:

[0083]

[0084]

[0085] Where b1 represents the bandwidth, θ represents the direction of the Gaussian function, and σ represents the standard deviation of the Gaussian function.

[0086] By processing the image using Gabor functions with different parameters, different spatial features can be obtained. In this embodiment, b1 is set to 1, wavelength λ is set to 2, and θ takes values ​​of 0°, 30°, 60°, and 90° respectively. With the phase offset set to 0 and γ set to 1, after Gaussian smoothing and normalization, the Gabor amplitude response is extracted as a new classification feature to form the final feature representation of the pixel, which is then input into the target anomaly detection network.

[0087] In this embodiment, in step S5, all images obtained in steps S1-S4 are labeled to construct a training set of normal samples.

[0088] In this embodiment, the deep neural network used in step S6 is an end-to-end convolutional neural network. By introducing the CBAM attention mechanism module, image features are automatically extracted and assigned adaptive weights, enabling greater focus on features and information of interest. Figure 3 As shown in the schematic diagram of the neural network structure provided in this embodiment of the invention, solid dots represent normal targets, hollow dots represent abnormal targets, χ represents the original feature space, η represents the feature space after neural network feature mapping, and δ(·) represents the neural network mapping. The principle is to map the features of normal targets into a clustering circle, separating them from the features of abnormal targets. After iterative training in step S7, the features automatically extracted by the neural network are input into the fully connected layer to obtain the classification labels of the image samples.

[0089] like Figure 4 The diagram shows the specific network structure of this embodiment, where the convolutional layer is represented as "Conv.(number of feature maps)@(kernel size)", the global pooling layer is represented as "Max-Pool", the average pooling layer is represented as "Avg-Pool", and the fully connected layer is represented as "FC".

[0090] In this embodiment, the image features extracted by S4 are used as input. The output of this input is summed with the output of a convolutional layer and a pooling layer, and then with the output of two more convolutional layers. This structure is based on the shortcut method proposed by ResNet, aiming to prevent degradation. The designed neural network has two shortcut connections, and then a CBAM attention mechanism module is introduced in the middle. Figure 5The diagram shows the specific network structure of this module, which aims to assign adaptive weights to both channels and positions of the feature map to enhance the features of interest. The CBAM attention mechanism module consists of two sub-modules: CAM and SAM. In CAM, the input undergoes max pooling and average pooling, then passes through a fully connected layer. The "+" sign indicates addition at corresponding positions, and finally, the output is processed by a sigmoid activation function. In SAM, the input also undergoes max pooling and average pooling, but this module concatenates the channel numbers together before passing them through a convolutional layer and a sigmoid activation function. It's important to note that ReLU activation and BN normalization are used in each neural network layer, resulting in a one-channel feature output map. The classification result is then obtained through a fully connected layer and a softmax function.

[0091] In this embodiment, in step S7, the normal sample training set obtained in step S5 is input into the deep convolutional neural network constructed in step S6 for forward propagation, the loss function value is calculated, and the parameters of the deep convolutional neural network are updated using a gradient descent-based backpropagation algorithm. The forward and backward propagation is performed iteratively until the loss function converges. The specific steps are as follows:

[0092] S71, Forward Propagation

[0093] by Let represent the t-th feature map of the l-th layer (l≥2). If the l-th layer is a convolutional layer, then:

[0094]

[0095] in, This represents the convolution kernel that connects the s-th input feature map and the t-th output feature map. φ represents the bias term, φ(·) represents the nonlinear activation function, and "*" represents the convolution operation.

[0096] If the l-th layer is a max-pooling layer, then:

[0097]

[0098] in, This represents the t-th feature map of the l-th layer (l≥2). The coordinates of the position, r1 and r2 represent the size of the pooling window, and u1 and u2 represent the variables pre-set within the pooling window.

[0099] If the l-th layer is an average pooling layer, then:

[0100]

[0101] Where r1 and r2 represent the size of the pooling window;

[0102] If the l-th layer is a fully connected layer, then:

[0103] ν (l) =φ(w (l) ν (l-1) +b (l) (8)

[0104] Where, ν (l) w represents the feature map of the l-th layer. (l) Let b represent the weights of the l-th layer. (l) This is the bias term for this layer.

[0105] If the Lth layer is the output layer, then the posterior probability p(i|z) of the current sample belonging to the i-th class is... (L) )for:

[0106]

[0107] Among them, z (L) This represents the input to the classification output layer. This represents the input of the i-th category in the classification output layer, where c is a pre-defined variable. This represents the input of the c-th category in the classification output layer, where C represents the total number of categories.

[0108] S72. Calculate the loss function value.

[0109] The loss function is constructed based on the binary cross-entropy function, and its calculation formula is as follows:

[0110]

[0111] Where LF(w) represents the loss function, w represents the set of weights in the network, n represents the total number of samples, and y i The sample categories are represented by 0 for normal samples and 1 for abnormal samples. A represents the output feature map matrix, and its 1-norm is defined as follows: λ1 and λ2 represent the size of the output feature map matrix.

[0112] S73. The backpropagation algorithm based on gradient descent updates the network parameters, as shown in the following formula:

[0113]

[0114]

[0115] Where α is the learning rate, LF represents the loss function, and w old w new Let b represent the set of old and new weights in the network. old b new This represents the set of old and new values ​​of the bias term in the network.

[0116] In this embodiment, the anomaly identification method further includes step S8, which involves testing the detection performance of the network trained in step S7, inputting the test sample into the network trained in S7 for forward propagation to obtain the final feature map of the test sample, determining whether the sample is a normal sample or an abnormal sample based on whether there is an abnormal score, and then locating the abnormal region by upsampling back to the original image size.

[0117] Table 1 shows the number of images used for network training and testing in the embodiments, and Table 2 shows the identification results of the test set obtained in the embodiments. Figure 6 This diagram illustrates the anomaly detection results provided in an embodiment of the present invention. The quantitative results show that the method of the present invention can be trained using normal samples to achieve the target anomaly detection task, with an average target discrimination rate of up to 99.08%. The results also demonstrate that the method of the present invention can effectively detect target anomalies.

[0118] Table 1

[0119]

[0120] Table 2

[0121] Real / Predicted normal samples Abnormal samples normal samples 99.55％ 0.45％ Abnormal samples 1.40％ 98.60％

[0122] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for identifying abnormal bursting beads in a high-speed moving filter rod, the specific steps of which are as follows: S1. Acquire images of high-speed moving filter rods bursting with beads; In step S1, a high-speed camera is used to capture RGB images of the high-speed moving filter rod popping beads. S2. Preprocess the high-speed moving filter rod bursting bead image acquired in step S1. In step S2, the collected RGB images are converted to grayscale using the averaging method. S3. Perform binarization on the image obtained in step S2 to separate the foreground and background; In step S3, a threshold is set based on the background and foreground, and the pixel position of the foreground is obtained by binarization processing. The original image is then returned to obtain an RGB image containing only the target. S4. The target location information and the extracted target texture features are fused together and used as the input to the target anomaly detection network; In step S4, different parameters are set to extract the texture features of the image target using a Gabor filter, and the target's location information is used as the input to the target anomaly detection network. S5. Collect and label normal samples to generate a normal sample training set; In step S5, all images obtained in steps S1-S4 are labeled to construct a training set of normal samples. S6. Combine the CBAM self-attention mechanism module to construct a deep convolutional neural network; In step S6, after obtaining the target feature matrix from step S4, it passes through a pooling layer and a convolutional layer, and then through the CBAM attention mechanism module, which assigns adaptive weights to each position and channel of the feature map. Finally, it passes through two convolutional layers to obtain the final feature map. S7. Training a deep neural network; The specific steps in step S7 are as follows: The normal sample training set obtained in step S5 is input into the deep convolutional neural network constructed in step S6 for forward propagation, the loss function value is calculated, and the parameters of the deep convolutional neural network are updated using a gradient descent-based backpropagation algorithm. The forward and backward propagation is iterated until the loss function converges. After iterative training in step S7, the features automatically extracted by the neural network are input into the fully connected layer to obtain the classification labels of the image samples.

2. The method for identifying abnormal bursting beads in a high-speed moving filter rod according to claim 1, characterized in that, In step S3, the specific steps are as follows: Set the grayscale value of background pixels in the image to 0 and the grayscale value of foreground pixels to 1. Then, set a threshold based on the grayscale values ​​of the foreground and background to separate them. The expression is as follows: (1)； in, and These represent the old and new values ​​of the pixel, respectively, and threshold represents the pre-set threshold. After binarizing the image, the pixel positions of the foreground are obtained, and then the positions are returned to the original image to obtain an RGB color image containing only the foreground. Then, feature extraction is performed on this image.

3. The method for identifying abnormal bursting beads in a high-speed moving filter rod according to claim 1, characterized in that, In step S4, the specific steps are as follows: The general formula for the two-dimensional expression of the Gabor function is as follows: (2)； Where x and y represent two-dimensional coordinates, This is a two-dimensional Gabor function expression. The wavelength of the sine function is represented. Indicates phase shift, express and Aspect ratio in two directions Let represent the constants of the Gaussian envelope along the x and y directions, respectively. , Conditions to be met: (3)； (4)； in, Indicates bandwidth. Indicates the direction of the Gaussian function. This represents the standard deviation of the Gaussian function.

4. The method for identifying abnormal bursting beads in a high-speed moving filter rod according to claim 1, characterized in that, The specific steps in step S7 are as follows: S71, Forward Propagation by Indicates the first Layer Each feature map ; If the first If the layer is a convolutional layer, then: (5)； in, Indicates the connection of the first The input feature map and the first input feature map Convolutional kernels for each output feature map, Indicates the bias term. Represents a non-linear activation function, symbol " " indicates convolution operation; If the first If the layer is a max pooling layer, then: (6)； in, Indicates the first Layer Each feature map represents The coordinates of the location. , Indicates the size of the pooling window. This represents variables that are pre-defined within the pooling window; If the first If the layer is an average pooling layer, then: (7)； in, , Indicates the size of the pooling window; If the first If the layer is a fully connected layer, then: (8)； in, Indicates the first Feature map of the layer Indicates the first Layer weights, This is the bias term for this layer; If the first If the layer is the output layer, then the current sample belongs to the th layer. Posterior probability of class for: (9)； in, This represents the input to the classification output layer. This represents the input of the i-th category in the classification output layer. For a pre-defined variable, Indicates the classification output layer number 1 Input of each category, Indicates the total number of categories; S72. Calculate the loss function value. The loss function is constructed based on the binary cross-entropy function, and its calculation formula is as follows: (10)； in, Represents the loss function. Let n represent the set of weights and biases in the network, where n represents the total number of samples. This indicates the sample category, with 0 for normal samples and 1 for abnormal samples. The output feature map matrix is ​​defined as follows: , and Indicates the size of the output feature map matrix; S73. The backpropagation algorithm based on gradient descent updates the network parameters, as shown in the following formula: (11)； (12)； in, For learning rate, Represents the loss function. , This represents the set of old and new weights in the network. , This represents the set of old and new values ​​of the bias term in the network.

5. The method for identifying abnormal bursting beads in a high-speed moving filter rod according to claim 1, characterized in that, The anomaly detection method further includes step S8, which involves testing the detection performance of the network trained in step S7. The specific steps are as follows: The test sample is input into the network trained in step S7 for forward propagation to obtain the final feature map of the test sample. Based on whether there is an abnormal score, the sample is determined to be a normal sample or an abnormal sample. Then, the abnormal region is located by upsampling back to the original image size.

Images