Intelligent inspection method and system for water seal cutter of annular heating furnace

Abstract

The invention discloses an intelligent inspection method and system for a water seal cutter of an annular heating furnace, and the method comprises the following steps: collecting image data of the water seal cutter, carrying out the preprocessing of the image data, and transmitting the image data into a pre-trained surface abnormality detection model; and the pre-trained surface anomaly detection model detects whether there is an anomaly in the water seal cutter image based on an image identification mode and marks an abnormal part in the water seal cutter image in a visual form to present the water seal cutter image with an abnormal mark. According to the scheme, the water seal cutter can be automatically, accurately and reliably inspected.

Description

Technical Field

[0001] The invention relates to the field of railway wheel equipment tooling, and in particular to an intelligent inspection method for a water seal knife of an annular heating furnace. Background Art

[0002] The water seal system of the annular heating furnace includes a water seal knife, a water seal groove and medium water to isolate the gas inside and outside the heating furnace and reduce heat loss. The water seal knife is in a high temperature environment and soaked in water for a long time. It is easy to deform or rust and break after a period of use. If the water seal knife is deformed and damaged, it cannot well isolate the gas inside and outside the furnace, which will reduce the energy efficiency of the heating furnace and even cause safety hazards due to CO gas leakage. Therefore, it is necessary to detect abnormal surface conditions of the water seal knife.

[0003] The prior art does not have a clear automatic detection solution for abnormal detection of water seal knives. The general method is to detect the abnormality of the water seal knife through manual visual inspection when the water seal knife rotates slowly or when it is shut down. However, this detection method relies on visual inspection, which is limited by the fact that the water seal knife is always rotating and cannot be accurately judged. If it is shut down, it will affect the production progress, and in the field of industrial production, it is not possible to shut down at will. Therefore, the abnormal detection of water seal knives in the prior art has the defect that it cannot be accurately and reliably inspected. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an intelligent inspection method for the water seal knife of an annular heating furnace. The abnormal state of the water seal knife is checked by identifying the image of the water seal knife through machine vision, thereby realizing automatic, accurate and reliable inspection of the water seal knife.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: an intelligent inspection method for water seal knife of annular heating furnace, comprising the following steps:

[0006] The water seal knife image data is collected and preprocessed and then sent to the pre-trained surface anomaly detection model. The pre-trained surface anomaly detection model detects whether there is an anomaly in the water seal knife image based on image recognition and marks the abnormal part in the water seal knife image to present the water seal knife image with abnormal marks in a visual form.

[0007] The pre-trained surface anomaly detection model includes surface anomaly detection model structure construction, model loss function selection and training; wherein the structure of the surface anomaly model is built based on the surface anomaly detection model structure construction, and then training is performed after selecting the loss function to obtain the pre-trained surface anomaly detection model.

[0008] The constructed surface anomaly detection model includes a teacher network, a student network and an autoencoder network; both the teacher network and the student network use an efficient feature extractor PDN containing 4 convolutional layers; the autoencoder network includes an encoder layer and a decoder layer to output surface anomaly detection results.

[0009] The teacher network and the student network use a feature extractor of 4 convolutional layers, whose convolution kernel dimensions are 4*4*128, 4*4*256, 3*3*256, 4*4*256, and the convolution step size is 1*1; PDN achieves downsampling by using a 2*2*128 average pooling layer after the first and second convolutional layers; the output dimension of the fully connected layer after the 4 convolutional layers is 384. In the network output, each output contains a receptive field of 33X33 pixels.

[0010] The autoencoder network is composed of an encoder layer containing 6 convolutional layers and a decoder layer containing 8 convolutional layers. The convolution kernel dimensions of the encoder layer are 4*4*32, 4*4*32, 4*4*75, 4*4*64, 4*4*64, and 8*8*64 respectively; the convolution kernel dimension of the decoder layer is uniformly 4*4*64, and the output dimension of the fully connected layer after the decoder layer is 384.

[0011] The teacher network is distilled from a WideResNet-101 network on the ImageNet dataset, and the loss function is MSE;

[0012] The training of the student network and the autoencoder is implemented by a comprehensive loss function. The comprehensive loss function L total for

[0013] L total =L ST +L AE +L STAE ;

[0014] Where: L ST It is a measure of the difference between the output of the student network and the output of the teacher network;

[0015] L AE It measures the difference between the output of the autoencoder and the output of the teacher network;

[0016] L STAE is a measure of the difference between the output of the autoencoder and another student network.

[0017] After the surface anomaly detection model detects an abnormality in the water seal knife, it outputs an alarm signal to the heating furnace inspection robot, which is used to link the robot to the on-site re-inspection.

[0018] An intelligent inspection system for water seal blades of an annular heating furnace comprises a data acquisition module, a data preprocessing module, a surface anomaly detection model and a visualization module; wherein the data acquisition module monitors the surface of the water seal blade in real time and collects its image data, and its output end is connected to the data preprocessing module, the preprocessing module preprocesses the collected water seal blade image data and then sends it to the surface anomaly detection model, the surface anomaly detection model recognizes the water seal blade image, and sends the recognition result to the visualization module for visual display.

[0019] The output end of the surface anomaly detection model is connected to the heating furnace inspection robot, which is used to link the robot to go to the site to recheck the water seal knife anomaly after the anomaly is detected.

[0020] The surface anomaly detection model is a pre-built image recognition model and is used to identify water seal knife image anomalies after pre-training.

[0021] The advantages of the present invention are:

[0022] 1. This solution uses a highly efficient feature extraction network PDN, which has a faster reasoning speed and meets the detection speed requirements of this scenario.

[0023] 2. This solution introduces a difficult feature loss function during the training of the student network. Through this loss function, the student network can fully imitate the teacher on normal images while avoiding generalization to abnormal images. This reduces the algorithm's misrecognition rate for structural anomalies.

[0024] 3. This scheme not only designs a teacher-student network structure that can detect local feature differences for structural anomalies, but also designs a student-autoencoder network structure that can detect global feature differences for logical anomalies. This enables the algorithm to detect not only surface damage but also surface cracks.

[0025] 4. This solution provides for the first time an unmanned inspection method for the water seal system of an annular high-temperature heating furnace, which avoids workers from working in high-temperature environments on site and avoids personnel from being exposed to gas hazardous areas, thus achieving the inherent safety of equipment inspection.

[0026] 5. This solution provides an automated status detection method for the furnace bottom water seal system. Through statistical analysis of historical data, more accurate maintenance standards for the heating furnace water seal system can be formulated, providing a basis for heating furnace maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The following is a brief description of the contents expressed in the drawings of the present specification and the marks in the drawings:

[0028] Figure 1 It is a structural diagram of the water seal knife surface abnormality recognition model of the present invention. DETAILED DESCRIPTION

[0029] The specific implementation of the present invention will be further explained in detail below by describing the optimal embodiment with reference to the accompanying drawings.

[0030] Embodiment 1:

[0031] In order to realize the intelligent inspection of the water seal knife of the annular heating furnace and check the deformation and damage of the water seal knife surface, this scheme adopts the image recognition method, and the specific scheme is as follows:

[0032] An intelligent inspection system for water seal cutters of an annular heating furnace, comprising a data acquisition module, a data preprocessing module, a surface anomaly detection model, and a visualization module;

[0033] The data acquisition module monitors the surface of the water seal knife in real time and collects its image data, and its output end is connected to the data preprocessing module; the data acquisition module includes a camera, and a camera is installed in the inner ring of the annular furnace to monitor the water seal knife in real time, so as to obtain the surface image data of the water seal knife;

[0034] The preprocessing module preprocesses the collected water seal knife image data and then sends it to the surface anomaly detection model; the preprocessing module preprocesses the collected image data, including noise reduction, contrast enhancement, color correction, etc., to improve the image quality and facilitate subsequent anomaly detection.

[0035] The surface anomaly detection model recognizes the water seal knife image and sends the recognition result to the visualization module for visualization. The surface anomaly detection model is an easy-to-use image recognition model. The image recognition analysis model algorithm is used to analyze and recognize the surface graphics of the water seal knife to obtain the recognition result to determine whether there are abnormal conditions such as deformation and damage of the water seal knife. The surface anomaly detection model is an image recognition model. After the image recognition model is pre-built, the model is constructed through training. The trained model can be used to identify abnormal conditions of the water seal knife.

[0036] The visualization module can display the abnormal water seal knife surface image in various ways such as a display screen, and can mark the abnormal position identified in the image or visualize it in the form of a heat map.

[0037] In a preferred solution, the output end of the surface anomaly detection model is connected to the heating furnace inspection robot, which is used to link the robot to the site to review the abnormality of the water seal knife after detecting the abnormality. The surface anomaly detection model can be integrated into the heating furnace inspection robot, and the uploaded water seal knife surface image data can be recognized to automatically determine whether the water seal knife is deformed, and an alarm can be issued in time if there is an abnormality. At the same time, the robot can also be linked to the site to review the deformation of the water seal knife.

[0038] When the water seal knife surface is photographed through the image acquisition module, since the water seal knife rotates in real time, in order to capture clearer image data, it is necessary to collect the rotation speed of the water seal knife, collect the rotation speed data of the water seal knife through the rotation speed sensor, and set the shutter speed of the camera or video camera based on the comparison table of the water seal knife rotation speed data and shutter data, thereby improving the clarity of the image capture. Since it is impossible to capture a clear surface image when the shutter speed is low, it is necessary to dynamically adjust the photography, obtain the rotation speed of the water seal knife through the rotation speed sensor, etc., pre-calibrate the comparison table between the speed and shutter time, and when collecting image data, obtain the rotation speed data of the water seal knife, search the speed and shutter time comparison table based on the rotation speed data to obtain the shutter time corresponding to the rotation speed data, and collect the shutter time to capture the image to obtain clear surface image data.

[0039] Embodiment 2:

[0040] This embodiment provides an image recognition model based on the first embodiment, namely, a water seal knife surface abnormality detection model. Based on the model provided in this embodiment, the recognition and judgment of the water seal knife surface abnormality can be realized more accurately. The specific introduction is as follows:

[0041] An intelligent inspection method for a ring heating furnace water seal knife comprises the following steps: collecting water seal knife image data, pre-processing the data and then sending the data to a pre-trained surface anomaly detection model; the pre-trained surface anomaly detection model detects whether there is an anomaly in the water seal knife image based on image recognition and marks the abnormal part in the water seal knife image to present the water seal knife image with the abnormal mark in a visual form.

[0042] The pre-trained surface anomaly detection model includes surface anomaly detection model structure construction, model loss function selection and training; wherein the structure of the surface anomaly model is built based on the surface anomaly detection model structure construction, and then training is performed after selecting the loss function to obtain the pre-trained surface anomaly detection model.

[0043] The constructed surface anomaly detection model includes a teacher network, a student network and an autoencoder network; both the teacher network and the student network use an efficient feature extractor PDN containing 4 convolutional layers; the autoencoder network includes an encoder layer and a decoder layer to output surface anomaly detection results.

[0044] The teacher network and the student network use a feature extractor of 4 convolutional layers, whose convolution kernel dimensions are 4*4*128, 4*4*256, 3*3*256, 4*4*256, and the convolution step size is 1*1; PDN achieves downsampling by using a 2*2*128 average pooling layer after the first and second convolutional layers; the output dimension of the fully connected layer after the 4 convolutional layers is 384. In the network output, each output contains a receptive field of 33X33 pixels.

[0045] The autoencoder network is composed of an encoder layer containing 6 convolutional layers and a decoder layer containing 8 convolutional layers. The convolution kernel dimensions of the encoder layer are 4*4*32, 4*4*32, 4*4*75, 4*4*64, 4*4*64, and 8*8*64 respectively; the convolution kernel dimension of the decoder layer is uniformly 4*4*64, and the output dimension of the fully connected layer after the decoder layer is 384.

[0046] The teacher network is distilled from a WideResNet-101 network on the ImageNet dataset, and the loss function is MSE;

[0047] The training of the student network and the autoencoder is implemented by a comprehensive loss function. The comprehensive loss function L total for

[0048] L total =L ST +L AE +L STAE ;

[0049] Where: L ST It is a measure of the difference between the output of the student network and the output of the teacher network;

[0050] L AE It measures the difference between the output of the autoencoder and the output of the teacher network;

[0051] L STAE is a measure of the difference between the output of the autoencoder and another student network.

[0052] After the surface anomaly detection model detects an abnormality in the water seal knife, it outputs an alarm signal to the heating furnace inspection robot, which is used to link the robot to the on-site re-inspection.

[0053] This technology provides a visual detection algorithm for water seal knife surface anomalies. Through water seal knife image data collection, data preprocessing, surface anomaly detection model structure construction, model loss function selection and training, surface anomaly detection model reasoning, the detected anomalies are presented to the user in a visual way, including the location of the anomaly, severity and other information. Specifically, the visualization is performed in the form of a heat map.

[0054] The implementation objectives of this technology are: first, to improve the accuracy of the detection of the water seal system of the annular heating furnace and the health trend analysis, and to carry out preventive maintenance; second, to replace manual inspections of equipment, reduce the workload of personnel, and achieve less-manpowered operations; third, to reduce the frequency of personnel entering the annular furnace area, stay away from the gas area, and reduce the safety risks of personnel operations.

[0055] A camera is installed inside the annular furnace to monitor the water seal knife in real time. In order to detect the deformation of the water seal knife more intelligently, the data collected by the fixed camera on site is uploaded to the heating furnace inspection robot. The robot adds a built-in image recognition algorithm to recognize the uploaded image data, and automatically determines whether the water seal knife is deformed. If there is an abnormality, an alarm will be issued in time. At the same time, the robot can also be linked to the site to review the deformation of the water seal knife. The track-type robot moves to the designated inspection point to take photos for review.

[0056] The specific image analysis algorithm is as follows:

[0057] 1. Data acquisition: Use high-resolution image acquisition equipment to monitor the surface of the water seal knife in real time and collect its image data.

[0058] 2. Data preprocessing: The collected image data is preprocessed, including noise reduction, contrast enhancement, color correction, etc., to improve image quality and facilitate subsequent anomaly detection.

[0059] 3. Surface anomaly detection model structure construction

[0060] The surface anomaly detection model consists of three parts: teacher network, student network and autoencoder network.

[0061] Both the teacher network and the student network use an efficient feature extractor consisting of 4 convolutional layers, referred to as PDN.

[0062] The dimensions of the convolution kernels of the four convolutional layers of PDN are 4*4*128, 4*4*256, 3*3*256, 4*4*256, and the stride of the convolution is 1*1. PDN achieves downsampling by using a 2*2*128 average pooling layer after the first and second convolutional layers. The output dimension of the fully connected layer after the four convolutional layers is 384. In the network output, each output has a receptive field of 33X33 pixels.

[0063] The autoencoder network is composed of an encoder layer with 6 convolutional layers and a decoder layer with 8 convolutional layers. The dimensions of the convolution kernels of the encoder layer are 4*4*32, 4*4*32, 4*4*75, 4*4*64, 4*4*64, and 8*8*64 respectively; while the dimensions of the convolution kernels of the decoder layer are uniformly 4*4*64. The output dimension of the fully connected layer following the decoder layer is 384.

[0064] Model structure principle:

[0065] 1. Lightweight feature extractor: First, use a pre-trained lightweight neural network to extract image features, which are used to describe the local information of the image (such as texture, color, etc.).

[0066] 2. Student-teacher network structure: During the training phase, the teacher network extracts features of normal images, and the student network is trained to mimic the output of the teacher network on these normal images. During the inference phase, since the student network has never seen abnormal images, it cannot well predict the abnormal image features extracted by the teacher network, thereby achieving anomaly detection.

[0067] 3. Autoencoder: In order to detect more complex logical anomalies (such as object order errors), this model introduces an autoencoder to learn the global information of the image. This can help detect incorrect combinations or arrangements in the image.

[0068] How the model works:

[0069] 1. Image input: The input image is first sent to a lightweight feature extractor to extract local features.

[0070] 2. Local feature processing: These features are then passed to the student-teacher model, where the student network tries to imitate the output of the teacher network. If the student network fails to successfully imitate the output of the teacher network, it means that there are local anomalies in the image.

[0071] 3. Logical anomaly detection: For complex logical anomalies, the image is also passed to the autoencoder. The autoencoder detects logical anomalies by reconstructing the image, and the part where the reconstruction fails may represent anomalies.

[0072] 4. Model loss function selection and training process

[0073] For the training of the teacher network, its PDN is distilled from a WideResNet-101 network on the ImageNet dataset, and the loss function is MSE.

[0074] The training of the student network and the autoencoder is implemented by a comprehensive loss function, which contains the following three loss function sub-parts:

[0075] L total =L ST +L AE +L STAE ;

[0076] The first part, the loss function Lst, measures the difference between the output of the student network and the output of the teacher network. This allows the student network to fully imitate the teacher on normal images while avoiding generalization on abnormal images. Since both the student network and the teacher network extract local features, this part is mainly used to improve the model's ability to detect local anomalies, that is, structural anomalies.

[0077] L ST =(CWH) -1 ∑ c ||T(I) c -S(I) c || 2 F ; CWH represents the number of feature map channels,

[0078] The width and height of the feature map. T(I) represents the feature map obtained by the teacher network after the input image passes through

[0079] Feature map, S(I) represents the feature map obtained after the input image passes through the student network.

[0080] Training of the student network. In the standard ST framework, increasing the number of training images can improve the student's ability to imitate the teacher in abnormal situations, which will deteriorate the anomaly detection performance. At the same time, intentionally reducing the number of training images may suppress important information about normal images. Similar to Online Hard Example Mining, the design of this scheme is to limit the loss of the student network to the most relevant part of the image. Specifically, a difficult feature loss is used, and only the part of the output feature with the largest loss is calculated for backpropagation.

[0081] For example, the normal loss is calculated as Dc,w,h=(T(I)cwh-S(I)cwh). T represents the teacher network, S represents the student network, I represents the input image, and cwh represents the number of channels, image width, and image height. Here, a Phard~[0,1] is set, such as 0.999, and only the part of Dc,w,h greater than Phard, that is, Dc,w,h[D>Phard], is taken out to participate in the loss calculation.

[0082] The second part of the loss function Lae measures the difference between the output of the autoencoder and the output of the teacher network. Its expression is as follows:

[0083] CWH represents the number of feature map channels, feature map width and feature map height respectively. T(I) represents the feature map obtained by the input image after passing through the teacher network, and A(I) represents the feature map obtained after the input image passes through the autoencoder network. Among them, A(I) is the feature extraction of image I by the autoencoder, and T(I) is the feature extraction of image I by the teacher network.

[0084] The third part of the loss function Lstae measures the difference between the output of the self-encoder and another student network. Its expression is as follows:

[0085]

[0086] Among them, CWH represents the number of feature map channels, feature map width and feature map height respectively. S`(I) represents the feature map obtained by the input image after passing through another student network, and A(I) represents the feature map obtained after the input image passes through the autoencoder network.

[0087] The student network learns the systematic reconstruction errors of the autoencoder on normal images, such as blurred reconstructions. At the same time, it does not learn the reconstruction errors in abnormal cases, because these features are not in the training set. This makes the difference between the autoencoder output and the student output very suitable for computing the logistic anomaly feature map. Similar to the student-teacher pair, the anomaly feature map is the MSE between the two outputs, averaged across channels.

[0088] 5. Surface Anomaly Detection Model Reasoning

[0089] Since this scheme contains two modules, namely the local and global anomaly detection modules. Therefore, it is necessary to normalize the local and global anomaly maps to similar scales in order to calculate the combined anomaly map by channel average. Otherwise, the noise in one module may interfere with the detection accuracy in the other module. In order to estimate the scale of noise in normal images, this method uses the validation set images, that is, the images that did not appear in the training set, to calculate a possible maximum and minimum value for each anomaly module, and use this scale to align the outputs of the two modules. Use the validation image training set. For these two modules, calculate the set of all pixel anomaly scores on the feature map obtained by all validation images after passing through the above two modules. Then obtain the maximum value qb and the minimum value qa in the set respectively. Then design a linear transformation to map qa to an anomaly score of 0 and qb to an anomaly score of 1. During testing, the local and global anomaly maps are normalized by their respective linear transformations.

[0090] 6. Result output: The detected anomalies are presented to the user in a visual way, including the location, severity and other information of the anomaly. Specifically, the visualization is performed in the form of a heat map.

[0091] Conventional models only consider local anomaly detection, and only the Lst part is included in the loss function. Here, three parts are used to form the loss function. It takes into account both the detection of local anomalies and the detection of logical anomalies. The student network learns the systematic reconstruction errors of the autoencoder on normal images, such as blurred reconstruction. At the same time, it does not learn the reconstruction errors in abnormal cases because these features are not in the training set. This makes the difference between the autoencoder output and the student output very suitable for calculating the logical anomaly feature map.

[0092] The beneficial effects of the present invention are as follows:

[0093] 1) Promptly discover abnormal deformation and damage of the water seal knife to prevent heat loss from the heating furnace, affect the rolling steel temperature, and cause wheel rolling quality problems.

[0094] 2) Prevent CO leakage and potential safety hazards

[0095] 3) It can use data tracking to warn of water seal knife deterioration trends and perform preventive maintenance.

[0096] Obviously, the specific implementation of the present invention is not limited to the above-mentioned methods. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, they are all within the protection scope of the present invention.

Claims

1. An intelligent inspection method for water seal knife of annular heating furnace, characterized in that: The steps include: The water seal knife image data is collected and preprocessed and then sent to the pre-trained surface anomaly detection model. The pre-trained surface anomaly detection model detects whether there is an anomaly in the water seal knife image based on image recognition and marks the abnormal part in the water seal knife image to present the water seal knife image with abnormal marks in a visual form.

2. The intelligent inspection method for water seal blade of an annular heating furnace according to claim 1, characterized in that: The pre-trained surface anomaly detection model includes surface anomaly detection model structure construction, model loss function selection and training; wherein the structure of the surface anomaly model is built based on the surface anomaly detection model structure construction, and then training is performed after selecting the loss function to obtain the pre-trained surface anomaly detection model.

3. An intelligent inspection method for water seal blade of an annular heating furnace as claimed in claim 1 or 2, characterized in that: The constructed surface anomaly detection model includes a teacher network, a student network and an autoencoder network; both the teacher network and the student network use an efficient feature extractor PDN containing 4 convolutional layers; the autoencoder network includes an encoder layer and a decoder layer to output surface anomaly detection results.

4. The intelligent inspection method for water seal blade of an annular heating furnace according to claim 3, characterized in that: The teacher network and the student network use a feature extractor of 4 convolutional layers, whose convolution kernel dimensions are 4*4*128, 4*4*256, 3*3*256, 4*4*256, and the convolution step size is 1*1; PDN achieves downsampling by using a 2*2*128 average pooling layer after the first and second convolutional layers; the output dimension of the fully connected layer after the 4 convolutional layers is 384. In the network output, each output contains a receptive field of 33X33 pixels.

5. The intelligent inspection method for water seal blade of an annular heating furnace according to claim 3, characterized in that: The autoencoder network is composed of an encoder layer containing 6 convolutional layers and a decoder layer containing 8 convolutional layers. The convolution kernel dimensions of the encoder layer are 4*4*32, 4*4*32, 4*4*75, 4*4*64, 4*4*64, and 8*8*64 respectively; the convolution kernel dimension of the decoder layer is uniformly 4*4*64, and the output dimension of the fully connected layer after the decoder layer is 384.

6. An intelligent inspection method for water seal blade of an annular heating furnace according to any one of claims 2 to 5, characterized in that: The teacher network is distilled from a WideResNet-101 network on the ImageNet dataset, and the loss function is MSE; The training of the student network and the autoencoder is implemented by a comprehensive loss function. The comprehensive loss function L total for L total =L ST +L AE +L STAE ; Where: L ST It is a measure of the difference between the output of the student network and the output of the teacher network; L AE It measures the difference between the output of the autoencoder and the output of the teacher network; L STAE is a measure of the difference between the output of the autoencoder and another student network.

7. An intelligent inspection method for water seal blade of an annular heating furnace according to any one of claims 1 to 5, characterized in that: After the surface anomaly detection model detects an abnormality in the water seal knife, it outputs an alarm signal to the heating furnace inspection robot, which is used to link the robot to the on-site re-inspection.

8. An intelligent inspection system for water seal blade of annular heating furnace, characterized in that: It includes a data acquisition module, a data preprocessing module, a surface anomaly detection model, and a visualization module; the data acquisition module monitors the surface of the water seal knife in real time and collects its image data, and its output end is connected to the data preprocessing module. The preprocessing module preprocesses the collected water seal knife image data and sends it to the surface anomaly detection model. The surface anomaly detection model recognizes the water seal knife image and sends the recognition result to the visualization module for visual display.

9. The intelligent inspection system for water seal blade of annular heating furnace according to claim 8, characterized in that: The output end of the surface anomaly detection model is connected to the heating furnace inspection robot, which is used to link the robot to go to the site to recheck the water seal knife anomaly after the anomaly is detected.

10. An intelligent inspection system for water seal blade of an annular heating furnace as claimed in claim 8 or 9, characterized in that: The surface anomaly detection model is a pre-built image recognition model and is used to identify water seal knife image anomalies after pre-training.

Images