Method for identifying and detecting coking furnace shutter valve based on inspection robot

Abstract

A kind of identification detection method for coking furnace shutter valve based on inspection robot, specific steps are as follows: step one: increase the auxiliary image mark of image recognition on shutter valve;Step two: collect image data, carry out model training, establish coking furnace shutter valve dataset;Step three: collect shutter valve image data by inspection robot, judge whether state is normal, if yes, continue to check next shutter valve image, otherwise, send abnormal data to control center.The present application first adds two kinds of auxiliary marks to shutter valve, according to the detection requirement, corresponding installation is carried out on switch cover and telescopic rod, which is convenient for detection and recognition algorithm implementation.Then, track inspection robot is used for inspection and shooting, and the data shot is transmitted to detection and recognition module.The detection and recognition algorithm is optimized and improved on the basis of yolov3, the backbone network is replaced by MobileNet V2 to reduce network depth and improve speed.

Description

Technical Field

[0001] This invention relates to the field of valve inspection and control technology, specifically to a method for identifying and detecting coking furnace opener / closer valves based on an inspection robot. Background Technology

[0002] The status identification of coking furnace valves in a coking plant primarily relies on manual inspections. However, a coking plant typically has nearly a thousand valves, requiring inspections of various aspects, including the proper functioning of valve switches and the integrity of valve chain linkages. Manual inspections are labor-intensive, complex, and require significant manpower, while also being inefficient and prone to errors. Furthermore, the coking plant environment presents challenges, such as dust and particulate matter quickly covering valves and switches, and poor lighting in some indoor areas, all of which increase inspection costs and efficiency. Considering that valves are traditional equipment switches lack autonomous data transmission and retrieval capabilities, but given the increasing need for equipment status data in the era of industrial digitalization and intelligentization, there is an urgent need for an automated, unmanned inspection and identification method to extract valve status data and meet the demands of modern smart factories. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by proposing a simple structure that utilizes an inspection robot for accurate and efficient inspection via track movement. Furthermore, sales and overall machine operation data are uploaded to a server in real time, enabling scientific management. The specific technical solution is as follows:

[0004] A method for identifying and detecting valves on / off switches of a coking furnace based on an inspection robot, comprising the following steps:

[0005] Step 1: Add auxiliary image labels to the valve to facilitate image recognition;

[0006] Step 2: Collect image data, train the model, and establish a coking furnace on / off valve dataset;

[0007] Step 3: Collect image data of the valve opener / closer using the inspection robot, and determine whether the status is normal. If so, continue to check the next valve opener / closer image; otherwise, send the abnormal data to the control center.

[0008] As an optimization, step two specifically includes:

[0009] 2.1 Set the parameters of the inspection robot to collect image data of each valve opener / closer;

[0010] 2.2 Start the robot and begin the inspection to collect data;

[0011] 2.3 Data preprocessing: Using image data augmentation libraries, the collected data is augmented by flipping, rotating, scaling, cropping, translating, adding noise, etc., to improve the generalization ability and robustness of the model.

[0012] 2.4 Data annotation: Use graphic image annotation tools to annotate the auxiliary image labels separately, and finally generate an XML file for each image;

[0013] 2.5 Compiling the neural network framework;

[0014] 2.6. Store the data in the neural network framework;

[0015] 2.7 Convert the file format;

[0016] 2.8. Lightweight optimization of the backbone network, using MobileNet v2 as the basic network;

[0017] 2.9 Adjustment of the detection scale network;

[0018] 2.10 Activation function adjustment: The ReLU6 activation function at the beginning and end of the original MobileNet V2 network is changed to h-swish;

[0019] 2.11. The BN layer and CONV layer are merged. Therefore, the parameters of the BN layer of some network layers are merged into the convolutional layer to improve the model's computation speed.

[0020] 2.12. Adjustment of Box Regression Loss: To improve accuracy, the box regression loss function is optimized and adjusted using DIOU. The original box regression loss function is adjusted as follows:

[0021]

[0022] Where b, b gt ρ and c represent the center points of the predicted and ground truth boxes, respectively, where ρ represents the Euclidean distance between the two center points and c represents the diagonal distance of the smallest packet region that can contain both the predicted and ground truth boxes.

[0023] 2.13. Begin training.

[0024] As an optimization: Step three specifically involves;

[0025] 3.1 Load the pre-trained weight file;

[0026] 3.2 Image loading and detection: The images of the valves captured by the track inspection robot are input into the model network, and the model can output the corresponding detection information. The output information includes the confidence scores of single and double categories, the coordinates of the center point of the detection box and the width and height of the box, as well as the confidence score of the detection box.

[0027] 3.3 Switching Status Determination: Based on the output detection frame, i.e., the position of the auxiliary label and its center position y-coordinate information, the current status of the valve is determined. When the center position y-coordinate of the single label is greater than a threshold... When the center position y-coordinate of the double label is greater than the threshold y′2, it indicates that the telescopic rod is lifted by the chain; otherwise, it indicates that the telescopic rod is lowered.

[0028] 3.4 System status judgment: According to the requirement that a set of valves must keep one side open and the other side closed under normal circumstances, determine whether the corresponding switch status is normal.

[0029] The beneficial effects of this invention are as follows: By adding two types of auxiliary markers to the valve opener, the marker surfaces are treated and installed on the switch cover and telescopic rod according to the detection requirements, facilitating the implementation of the detection and recognition algorithm. Then, a track-based inspection robot performs inspection and photography, transmitting the captured data to the detection and recognition module. The detection and recognition algorithm is based on YOLOv3, with optimizations and improvements. The backbone network is replaced with MobileNet V2 to reduce network depth and improve speed. Because the auxiliary markers are relatively large in the image, the 52*52 detection scale network is cropped to remove the detection scale for small targets. The improvement of fusing the BN layer and CONV layer in the network reduces the network computation load and effectively solves the problem of high resource consumption in embedded environments. Finally, to improve detection accuracy, the activation function is changed to h-swish, and the bounding box loss function is adjusted to DIOU. Ultimately, the accuracy and time consumption of the detection and recognition algorithm meet the requirements of unmanned automatic monitoring and satisfy the data acquisition and monitoring needs of a smart factory in a coking plant. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the valve opening mechanism in this invention;

[0031] Figure 2 This is a schematic diagram of the valve closing in the opening and closing device of the present invention;

[0032] Figure 3 This is a schematic diagram of the unidirectional auxiliary marking assembly in this invention;

[0033] Figure 4 This is a schematic diagram of the bidirectional auxiliary marking assembly in this invention;

[0034] Figure 5 This is a schematic diagram of the model construction and training process in this invention;

[0035] Figure 6 This is a schematic diagram of the depthwise separable convolution algorithm in this invention;

[0036] Figure 7 This is a schematic diagram illustrating the number of depth-separable convolution parameters and computational complexity in this invention;

[0037] Figure 8 This is a schematic diagram of the ReLU and ReLU6 functions in this invention;

[0038] Figure 9 This is a schematic diagram of the expanded layer in this invention;

[0039] Figure 10 This is a schematic diagram of the reverse residual module in this invention;

[0040] Figure 11 This is a schematic diagram of the smallest unit of the MobileNet V2 algorithm in this invention;

[0041] Figure 12 This is a schematic diagram of the MobileNet V2 downsampling scale detection in this invention;

[0042] Figure 13 This is a schematic diagram of the sigmod function in this invention;

[0043] Figure 14 This is a schematic diagram of the swish function in this invention;

[0044] Figure 15 This is a schematic diagram of the h-swish function in this invention;

[0045] Figure 16 This is a schematic diagram of the IOU structure in this invention;

[0046] Figure 17 This is a schematic diagram of the structure of GIOU in this invention;

[0047] Figure 18 This is a schematic diagram of the structure of DIOU in this invention;

[0048] Figure 19 This is a flowchart of the detection and identification process in this invention. Detailed Implementation

[0049] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0050] A method for identifying and detecting valves on / off devices of coking furnaces based on inspection robots.

[0051] Step 1: Add auxiliary image labels to the valve to facilitate image recognition, specifically:

[0052] in Figure 1 and Figure 2 A group is referred to as a set, meaning that under normal circumstances, one side of a set of valves must be in the open state and the other side in the closed state.

[0053] a1 and b1 are the switch covers of the valve. Under normal circumstances, when the switch cover is in state a, the valve is open; when it is in state b, the valve is closed.

[0054] a2 and b2 are chains, one end of which is connected to the valve enabling structure of the opener / closer, and the other end is connected to the telescopic rod (a5, b5).

[0055] a3 and b3 are sector-shaped balance wheels; in reality, a3 is red and b3 is blue. The balance wheels are controlled by an on / off valve enable mechanism. When the balance wheel is in state a, the valve is open; when the balance wheel is in state b, the valve is closed.

[0056] A4 and B4 are one-way auxiliary markers, manually made and installed on the switch covers (A1 and B1) to assist in detecting the status of the switch covers. The position of the one-way auxiliary markers changes as the switch covers are opened and closed; the status of the switch covers can be determined by detecting the position of the one-way auxiliary markers. The surface of the auxiliary markers is frosted to reduce the impact of reflections from the auxiliary lighting during robot photography.

[0057] A5 and B5 are telescopic rods, one end of which is connected to a chain. The chain acts as a sling, lifting the telescopic rod. When the valve is in the open position, the telescopic rod descends, extending into the hole where it is being lowered. Figure 1 As shown in the diagram; when the valve is closed, the telescopic rod is stretched upwards by the chain, rising out of the hole, as if... Figure 2 The status is as shown.

[0058] A6 and B6 are bidirectional auxiliary markers, manually made and installed on the telescopic rods (A5 and B5) to assist in detecting the rod's status. As shown in the figure, the bidirectional auxiliary markers change position as the telescopic rod moves; by detecting the position of the bidirectional auxiliary markers, the chain status can be determined. The surface of the auxiliary markers is frosted to reduce the impact of glare from the auxiliary lighting during robot photography.

[0059] Figure 3This is an assembly diagram of the one-way auxiliary markings, where a4 and b4 are the one-way auxiliary markings, which are installed onto the switch covers a1 and b1 via mounting plates a7 and b7. Note that the lengths of a7 and b7 should be referenced when photographing from the actual front view. Figure 1 The size of the fan-shaped balance wheels a3 and b3 near the screen surface should be such that they do not obstruct the auxiliary markings a4 and b4 when the fan-shaped balance wheels are open or closed.

[0060] Figure 4 This is an assembly diagram of the bidirectional auxiliary labels, where a2 and b2 are chains connecting the a5 and b5 telescopic rods. The bidirectional auxiliary labels a6 and b6 are installed onto the telescopic rods via mounting plates a8 and b8. The installation position of the auxiliary labels must ensure that when the telescopic rod descends to its limit position, the auxiliary labels do not affect the descent depth.

[0061] Step 2: Collect image data, train the model, and establish a coking furnace on / off valve dataset;

[0062] Specifically, such as Figure 5 As shown, the preparation stage. After the auxiliary signs are installed, the robot inspection time and inspection point information are set, and the pan-tilt and camera parameters corresponding to the inspection points are set to ensure that each inspection point can capture a complete and clear set (or one side of a set) of valve image data.

[0063] Start the robot to begin inspection and data collection. During the model training phase, sufficient data should be collected to establish a dataset of coking furnace on / off valves.

[0064] Data preprocessing. Imgaug is used to perform data augmentation, such as flipping, rotating, scaling, cropping, translating, and adding noise, to improve the model's generalization ability and robustness.

[0065] Data annotation. The augmented data was annotated using LabelImg software. There were two types of target labels: one type is a unidirectional auxiliary label, denoted as single, and the other type is a bidirectional auxiliary label, denoted as double. Finally, an XML file was generated for each image.

[0066] Compiling the Darknet network framework. Download the Darknet network, and in the main directory, modify the Makefile configuration file, changing the values ​​of GPU, CUDNN, and OpenCV to 1. Save the file and then execute the `make` command to compile.

[0067] Data storage. Create a folder named MyData in the darknet directory, and then create three subfolders under MyData: Annotation, JPEGImage, and ImageSet. Annotation stores the XML files corresponding to the annotated images. JPEGImage stores the images before annotation. In the ImageSet folder, create a main folder containing TXT files named after the training, validation, and test sets, as well as the images for both training and validation: train.txt, val.txt, test.txt, and trainval.txt. All annotated data is stored in a randomly shuffled order. The ratio of train to trainval is 9:1, and trainval includes test and val in a 9:1 ratio.

[0068] File format conversion. In the `scripts / voc_label.py` file, YOLOv3 provides code to convert the VOC dataset to the format required for YOLO training. First, modify the `sets` list in the code to `[("MyData", "train"), ("MyData", "val"), ("MyData", "test")]`, and the `classes` list to `["single", "doubule"]`. Then, modify the `in_file` and `out_file` paths to the current Annotation and MyData / labels paths, and execute the code. This will generate a `labels` folder (containing the converted .txt files) and `myData_train.txt`, `myData_val.txt`, and `myData_test.txt` files in the `MyData` folder.

[0069] Lightweight Backbone Optimization. The original YOLOv3 backbone uses the DarkNet53 network, which is faster and more accurate than ResNet-152 and ResNet-101. However, in the coking furnace valve detection scenario, such a large and high-precision backbone is unnecessary. Considering that the auxiliary labels for the valves only have two types (single and double), the detection scenario is fixed and simple, and the auxiliary labels are of moderate size relative to the image, the original backbone was replaced with the MobileNet v2 network to optimize the network model, improve running speed and efficiency, and reduce system load.

[0070] The MobileNet series of networks, proposed by Google, is a lightweight network designed for embedded devices, saving significant resources on serial CPUs. The most significant contribution of MobileNet v1 is the use of Depthwise Separable Convolution, which can be divided into Depthwise and Pointwise convolutions. The key advantage of Depthwise Separable is that it drastically reduces the number of parameters and computational cost of the original model, to as low as 1 / 8 of the original. MobileNet v2 primarily combines residual networks with Depthwise Separable convolution. It improves the residual blocks by analyzing the manifold characteristics of single channels, including extending intermediate layers and linear activation of bottleneck layers.

[0071] Features of MobileNet v2 network:

[0072] like Figure 6 As shown, the parameters of depthwise separable convolution consist of two parts: depthwise convolution and pointwise convolution.

[0073] The kernel size D of depthwise convolution K ×D K ×M; the kernel size for pointwise convolution is 1×1×M, one

[0074] There are N in total, so the number of parameters for a depthwise separable convolution is:

[0075] Parameter quantity D K ×D K ×M+M×N

[0076] Computational cost of depthwise separable convolution:

[0077] The computational complexity of depthwise separable convolution also consists of two parts: depthwise convolution and pointwise convolution.

[0078] The kernel size of depthwise convolution is denoted as D. K ×D K ×M, a total of D needs to be done W ×D H The multiplication and addition operations are performed several times; the pointwise convolution kernel size is 1×1×M, there are N kernels, and a total of D operations are performed. W ×D H Therefore, the computational complexity of depthwise separable convolution is: (Number of multiplication and addition operations).

[0079] Computational quantity D K ×D K ×M×D W ×D H +M×N×D W ×DH

[0080] like Figure 7 As shown, the change can be obtained by dividing the number of parameters and computational cost of depthwise separable convolution by the number of parameters and computational cost of standard convolution.

[0081] It can be seen that the number of parameters and the computational complexity of multiply-accumulate operations have both decreased to the original level:

[0082] like Figure 8 As shown in the diagram, blue represents standard ReLU, which returns 0 when x is less than 0 and nothing when x is greater than 0. Orange represents ReLU6, which returns 6 when the input value is greater than 6. ReLU6 "has a boundary." As a non-linear activation function, ReLU6 is more robust in low-precision computation. Because the positive interval of the ReLU function does not impose any non-linear constraints, it can cause gradient explosion when the input is very large. Therefore, v1 uses the ReLU6 activation function.

[0083] However, in practice, V1 is prone to information loss when performing ReLU operations in low dimensions. Performing ReLU operations in high dimensions results in less information loss. Therefore, in the V2 network, the final ReLU6 is replaced with a linear activation function; this part is called the linear bottleneck.

[0084] like Figure 9 As shown, since depthwise convolution itself cannot change the number of channels, if the number of input channels is small, DW depthwise convolution can only work in low dimensions, so the dimensions need to be adjusted. The original authors used PW pointwise convolution to achieve dimensionality increase and decrease. Specifically, PW convolution is used before DW depthwise convolution to increase the dimensionality (the dimensionality increase factor is t, t=6), and then it works in a relatively higher dimension of 6 times.

[0085] like Figure 10 As shown, the inverted residual module borrows the reuse feature of the RestNet network and introduces a shortcut structure. Since the shortcut structure of RestNet is first reduced in dimension-convolution-incremented in dimension, while the shortcut structure of MobileNet v2 is increased in dimension-convolution-reduced in dimension, it is named Inverted residuals.

[0086] like Figure 11 As shown, this is a block in MobileNet V2;

[0087] MobileNet V2 network architecture

[0088]

[0089]

[0090] Table 1 Network Structure of MobileNet V2

[0091] in:

[0092] t is the multiplication factor (i.e., expansion factor) of the input channel.

[0093] n is the number of times this module is repeated.

[0094] c is the number of output channels.

[0095] s stands for stride (repeatedly as stride 1).

[0096] like Figure 12 As shown, the detection scale network was adjusted. The original YOLOv3 had three detection scales: 13*13, 26*26, and 52*52, corresponding to the detection of large, medium, and small targets, respectively. Since the scene background for detecting the coking furnace valve is fixed, the number of detection types is small, and the size of the auxiliary markers is relatively large in the image data, it was decided to trim the 52*52 detection scale network to improve the network detection speed.

[0097] The output of layer 6 is 26*26*96. Adding a branch and using a 1*1 convolution, we get 26*26*512. The output of layer 8 is 13*13*320. Adding a branch and using a 1*1 convolution, we get 13*13*256. After upsampling, we get 26*26*256. Then, we connect it to the original scale network to finally get a scale of 26*26*255.

[0098] The output at layer 9 is 13*13*1280. Adding a branch and using a 1*1 convolution results in a scale of 13*13*255.

[0099] Activation function adjustment. Borrowing from MobileNet V3, the ReLU6 activation function at the beginning and end of the original MobileNet V2 network was modified to h-swish. The h-swish function is an evolution of the swish function. The swish function is characterized by smoothness, non-monotonicity, and no upper bound but a lower bound. It performs better than ReLU in deep models, improving accuracy by approximately 0.75%.

[0100] The exponent calculation in the Swish function is particularly time-consuming, especially in embedded environments, and is relatively resource-intensive. Therefore, the Swish function is improved to h-swish. The sigmoid in the Swish function is replaced with ReLU6(x+3) / 6. The h-swish function is equivalent to a rigid Swish, which effectively reduces the resource consumption of exponent operations in embedded environments.

[0101] like Figure 13 As shown, the Sigmoid function is:

[0102]

[0103] like Figure 14 As shown, the Swish function,

[0104]

[0105] like Figure 15 As shown, the H-Swish function,

[0106]

[0107] 1) Merging BN and CONV layers. Batch Normalization (BN) layers are typically placed after convolutional layers and before activation layers to accelerate network convergence and reduce overfitting. However, BN layers also increase the computational load as the network progresses, consuming more memory or GPU storage. Therefore, the parameters of some BN layers in the model are merged into the convolutional layers to improve the model's computational speed.

[0108] Batch normalization layer and convolutional layer fusion:

[0109]

[0110] Among them, f i，j and For input and output feature maps; W BN W is the weight matrix of the BN layer. conv Let b be the weight matrix of the convolutional layer. BN For the bias of the BN layer, b conv This is the bias of the convolutional layer.

[0111] The fused convolutional layer is then represented as:

[0112] The weights after fusion W = W BN ·W conv

[0113] The bias after fusion is b = W BN ·b conv +b BN

[0114] After removing the BN layer, using new weights and biases, you can get the same result as the original convolutional layer plus the BN layer.

[0115] 2) Adjustment of Box Regression Loss. To improve accuracy, the box regression loss function is optimized and adjusted using DIOU. The original box regression loss function is adjusted as follows:

[0116]

[0117] IOU (Intersection over Union) effectively reflects the detection performance of predicted and ground truth bounding boxes, and exhibits scale invariance. However, it fails to reflect the overlap between two non-intersecting boxes and lacks gradient backpropagation, making it unsuitable for training. GIOU, unlike IOU, measures both overlapping and non-overlapping regions, providing a better picture of overlap. However, its lack of consideration for factors such as distance between target anchors, overlap rate, and scale can lead to divergence issues during training. DIOU (Digitalization over Union) better aligns with the bounding box regression mechanism, resulting in more stable bounding boxes.

[0118] like Figure 16 As shown,

[0119]

[0120] like Figure 17 As shown,

[0121]

[0122] Where C is the area of ​​the minimum closure region between the two boxes;

[0123] like Figure 18 As shown,

[0124]

[0125] Where b, b gt ρ and c represent the center points of the predicted and ground truth boxes, respectively, where ρ represents the Euclidean distance between the two center points and c represents the diagonal distance of the smallest packet region that can simultaneously contain both the predicted and ground truth boxes.

[0126] The corresponding loss function is: L DIOU =1-DIOU

[0127] Start training. After adjusting the above settings, execute the `train.py` file. Observe the training logs. During training, the model will output the corresponding weight file based on the number of iterations. When the loss value stops decreasing or oscillates within a small range, training can be stopped. Save the weight file (.weight) generated by the model.

[0128] Step 3: Collect image data of the valve opener / closer using the inspection robot, and determine whether the status is normal. If so, continue to check the next valve opener / closer image; otherwise, send the abnormal data to the control center.

[0129] Specifically, such as Figure 19 As shown, load the already trained weight file.

[0130] 3.1 Image Loading and Detection. Images of the valves captured by the track-mounted inspection robot are input into the model network, and the model outputs corresponding detection information. The output information includes the confidence scores for single and double categories, the coordinates of the center point of the detection box, the width and height of the box, and the confidence score of the detection box.

[0131] 3.2 Switch Status Determination. The current valve status is determined based on the y-coordinate of the center position of the output detection box (i.e., the position of the auxiliary label). When the y-coordinate of the center position of the single label is greater than a threshold... When the value is equal to 0, it indicates that the cover is open; otherwise, it indicates that the cover is closed. When the y-coordinate of the center position of the double label is greater than the threshold... This indicates that the telescopic rod is in a state where it is lifted by the chain, and vice versa, it indicates that the telescopic rod is in a state where it is lowered.

[0132] 3.3 System Status Determination. Based on the requirement that a set of on / off valves must maintain one side in the open state and the other side in the closed state under normal circumstances, determine whether the corresponding switch status is normal.

Claims

1. A method for identifying and detecting valves of a coking furnace opener / closer based on an inspection robot, characterized in that, The specific steps are as follows: Step 1: Add auxiliary image labels to the valve to facilitate image recognition; Step 2: Collect image data, train the model, and establish a coking furnace on / off valve dataset; 2.1 Set the parameters of the inspection robot to collect image data of each valve opener / closer; 2.2 Start the robot and begin the inspection to collect data; 2.3 Data preprocessing: Using an image data augmentation library, the collected data is augmented by flipping, rotating, scaling, cropping, translating, and adding noise to improve the model's generalization ability and robustness. 2.4 Data annotation: Use graphic image annotation tools to annotate the auxiliary image labels separately, and finally generate an XML file for each image; 2.5 Compiling the Neural Network Framework: Download the Darknet network. In the main directory, modify the Makefile configuration file, changing the values ​​of GPU, CUDNN, and OPENCV to 1. Save the file and execute the command `make` to compile it. 2.

6. Store the data in the neural network framework. Create a folder named MyData under the darknet directory, and then create three folders under MyData: Annotation, JPEGImage, and ImageSet. Annotation stores the XML files corresponding to the previously annotated images, JPEGImage stores the images before annotation, and ImageSet stores the main folder within the ImageSet folder. This main folder contains the training set, validation set, test set, and the TXT files containing the names of the training and validation sets: train.txt, val.txt, test.txt, and trainval.txt. All annotated data is stored in a randomly shuffled order, with the ratio of train to trainval being 9:

1. Trainval includes test and val in a 9:1 ratio. 2.7 Convert the file format; 2.

8. Lightweight optimization of the backbone network, using MobileNet v2 as the basic network; 2.9 Adjustment of the detection scale network; 2.10 Activation function adjustment: The ReLU6 activation function at the beginning and end of the original MobileNet V2 network is changed to h-swish; 2.

11. The BN layer and CONV layer are merged. Therefore, the parameters of the BN layer of some network layers are merged into the convolutional layer to improve the model's computation speed. 2.

12. Adjustment of Box Regression Loss: To improve accuracy, the box regression loss function is optimized and adjusted using DIOU. The original box regression loss function is adjusted as follows: in These represent the center points of the predicted bounding box and the ground truth bounding box, respectively. This represents the calculation of the Euclidean distance between two center points. It represents the diagonal distance of the smallest packet region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; 2.

13. Begin training; Step 3: Collect image data of the valve opener / closer through the inspection robot, and determine whether the status is normal. If it is, continue to check the next valve opener / closer image; otherwise, send the abnormal data to the control center. 3.1 Load the pre-trained weight file; 3.2 Image loading and detection: The images of the valves captured by the track inspection robot are input into the model network, and the model can output the corresponding detection information. The output information includes the confidence scores of single and double categories, the coordinates of the center point of the detection box and the width and height of the box, as well as the confidence score of the detection box. 3.3 Switching Status Determination: Based on the output detection frame, i.e., the position of the auxiliary label and its center position y-coordinate information, the current status of the valve is determined. When the center position y-coordinate of the single label is greater than a threshold... When the value is equal to 0, it indicates that the cover is open; otherwise, it indicates that the cover is closed. When the y-coordinate of the center position of the double label is greater than the threshold... If the value is 0, it means the telescopic rod is in a state of being lifted by the chain; otherwise, it means the telescopic rod is in a state of being lowered. 3.4 System status judgment: According to the requirement that a set of valves must keep one side open and the other side closed under normal circumstances, determine whether the corresponding switch status is normal.

Images