Image-laser collaborative nuclear power valve sealing surface automatic detection system and method

Abstract

The present invention discloses an image-laser collaborative nuclear power valve sealing surface automatic detection system and method, the technical field of which is computer vision and nuclear power maintenance, including: collecting image data of the nuclear power valve sealing surface based on an image acquisition unit to generate a picture data set; scanning the nuclear power valve sealing surface based on a laser to obtain a scan data set; constructing a deep learning network model, calculating the picture data set and the scan data set based on the deep learning model to obtain a damage identification data set; processing the damage identification data set and extracting information to obtain a detection result. Through the present invention, automatic detection of nuclear power valve sealing surface defects can be achieved, and detection efficiency and accuracy can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and nuclear power maintenance, and in particular relates to an image-laser collaborative nuclear power valve sealing surface automatic detection system and method. Background Art

[0002] The integrity of the sealing surface of nuclear power valves is crucial to the safe operation of nuclear power plants. At present, the detection of the sealing surface of nuclear power valves mainly relies on manual visual inspection, which has the problems of low efficiency and low accuracy. Therefore, an automatic detection system and method are needed to improve the efficiency and accuracy of nuclear power valve sealing surface detection.

[0003] Regarding the detection of nuclear power valve sealing surfaces, in addition to the currently used visual inspection that requires manpower, ultrasonic, magnetic powder, eddy current and other detection technologies are used to detect the sealing surfaces of nuclear power valves; by applying a certain pressure to the sealing surface of the nuclear power valve, observe whether there is leakage to determine whether the sealing surface has defects; use infrared thermal imagers to scan the sealing surface of nuclear power valves, and observe the thermal map to detect whether there are abnormalities on the sealing surface, which is also often used to detect damage to the sealing surface. Due to the complexity of nuclear power working conditions, manual detection has high labor costs and dangers, and cannot guarantee the accuracy and precision of the detection; using infrared thermal imagers to detect the sealing surface of nuclear power valves has stringent requirements on the environment, and the cumbersome installation and layout lead to inconvenience in operation.

[0004] Therefore, how to perform high-precision, high-accuracy and high-speed detection of valve sealing surface damage while reducing labor costs has become an urgent problem that needs to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide an image-laser collaborative nuclear power valve sealing surface automatic detection system and method to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides an image-laser collaborative nuclear power valve sealing surface automatic detection system, comprising:

[0007] An image acquisition unit, used to acquire image data of the sealing surface of a nuclear power valve and generate a picture data set;

[0008] A laser scanning unit, used to scan the sealing surface of a nuclear power valve based on a laser to obtain a scanning data set;

[0009] A damage identification unit, connected to the image acquisition unit and the laser scanning unit, and used to construct a deep learning network model, and to calculate the image data set and the scanning data set based on the deep learning model to obtain a damage identification data set;

[0010] The information extraction unit is connected to the damage identification unit and is used to process the damage identification data set and extract information to obtain a detection result.

[0011] Preferably, the image acquisition unit acquires images of the sealing surface of the nuclear power valve based on different illumination modes, synthesizes the acquired images, and obtains the image data set.

[0012] Preferably, the laser scanning unit includes a scanning subunit and a linear subunit; the scanning subunit is used to scan the static state of the sealing surface of the nuclear power valve to obtain a static state data set; the linear subunit is used to scan the motion state of the sealing surface of the nuclear power valve to obtain a motion state data set.

[0013] Preferably, the laser scanning unit obtains the height difference between two points by identifying the height of a specified point in the sealing surface of a nuclear power valve, obtains scanning data based on the height difference between the two points, and then integrates the data with the static state data set and the moving state data set to obtain the scanning data set.

[0014] Preferably, the damage identification unit uses ABUNet++ network to construct the deep learning network model, trains the deep learning network model through NEU-CLS data set and KolektorSDD data set to obtain a damage detection model, inputs the scanning data set into the damage detection model for calculation, and obtains the damage identification data set.

[0015] Preferably, the information extraction unit extracts the information of the damage identification data set using a digital image processing algorithm and then performs a comprehensive analysis, and evaluates the degree of damage based on the comprehensive analysis result to obtain the detection result.

[0016] Preferably, the damage identification data set includes: chemical ionization corrosion data caused by medium scouring, fluid erosion shadow data in the medium, operation squeezing data between valves, and valve structure damage data caused by hot and cold changes.

[0017] To achieve the above object, the present invention also provides an image-laser collaborative nuclear power valve sealing surface automatic detection method, comprising:

[0018] The image data of the sealing surface of the nuclear power valve is collected based on the image acquisition unit to generate a picture data set;

[0019] Scan the sealing surface of nuclear power valves based on laser to obtain a scanning data set;

[0020] Constructing a deep learning network model, and calculating the image dataset and the scan dataset based on the deep learning model to obtain a damage identification dataset;

[0021] The damage identification data set is processed and information is extracted to obtain a detection result.

[0022] Preferably, the process of obtaining the damage identification data set includes:

[0023] The deep learning network model is constructed using the ABUNet++ network and attention mechanism;

[0024] The deep learning network model is trained by using the NEU-CLS data set and the KolektorSDD data set to obtain a damage detection model;

[0025] The scanning data set is input into the damage detection model for calculation to obtain the damage identification data set.

[0026] Preferably, the expression of the attention mechanism is:

[0027]

[0028] Where F is the input tensor feature, and AvgPool represents average pooling, and MaxPool represent maximum pooling, W0 is the first fully connected multilayer perceptron, W1 is the second fully connected multilayer perceptron, and σ is the sigmoid function used for normalization.

[0029] The technical effects of the present invention are:

[0030] The present invention can realize automatic detection of sealing surface defects of nuclear power valves, thereby improving detection efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0032] Figure 1 System modeling and physical diagrams for embodiments of the present invention;

[0033] Figure 2 is a deep learning network structure diagram in an embodiment of the present invention;

[0034] Figure 3 is a diagram of the attention mechanism in an embodiment of the present invention;

[0035] Figure 4 This is a flow chart of a damage information extraction algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] Embodiment 1

[0039] like Figure 1 As shown, in this embodiment, an image-laser collaborative nuclear power valve sealing surface automatic detection system is provided, including:

[0040] The system is mounted on nuclear power maintenance equipment, and the image-laser sensor provides it with digital images and depth information, including:

[0041] The image acquisition unit includes an image sensor. The image sensor has a reference distance of 55 mm, a measuring distance of 45-65 mm, a detection range of 27.5 × 20.5 mm or more for a measuring distance of 45 mm, 32.5 × 24.5 mm or more for a measuring distance of 55 mm, and 37.5 × 28.5 mm (X direction) × (Y direction) * 1 for a measuring distance of 65 mm, a sampling period of 120 ms or more, an image receiving element of a CMOS image sensor, and a weight of 190 g.

[0042] The laser scanning unit includes a laser generator and a receiver. The laser scanning device uses a red semiconductor laser with a wavelength of 660nm (visible light) and a sampling period of more than 120ms.

[0043] The damage identification unit includes a deep learning network model and a data set. The network model uses deep learning to centrally train a data set of images collected from the sealing surface of a nuclear power valve in a host computer to obtain a weight coefficient that can segment and identify the damaged area and the background area.

[0044] The information extraction unit includes an image preprocessing algorithm and an information calculation algorithm. The image preprocessing algorithm and the information calculation algorithm use digital image processing to enhance the segmented image based on damage identification and record the size, location, quantity and depth information of the damage to guide subsequent maintenance work.

[0045] Further optimization scheme, the image acquisition unit has a built-in lighting function, which can capture clear images. By combining the images captured in 4 lighting modes into one image, the halo phenomenon can be reduced. The image acquisition unit is suitable for different installation environments and is convenient for subsequent installation. The protection level of the sensor head housing meets the IP67 standard and can achieve dustproof and waterproof functions. In addition, the connector of the connecting cable connection can be rotated 330° to freely adjust the assembly position. The image acquisition unit is also equipped with an FTP client function for traceability. Using the FTP client function, image data can be transferred to an FTP server (such as a computer, NAS, PLC). At the same time, users can also confirm the measurement results and shooting time through this function. In addition, according to the detailed results of the run, each file can be automatically assigned to the corresponding folder.

[0046] According to a further optimization scheme, the laser scanning unit has two working modes: scanning mode and linear mode. In scanning mode, the laser scanning unit is used to detect stationary targets in an area. In linear mode, the laser scanning unit uses linear laser to detect targets, and can effectively detect targets even when they are in motion.

[0047] In a further optimization scheme, the laser scanning unit can identify the height of a specified point, and can also identify the height difference between two specified points. This height difference detection method can achieve stable detection results and can effectively track the target even if the overall height of the target changes. In addition, the laser scanning unit can also identify the average height of a specified area. Even if the target is very small or has a complex shape, the scanning unit can still stably perform height detection.

[0048] To further optimize the solution, the damage identification unit uses the ABUNet++ network to perform deep learning on the damage image dataset. The dataset used is composed of NEU-CLS and KolektorSDD. In order to improve the network's adaptive ability to identify different damages under different conditions and on different surfaces, it is necessary to perform partial training on these images in multiple categories. This ensures that the network can accurately identify various types of damage in practical applications.

[0049] The ABUNet++ network is based on the U-net network structure. It improves the downsampling part (adds attention mechanism) and comprehensively considers the impact of each downsampling number on the training effect. It is a self-developed network.

[0050] To further optimize the scheme, the dataset contains four different types of damage: medium scouring caused by chemical ionization corrosion, fluid erosion in the medium, operational squeezing between valves, and valve structural damage caused by hot and cold changes. In order to train the ABUNet++ network to identify these different types of damage, 50 images that can fully represent each type of damage were selected from the dataset and input into the network for cyclic training to update each parameter weight of the network. Then, we selected 10 images from each type as a test set to evaluate the accuracy of the network in different damage types.

[0051] Further optimizing the scheme, the left half of the ABUNet++ network structure is responsible for feature extraction, that is, extracting high-level semantic information of the damaged location in the image. The right half of the network structure is responsible for image restoration, which combines and restores the extracted high-level semantic information. This network structure, called ABUNet++, connects information at different semantic levels through "skip connections" and "concatenation", thereby more effectively restoring the image and retaining high-level semantic information during the downsampling process. The downsampling part of ABUNet++ also introduces channel attention mechanism and spatial attention mechanism to consider the impact of different channels on the final detection results, thereby improving the accuracy of the detection results.

[0052] To further optimize the scheme, the digital image processing algorithm performs preprocessing operations such as grayscale, binarization and morphological processing on the recognized image to obtain identifiable connected domains. The information calculation algorithm uses the regionprops function to obtain the location, quantity and area information of the damage, thereby realizing automated, high-speed and high-precision detection of photovoltaic dust accumulation. By quantitatively feedback the degree of damage, the grinding position, force, time and grinding disk of the grinding equipment can be adjusted. A damage degree assessment system is constructed to establish a quantitative connection between the degree of damage and the grinding operation.

[0053] Embodiment 2

[0054] An image-laser collaborative nuclear power valve sealing surface automatic detection system and method, the system is applied to the image acquisition device and central processing unit of the nuclear power valve maintenance robot, and the nuclear power maintenance robot is used as a carrier to complete the corresponding detection items, including:

[0055] The image acquisition unit uses an image sensor to acquire images of valve sealing surface damage for detection.

[0056] The laser scanning unit uses a laser generator and a receiver. The laser scanning device uses a red semiconductor laser with a wavelength of 660nm (visible light) and a sampling period of more than 120ms.

[0057] The damage identification unit uses deep learning to conduct centralized training on the images collected from the sealing surfaces of nuclear power valves in the host computer, and obtains weight coefficients that can segment and identify damaged areas and background areas.

[0058] The information extraction unit uses digital image processing to enhance the segmented image based on damage identification and records the size, location, quantity, and depth information of the damage to guide subsequent maintenance work.

[0059] Further optimization scheme, the image acquisition unit can take clear images only by built-in lighting, and the images taken in 4 lighting modes are combined into 1 to reduce halo. Suitable for different installation environments, easy for subsequent installation, the shell protection level of the sensor head adopts IP67 standard, which is dustproof and waterproof. In addition, the connector of the cable connection part can be rotated 330°, and the assembly can be adjusted freely. Equipped with FTP client function, with traceability, using FTP client function, image data can be transferred to FTP server (computer, NAS, PLC). The measurement results and shooting time can also be confirmed. Each file can also be assigned to the corresponding folder according to the detailed results of the operation.

[0060] In a further optimized solution, the laser scanning unit is divided into a scanning mode and a linear mode. The scanning mode detects the target object in the area and is used for stationary targets. The linear mode detects the target object with a line laser, and can detect the target object even if it is moving.

[0061] Further optimization scheme, the laser scanning unit can identify the height of the specified point; identify the height difference between the specified two points. Always detect with the height difference, even if the overall height of the target changes, it can be tracked to achieve stable detection; identify the average height of the specified area. Even if the target is small or complex in shape, it can be stably detected.

[0062] To further optimize the scheme, the damage recognition unit uses the ABUNet++ network to perform deep learning on the damage image dataset, and the dataset used is composed of NEU-CLS and KolektorSDD. KolektorSDD is a dataset consisting of metal surface loss images, which is commonly used for defect detection. The dataset is constructed from images of defective production items provided and annotated by the Kolektor Group doo. NEU-CLS is a metal loss image dataset that includes 1,800 grayscale images of typical surface defect samples. In order to make the network more adaptive to identify different damages under different conditions and on different surfaces, it is necessary to use different types of parts of these images for network training.

[0063] To further optimize the solution, the data set is divided into four different types, namely, chemical ionization corrosion data caused by medium scouring, fluid erosion shadow data in the medium, operation squeezing data between valves, and valve structure damage data caused by hot and cold changes.

[0064] To further optimize the scheme, the ABUNet++ network selects 50 images that can characterize each lesion in the dataset, and then inputs the images into the network for cyclic training to update each parameter weight. Then 10 images from each type are selected as the test set. Finally, we get different detection difficulties for four different types of images.

[0065] Further optimization scheme, the left half of the ABUNet++ network structure is used for feature extraction, that is, high-level semantic extraction of the damaged position in the image. The function of the right half of the network structure is image restoration, which combines and restores the extracted high-level semantics. This network structure called ABUNet++ performs splicing between different semantics through "skip connection" and "concatenation", which can more effectively restore the image and give high-level semantic information in downsampling.

[0066] To further optimize the scheme, the downsampling part of ABUNet++ introduces a channel attention mechanism and a spatial attention mechanism, so that the impact of different channels on the final detection results can be considered during the downsampling process, making the detection results more accurate.

[0067] According to a further optimization scheme, the information extraction unit uses digital image processing technology to extract the size, position and quantity of damage on the valve sealing surface in the image and combines and records them accordingly with the depth information collected by the laser scanning unit.

[0068] To further optimize the scheme, the digital image processing technology performs image preprocessing operations such as grayscale, binarization and morphological processing on the recognized image to obtain an identifiable connected domain, and then obtains the location information, quantity information and area information of the damage through the regionprops function to achieve the purpose of automatic, high-speed and high-precision detection of photovoltaic dust accumulation.

[0069] To further optimize the solution, the information extraction unit extracts the size, position, depth and quantity of damage on the valve sealing surface, conducts a comprehensive analysis of the damage degree based on the above information, and acts on the grinding position, grinding force, grinding time and grinding disc selection of the grinding equipment through quantitative feedback of the damage degree. A damage degree assessment system is constructed to establish a quantitative connection between the damage degree and the grinding operation.

[0070] Further optimization of the scheme, in the damage degree detection, mainly includes: defect classification detection; damage degree assessment; damage degree determination. Through damage degree detection, the severity of defects on the valve sealing surface can be effectively evaluated, providing a reference for subsequent repair and maintenance work.

[0071] Embodiment 3

[0072] This embodiment provides an image-laser collaborative nuclear power valve sealing surface automatic detection method, including:

[0073] The image data of the sealing surface of the nuclear power valve is collected based on the image acquisition unit to generate a picture data set;

[0074] Scan the sealing surface of nuclear power valves based on laser to obtain a scanning data set;

[0075] Constructing a deep learning network model, and calculating the image dataset and the scan dataset based on the deep learning model to obtain a damage identification dataset;

[0076] The damage identification data set is processed and information is extracted to obtain a detection result.

[0077] See also Figure 2 As shown in the figure, the size of the input RGB image of the ABUNet++ network is 256×256 for the next data processing. In the process of extracting high-level semantic information, it will undergo four downsamplings combined with the attention mechanism. In the process of downsampling, the tensor representing the image features will be pooled or convolved to achieve the purpose of compressing the semantics of the target to be tested, which will make the length and width of the tensor smaller. After each downsampling, a convolution operation with a convolution kernel size of 3 is performed, which will increase the number of channels of the tensor. It can be seen that in the image enhancement part, the number of channels will continue to increase and the image size will continue to decrease, so that the semantic information of the image will be compressed into a long tensor of 16×16×1024. On the contrary, in the image restoration part, the size of the tensor changes in the opposite way, which means that the tensor will be restored to its original size and number of channels through this part with high-level semantic annotation information.

[0078] Further optimizing the scheme, the process of obtaining the damage identification data set includes:

[0079] The deep learning network model is constructed using the ABUNet++ network and attention mechanism;

[0080] The deep learning network model is trained by using the NEU-CLS data set and the KolektorSDD data set to obtain a damage detection model;

[0081] The scanning data set is input into the damage detection model for calculation to obtain the damage identification data set.

[0082] See also Figure 3 As shown in Figure 1, introducing the attention mechanism in the network structure can better retain valid semantic information and discard invalid semantic information. The attention mechanism is applied to each downsampling, and each downsampling is a convolutional block attention module, including channel downsampling and spatial downsampling. The channel attention module can be expressed as formula:

[0083]

[0084] The input F of the attention mechanism is average pooled and max pooled to obtain different output tensors, and then two fully connected multi-layer perceptrons are executed. Finally, the two different tensors are added and normalized by the sigmoid function. In the formula, F is the input tensor feature, and AvgPool represents average pooling, and MaxPool represent maximum pooling, W0 is the first fully connected multilayer perceptron, W1 is the second fully connected multilayer perceptron, and σ is the sigmoid function used for normalization. The spatial attention module can be expressed as the formula:

[0085]

[0086] In the formula, the input F is output as a two-layer tensor after average pooling and maximum pooling, and then an n-order full connection is performed, and finally the result is normalized by the sigmoid function. and AvgPool represents average pooling, and MaxPool represents the maximum pooling, f n×n represents n-order full connection, and σ represents the sigmoid function.

[0087] See also Figure 4 As shown in the figure, after obtaining the recognized image, the particle image is preprocessed before the formal image processing, including: grayscale, binarization, corrosion, filling, expansion, and damage information calculation.

[0088] The grayscale conversion performs mean compression on the information of the RGB space three-dimensional matrix, and each pixel is divided into 256 levels of brightness ranging from 0 to 255.

[0089] The binarization divides the grayscale image of 256 levels of brightness into two at a fixed threshold value, wherein values ​​less than the threshold value become 0, and values ​​greater than the threshold value become 1.

[0090] The erosion reduces the area of ​​the white connected domain, the filling turns the small white connected domain black, and the collision increases the area of ​​the white connected domain. Such operations can effectively reduce the too small connected domains and the large connected domains divided by the small connected domains.

[0091] The damage information calculation uses the function STATS=regionprops(L, properties) to measure the total number of pixels in each region of the image and the minimum rectangle containing the corresponding region. A series of properties of each annotated region in the annotation matrix L are measured. Different positive integer elements in L correspond to different regions. For example, elements in L equal to the integer 1 correspond to region 1, elements in L equal to the integer 2 correspond to region 2, and so on. The return value STATS is a structure array of length L. The corresponding fields of the structure array define the metrics under the corresponding attributes of each region. properties can be a list of strings separated by commas, a cell array containing strings, a single string all or basic. If properties is equal to the string all, all the metric data in the following string list will be calculated. If properties is not specified or equal to basic, the attributes Area, Centroid and BoundingBox will be calculated.

[0092] Then use the cat function to re-divide the rectangle's four corners into an array. Use the for function to loop the rectangles for a certain number of times, and use the rectangle function in each loop to draw a rectangle with a red border color and the rectangle information as the starting point, length and width.

[0093] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An image-laser collaborative nuclear power valve sealing surface automatic detection system, characterized in that: include: An image acquisition unit, used to acquire image data of the sealing surface of a nuclear power valve and generate a picture data set; A laser scanning unit, used to scan the sealing surface of a nuclear power valve based on a laser to obtain a scanning data set; A damage identification unit, connected to the image acquisition unit and the laser scanning unit, is used to construct a deep learning network model, and calculate the image data set and the scanning data set based on the deep learning model to obtain a damage identification data set; An information extraction unit, connected to the damage identification unit, for processing the damage identification data set and extracting information to obtain a detection result; The image acquisition unit acquires images of the sealing surface of the nuclear power valve based on different illumination modes, synthesizes the acquired images, and obtains the image data set; The laser scanning unit includes a scanning subunit and a linear subunit; the scanning subunit is used to scan the static state of the nuclear power valve sealing surface to obtain a static state data set; the linear subunit is used to scan the motion state of the nuclear power valve sealing surface to obtain a motion state data set; The damage identification unit uses the ABUNet++ network to construct the deep learning network model, trains the deep learning network model through the NEU-CLS data set and the KolektorSDD data set to obtain a damage detection model, and inputs the scanning data set into the damage detection model for calculation to obtain the damage identification data set; The laser scanning unit obtains the height difference between two points by identifying the height of a designated point on the sealing surface of a nuclear power valve, obtains scanning data based on the height difference between the two points, and integrates the scan data with the static state data set and the motion state data set to obtain the scan data set; The information extraction unit extracts information from the damage identification data set using a digital image processing algorithm and then performs a comprehensive analysis, and evaluates the degree of damage based on the comprehensive analysis result to obtain the detection result.

2. The image-laser collaborative nuclear power valve sealing surface automatic detection system according to claim 1 is characterized in that: The damage identification data set includes: chemical ionization corrosion data caused by medium scouring, fluid erosion shadow data in the medium, operation squeezing data between valves, and valve structure damage data caused by hot and cold changes.

Images