Power plant boiler water wall defect detection method, device, equipment and medium

Abstract

Embodiments of the present application disclose a kind of power station boiler water wall defect detection method, device, equipment and medium.Therein, the method includes: obtaining the defect image to be measured of water wall to be measured, at least two groups of image feature information of defect image to be measured are determined according to multi-scale attention network;Wherein, multi-scale attention network includes at least two feature extraction links, at least two same number of feature extraction groups are included in each feature extraction link;At least two candidate defect images corresponding to at least two groups of image feature information are determined according to path aggregation network;Candidate defect information of candidate defect image is determined according to target detection network;According to candidate defect detection accuracy, target defect information is determined from candidate defect information, and defect detection result of defect image to be measured is determined according to target defect information.The technical scheme can effectively solve the problems of low efficiency and low economic efficiency of power station boiler water wall defect detection precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, in particular to a defect detection method, device, equipment and medium for a water-cooled wall of a power station boiler. BACKGROUND

[0002] In China, thermal power bears the important role of consuming new energy and energy supply guarantee. As the core equipment of a thermal power plant, the water-cooled wall detection effect is of great significance to the safe, reliable and economic operation of the unit. Therefore, scientific and effective water-cooled wall detection is an important means to ensure the safe operation of coal-fired units.

[0003] In the prior art, the defect detection of the water-cooled wall of the power station boiler is mainly completed by manual detection. This method not only has a large amount of work, heavy tasks, limited coverage, but also has great subjectivity, resulting in low detection efficiency, large detection result error, high omission rate and other problems. SUMMARY

[0004] The present application provides a defect detection method, device, equipment and medium for a water-cooled wall of a power station boiler, which realizes multi-dimensional extraction and fusion of image defect features based on a pre-trained network model, and can effectively solve the problems of low defect detection precision, low efficiency and low economy of the water-cooled wall of the power station boiler.

[0005] According to one aspect of the present application, a defect detection method for a water-cooled wall of a power station boiler is provided, the method comprising:

[0006] Obtaining a defect image to be detected of a water-cooled wall to be detected, and determining at least two groups of image feature information of the defect image to be detected according to a multi-scale attention network; wherein the multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit;

[0007] Determining at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network; wherein the path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension;

[0008] Determining candidate defect information of the candidate defect image according to a target detection network; wherein the candidate defect information comprises a candidate defect type, a candidate defect position and a candidate defect detection precision;

[0009] Determining target defect information from the candidate defect information according to the candidate defect detection precision, and determining a defect detection result of the defect image to be detected according to the target defect information.

[0010] According to another aspect of the present application, there is provided a defect detection device for a water-cooled wall of a power plant boiler, comprising:

[0011] An image feature information determination module is configured to acquire a to-be-detected defect image of a to-be-detected water-cooled wall, and determine at least two groups of image feature information of the to-be-detected defect image according to a multi-scale attention network; wherein the multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit;

[0012] A candidate defect image determination module is configured to determine at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network; wherein the path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension;

[0013] A candidate defect information determination module is configured to determine candidate defect information of the candidate defect image according to a target detection network; wherein the candidate defect information comprises a candidate defect type, a candidate defect position, and a candidate defect detection accuracy;

[0014] A defect detection result determination module is configured to determine target defect information from the candidate defect information according to the candidate defect detection accuracy, and determine a defect detection result of the to-be-detected defect image according to the target defect information.

[0015] According to another aspect of the present application, there is provided an electronic device, comprising:

[0016] at least one processor; and

[0017] a memory in communication with the at least one processor; wherein

[0018] the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the defect detection method for a water-cooled wall of a power plant boiler according to any one of the embodiments of the present application.

[0019] According to another aspect of the present application, there is provided a computer readable storage medium storing computer instructions for enabling a processor to execute the defect detection method for a water-cooled wall of a power plant boiler according to any one of the embodiments of the present application when executed by the processor.

[0020] The technical scheme of the embodiment of the present application obtains a to-be-tested defect image of a to-be-tested water cooling wall, determines at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network; wherein the multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit; determines at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network; wherein the path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension; determines candidate defect information of the candidate defect image according to a target detection network; wherein the candidate defect information comprises a candidate defect type, a candidate defect position and a candidate defect detection accuracy; determines target defect information from the candidate defect information according to the candidate defect detection accuracy, and determines a defect detection result of the to-be-tested defect image according to the target defect information. The technical scheme is based on a pre-trained network model to realize multi-dimensional extraction and fusion of image defect features, and can effectively solve the problems of poor defect detection accuracy, low efficiency and low economy of the water cooling wall of the power station boiler.

[0021] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present application, nor is it used to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0023] Figure 1 is a flowchart of a defect detection method of a water cooling wall of a power station boiler according to the first embodiment of the present application;

[0024] Figure 2 is a flowchart of defect detection model training according to the first embodiment of the present application;

[0025] Figure 3 is a flowchart of data sample enhancement and expansion according to the first embodiment of the present application;

[0026] Figure 4 is a network structure diagram of a defect detection model according to the first embodiment of the present application;

[0027] Figure 5 is a flowchart of a defect detection method of a water cooling wall of a power station boiler according to the second embodiment of the present application;

[0028] Figure 6 Fig. 3 is a structural schematic diagram of a defect detection device of a water wall of a power station boiler according to an embodiment of the present application;

[0029] Figure 7 Fig. 4 is a structural schematic diagram of an electronic device for implementing a defect detection method of a water wall of a power station boiler according to an embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts should fall within the scope of the present application.

[0031] It should be noted that the terms "first", "second", "target" and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to include only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product or device.

[0032] Embodiment one

[0033] Figure 1 Fig. 1 is a flowchart of a defect detection method of a water wall of a power station boiler according to an embodiment of the present application. The embodiment can be applicable to the case of quickly and accurately detecting defects of a water wall of a power station boiler. The method can be executed by a defect detection device of a water wall of a power station boiler. The defect detection device of a water wall of a power station boiler can be realized in the form of hardware and / or software. The defect detection device of a water wall of a power station boiler can be configured in an electronic device having data processing capability. As shown in the figure, the method comprises the following steps. Figure 1

[0034] S110, acquiring a to-be-detected defect image of a to-be-detected water wall, and determining at least two groups of image feature information of the to-be-detected defect image according to a multi-scale attention network.

[0035] ​The multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit.

[0036] The technical scheme of the embodiment can detect defects of a to-be-tested water cooling wall based on a pre-trained defect detection model. The defect types can include thickening, corrosion, coking, cracking, breaking and the like. The defect detection model can include a multi-scale attention network, a path aggregation network and a target detection network. The multi-scale attention network can realize multi-dimensional extraction and fusion of image defect features by introducing a lightweight feature extraction unit and a double attention mechanism (position attention mechanism and channel attention mechanism), and can enhance the learning efficiency and generalization ability of the algorithm for multiple types of complex features. The path aggregation network can realize multi-level cross fusion of different dimensional features, avoid inconsistent learning accuracy caused by different image sizes and image qualities, and ensure the completeness of the model in learning defect features. The target detection network can be used for type identification and position calibration of image defects.

[0037] Figure 2 A flowchart of a defect detection model training process is provided for the embodiment one of the application. As shown in the figure, Figure 2 the model training process is as follows:

[0038] 1. Raw data sample collection. In the raw data sample collection, the collected defect data samples mainly include multiple types such as thickening, corrosion, coking, cracking and breaking, and the number of images of each type can be 50 or more.

[0039] 2. Data sample enhancement and expansion. In the data sample enhancement and expansion, the complex background data samples in the weak light and dust environment are optimized by brightness automatic calculation and data enhancement techniques (such as image denoising, contrast enhancement, brightness enhancement, image flipping and saturation adjustment), and a batch of samples are generated for algorithm training, so as to realize data set sample expansion.

[0040] Figure 3 A flowchart of data sample enhancement and expansion is provided for the embodiment one of the application. As shown in the figure, Figure 3 first, the collected sample data is converted from RGB to HSV, and the principle is as follows: R'=R / 255, G'=G / 255, B'=B / 255, Cmax=max(R', G', B'), Cmin=min(R', G', B'), and Δ=Cmax-Cmin. The RGB channel is mainly normalized, and the difference is obtained according to the maximum value and the minimum value, so as to be converted into the HSV format. The specific conversion formula is as follows:

[0041]

[0042]

[0043] V = Cmax.

[0044] Then the brightness automation calculation is carried out, and the image below the threshold T (which can be set according to the situation, such as T = 130) is enhanced in brightness, and the principle is as follows: the brightness L = all the values of the third channel V in the HSV image are averaged, if L ≥ T, the data enhancement operations such as translation, flip, filter sharpening, mirror, scaling, etc. are carried out on the image; if L < T, first the brightness of the image is enhanced through the formula Out = L + (In - L) * (1 + percent), wherein In represents the original pixel point brightness, L represents the average brightness of the whole image, Out represents the adjusted brightness, and percent is the adjustment range [-1, 1], and then the data enhancement operations such as translation, flip, filter sharpening, mirror, scaling, etc. are carried out on the image. Then the image after the data enhancement operation is converted from HSV to GRB, and thus the expanded enhanced image sample can be obtained.

[0045] 3. Data sample classification calibration and division. In the data sample classification calibration and division, a calibration tool (such as labelimg, a visual image calibration tool) can be used to classify and position calibrate the target object in the sample, and save it as an xml file, and at the same time, the randomly shuffled data set can be divided into a training set, a validation set and a test set according to 8:1:1.

[0046] 4. A defect detection model is constructed, including a multi-scale attention network, a path aggregation network, and a target detection network. For example, the defect detection model can be a model obtained by improving the YOLO algorithm. YOLO (You Only Look Once) is an algorithm for single target detection using a convolutional neural network, which has a fast detection speed. Specifically, the multi-scale attention network includes at least two feature extraction links, each of which includes at least two feature extraction groups of the same number, and each feature extraction group includes a feature extraction unit and an attention unit. For example, the feature extraction unit can be a Bottleneck in MobileNetV3, or a combination of CBS and CSPLayer provided by YOLO itself, or a combination of CBS, SPPBottleneck, and CSPLayer. Bottleneck contains CBR with Relu activation function, global average pooling layer, Conv2D layer, and BN layer. CSPLayer is a multi-residual superposition branch structure, the right branch is a cyclic residual structure, and the left branch is a single residual edge. The output of the two branches can be combined to obtain the residual features of the image sample. SPPBottleneck contains three maximum pooling layers with pool sizes of 5x5, 9x9, and 13x13 for feature value extraction. The attention unit can be a position attention unit (corresponding to a position attention mechanism) or a channel attention unit (corresponding to a channel attention mechanism). The path aggregation network (PANet++) can be composed of CSPLayer, convolution operation, and sampling operation. For example, the target detection network can be YoloHead provided by YOLO itself, which processes the features through CBS to obtain three types of prediction results (defect position, defect type, and defect detection accuracy).

[0047] 5. Algorithm model optimization training. In the model optimization training process, a two-stage training mode of freezing and unfreezing can be used. Specifically, in the freezing stage, the Mosaic data augmentation method can be used to randomly splice four images into the defect detection model for iterative training; in the unfreezing stage, the multi-scale attention network is unfreezed, and the same data set and training parameters in the freezing stage are used for optimization training. The batchsize, learning rate Lr, and other parameters need to be set throughout the process, wherein the learning rate adopts the cosine annealing method. The evaluation criteria are accuracy, loss value, and map. By training a higher accuracy, map value, and smaller loss value, the performance improvement effect is ensured. When the loss function converges, the optimal model weight parameter is obtained.

[0048] 6. Algorithm model optimization and testing. In the algorithm model optimization and testing, the optimal weight obtained by each iteration training is introduced into the defect detection model, and the accuracy, loss and other indicators of the algorithm are tested and counted using the verification set, and then the model parameter adjustment is used to realize the optimization of the indicators, and then the model optimization training is completed. In addition, for the performance test of the defect detection model, the final optimal model weight parameter in the training can be introduced into the model, and the generalization ability verification is carried out through the test set, and then the index test and performance detection of the algorithm in various defects are completed.

[0049] In the embodiment, optionally, the multi-scale attention network comprises a first feature extraction link and a second feature extraction link; the first feature extraction link comprises four first link feature extraction groups, each first link feature extraction group comprising a first link feature extraction unit and a position attention unit; the second feature extraction link comprises four second link feature extraction groups, each second link feature extraction group comprising at least two second link feature extraction units and a channel attention unit.

[0050] Among them, the first link feature extraction group and the second link feature extraction group can respectively refer to the feature extraction groups in the first feature extraction link and the second feature extraction link. The first link feature extraction unit and the second link feature extraction unit can respectively refer to the feature extraction units in the first link feature extraction group and the second link feature extraction group. It should be noted that the first link feature extraction group can be the same or different, and the second link feature extraction group can also be the same or different. Exemplarily, the first link feature extraction group can all be composed of a bottleneck and a position attention unit, and the second link feature extraction group can be composed of a CBS, a CSPLayer and a channel attention unit, or composed of a CBS, a SPPBottleneck, a CSPLayer and a channel attention unit.

[0051] In this embodiment, first, the image of the defect to be measured of the water-cooled wall to be measured is acquired, and at least two groups of image feature information of the defect to be measured are determined according to the multi-scale attention network. Optionally, the at least two groups of image feature information of the defect to be measured are determined according to the multi-scale attention network, including: determining first link first feature information of the defect to be measured on the first feature extraction link according to a first link first feature extraction group of the first feature extraction link; determining second link first feature information of the defect to be measured on the second feature extraction link according to a second link first feature extraction group of the second feature extraction link; fusing the first link first feature information and the second link first feature information to obtain first fusion feature information of the defect to be measured; determining first link second feature information of the first fusion feature information on the first feature extraction link according to a first link second feature extraction group of the first feature extraction link; determining second link second feature information of the first fusion feature information on the second feature extraction link according to a second link second feature extraction group of the second feature extraction link; fusing the first link second feature information and the second link second feature information to obtain second fusion feature information of the defect to be measured; determining first link third feature information of the second fusion feature information on the first feature extraction link according to a first link third feature extraction group of the first feature extraction link; determining second link third feature information of the second fusion feature information on the second feature extraction link according to a second link third feature extraction group of the second feature extraction link; fusing the first link third feature information and the second link third feature information to obtain third fusion feature information of the defect to be measured; determining first link fourth feature information of the third fusion feature information on the first feature extraction link according to a first link fourth feature extraction group of the first feature extraction link; determining second link fourth feature information of the third fusion feature information on the second feature extraction link according to a second link fourth feature extraction group of the second feature extraction link; fusing the first link fourth feature information and the second link fourth feature information to obtain fourth fusion feature information of the defect to be measured; and taking the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information as the image feature information of the defect to be measured.

[0052] In this embodiment, the multi-dimensional feature extraction of the defect to be measured can be directly performed through the first feature extraction link and the second feature extraction link in the multi-scale attention network, so as to obtain the multi-dimensional feature information of the defect to be measured, or the feature rough extraction of the defect to be measured can be performed first to obtain the global feature information of the defect to be measured, and then the multi-dimensional feature extraction of the global feature information is performed through the first feature extraction link and the second feature extraction link to obtain the partial feature information of the defect to be measured. This embodiment does not make specific limitation, and can be set according to actual needs.

[0053] In the embodiment, before determining the first-link first feature information of the to-be-tested defect image on the first feature extraction link according to the first-link first feature extraction group of the first feature extraction link, the method further includes: performing feature preprocessing on the to-be-tested defect image based on a preset convolution template to obtain reference feature information, wherein the preset convolution template has one zero value in adjacent elements, and the feature preprocessing includes image feature extraction and image channel expansion; and correspondingly, determining the first-link first feature information of the to-be-tested defect image on the first feature extraction link according to the first-link first feature extraction group of the first feature extraction link includes: determining the first-link first feature information of the reference feature information on the first feature extraction link according to the first-link first feature extraction group of the first feature extraction link; and determining the second-link first feature information of the reference feature information on the second feature extraction link according to the second-link first feature extraction group of the second feature extraction link.

[0054] The preset convolution template can be a preset convolution template, which has one zero value in adjacent elements, that is, one zero value is set every interval in the preset convolution template, so that a value can be taken every interval of a pixel point for the to-be-tested defect image, thereby avoiding the problem of image feature information redundancy.

[0055] Figure 4 A network structure diagram of a defect detection model provided for the first embodiment of the application. Among them, Double-linked multi-scale attention network refers to a double-chain multi-scale attention network, PANet++ refers to a path aggregation network, and YoloHead refers to a target detection network. Among them, CBS includes a normal convolution layer (Conv2D), a standardization layer (BN) and an activation function (SiLU). As shown in Figure 4 In the double-chain multi-scale attention network, the upper link is composed of four feature extraction groups, and each feature extraction group is composed of a bottleneck and a position attention unit; the lower link is also composed of four feature extraction groups, and the feature extraction groups are composed of CBS, CSPLayer and a channel attention unit, and CBS, SPPBottleneck, CSPLayer and a channel attention unit, respectively.

[0056] For example, Figure 4As shown, first, the to-be-tested defect image with a dimension of 640x640x3 is input into the defect detection model, and the to-be-tested defect image is subjected to convolution operation based on a preset convolution template through Fcous in the double-chain multi-scale attention network, to obtain a feature group P1 with a dimension of 320x320x12, so as to expand the image channel number from 3 to 12. Then, the feature group P1 with a dimension of 320x320x12 is converted into a feature group P2 with a dimension of 320x320x64 (i.e., reference feature information) through CBS for feature extraction and channel depth increase, so as to realize feature preprocessing (coarse feature extraction) of the to-be-tested defect image. The feature group P2 is used to represent the global feature information of the to-be-tested defect image.

[0057] Next, the reference feature information (feature group P2) obtained after feature preprocessing can be subjected to further multi-dimensional feature extraction through the two feature extraction links in the double-chain multi-scale attention network. Specifically, the feature group P2 with a dimension of 320x320x64 is subjected to feature extraction through the first feature extraction group in the upper link, to obtain a feature group P3 with a dimension of 160x160x64; and the feature group P2 with a dimension of 320x320x64 is subjected to feature extraction through the first feature extraction group in the lower link, to obtain a feature group P3' with a dimension of 160x160x64; the feature groups P3 and P3' are fused to obtain a fused feature group P3" with a dimension of 160x160x64 (i.e., first fused feature information). Then, the fused feature group P3" is subjected to feature extraction through the second feature extraction group in the upper link, to obtain a feature group P4 with a dimension of 80x80x128; and the fused feature group P3" is subjected to feature extraction through the second feature extraction group in the lower link, to obtain a feature group P4' with a dimension of 80x80x128; the feature groups P4 and P4' are fused to obtain a fused feature group P4" with a dimension of 80x80x128 (i.e., second fused feature information). Further, the fused feature group P4" is subjected to feature extraction through the third feature extraction group in the upper link, to obtain a feature group P5 with a dimension of 40x40x256; and the fused feature group P4" is subjected to feature extraction through the third feature extraction group in the lower link, to obtain a feature group P5' with a dimension of 40x40x256; the feature groups P5 and P5' are fused to obtain a fused feature group P5" with a dimension of 40x40x256 (i.e., third fused feature information). Again, the fused feature group P5" is subjected to feature extraction through the fourth feature extraction group in the upper link, to obtain a feature group P6 with a dimension of 20x20x512; and the fused feature group P5" is subjected to feature extraction through the fourth feature extraction group in the lower link, to obtain a feature group P6' with a dimension of 20x20x512; the feature groups P6 and P6' are fused to obtain a fused feature group P6" with a dimension of 20x20x512 (i.e., fourth fused feature information). Finally, the fused feature groups P3", P4", P5" and P6" can be taken as the image feature information of the to-be-tested defect image.

[0058] S120, determining at least two candidate defect images corresponding to the at least two groups of image feature information according to the path aggregation network.

[0059] The path aggregation network includes a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension.

[0060] In this embodiment, after determining the at least two groups of image feature information of the to-be-tested defect image, at least two candidate defect images corresponding to the at least two groups of image feature information can be determined through the path aggregation network. Optionally, determining the at least two candidate defect images corresponding to the at least two groups of image feature information according to the path aggregation network includes: determining a first candidate defect image corresponding to the first fusion feature information according to the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information; determining a second candidate defect image corresponding to the second fusion feature information according to the first candidate defect image, the second fusion feature information, the third fusion feature information and the fourth fusion feature information; determining a third candidate defect image corresponding to the third fusion feature information according to the second candidate defect image, the third fusion feature information and the fourth fusion feature information; and determining a fourth candidate defect image corresponding to the fourth fusion feature information according to the second candidate defect image, the third candidate defect image, the third fusion feature information and the fourth fusion feature information.

[0061] In this embodiment, a first candidate defect image corresponding to the first fusion feature information is first determined. Optionally, determining the first candidate defect image corresponding to the first fusion feature information according to the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information includes: performing convolution and upsampling operations on the fourth fusion feature information, and then performing fusion with the third fusion feature information to obtain first-dimension fusion information; performing convolution and upsampling operations on the first-dimension fusion information, and then performing fusion with the second fusion feature information to obtain second-dimension fusion information; performing convolution and upsampling operations on the second-dimension fusion information, and then performing fusion with the first fusion feature information to obtain the first candidate defect image corresponding to the first fusion feature information.

[0062] For example, Figure 4For example, after performing convolution and upsampling operations on the fusion feature group P6″, it is fused with the fusion feature group P5″ through CSPLAYE to obtain the first dimension fusion information; after performing convolution and upsampling operations on the first dimension fusion information, it is fused with the fusion feature group P4″ through CSPLAYE to obtain the second dimension fusion information; after performing convolution and upsampling operations on the second dimension fusion information, it is fused with the fusion feature group P3″ through CSPLAYE to obtain the first candidate defect image corresponding to the fusion feature group P3″. The dimensions of the first candidate defect image are 160×160×64.

[0063] Then, a second candidate defect image corresponding to the second fusion feature information is determined. Optionally, the second candidate defect image corresponding to the second fusion feature information is determined based on the first candidate defect image, the second fusion feature information, the third fusion feature information, and the fourth fusion feature information, including: performing convolution and upsampling operations on the fourth fusion feature information, and then fusing it with the third fusion feature information to obtain first-dimensional fusion information; performing convolution and upsampling operations on the first-dimensional fusion information, and then fusing it with the second fusion feature information to obtain second-dimensional fusion information; performing downsampling operations on the first candidate defect image to obtain a first reference image; and determining the second candidate defect image corresponding to the second fusion feature information based on the first reference image and the second-dimensional fusion information.

[0064] For example, with Figure 4 For example, after performing convolution and upsampling operations on the fusion feature group P6″, it is fused with the fusion feature group P5″ through CSPLAYE to obtain the first dimension fusion information; after performing convolution and upsampling operations on the first dimension fusion information, it is fused with the fusion feature group P4″ through CSPLAYE to obtain the second dimension fusion information; the first candidate defect image is downsampled to obtain the first reference image; after performing convolution operations on the second dimension fusion information, it is fused with the first reference image through CSPLAYE to obtain the second candidate defect image corresponding to the fusion feature group P4″. The dimensions of the second candidate defect image are 80×80×128.

[0065] Then, a third candidate defect image corresponding to the third fusion feature information is determined. Optionally, the third candidate defect image corresponding to the third fusion feature information is determined based on the second candidate defect image, the third fusion feature information, and the fourth fusion feature information, including: performing convolution and upsampling operations on the fourth fusion feature information, and then fusing it with the third fusion feature information to obtain first-dimensional fusion information; performing a first downsampling operation on the second candidate defect image to obtain a second reference image; performing a second downsampling operation on the second candidate defect image to obtain a third reference image; and determining the third candidate defect image corresponding to the third fusion feature information based on the second reference image, the third reference image, and the first-dimensional fusion information.

[0066] For example, with Figure 4 For example, after performing convolution and upsampling operations on the fusion feature group P6″, it is fused with the fusion feature group P5″ through CSPLAYE to obtain the first dimension fusion information; the second candidate defect image is subjected to a first downsampling operation to obtain the second reference image; the second candidate defect image is subjected to a second downsampling operation to obtain the third reference image; after performing convolution operations on the first dimension fusion information, it is first fused with the second reference image through CSPLAYE, and then the fused information is fused with the third reference image to obtain the third candidate defect image corresponding to the fusion feature group P5″. The dimension of the third candidate defect image is 40×40×256.

[0067] Finally, the fourth candidate defect image corresponding to the fourth fusion feature information is determined. Optionally, the fourth candidate defect image corresponding to the fourth fusion feature information is determined based on the second candidate defect image, the third candidate defect image, the third fusion feature information, and the fourth fusion feature information, including: performing convolution and upsampling operations on the fourth fusion feature information, and then fusing it with the third fusion feature information to obtain first-dimensional fusion information; performing downsampling operations on the second candidate defect image to obtain a second reference image; performing convolution operations on the first-dimensional fusion information and then fusing it with the second reference image, and performing downsampling operations on the fused information to obtain a fourth reference image; performing a third downsampling operation on the third candidate defect image to obtain a fifth reference image; performing a fourth downsampling operation on the third candidate defect image to obtain a sixth reference image; and determining the fourth candidate defect image corresponding to the fourth fusion feature information based on the fourth reference image, the fifth reference image, the sixth reference image, and the fourth fusion feature information.

[0068] For example, with Figure 4For example, after the convolution and up-sampling operations are performed on the fusion feature group P6", the first dimension fusion information is obtained by fusing the fusion feature group P5" through the CSPLayer; the second candidate defect image is down-sampled to obtain a second reference image; after the convolution operation is performed on the first dimension fusion information, the second reference image is fused through the CSPLayer, and the down-sampling operation is performed on the fused information to obtain a fourth reference image; the third candidate defect image is down-sampled to obtain a fifth reference image; the third candidate defect image is down-sampled to obtain a sixth reference image; after the convolution operation is performed on the fusion feature group P6", the fourth reference image is first fused through the CSPLayer, then the information obtained after the first fusion is secondly fused with the fifth reference image, and then the information obtained after the second fusion is thirdly fused with the sixth reference image, and the information obtained after the third fusion is determined as the fourth candidate defect image corresponding to the fusion feature group P6".

[0069] In S130, candidate defect information of the candidate defect image is determined according to the target detection network.

[0070] The candidate defect information includes candidate defect type, candidate defect position and candidate defect detection accuracy. For example, the defect position can be represented by the position information of the prediction box, which can include the coordinate value of the upper left corner of the prediction box in the image and the width and height of the prediction box. The defect detection accuracy can be represented by the object ratio contained in the prediction box. The ratio range can be set to 0-1, and the larger the value, the higher the defect detection accuracy. For example, the target detection network can use YoloHead provided by YOLO itself. As shown in FIG. 2, the output of YoloHead includes three types of prediction results, namely Reg, Obj and Cls. Reg can be used to judge the regression parameters of each feature point, i.e., the position of the prediction box; Obj can be used to judge whether each feature point contains an object and the object ratio contained; and Cls can be used to judge the object category contained in each feature point, i.e., the defect type. Figure 4

[0071] ​In this embodiment, after determining the candidate defect image, the candidate defect information of the candidate defect image can be determined by the target detection network. Optionally, determining the candidate defect information of the candidate defect image according to the target detection network includes: determining the candidate defect type of the candidate defect image according to the first prediction unit in the target detection network; wherein the first prediction unit is used to predict the image defect type; determining the candidate defect location of the candidate defect image according to the second prediction unit in the target detection network; wherein the second prediction unit is used to predict the image defect location; and determining the candidate defect detection accuracy of the candidate defect image according to the third prediction unit in the target detection network; wherein the third prediction unit is used to predict the image defect detection accuracy.

[0072] For example, with Figure 5 For example, the first prediction unit is Cls, the second prediction unit is Reg, and the third prediction unit is Obj. Specifically, the four candidate defect images are input into different YoloHeads (YoloHead1-YoloHead4), and defect detection can be performed on the four candidate defect images by the YoloHeads respectively. YoloHead1-YoloHead4 have the same structure. Taking one candidate defect image as an example, the candidate defect type can be determined based on the output of Cls, the candidate defect location can be determined based on the output of Reg, and the candidate defect detection accuracy can be determined based on the output of Obj. The candidate defect type, candidate defect location, and candidate defect detection accuracy are then used as the candidate defect information for that candidate defect image, thus obtaining the candidate defect information corresponding to each candidate defect image.

[0073] S140, determine the target defect information from the candidate defect information based on the candidate defect detection accuracy, and determine the defect detection result of the image to be tested based on the target defect information.

[0074] In this embodiment, after obtaining the candidate defect information corresponding to the candidate defect image, the target defect information can be determined from the candidate defect information based on the candidate defect detection accuracy. Optionally, determining the target defect information from the candidate defect information based on the candidate defect detection accuracy includes: determining the maximum value among the candidate defect detection accuracies as the target defect detection accuracy; and determining the candidate defect information corresponding to the target defect detection accuracy as the target defect information.

[0075] Specifically, the candidate defect detection precisions are sorted in descending order, and the maximum value is selected as the target defect detection precision, and then the candidate defect information corresponding to the target defect detection precision is determined as the target defect information. Further, the defect detection result of the to-be-tested defect image can be determined according to the target defect information. For example, the target defect type and the target defect position in the target defect information can be used as the defect detection result of the to-be-tested defect image, or all the target defect information can be used as the defect detection result of the to-be-tested defect image, which is not limited in the embodiment.

[0076] The technical scheme of the embodiment of the application obtains a to-be-tested defect image of a to-be-tested water cooling wall, determines at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network, wherein the multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit; determines at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network, wherein the path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension; determines candidate defect information of the candidate defect image according to a target detection network, wherein the candidate defect information comprises a candidate defect type, a candidate defect position and a candidate defect detection precision; determines target defect information from the candidate defect information according to the candidate defect detection precision, and determines a defect detection result of the to-be-tested defect image according to the target defect information. The technical scheme can realize multi-dimensional extraction and fusion of image defect features based on a pre-trained network model, and can effectively solve the problems of low defect detection precision, low efficiency and low economy of the water cooling wall defect detection of the power station boiler.

[0077] Embodiment two

[0078] Figure 5 A flowchart of a defect detection method of a water cooling wall of a power station boiler provided by the embodiment two of the application, which is optimized based on the above-mentioned embodiment.

[0079] As Figure 6 shown, the method of the embodiment specifically comprises the following steps:

[0080] S210, obtaining a to-be-tested defect image of a to-be-tested water cooling wall, and determining at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network.

[0081] The multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit.

[0082] S220, determine at least two candidate defect images corresponding to the at least two groups of image feature information according to the path aggregation network.

[0083] The path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension.

[0084] S230, determine candidate defect information of the candidate defect image according to the target detection network.

[0085] The candidate defect information comprises a candidate defect type, a candidate defect position and a candidate defect detection accuracy.

[0086] S240, determine the maximum value in the candidate defect detection accuracy as a target defect detection accuracy.

[0087] S250, determine the candidate defect information corresponding to the target defect detection accuracy as target defect information.

[0088] S260, determine a defect detection result of the to-be-tested defect image according to the target defect information.

[0089] The technical scheme of the embodiment of the application can realize multi-dimensional extraction and fusion of image defect features based on a pre-trained network model, and can effectively solve the problems of poor defect detection accuracy, low efficiency and low economy of the water-cooled wall of the power plant boiler.

[0090] Embodiment three

[0091] Figure 6 A structure diagram of a defect detection device of a water-cooled wall of a power plant boiler is provided for the third embodiment of the application. The device can execute the defect detection method of the water-cooled wall of the power plant boiler provided by any embodiment of the application, and has the corresponding function modules and beneficial effects of the execution method. As shown in the figure, the device comprises: Figure 7

[0092] The image feature information determination module 310 is configured to obtain a to-be-tested defect image of a to-be-tested water-cooled wall, and determine at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network. The multi-scale attention network comprises at least two feature extraction links, each feature extraction link comprises at least two feature extraction groups of the same number, and each feature extraction group comprises a feature extraction unit and an attention unit.

[0093] The candidate defect image determination module 320 is configured to determine at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network. The path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension.

[0094] ​The candidate defect information determination module 330 is used to determine the candidate defect information of the candidate defect image based on the target detection network; wherein, the candidate defect information includes the candidate defect type, the candidate defect location, and the candidate defect detection accuracy;

[0095] The defect detection result determination module 340 is used to determine target defect information from the candidate defect information based on the candidate defect detection accuracy, and to determine the defect detection result of the image to be tested based on the target defect information.

[0096] Optionally, the multi-scale attention network includes a first feature extraction link and a second feature extraction link; the first feature extraction link includes four first link feature extraction groups, each first link feature extraction group including one first link feature extraction unit and one position attention unit; the second feature extraction link includes four second link feature extraction groups, each second link feature extraction group including at least two second link feature extraction units and one channel attention unit.

[0097] Optionally, the image feature information determination module 310 is specifically used for:

[0098] Based on the first feature extraction group of the first feature extraction link, the first feature information of the defect image to be tested on the first feature extraction link is determined;

[0099] Based on the second link first feature extraction group of the second feature extraction link, the second link first feature information of the defect image to be tested on the second feature extraction link is determined;

[0100] The first feature information of the first link and the first feature information of the second link are fused to obtain the first fused feature information of the defect image to be tested;

[0101] Based on the first link second feature extraction group of the first feature extraction link, the first fused feature information is determined as the first link second feature information on the first feature extraction link;

[0102] Based on the second feature extraction group of the second feature extraction link, the second feature information of the first fused feature information on the second feature extraction link is determined;

[0103] The second feature information of the first link and the second feature information of the second link are fused to obtain the second fused feature information of the defect image to be tested;

[0104] Based on the first link third feature extraction group of the first feature extraction link, the second fused feature information is determined as the first link third feature information on the first feature extraction link;

[0105] determine second link third feature information of the second fusion feature information on the second feature extraction link according to a second link third feature extraction group of the second feature extraction link;

[0106] fuse the first link third feature information and the second link third feature information to obtain third fusion feature information of the to-be-tested defect image;

[0107] determine first link fourth feature information of the third fusion feature information on the first feature extraction link according to a first link fourth feature extraction group of the first feature extraction link;

[0108] determine second link fourth feature information of the third fusion feature information on the second feature extraction link according to a second link fourth feature extraction group of the second feature extraction link;

[0109] fuse the first link fourth feature information and the second link fourth feature information to obtain fourth fusion feature information of the to-be-tested defect image;

[0110] use the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information as image feature information of the to-be-tested defect image.

[0111] Optionally, the candidate defect image determination module 320 is specifically used for:

[0112] determine a first candidate defect image corresponding to the first fusion feature information according to the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information;

[0113] determine a second candidate defect image corresponding to the second fusion feature information according to the first candidate defect image, the second fusion feature information, the third fusion feature information and the fourth fusion feature information;

[0114] determine a third candidate defect image corresponding to the third fusion feature information according to the second candidate defect image, the third fusion feature information and the fourth fusion feature information;

[0115] determine a fourth candidate defect image corresponding to the fourth fusion feature information according to the second candidate defect image, the third candidate defect image, the third fusion feature information and the fourth fusion feature information.

[0116] Optionally, the image feature information determination module 310 is further used for:

[0117] Before determining the first-link first feature information of the to-be-tested defect image on the first feature extraction link according to the first-link first feature extraction group of the first feature extraction link, the to-be-tested defect image is subjected to feature preprocessing based on a preset convolution template to obtain reference feature information, wherein a zero value is arranged in adjacent elements of the preset convolution template, and the feature preprocessing comprises image feature extraction and image channel expansion.

[0118] Correspondingly, the image feature information determination module 310 is further configured to:

[0119] Determine the first-link first feature information of the reference feature information on the first feature extraction link according to the first-link first feature extraction group of the first feature extraction link.

[0120] Determine the second-link first feature information of the reference feature information on the second feature extraction link according to the second-link first feature extraction group of the second feature extraction link.

[0121] Optionally, the candidate defect information determination module 330 is specifically configured to:

[0122] Determine the candidate defect type of the candidate defect image according to a first prediction unit in the target detection network, wherein the first prediction unit is used to predict an image defect type.

[0123] Determine the candidate defect position of the candidate defect image according to a second prediction unit in the target detection network, wherein the second prediction unit is used to predict an image defect position.

[0124] Determine the candidate defect detection precision of the candidate defect image according to a third prediction unit in the target detection network, wherein the third prediction unit is used to predict an image defect detection precision.

[0125] Optionally, the defect detection result determination module 340 is specifically configured to:

[0126] Determine the maximum value in the candidate defect detection precision as a target defect detection precision.

[0127] Determine the candidate defect information corresponding to the target defect detection precision as target defect information.

[0128] The defect detection device for a power station boiler water cooling wall provided in the embodiment can execute the defect detection method for a power station boiler water cooling wall provided in any embodiment of the application, and has the corresponding function modules and beneficial effects of the execution method.

[0129] Embodiment Four

[0130] Figure 7A structural diagram of an electronic device 10 that can be used to implement embodiments of the present application is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices (e.g., headsets, glasses, watches, etc.), and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application described and / or claimed in this document.

[0131] As shown in ​ The electronic device 10 includes at least one processor 11, and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., communicatively connected to the at least one processor 11, where the memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes according to the computer programs stored in the read-only memory (ROM) 12 or loaded into the random access memory (RAM) 13 from the storage unit 18. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0132] Various components in the electronic device 10 are connected to the I / O interface 15, including an input unit 16, such as a keyboard, a mouse, etc., an output unit 17, such as various types of displays, speakers, etc., a storage unit 18, such as a magnetic disk, an optical disk, etc., and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0133] The processor 11 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 performs various methods and processes described above, such as the method of defect detection of a power plant boiler water wall.

[0134] In some embodiments, the method of defect detection of a power plant boiler water wall can be implemented as a computer program tangibly embodied in a computer readable storage medium, e.g., storage unit 18. In some embodiments, parts or all of the computer program can be loaded and / or installed onto electronic device 10 via, e.g., ROM 12 and / or communication unit 19. When the computer program is loaded onto RAM 13 and executed by processor 11, one or more steps of the method of defect detection of a power plant boiler water wall described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the method of defect detection of a power plant boiler water wall by way of other means (e.g., by way of firmware).

[0135] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a programmable logic device (PLD), a computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0136] Computer programs used to implement the methods of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the computer program, when executed by the processor of the machine, implements the functions / acts specified in the flowcharts and / or block diagrams. The computer program can be executed entirely on a machine, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0137] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. A computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable signal medium. More specific examples of a machine-readable storage medium will include one or more lines of a program of instructions in a transitory signal, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0138] To provide for interaction with a user, the systems and techniques described here can be implemented on an electronic device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the electronic device. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0139] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0140] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system, to solve the defects of large management difficulty and weak business scalability in traditional physical host and VPS service.

[0141] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, each step described in the present application can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions of the present application can be achieved, which is not limited herein.

[0142] The above detailed description does not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for detecting defects in a water wall of a utility boiler, characterized by, The method comprises: acquiring a to-be-tested defect image of a to-be-tested water-cooled wall, and determining at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network; wherein the multi-scale attention network comprises a first feature extraction link and a second feature extraction link; the first feature extraction link comprises four first-link feature extraction groups, and each first-link feature extraction group comprises a first-link feature extraction unit and a position attention unit; the second feature extraction link comprises four second-link feature extraction groups, and each second-link feature extraction group comprises at least two second-link feature extraction units and a channel attention unit; determining at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network; wherein the path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension; determining candidate defect information of the candidate defect image according to a target detection network; wherein the candidate defect information comprises a candidate defect type, a candidate defect position and a candidate defect detection precision; determining target defect information from the candidate defect information according to the candidate defect detection precision, and determining a defect detection result of the to-be-tested defect image according to the target defect information.

2. The method of claim 1, wherein, Determining at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network comprises: determining first-link first feature information of the to-be-tested defect image on the first feature extraction link according to a first-link first feature extraction group of the first feature extraction link; determining second-link first feature information of the to-be-tested defect image on the second feature extraction link according to a second-link first feature extraction group of the second feature extraction link; fusing the first-link first feature information and the second-link first feature information to obtain first fusion feature information of the to-be-tested defect image; determining first-link second feature information of the first fusion feature information on the first feature extraction link according to a first-link second feature extraction group of the first feature extraction link; determining second-link second feature information of the first fusion feature information on the second feature extraction link according to a second-link second feature extraction group of the second feature extraction link; fusing the first-link second feature information and the second-link second feature information to obtain second fusion feature information of the to-be-tested defect image; determining first-link third feature information of the second fusion feature information on the first feature extraction link according to a first-link third feature extraction group of the first feature extraction link; determining second-link third feature information of the second fusion feature information on the second feature extraction link according to a second-link third feature extraction group of the second feature extraction link; fusing the first-link third feature information and the second-link third feature information to obtain third fusion feature information of the to-be-tested defect image; determine, according to the first-link fourth feature extraction group of the first feature extraction link, first-link fourth feature information of the third fusion feature information on the first feature extraction link; determine, according to the second-link fourth feature extraction group of the second feature extraction link, second-link fourth feature information of the third fusion feature information on the second feature extraction link; fuse the first-link fourth feature information and the second-link fourth feature information to obtain fourth fusion feature information of the to-be-tested defect image; use the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information as image feature information of the to-be-tested defect image.

3. The method of claim 2, wherein, determine at least two candidate defect images corresponding to the at least two groups of image feature information according to the path aggregation network, including: determine, according to the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information, a first candidate defect image corresponding to the first fusion feature information; determine, according to the first candidate defect image, the second fusion feature information, the third fusion feature information and the fourth fusion feature information, a second candidate defect image corresponding to the second fusion feature information; determine, according to the second candidate defect image, the third fusion feature information and the fourth fusion feature information, a third candidate defect image corresponding to the third fusion feature information; determine, according to the second candidate defect image, the third candidate defect image, the third fusion feature information and the fourth fusion feature information, a fourth candidate defect image corresponding to the fourth fusion feature information.

4. The method of claim 2, wherein, Before determining, according to the first-link first feature extraction group of the first feature extraction link, first-link first feature information of the to-be-tested defect image on the first feature extraction link, the method further includes: perform feature preprocessing on the to-be-tested defect image based on a preset convolution template to obtain reference feature information; wherein one zero value is arranged in adjacent elements of the preset convolution template, and the feature preprocessing includes image feature extraction and image channel expansion; Correspondingly, determining, according to the first-link first feature extraction group of the first feature extraction link, first-link first feature information of the to-be-tested defect image on the first feature extraction link includes: determine, according to the first-link first feature extraction group of the first feature extraction link, first-link first feature information of the reference feature information on the first feature extraction link; determine, according to the second-link first feature extraction group of the second feature extraction link, second-link first feature information of the reference feature information on the second feature extraction link.

5. The method according to any of claims 1 to 4, characterized in that, determine candidate defect information of the candidate defect image according to the target detection network, including: determine a candidate defect type of the candidate defect image according to a first prediction unit in the target detection network; wherein the first prediction unit is used for predicting an image defect type; determine a candidate defect position of the candidate defect image according to a second prediction unit in the target detection network, wherein the second prediction unit is configured to predict an image defect position; determine a candidate defect detection precision of the candidate defect image according to a third prediction unit in the target detection network, wherein the third prediction unit is configured to predict an image defect detection precision.

6. The method of claim 5, wherein, determine target defect information from the candidate defect information according to the candidate defect detection precision, comprising: determine a maximum value in the candidate defect detection precision as a target defect detection precision; determine candidate defect information corresponding to the target defect detection precision as the target defect information.

7. An apparatus for detecting defects in a water-cooled wall of a utility boiler, characterized by The apparatus comprises: an image feature information determination module configured to acquire a to-be-tested defect image of a to-be-tested water-cooled wall, and determine at least two groups of image feature information of the to-be-tested defect image according to a multi-scale attention network; wherein the multi-scale attention network comprises a first feature extraction link and a second feature extraction link; the first feature extraction link comprises four first-link feature extraction groups, and each first-link feature extraction group comprises a first-link feature extraction unit and a position attention unit; the second feature extraction link comprises four second-link feature extraction groups, and each second-link feature extraction group comprises at least two second-link feature extraction units and a channel attention unit; a candidate defect image determination module configured to determine at least two candidate defect images corresponding to the at least two groups of image feature information according to a path aggregation network; wherein the path aggregation network comprises a convolution unit and a sampling unit, and the image feature information and the candidate defect image have the same dimension; a candidate defect information determination module configured to determine candidate defect information of the candidate defect image according to a target detection network; wherein the candidate defect information comprises a candidate defect type, a candidate defect position, and a candidate defect detection precision; a defect detection result determination module configured to determine target defect information from the candidate defect information according to the candidate defect detection precision, and determine a defect detection result of the to-be-tested defect image according to the target defect information.

8. The apparatus of claim 7, wherein, The image feature information determination module is specifically configured to: determine first-link first feature information of the to-be-tested defect image on the first feature extraction link according to a first-link first feature extraction group of the first feature extraction link; determine second-link first feature information of the to-be-tested defect image on the second feature extraction link according to a second-link first feature extraction group of the second feature extraction link; fuse the first-link first feature information and the second-link first feature information to obtain first fusion feature information of the to-be-tested defect image; determine first-link second feature information of the first fusion feature information on the first feature extraction link according to a first-link second feature extraction group of the first feature extraction link; determine second-link second feature information of the first fusion feature information on the second feature extraction link according to a second-link second feature extraction group of the second feature extraction link; The first link second feature information and the second link second feature information are fused to obtain second fusion feature information of the to-be-tested defect image; First link third feature information of the second fusion feature information on the first feature extraction link is determined according to a first link third feature extraction group of the first feature extraction link; Second link third feature information of the second fusion feature information on the second feature extraction link is determined according to a second link third feature extraction group of the second feature extraction link; The first link third feature information and the second link third feature information are fused to obtain third fusion feature information of the to-be-tested defect image; First link fourth feature information of the third fusion feature information on the first feature extraction link is determined according to a first link fourth feature extraction group of the first feature extraction link; Second link fourth feature information of the third fusion feature information on the second feature extraction link is determined according to a second link fourth feature extraction group of the second feature extraction link; The first link fourth feature information and the second link fourth feature information are fused to obtain fourth fusion feature information of the to-be-tested defect image; The first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information are taken as image feature information of the to-be-tested defect image.

9. The apparatus of claim 8, wherein, The candidate defect image determination module is specifically configured to: A first candidate defect image corresponding to the first fusion feature information is determined according to the first fusion feature information, the second fusion feature information, the third fusion feature information and the fourth fusion feature information; A second candidate defect image corresponding to the second fusion feature information is determined according to the first candidate defect image and the second fusion feature information; A third candidate defect image corresponding to the third fusion feature information is determined according to the second candidate defect image and the third fusion feature information; A fourth candidate defect image corresponding to the fourth fusion feature information is determined according to the second candidate defect image, the third fusion feature information and the fourth fusion feature information.

10. The apparatus of claim 8, wherein, The image feature information determination module is further configured to: Before determining the first link first feature information of the to-be-tested defect image on the first feature extraction link according to the first link first feature extraction group of the first feature extraction link, the to-be-tested defect image is subjected to feature preprocessing based on a preset convolution template to obtain reference feature information; wherein one zero value is arranged in adjacent elements of the preset convolution template, and the feature preprocessing includes image feature extraction and image channel amplification; Correspondingly, the image feature information determination module is further configured to: The first link first feature information of the reference feature information on the first feature extraction link is determined according to the first link first feature extraction group of the first feature extraction link; The second link first feature information of the reference feature information on the second feature extraction link is determined according to the second link first feature extraction group of the second feature extraction link.

11. The apparatus of any of claims 7-10, wherein, The candidate defect information determination module is specifically configured to: determine a candidate defect type of the candidate defect image according to a first prediction unit in the target detection network, wherein the first prediction unit is used for predicting an image defect type; determine a candidate defect position of the candidate defect image according to a second prediction unit in the target detection network, wherein the second prediction unit is used for predicting an image defect position; determine a candidate defect detection precision of the candidate defect image according to a third prediction unit in the target detection network, wherein the third prediction unit is used for predicting an image defect detection precision.

12. The apparatus of claim 11, wherein, The defect detection result determination module is specifically configured to: determine a maximum value in the candidate defect detection precision as a target defect detection precision; determine candidate defect information corresponding to the target defect detection precision as target defect information.

13. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the defect detection method of the water wall of the power station boiler according to any one of claims 1-6.

14. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are used to enable the processor to execute the defect detection method of the water wall of the power station boiler according to any one of claims 1-6 when executed.

Images