Magnetic flux leakage field scanning steel bar corrosion image recognition system and method thereof

Abstract

The application discloses a magnetic flux leakage field scanning steel bar corrosion image recognition system and method, and belongs to the technical field of nondestructive testing and computer vision technology. The system rapidly acquires a two-dimensional distribution image of a magnetic flux leakage field through a magnetic sensor array, adopts a semantic segmentation network based on an encoder-decoder architecture to perform pixel-by-pixel classification on the image, and divides healthy, mild, moderate and severe corrosion areas. Spatial attention and channel attention mechanisms are introduced to identify severe corrosion features such as magnetic flux leakage field gradient mutation and polarity reversal, and to enhance detection accuracy. A domain self-adaptive strategy of adversarial training is adopted to reduce feature distribution differences under different detection scenarios and to improve generalization ability. A three-dimensional reconstruction is performed on the corrosion spatial distribution model by combining reinforcement information through coordinate registration and spatial interpolation. A repair priority list and a bill of quantities are automatically generated. The application has high detection efficiency, accurate recognition accuracy, strong adaptability and visual intuition, and provides an intelligent technical means for concrete structure health monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-destructive testing and computer vision, in particular to a magnetic flux leakage field scanning steel bar corrosion image recognition system and method based on deep learning, which is suitable for automatic detection and evaluation of steel bar corrosion degree in concrete structure. BACKGROUND

[0002] Reinforced concrete structures are widely used in bridges, buildings and infrastructure construction. However, steel bar corrosion is one of the main reasons for the degradation of the durability of concrete structures and the decrease of bearing capacity. Factors such as chloride ion erosion, carbonation and electrochemical corrosion can cause different degrees of corrosion of steel bars, which in turn causes the reduction of steel bar cross section, the cracking and spalling of concrete protective layer, and the degradation of the bonding performance of steel bars and concrete, seriously threatening the safety and service life of the structure. Therefore, accurate detection and evaluation of the degree of steel bar corrosion is of great significance for developing reasonable maintenance strategies and ensuring the safety of the structure.

[0003] Traditional steel bar corrosion detection methods mainly include half-cell potential method, resistivity method and visual inspection. The half-cell potential method measures the electrochemical potential of the steel bar surface to determine the corrosion probability, but this method can only provide qualitative evaluation and cannot accurately quantify the corrosion degree, and is easily affected by environmental factors. The resistivity method measures the concrete resistivity to indirectly infer the corrosion risk, but its detection accuracy is low and it is difficult to identify local corrosion. Visual inspection relies on human experience and is low in efficiency and subjective, and cannot detect early corrosion inside the concrete protective layer.

[0004] Chinese utility model patent CN212904661U discloses a non-uniform corrosion non-destructive monitoring sensor for steel bars of existing reinforced concrete structures, which detects steel bar corrosion by external sensors and surface-mounted sensors using magnetic induction principle. This method measures the magnetic flux leakage field intensity by Hall sensors and determines the corrosion degree according to the change of the magnetic field. However, this method has the following shortcomings: first, it uses point magnetic induction detection, which requires point-by-point measurement of the magnetic flux leakage field signal, resulting in low detection efficiency and difficulty in obtaining complete images of the spatial distribution of corrosion; second, it relies on a single Hall voltage threshold to determine the corrosion level, which cannot accurately identify the complex spatial morphology and connectivity characteristics of the corrosion area; third, it lacks intelligent image analysis methods, and has poor adaptability to the texture differences of magnetic flux leakage field images under different concrete carbonation degrees, steel bar protective layer thicknesses and steel bar arrangement forms; fourth, it cannot directly present the three-dimensional spatial distribution of steel bar corrosion, making it difficult to provide visual support for maintenance decisions.

[0005] With the rapid development of computer vision and deep learning technologies, intelligent detection methods based on image semantic segmentation have made significant progress in industrial defect detection, medical image analysis, and other fields. Semantic segmentation networks such as U-Net, DeepLabV3+, and others use an encoder-decoder architecture that can classify images pixel by pixel, accurately identifying the boundaries and morphology of target regions. Attention mechanisms learn the importance weights of features, enhancing the network's ability to extract key features and improving recognition accuracy in complex scenarios. Domain adaptive learning uses adversarial training to reduce the feature distribution difference between the source and target domains, improving the model's generalization ability. However, current research has not yet applied these advanced techniques to the intelligent recognition of magnetic field scanning steel reinforcement corrosion images, and existing methods still have significant limitations in detection efficiency, recognition accuracy, and visual presentation.

[0006] Therefore, there is an urgent need to develop an efficient, accurate, and intelligent magnetic field scanning steel reinforcement corrosion image recognition system to overcome the shortcomings of traditional methods and achieve rapid detection, accurate identification, and intuitive visualization of steel reinforcement corrosion. SUMMARY

[0007] The purpose of the present application is to provide a magnetic field scanning steel reinforcement corrosion image recognition system and method, which rapidly acquires a two-dimensional distribution image of the magnetic field by using a magnetic sensor array, realizes automatic identification and quantitative evaluation of the corrosion area by using a deep learning semantic segmentation network, enhances the detection ability of severe corrosion by combining an attention mechanism, improves the generalization performance in different detection scenarios by using domain adaptive learning, and intuitively presents the spatial distribution of steel reinforcement corrosion through three-dimensional reconstruction, providing a scientific basis for concrete structure maintenance decisions.

[0008] In order to achieve the above-mentioned purpose, the present application provides a magnetic flux leakage field scanning steel bar corrosion image recognition system, which comprises a magnetic sensor array, an image preprocessing module, a semantic segmentation module, an attention enhancement module, a domain self-adaption module, a three-dimensional reconstruction module and an output module. The magnetic sensor array rapidly scans the concrete surface to obtain a magnetic flux leakage field two-dimensional distribution image of the steel bar corrosion area. The image preprocessing module performs normalization and enhancement processing on the magnetic flux leakage field image to generate a standardized magnetic flux leakage image. The semantic segmentation module classifies the standardized magnetic flux leakage image pixel by pixel using an encoder-decoder architecture. The encoder extracts hierarchical features through multi-scale convolution, and the decoder restores the spatial resolution through upsampling and generates a corrosion segmentation mask to divide the magnetic flux leakage image into healthy areas, mild corrosion areas, moderate corrosion areas and severe corrosion areas. The attention enhancement module identifies the local patterns strongly related to severe corrosion through a spatial attention unit and a channel attention unit. The spatial attention unit generates a spatial weight map based on the magnetic flux leakage field gradient features, and the channel attention unit generates a channel weight vector according to the importance of the feature channels. The high-risk corrosion area is highlighted and the background noise is suppressed through weighted fusion. The domain self-adaption module adopts an adversarial training strategy to reduce the difference in feature distribution of the magnetic flux leakage images in different detection scenarios through a feature extractor, a domain discriminator and a gradient reversal layer, and realizes domain invariant feature learning. The three-dimensional reconstruction module combines the reinforcement information in the structural design drawings, maps the two-dimensional magnetic flux leakage scanning image to the three-dimensional space through a coordinate registration algorithm, and generates a three-dimensional spatial distribution model of the steel bar corrosion. The output module generates a corrosion evaluation report, including a repair priority list sorted by corrosion severity and a bill of quantities.

[0009] The present application also provides a magnetic flux leakage field scanning steel bar corrosion image recognition method, comprising the following steps: obtaining a magnetic flux leakage field two-dimensional distribution image through a magnetic sensor array; preprocessing the image to generate a standardized magnetic flux leakage image; performing pixel-by-pixel classification using a semantic segmentation network to divide the corrosion level; identifying high-risk corrosion areas using an attention mechanism; improving the generalization ability using a domain self-adaption strategy; combining reinforcement information for three-dimensional reconstruction; and generating a corrosion evaluation report and a repair priority list.

[0010] Compared with the prior art, the present application has the following beneficial effects:

[0011] (1) The detection efficiency is significantly improved. The present application uses a magnetic sensor array to obtain a magnetic flux leakage field two-dimensional distribution image in one scan, and the point-by-point detection method of the prior art improves the detection efficiency by more than 80%, which can quickly obtain complete spatial distribution information of the corrosion area.

[0012] (2) Recognition accuracy is greatly improved. The present application realizes pixel-by-pixel classification of the magnetic flux leakage image based on a deep learning semantic segmentation network, can accurately identify the boundary, morphology and connectivity characteristics of the corrosion area, and compared with the traditional threshold segmentation method, the corrosion area recognition accuracy is improved by 25%, and the average intersection over union (IoU) reaches 0.89.

[0013] (3) Strong adaptability to complex scenes. The present application introduces an attention mechanism to automatically focus on the magnetic flux leakage field gradient mutation and polarity reversal and other serious corrosion features, enhances the detection ability of the key area, and still maintains high-precision recognition under complex background and noise interference. Through the synergistic effect of spatial attention and channel attention, the recall rate of the serious corrosion area reaches 94%.

[0014] (4) Generalization ability is significantly enhanced. The present application adopts a domain adaptive learning strategy, uses adversarial training to reduce the feature distribution difference under different concrete carbonation degrees and steel bar arrangement forms, so that the model can adapt to multiple detection scenes without a large amount of re-labeled training data. In the cross test of different scenes, the average recognition accuracy decreases by no more than 3%.

[0015] (5) The visual effect is intuitive. The present application maps the two-dimensional magnetic flux leakage scanning image to the three-dimensional space through the three-dimensional reconstruction technology, and intuitively presents the spatial distribution of the steel bar corrosion in combination with the virtual reality technology. Engineers can immerse themselves in the virtual reality device to view the internal corrosion condition of the structure, which significantly improves the scientificity and operability of the maintenance decision.

[0016] (6) Strong practicability. The present application automatically generates a maintenance priority list and a detailed bill of quantities according to the corrosion severity, provides a quantitative basis for the preventive maintenance and precise repair of the concrete structure, effectively reduces the maintenance cost and prolongs the service life of the structure. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the overall architecture schematic diagram of the magnetic flux leakage field scanning steel bar corrosion image recognition system of the present application;

[0018] Figure 2 is the structure schematic diagram of the attention enhancement module of the present application;

[0019] Figure 3 is the adversarial training schematic diagram of the domain adaptive module of the present application;

[0020] Figure 4 is the coordinate registration process schematic diagram of the three-dimensional reconstruction module of the present application;

[0021] Figure 5 is the flowchart of the magnetic flux leakage field scanning steel bar corrosion image recognition method of the present application. DETAILED DESCRIPTION

[0022] Please refer to the attachedFigures 1-5 The specific embodiments of the present application are described in detail below with reference to the accompanying drawings, so that the technical solutions of the present application are clearer. It should be noted that the protection scope of the present application is not limited to the following specific embodiments, and any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0023] Referring to Figure 1 The present application provides a magnetic flux leakage field scanning steel bar corrosion image recognition system, which comprises a magnetic sensor array 1, an image preprocessing module 2, a semantic segmentation module 3, an attention enhancement module 4, a domain self-adaptive module 5, a three-dimensional reconstruction module 6 and an output module 7.

[0024] The magnetic sensor array 1 is used for quickly scanning the concrete surface to obtain the two-dimensional distribution image of the magnetic flux leakage field of the steel bar corrosion area. The magnetic sensor array 1 comprises 16x16 magnetic sensor units, the magnetic sensor units adopt Hall effect sensors, the sensitivity is 140mV / mT, and the measurement range is -50mT to +50mT. The spacing between the magnetic sensor units is set to 5mm, the single scanning coverage area is 80mmx80mm, and the magnetic flux leakage field distribution of the local area can be quickly obtained. The sampling frequency is set to 100Hz to ensure that high-density data sampling points are obtained during scanning. The magnetic sensor array 1 moves along the concrete surface through a mechanical scanning device, and obtains the magnetic flux leakage field image region by region, the scanning speed is 10mm / s, and the distance between the magnetic sensor array and the concrete surface is kept constant at 3mm±0.5mm during scanning. The principle of generating the magnetic flux leakage field is based on the ferromagnetism of the steel bar. When the steel bar is corroded, the magnetic permeability of the corrosion part decreases, causing the magnetic force line to be distorted and leaked to the concrete surface, forming a detectable magnetic flux leakage field signal. The magnetic flux leakage field intensity of the healthy steel bar area is low and uniformly distributed, while the magnetic flux leakage field intensity of the corrosion area is significantly increased and presents obvious gradient change and polarity inversion characteristics.

[0025] The image preprocessing module 2 is connected with the magnetic sensor array 1, and is used for normalizing and enhancing the two-dimensional distribution image of the magnetic flux leakage field to generate a standardized magnetic flux leakage image. The normalization processing adopts a minimum-maximum normalization method, linearly maps the magnetic flux leakage field intensity value to the 0-1 interval, and the mapping formula is:

[0026] ,

[0027] Among them, is the normalized magnetic flux leakage field intensity value, is the original magnetic flux leakage field intensity value, and The minimum and maximum magnetic flux leakage field intensity values in the image, respectively. Normalization eliminates the numerical range differences between different scanning batches due to sensor sensitivity differences, environmental magnetic field interference, etc., ensuring the consistency of subsequent processing.

[0028] The enhancement processing includes two steps of denoising and contrast enhancement. The denoising adopts a Gaussian filtering method, and the filter kernel size is 5x5, and the standard deviation is set to 1.2, effectively suppressing sensor noise and environmental electromagnetic interference. The contrast enhancement adopts an adaptive histogram equalization method (CLAHE), which divides the image into 8x8 sub-regions, independently performs histogram equalization on each sub-region, and then smoothes the boundaries of adjacent sub-regions through bilinear interpolation. The clipping limit parameter is set to 2.0 to avoid noise amplification caused by excessive enhancement.

[0029] The semantic segmentation module 3 is connected with the image preprocessing module 2, and is used for classifying the standardized magnetic flux leakage image pixel by pixel, and dividing the magnetic flux leakage image into healthy area, mild corrosion area, moderate corrosion area and severe corrosion area. The semantic segmentation module 3 adopts an improved U-Net encoder-decoder architecture combined with the DeepLabV3+ Atrous Spatial Pyramid Pooling (ASPP) module, which expands the receptive field while maintaining high-resolution features.

[0030] The encoder unit includes 4 levels of down-sampling layers, and each level of down-sampling layer includes two convolution layers and one max-pooling layer. The number of convolution kernels of the first level of down-sampling layer is 64, the second level is 128, the third level is 256, and the fourth level is 512. The convolution kernel size of all convolution layers is 3x3, the step is 1, the padding is 1, and the ReLU activation function is adopted. The window size of the pooling layer is 2x2, the step is 2, and the spatial size of the input image is gradually reduced from 256x256 to 16x16 through step-by-step down-sampling, while extracting multi-scale representations from low-level edge texture to high-level semantic features.

[0031] The ASPP module is introduced at the bottom of the encoder, including a 1x1 convolution, three 3x3 atrous convolutions with different dilation rates (dilation rates are 6, 12, and 18 respectively), and a global average pooling branch. The atrous convolution expands the receptive field by inserting holes in the convolution kernel, capturing multi-scale context information without increasing the number of parameters and computational complexity. The global average pooling branch globally pools the feature map and then restores it to the same channel number through a 1x1 convolution, and then up-samples to the original size to encode the global context. The outputs of the five branches of the ASPP module are concatenated in the channel dimension, and then fused into a unified feature representation through a 1x1 convolution.

[0032] The decoder unit includes 4 up-sampling layers, each of which expands the spatial size of the feature map by 2 times through a deconvolution operation, and the number of convolution kernels is 256, 128, 64, and 32 in turn. In order to restore the spatial detail information lost in the down-sampling process, each level of the decoder is connected with the feature map of the corresponding level of the encoder through a jump connection, the encoder feature map is spliced with the up-sampled decoder feature map in the channel dimension, and then further fused through two 3x3 convolution layers. The jump connection mechanism effectively combines the high-resolution low-level features of the encoder and the low-resolution high-level semantic information of the decoder, significantly improving the positioning accuracy of the corrosion area boundary.

[0033] The last layer of the decoder maps the feature map to 4 channels through a 1x1 convolution, corresponding to 4 corrosion levels (healthy, mild, moderate, and severe), and then applies a softmax function to generate a probability distribution of each pixel belonging to each corrosion level. The class with the highest probability is selected as the predicted label of the pixel to form the final corrosion segmentation mask.

[0034] In the model training stage, the loss function adopts a combination of weighted cross-entropy loss and Dice loss. The weighted cross-entropy loss assigns different weights to different classes, with a weight of 1.0 for the healthy area, 1.5 for the mild corrosion, 2.0 for the moderate corrosion, and 2.5 for the severe corrosion, to address the class imbalance problem. The Dice loss directly optimizes the overlap between the predicted mask and the true label, improving the segmentation accuracy. The total loss function is defined as:

[0035] ,

[0036] wherein, is the weighted cross-entropy loss, is the Dice loss, is a balance coefficient, which is set to 0.6 in this embodiment. The Adam optimizer is used for training, with an initial learning rate of 0.001, a batch size of 16, a training epoch number of 150, and a learning rate decayed to 0.1 times of the original at the 100th and 130th epochs.

[0037] The attention enhancement module 4 is connected with the semantic segmentation module 3, and is used for identifying the local patterns strongly related to severe corrosion in the magnetic leakage field image, highlighting the high-risk corrosion area and suppressing background noise. Referring to Figure 2 , the attention enhancement module 4 includes a spatial attention unit and a channel attention unit.

[0038] The spatial attention unit generates a spatial weight map based on the magnetic leakage field gradient features. The magnetic leakage field gradient includes a horizontal gradient and a vertical gradient , which is calculated by Sobel operator. The leakage magnetic field intensity of severe corrosion area changes dramatically, which is characterized by large gradient amplitude and fast direction change. Gradient amplitude and gradient direction are calculated as follows:

[0039] ,

[0040] ,

[0041] The spatial attention unit concatenates the gradient amplitude map and the gradient direction map in the channel dimension, extracts spatial features through a 7x7 convolution layer, and then applies a sigmoid activation function to generate a spatial attention map , whose value ranges from 0 to 1. The position with a value close to 1 corresponds to the area with significant changes in the leakage magnetic field gradient, i.e., the potential severe corrosion location. The is multiplied element-wise with the feature map output by the semantic segmentation module to enhance the features of high-risk areas:

[0042] ,

[0043] where, is the output feature map of the semantic segmentation module, denotes element-wise multiplication, is the enhanced feature map.

[0044] The channel attention unit generates a channel weight vector based on the importance of each feature channel. For a feature map with channels, where and are the height and width of the feature map, respectively, the channel attention unit first performs global average pooling (GAP) and global maximum pooling (GMP) on each channel to obtain two feature vectors with a length of :

[0045] ,

[0046] ,

[0047] where, and are the global average pooling and global maximum pooling results of the th channel, respectively. Global average pooling captures the overall response intensity of the channel, and global maximum pooling captures the most significant features in the channel.

[0048] and are concatenated to form a channel weight vector A shared multi-layer perceptron (MLP) is inputted respectively, which includes a hidden layer compressing the number of channels to (wherein is a compression ratio, which is set to 16 in the embodiment), and an activation function of ReLU, and then is restored to channels. The outputs of the two MLPs are added and then passed through a sigmoid activation function to generate a channel attention weight vector :

[0049] ,

[0050] wherein is a sigmoid function, denotes a multi-layer perceptron. The channel attention weight characterizes the relevance of each feature channel to the corrosion recognition, and a channel with a larger value corresponds to a channel with a strong correlation with corrosion features. The channel attention weight is broadcasted with the feature map in the channel dimension to realize channel-level feature re-labeling:

[0051] ,

[0052] Finally, the outputs of the spatial attention and the channel attention are added to the original feature map through a residual connection to form an attention-enhanced feature representation:

[0053] ,

[0054] In a preferred embodiment, the attention enhancement module 4 can be embedded into multiple levels of the semantic segmentation module 3, and the attention mechanism is introduced at the 3rd and 4th levels of the encoder and the 1st level of the decoder to hierarchically enhance the expression ability of features at different scales.

[0055] The domain adaptation module 5 is connected to the attention enhancement module 4, which is used to reduce the difference in feature distribution of the magnetic leakage image under different carbonation degrees of concrete and different reinforcement arrangement forms, and to improve the generalization ability of the model in different detection scenarios. Referring to Figure 3 , the domain adaptation module 5 adopts an unsupervised domain adaptation strategy based on adversarial training, including a feature extractor, a domain discriminator, and a gradient reversal layer.

[0056] The feature extractor corresponds to the encoder part of the semantic segmentation module 3, and maps the input magnetic leakage image into a high-level feature representation. The source domain data is a labeled magnetic leakage image of corrosion grade, and the target domain data is an unlabeled magnetic leakage image to be detected, which may come from different detection environments (such as concrete with different carbonation degrees, different protection layer thicknesses, different reinforcement spacings, etc.), resulting in differences in feature distribution.

[0057] The domain discriminator is a multi-layer perceptron, which takes the high-level features output by the feature extractor as input and outputs the probability that the feature comes from the source domain or the target domain. The domain discriminator includes three fully connected layers, with the number of hidden layer neurons being 1024 and 512 respectively, the activation function being ReLU, and the output layer being a single neuron with the activation function being sigmoid. The training objective of the domain discriminator is to accurately distinguish between the source domain features and the target domain features:

[0058] ,

[0059] wherein, is the domain discrimination loss, denotes the feature extractor, denotes the domain discriminator, and are the i-th source domain sample and the j-th target domain sample respectively, and are the number of source domain samples and target domain samples respectively. and are the number of source domain samples and target domain samples respectively.

[0060] The gradient reversal layer (GRL) is located between the feature extractor and the domain discriminator, which keeps the features unchanged during forward propagation and multiplies the gradient of the domain discriminator by a negative scalar during backward propagation and then passes it to the feature extractor. The gradient reversal operation makes the optimization objective of the feature extractor opposite to that of the domain discriminator, i.e., the feature extractor tries to learn domain-invariant feature representations so that the domain discriminator cannot distinguish whether the features come from the source domain or the target domain. The mathematical definition of the gradient reversal layer is:

[0061] ,

[0062] ,

[0063] wherein, is the identity matrix, is the gradient reversal coefficient, which is dynamically adjusted according to the training progress:

[0064] ,

[0065] wherein, is the training progress (taking values between 0 and 1), is the adjustment rate, which is set to 10 in this embodiment. In the early stage of training, is small, and the feature extractor mainly focuses on the rust classification task of the source domain; as the training progresses, increases gradually, and the role of domain adaptation gradually increases.

[0066] The total loss function of the domain adaptation module combines the rust classification loss of the source domain and the domain discrimination loss:

[0067] ,

[0068] wherein, is the semantic segmentation loss of the source domain sample, is a balance coefficient, which is set to 0.5 in the embodiment. By jointly optimizing the two loss functions, the model learns accurate corrosion classification ability on the source domain, while learning a feature representation robust to domain changes, so that the model can be directly applied to the target domain without re-labeling and training.

[0069] In one specific embodiment, the source domain data is a standard working condition with a carbonization depth of 5 mm, a protective layer thickness of 30 mm, and a steel bar spacing of 100 mm, and the target domain data is an actual detection working condition with a carbonization depth of 15 mm, a protective layer thickness of 25 mm, and a steel bar spacing of 80 mm. Through domain adaptive training, the corrosion recognition accuracy of the model in the target domain is improved from 72% when not adapted to 87%, close to the effect of supervised training in the target domain (89%).

[0070] The three-dimensional reconstruction module 6 is connected with the semantic segmentation module 3 and the domain adaptive module 5, and is used to combine the reinforcement information in the structural design drawing, map the two-dimensional magnetic flux leakage scanning image to the three-dimensional space through a coordinate registration algorithm, and generate a three-dimensional spatial distribution model of steel bar corrosion. Referring to Figure 4 , the three-dimensional reconstruction module 6 includes a coordinate registration unit and a spatial interpolation unit.

[0071] The coordinate registration unit establishes a mapping relationship between the two-dimensional magnetic flux leakage scanning image coordinate system and the three-dimensional structure coordinate system based on the control points in the structural design drawing. Before detection, at least four non-coplanar control points are set on the concrete surface, and the coordinates of the control points in the three-dimensional structure coordinate system are accurately obtained through a measuring instrument , and the pixel coordinates of the control points in the two-dimensional magnetic flux leakage scanning image are recorded . A perspective transformation model is used to establish the mapping relationship from two-dimensional to three-dimensional:

[0072] ,

[0073] wherein, is a 3x4 transformation matrix containing 12 unknown parameters. The transformation matrix is solved by using the control point coordinates through the least squares method:

[0074] ,

[0075] wherein, is the number of control points (usually 8-12 to improve accuracy), is the first 3 rows and 3 columns of , is the first 3 rows and 3 columns of the 4th row of the transformation matrix, any pixel point in the two-dimensional magnetic flux leakage image can be mapped to the three-dimensional space coordinate . .

[0076] The spatial interpolation unit generates a continuous corrosion degree distribution field by three-dimensional spatial interpolation of discrete corrosion detection points. For multiple two-dimensional magnetic flux leakage images obtained by multiple scans, each image corresponds to a scanning position, and after being mapped to the three-dimensional space through coordinate registration, a discrete corrosion distribution point cloud is formed. A three-dimensional Kriging interpolation method is used to perform spatial interpolation on the point cloud data. Kriging interpolation is an optimal linear unbiased estimation method based on spatial autocorrelation. For the corrosion degree of any to-be-estimated point in the space , the corrosion degree is calculated by weighted average of surrounding known points as follows:

[0077] ,

[0078] wherein, is the corrosion degree of the th known point, is the corresponding weight, is the number of known points participating in interpolation. The weight is obtained by solving Kriging equations as follows:

[0079] ,

[0080] wherein, is the semi-variogram value corresponding to the distance , is the Euclidean distance between the th point and the th point in the space, is the Lagrange multiplier. The semi-variogram adopts an exponential model as follows:

[0081] ,

[0082] wherein, is the nugget value, is the sill value, is the range, and these parameters are determined by experimental data fitting. In the embodiment, is set to 0.05, is set to 0.95, is set to 150 mm.

[0083] The corrosion degree estimation value of the regular grid point in the three-dimensional space is generated by Kriging interpolation at an interval of 5 mm, forming a three-dimensional corrosion distribution field. By superimposing the corrosion distribution field and the steel bar position information in the structural design drawing, the corrosion degree distribution of each steel bar can be directly displayed.

[0084] In a preferred embodiment, the three-dimensional reconstruction module 6 can also combine the finite element model to evaluate the steel cross-section loss rate according to the corrosion degree, and then calculate the degradation degree of the structural bearing capacity, to provide quantitative basis for maintenance and reinforcement design.

[0085] The output module 7 is connected with the three-dimensional reconstruction module 6, for generating a corrosion evaluation report. The corrosion evaluation report includes the following contents:

[0086] (1) Corrosion grading statistics. According to the semantic segmentation results, the area proportions of healthy area, mild corrosion area, moderate corrosion area and severe corrosion area are calculated. For the entire detection area, the area percentage of each corrosion grade is calculated; for each steel element, the corrosion grade distribution of the steel is calculated.

[0087] (2) Maintenance priority list. The maintenance priority is determined based on the corrosion severity and spatial distribution. The specific determination criteria are as follows: the steel element with severe corrosion area proportion exceeding 15% is marked as a high-priority maintenance object and needs to be repaired immediately; the steel element with moderate corrosion area proportion exceeding 25% is marked as a medium-priority maintenance object and is recommended to be repaired within 6 months; the steel element with mild corrosion area proportion exceeding 40% is marked as a low-priority maintenance object and is recommended to be repaired within 1 year. For the high-priority maintenance object, the connectivity of corrosion is further evaluated, and if there is a continuous severe corrosion belt with a length exceeding 30 cm, it is upgraded to an urgent maintenance object.

[0088] (3) Bill of quantities. According to the three-dimensional reconstruction results, the concrete volume to be repaired, the steel length to be replaced, the anti-corrosion material area to be painted, etc. are calculated to provide quantitative data for maintenance cost estimation. For example, for the steel to be replaced, the cutting range is determined according to the corrosion distribution, and the total replacement length is calculated by extending 30 cm on both ends of the corrosion area as the lap length according to the specification requirements. For the concrete to be repaired, the chiseling depth (usually the thickness of the protective layer plus 20 mm) is determined according to the corrosion depth and range, and the repair volume is calculated.

[0089] (4) Three-dimensional visualization model. The data file of the three-dimensional corrosion spatial distribution model is output, supporting the import of virtual reality equipment for immersive viewing. In the virtual reality environment, engineers can rotate, zoom and section the structure model through the handle controller to intuitively view the corrosion condition of the internal steel. Different corrosion grades are marked with different colors: healthy area is green, mild corrosion is yellow, moderate corrosion is orange, and severe corrosion is red. Clicking on any steel element can view the detailed corrosion data of the element, including corrosion grade proportion, maximum corrosion depth, average corrosion rate, etc.

[0090] In one specific application case, the magnetic flux leakage field scanning detection was performed on the main girder of a bridge, and 150 magnetic flux leakage field images were obtained in the scanning area of 6m×2m. After semantic segmentation, attention enhancement and domain adaptation processing, 27 steel bars with different degrees of corrosion were identified, of which 5 were high priority (serious corrosion accounted for 18%-32%), 11 were medium priority (moderate corrosion accounted for 28%-45%), and 11 were low priority (mild corrosion accounted for 42%-67%). The generated engineering quantity list shows that the total length of the steel bars to be replaced is 18.6m, the concrete chiseling and repairing volume is 2.3m³, and the anti-corrosion coating brushing area is 12.5m². By viewing the three-dimensional model through the virtual reality device, it is found that there are three serious corrosion zones of steel bars in the main girder midspan area, which are interconnected and form a weak structure area. The engineer adjusts the reinforcement scheme accordingly and adds the outer steel plate reinforcement measure.

[0091] Reference Figure 5 The application also provides a magnetic flux leakage field scanning steel bar corrosion image recognition method, comprising the following steps:

[0092] Step S1: rapidly scan the concrete surface by a magnetic sensor array to obtain the two-dimensional distribution image of the magnetic flux leakage field of the steel bar corrosion area. The magnetic sensor array moves along the concrete surface, and the magnetic flux leakage field data is collected region by region. The sampling frequency is 100Hz, the scanning speed is 10mm / s, and the distance between the sensor array and the concrete surface is maintained at 3mm. For large-area detection objects, a partition scanning mode is adopted, each scanning area is 80mm×80mm, and the adjacent scanning areas overlap by 20mm to ensure data continuity.

[0093] Step S2: normalize and enhance the two-dimensional distribution image of the magnetic flux leakage field to generate a standardized magnetic flux leakage image. The normalization adopts the minimum-maximum normalization method to map the magnetic flux leakage intensity value to the interval of 0-1. The noise removal adopts Gaussian filtering with a filter kernel of 5×5 and a standard deviation of 1.2. The contrast enhancement adopts adaptive histogram equalization with a sub-region division of 8×8 and a clipping limit of 2.0.

[0094] Step S3: a semantic segmentation network is used to classify the standardized magnetic flux leakage image pixel by pixel, and the magnetic flux leakage image is divided into healthy area, mild corrosion area, moderate corrosion area and serious corrosion area. The semantic segmentation network adopts the U-Net encoder-decoder architecture, the encoder extracts multi-scale features through 4-level downsampling, introduces the ASPP module to capture multi-scale context information, the decoder restores the spatial resolution through 4-level upsampling, fuses the encoder features through the skip connection, and finally generates the corrosion segmentation mask through softmax.

[0095] Step S4: Attention mechanism is used to identify local patterns strongly related to severe corrosion, highlighting high-risk corrosion areas and suppressing background noise. The spatial attention unit calculates the horizontal and vertical gradients of the magnetic flux leakage field image using the Sobel operator, extracts gradient amplitude and gradient direction features, and generates a spatial attention map through 7x7 convolution and sigmoid activation. The channel attention unit performs global average pooling and global maximum pooling on the feature map, and generates channel attention weights through a shared MLP. Apply spatial and channel attention to the semantic segmentation feature map to achieve multi-dimensional feature enhancement.

[0096] Step S5: Domain adaptation strategy is used to reduce the difference in feature distribution of magnetic flux leakage images under different detection scenarios, and to improve the generalization ability of the model. The labeled source domain data and unlabeled target domain data are input into the feature extractor to extract high-level feature representations. The domain discriminator attempts to distinguish between source domain features and target domain features. The gradient reversal layer reverses the gradient of the domain discriminator during backpropagation, allowing the feature extractor to learn corrosion features that are invariant to the domain. Through adversarial training, the corrosion classification loss and domain discrimination loss are jointly optimized to achieve unsupervised domain adaptation.

[0097] Step S6: Combine the reinforcement information in the structural design drawings, and map the two-dimensional magnetic flux leakage scan images to the three-dimensional space through the coordinate registration algorithm to generate a three-dimensional spatial distribution model of steel reinforcement corrosion. Set control points on the concrete surface, measure the three-dimensional coordinates and two-dimensional image coordinates of the control points, and establish a coordinate mapping relationship through perspective transformation. Map multiple two-dimensional magnetic flux leakage images to three-dimensional space to form a discrete corrosion distribution point cloud. Use three-dimensional Kriging interpolation to spatially interpolate the point cloud data to generate a continuous corrosion severity distribution field. Superimpose the corrosion distribution field and the reinforcement position information to form a three-dimensional visualization model.

[0098] Step S7: Generate a corrosion evaluation report based on the semantic segmentation results and three-dimensional reconstruction results. Calculate the area proportion of each corrosion level to determine the repair priority: severe corrosion area proportion exceeding 15% is high priority, moderate corrosion area proportion exceeding 25% is medium priority, and mild corrosion area proportion exceeding 40% is low priority. Calculate the volume of concrete that needs to be repaired, the length of reinforcement that needs to be replaced, and other engineering quantities. Generate a three-dimensional visualization model data file to support immersive viewing with virtual reality devices.

[0099] In one specific embodiment, the method of the present application is used to detect a concrete pier that has been in service for 15 years. The pier has a size of 1.5m x 1.5m x 6m, and is reinforced with 16 main bars each having a diameter of 25mm, and with stirrups having a spacing of 200mm. The detection area is selected to be a range of 2m in height at the lower part of the pier, and a total of 320 magnetic flux leakage images are obtained by scanning. The image preprocessing takes about 5 minutes, the semantic segmentation and attention enhancement processing takes about 15 minutes, the domain adaptive model has been pre-trained and can be used directly, and the three-dimensional reconstruction takes about 10 minutes. The total detection time is about 30 minutes, which is significantly more efficient than the traditional point detection method (which takes 2-3 days).

[0100] The detection results show that 12 of the 16 main bars are corroded to varying degrees. Among them, 3 main bars (located on the water-facing surface of the pier) are severely corroded with a percentage of 22%-28%, and are determined to be high-priority maintenance objects. The corrosion mainly occurs in the humidity cycle area at a distance of 0.3m-1.2m from the ground. 5 main bars (located on the side of the pier) are moderately corroded with a percentage of 30%-38%, and are determined to be medium-priority maintenance objects. 4 main bars (located on the backwater surface of the pier) are lightly corroded with a percentage of 45%-55%, and are determined to be low-priority maintenance objects. The generated maintenance recommendations include: local removal and replacement of the 3 high-priority main bars with a total replacement length of 15.8m; removal and repair of the corroded area of the concrete with a repair volume of about 1.8m³; and brushing of a corrosion-resistant coating on the surface of all steel bars with a brushing area of about 18m². After implementation of the maintenance scheme, re-detection and verification show that the repair effect is good, and the structural bearing capacity is restored to more than 95% of the design level.

[0101] The present application quickly obtains two-dimensional distribution images of the magnetic flux leakage field through a magnetic sensor array, realizes automatic identification and accurate classification of the corrosion area by combining a deep learning semantic segmentation network, enhances the detection ability for severe corrosion by using an attention mechanism, improves the generalization performance in different scenes by using domain adaptive learning, and intuitively presents the spatial distribution of corrosion by three-dimensional reconstruction, thereby providing an efficient, accurate and intelligent technical means for health monitoring and maintenance decision-making of concrete structures, and having significant engineering application value and broad popularization prospects.

[0102] The above merely describes preferred embodiments of the present application, but should not be used to limit the protection scope of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A magnetic flux leakage field scanning rebar corrosion image recognition system, characterized by, The method comprises the steps of: a magnetic sensor array for rapidly scanning the surface of the concrete to obtain a two-dimensional distribution image of the magnetic leakage field of the steel bar corrosion area; an image preprocessing module connected with the magnetic sensor array, for normalizing and enhancing the two-dimensional distribution image of the magnetic leakage field to generate a standardized magnetic leakage image; a semantic segmentation module connected with the image preprocessing module, for classifying the standardized magnetic leakage image pixel by pixel, including an encoder unit and a decoder unit, the encoder unit extracts hierarchical features of the magnetic leakage image through multi-scale convolution, and the decoder unit restores the spatial resolution through upsampling and generates a corrosion segmentation mask to divide the magnetic leakage image into healthy area, mild corrosion area, moderate corrosion area and severe corrosion area; an attention enhancement module connected with the semantic segmentation module, for identifying local patterns strongly related to severe corrosion in the magnetic leakage field image, including a spatial attention unit and a channel attention unit, the spatial attention unit generates a spatial weight map based on the magnetic leakage field gradient features, and the channel attention unit generates a channel weight vector according to the importance of the feature channels to highlight the high-risk corrosion area and suppress the background noise; a domain adaptation module connected with the attention enhancement module, for reducing the difference in feature distribution of the magnetic leakage image under different concrete carbonization degrees and steel bar arrangement forms, including a feature extractor, a domain discriminator and a gradient reversal layer, the domain discriminator is used to distinguish source domain features and target domain features, and the gradient reversal layer reverses the gradient direction during back propagation to realize domain-invariant feature learning; a three-dimensional reconstruction module connected with the semantic segmentation module and the domain adaptation module, for combining the reinforcement information in the structural design drawing to map the two-dimensional magnetic leakage scanning image to the three-dimensional space through a coordinate registration algorithm to generate a three-dimensional spatial distribution model of the steel bar corrosion; an output module connected with the three-dimensional reconstruction module, for generating a corrosion evaluation report, the corrosion evaluation report includes a repair priority list sorted by corrosion severity and a bill of quantities.

2. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1, characterized in that, The encoder unit in the semantic segmentation module includes 4 down-sampling layers, each down-sampling layer includes a convolution layer and a pooling layer, the convolution kernel size is 3x3, the step is 1, and the pooling window size is 2x2, the magnetic leakage image features from low level to high level are extracted through step-by-step down-sampling, and the decoder unit includes 4 up-sampling layers, each up-sampling layer restores the spatial size of the feature map through deconvolution operation and performs jump connection fusion with the feature map of the corresponding layer of the encoder.

3. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, The spatial attention unit in the attention enhancement module identifies the magnetic field mutation and polarity inversion position by calculating the horizontal gradient and vertical gradient of the magnetic leakage field image to generate a spatial attention map, the position with larger value in the spatial attention map corresponds to the severe corrosion area, and the channel attention unit performs global average pooling and global maximum pooling on each channel of the feature map, generates a channel weight vector through a fully connected layer, and is used to strengthen the channel related to the corrosion feature and weaken the noise channel.

4. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, The domain adaptation module adopts an adversarial training strategy, the domain discriminator is a multi-layer perceptron structure, the input is high-level features output by the feature extractor, and the output is the probability that the features come from the source domain or the target domain, the gradient inversion layer keeps the features unchanged during forward propagation, and transmits the gradient of the domain discriminator multiplied by a negative number to the feature extractor during backward propagation, so that the feature extractor learns the corrosion feature representation that is invariant to the domain.

5. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, The three-dimensional reconstruction module includes a coordinate registration unit and a spatial interpolation unit, the coordinate registration unit establishes a mapping relationship between a two-dimensional magnetic flux leakage scanning image coordinate system and a three-dimensional structure coordinate system based on control points in a structural design drawing, and the spatial interpolation unit performs three-dimensional spatial interpolation on discrete corrosion detection points to generate a continuous corrosion degree distribution field.

6. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, When the image preprocessing module performs normalization processing on the two-dimensional magnetic flux leakage field distribution image, the magnetic flux leakage field intensity value is linearly mapped to the 0-1 interval, and the mapping formula is y=(x-x min) / (x max-x min), wherein x is the original magnetic flux leakage field intensity value, and x min and x max are the minimum and maximum magnetic flux leakage field intensity values in the image, respectively.

7. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, The output module calculates the area ratio and spatial distribution of each corrosion grade according to the semantic segmentation result, marks a steel bar unit with a severe corrosion area ratio exceeding 15% as a high-priority repair object, marks a steel bar unit with a moderate corrosion area ratio exceeding 25% as a medium-priority repair object, and marks a steel bar unit with a slight corrosion area ratio exceeding 40% as a low-priority repair object.

8. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, The magnetic sensor array includes 16*16 magnetic sensor units, the spacing between the magnetic sensor units is 5 mm, the single scanning coverage area is 80 mm*80 mm, and the sampling frequency is 100 Hz.

9. The magnetic flux leakage field scanning steel bar corrosion image recognition system according to claim 1 or 2, characterized by, The system further includes a virtual reality display module connected with the three-dimensional reconstruction module, configured to present the three-dimensional corrosion spatial distribution model in an immersive manner, and support engineers to interactively view and analyze through a virtual reality device.

10. A method for identifying the corrosion image of steel bars by magnetic flux leakage scanning, using the system for identifying the corrosion image of steel bars by magnetic flux leakage scanning according to any one of claims 1-9, characterized in that, Comprising: S1: rapidly scanning the concrete surface through the magnetic sensor array to obtain a two-dimensional magnetic flux leakage field distribution image of the steel bar corrosion area; S2: performing normalization processing and enhancement processing on the two-dimensional magnetic flux leakage field distribution image to generate a standardized magnetic flux leakage image; S3: performing pixel-by-pixel classification on the standardized magnetic flux leakage image by using a semantic segmentation network, extracting multi-scale hierarchical features by using an encoder, restoring the spatial resolution by upsampling through a decoder, and generating a corrosion segmentation mask to divide the magnetic flux leakage image into a healthy area, a slight corrosion area, a moderate corrosion area, and a severe corrosion area; S4: identifying a local pattern strongly related to severe corrosion by using an attention mechanism, generating a spatial weight map based on the magnetic flux leakage field gradient features through a spatial attention unit, generating a channel weight vector according to the importance of the feature channels through a channel attention unit, and highlighting the high-risk corrosion area and suppressing the background noise after weighted fusion; S5: reducing the difference in feature distribution of the magnetic flux leakage image under different detection scenes by using a domain adaptation strategy, distinguishing the source domain features and the target domain features through a domain discriminator, and reversing the gradient direction during backward propagation through a gradient inversion layer, so that the feature extractor learns the corrosion feature representation that is invariant to the domain. S6: Combine the reinforcement information in the structural design drawing, map the two-dimensional magnetic flux leakage scanning image to the three-dimensional space through the coordinate registration algorithm, perform spatial interpolation on the discrete detection points, and generate a three-dimensional spatial distribution model of steel bar corrosion; S7: Generate a corrosion evaluation report according to the semantic segmentation result and the three-dimensional reconstruction result, wherein the corrosion evaluation report includes a maintenance priority list sorted by corrosion severity and a bill of quantities.

Images