Cable bent tower construction progress monitoring method and system

Abstract

The invention relates to a cable bent tower construction progress monitoring method and system. The method comprises the following steps that image data of cable bent tower construction height is collected; a UKAN-Transformer model based on a learnable activation function is constructed, wherein the UKAN-Transformer model comprises an encoder, a decoder and a Colmogolov-Arnodel converter model structure; training a UKAN-Transform model based on a learnable activation function by using the data set, segmenting the cable bent tower construction image by using the trained UKAN-Transform model based on the learnable activation function, and obtaining a cable bent tower construction segmented image marked out of a cable bent tower construction progress area; the pixel area of a cable bent tower construction progress area in the cable bent tower construction segmentation image and the pixel width of the bottom of the area are counted, and the actual height of cable bent tower construction is obtained; and judging the progress of the cable bent tower construction by comparing the actual height of the cable bent tower construction with the expected completed height of the cable bent tower construction. The problems that traditional cable bent tower construction progress monitoring is inaccurate, the timeliness is poor and the workload is huge can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent construction of cable tower construction, and in particular to a cable tower construction progress monitoring method and system. Background Art

[0002] With the rapid development of modern cable tower bridge engineering technology, cable towers are key load-bearing components of structures such as suspension bridges and cable-stayed bridges. However, during their construction, due to various reasons, there will be various difficulties in monitoring the progress of cable tower construction. The cable tower is located in the water area, which greatly restricts the monitoring of construction progress, and manual observation is usually used. The use of manual observation methods to judge the progress monitoring is inaccurate, which will cause the results to differ greatly from the actual situation. At the same time, the observation results cannot be quantified, and the specific height of the cable tower construction cannot be determined. Therefore, the significance of monitoring and measuring the progress of cable tower construction is extremely important.

[0003] The height, construction size and progress of the cable tower can be monitored through the layout measurement, but it increases the safety risk of the staff and is prone to falling from height. At the same time, the use of manual methods cannot achieve real-time monitoring and is prone to lag in progress information. These problems will not only affect the progress of construction, increase the risk of accidents during construction, affect the life safety of workers, but may also have a negative impact on the social economy and social life. With the continuous maturity of technologies such as the Internet of Things and artificial intelligence, the progress monitoring of cable tower construction is gradually developing in the direction of intelligence. By applying deep learning methods to realize automatic collection, intelligent analysis and early warning of monitoring data, it is an effective way to achieve real-time and accurate monitoring of the progress of the cable tower. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for monitoring the progress of cable tower construction, which can realize automatic, accurate and real-time monitoring of the progress of cable tower construction.

[0005] To achieve the above object, the technical solution of the present invention is:

[0006] In a first aspect, the present invention provides a method for monitoring the progress of cable tower construction, the method comprising the following steps:

[0007] Step 1: Image data collection of cable tower construction height:

[0008] Use the image capturing machine on the tower crane during the construction of the cable tower to take pictures. Fix the camera on the top of the cable tower. The image capturing machine can capture the construction object of the cable tower. During the data collection process, the horizontal distance between the image capturing machine and the side of the construction object remains unchanged. During the construction process, the image capturing machine can always capture pictures of the construction object at all heights. Collect all pictures during the entire construction process at regular intervals to build a data set.

[0009] Step 2: Build a UKAN-Transformer model based on a learnable activation function:

[0010] The UKAN-Transformer model adopts the framework of the UNet-Transformer model, including the encoder, decoder, and Kolmogorov-Arnold transformer model structure;

[0011] The encoder includes a preprocessing module, three 3×3 U-Net convolution modules and two KAN modules which are sequentially connected from top to bottom; the decoder includes three 3×3 U-Net convolution modules and two KAN modules which are sequentially connected from top to bottom; each U-Net convolution module and KAN module of the encoder is connected to the corresponding U-Net convolution module and KAN module of the decoder using a skip connection at the same time;

[0012] The output of the U-Net convolution module of the top layer of the decoder is used as the output feature of the decoder, and the output feature of the decoder is input into the Kolmogorov-Arnold converter model structure; the Kolmogorov-Arnold converter model structure includes a bridge layer module, a Kolmogorov-Arnold feature embedding layer KAN-Feature Embedding module, a Kolmogorov-Arnold converter KAN-Transformer module and a Kolmogorov-Arnold feedforward network KAN-Feed-Forward Net module connected in sequence from top to bottom; KAN modules are added to the Kolmogorov-Arnold feature embedding layer KAN-Feature Embedding module, the Kolmogorov-Arnold converter KAN-Transformer module and the Kolmogorov-Arnold feedforward network KAN-Feed-ForwardNet module;

[0013] So far, the UKAN-Transformer model based on the learnable activation function is obtained;

[0014] The KAN module includes a Tokenization layer, a 3×3 convolution layer, a KAN structure, a deep convolution layer and layer normalization. After the input feature sequence is processed by Tokenization, a series of flattened two-dimensional patches are obtained. The output structure of the 3×3 convolution layer, the KAN structure and the deep convolution layer is residually connected with the feature sequence before Tokenization. Finally, the output of the KAN module is obtained after layer normalization.

[0015] The preprocessing module includes a 5×5 standardized convolution, a group normalization, a ReLU activation function and a pooling layer connected in sequence;

[0016] Step 3: Use the data set constructed in step 1 to train the UKAN-Transformer model based on the learnable activation function, and use the trained UKAN-Transformer model based on the learnable activation function to segment the cable tower construction image to obtain the cable tower construction segmentation image that marks the cable tower construction progress area;

[0017] Step 4: Count the pixel area of ​​the tower construction progress area in the tower construction segmentation image and the pixel width at the bottom of the area, and obtain the actual height of the tower construction based on the pixel area and pixel width; and judge the progress of the tower construction by comparing the actual height of the tower construction with the expected height of the tower construction.

[0018] Furthermore, in the Kolmogorov-Arnold feature embedding layer module, the feature sequence is fused and flattened by 3×3 convolution before entering the KAN module, and then entering the Kolmogorov-Arnold transformer KAN-Transformer module;

[0019] In the Kolmogorov-Arnold transformer KAN-Transformer module, the feature sequence is first standardized and then enters the multi-head attention layer. The output of the multi-head attention layer is spliced ​​with the standardized input and then enters a KAN module. The output of the KAN module is spliced ​​with the input of the KAN module after the second standardization process to obtain the output of the Kolmogorov-Arnold transformer KAN-Transformer module;

[0020] Afterwards, in the Kolmogorov-Arnold feedforward network KAN-Feed-ForwardNet module, the feature sequence is reshaped and upgraded, and then processed by the KAN module and the CBR module to obtain the final prediction map.

[0021] Furthermore, during the UKAN-Transformer model training process,

[0022] The formula for the change of the learning rate of the i-th generation is:

[0023] N1=a*T,

[0024]

[0025] Among them, T represents the number of training generations, N1 represents the number of generations with increased learning rate, a represents the proportion of generations with increased learning rate, and b represents the maximum learning rate V max With the minimum learning rate V min The ratio of V i is the learning rate of the i-th round, and β represents the coefficient of increase or decrease of the learning rate;

[0026] The loss function is the weighted cross entropy loss function E, and the formula is:

[0027]

[0028] Where Ω is the true label of each pixel, is the activation value of the class output given by the label corresponding to point x, and ω is the weight added to each pixel during training.

[0029] Furthermore, T = 100-150, V max The value range of is between [0.005-0.01], and the value of b is greater than 1.

[0030] Furthermore, the process of the bridge layer module is expressed as:

[0031]

[0032] Among them, H0 is the height of the feature map, Flatten() is the flattening operation, Represents a self-similar convolution operation, that is, the same convolution layer C b Convolve its own output again; F f is the output feature of the bridge layer module; F b is the input feature of the bridge layer module, that is, the output feature of the decoder after upsampling.

[0033] Furthermore, the size of the convolution kernel F of the self-similar convolution operation is 3, the stride D is 4, the padding P is 1, and the output resolution of the bridge layer module is calculated as:

[0034]

[0035] Among them, H0 is the height of the input feature map of the bridge layer module, H out and W out is the height and width of the output feature map of the bridge layer module.

[0036] In a second aspect, the present invention provides a cable tower construction progress monitoring system, the system comprising:

[0037] An image acquisition module is used to obtain images of the height direction of the construction object during the cable tower construction process;

[0038] The data processing module is used to mark the tower construction progress area on the pictures collected by the image acquisition module;

[0039] The cable tower construction progress area segmentation model is used to obtain a segmented image of the cable tower construction progress area according to the input image;

[0040] A statistical calculation module is used to count the pixel area of ​​the cable tower construction progress area in the segmented image and the pixel width of the bottom of the area, and further calculate and convert to obtain the actual height of the cable tower construction progress area;

[0041] The progress monitoring module is used to compare the actual height monitored by the statistical calculation module with the preset expected height of the cable tower construction to determine the real-time progress R0 of the cable tower construction; compare the real-time progress with the expected target progress R. If R0 is less than R, the expected progress of the cable tower construction has not been completed and acceleration measures should be taken; if R0 is greater than R, the expected progress of the cable tower construction has been exceeded and the construction is expected to be completed ahead of schedule.

[0042] Furthermore, the tower construction progress area segmentation model is implemented using the UKAN-Transformer model based on a learnable activation function.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention creatively adds a KAN module with a learnable activation function to the UNet-Transformer model, and constructs a new UKAN-Transformer model based on the learnable activation function, which can solve the problems of inaccurate, poor timeliness and huge workload of traditional cable tower construction progress monitoring, replaces manual monitoring with machine monitoring, avoids life safety problems that may occur in manual operations, avoids the situation where workers cannot work in severe weather conditions, avoids the situation where workers cannot monitor the construction progress due to environmental factors, and avoids the situation where instruments are damaged due to improper operation.

[0045] The present invention realizes active model learning, reduces the complexity of cable tower construction progress monitoring, can reduce workload, and can accurately and efficiently monitor the actual height of the cable tower construction, and then judge the specific construction progress according to the actual height, thereby realizing real-time monitoring and providing an effective means for cable tower construction progress monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a structural diagram of the UKAN-Transformer model in the present invention.

[0047] Figure 2 It is a structural schematic diagram of the preprocessing module in the present invention.

[0048] Figure 3 It is a structural schematic diagram of the KAN structure in the present invention.

[0049] Figure 4 It is a structural schematic diagram of the KAN module in the present invention.

[0050] Figure 5 It is a structural schematic diagram of the Kolmogorov-Arnold converter model structure in the present invention.

[0051] Figure 6 It is a schematic flow diagram of the method of the present invention. DETAILED DESCRIPTION

[0052] The present invention is further explained below in conjunction with the embodiments and drawings, but this is not intended to limit the scope of protection of the present application.

[0053] Example 1

[0054] The overall method of the cable tower construction progress monitoring method of the present invention is: cable tower construction image data collection, using the UKAN-Transformer model based on the learnable activation function to train the cable tower construction image, obtain information about the cable tower construction, calculate the height of the cable tower construction, and finally judge the progress of the cable tower construction by comparing the height of the cable tower construction at the current stage with the height of the cable tower construction expected to be completed. The method comprises the following steps:

[0055] Step 1: Image data collection of cable tower construction height:

[0056] The image capturing machine on the tower crane during the construction of the cable tower is used for shooting. The camera is fixed at the bottom of the cable tower. There is a fixed horizontal distance between the cable tower and the image capturing machine. During the construction of the cable tower, the cable tower can be raised layer by layer. The image capturing machine shoots in the height direction of the side of the construction object. The captured picture contains the cable tower and the construction object. In this embodiment, the image capturing machine uses a high-precision camera. The horizontal distance between the image capturing machine and the cable tower is about 150m, which can capture the size of the cable tower construction site. The image capturing machine will take pictures every 60 minutes and store the captured images in the memory. 24 pictures are taken every day. After 3 months of such a cycle, a total of 2160 pictures can be obtained, and the captured images are all PNG pictures of 512×512 pixels.

[0057] The captured images were randomly shuffled, and the tower construction progress areas were marked on the images. Then, the images were divided into training set, test set and validation set in a ratio of 8:1:1, with the number of images in each set being 1728, 216 and 216 respectively.

[0058] Step 2: Build a UKAN-Transformer model based on a learnable activation function:

[0059] The UKAN-Transformer model adopts the framework of the UNet-Transformer model. The overall structure is as follows Figure 1As shown in the figure, it includes three main modules, namely encoder (Encoder), decoder (Decoder) and Kolmogorov-Arnold transformer (KAN-Transformer) model structure.

[0060] The encoder includes 6 sub-modules, which are, from top to bottom, a preprocessing module of the first layer, a 3×3 U-Net convolution module of the second layer, a 3×3 U-Net convolution module of the third layer, a 3×3 U-Net convolution module of the fourth layer, a KAN module of the fifth layer, and a KAN module of the sixth layer.

[0061] After downsampling, the input image input preprocessing module passes through three U-Net convolution modules to obtain a feature extraction map. The feature extraction map enters the first KAN module. In the first KAN module, it will first be tokenized, that is, the features will be reshaped to obtain a series of flattened two-dimensional patches. After passing through the 3×3 convolution layer and the KAN structure, the structure output by the deep convolution processing is residually connected with the feature sequence before tokenization, and the output of the first KAN module is obtained after layer standardization. Then it enters the second KAN module. Each U-Net convolution module and KAN module of the encoder is connected to the corresponding U-Net convolution module and KAN module of the decoder using jump connections at the same time.

[0062] The decoder includes 5 submodules, which are the 3×3 U-Net convolution module of the first layer, the 3×3 U-Net convolution module of the second layer, the 3×3 U-Net convolution module of the third layer, the KAN module of the fourth layer, and the KAN module of the fifth layer from top to bottom. In terms of feature input, except for the KAN module of the fifth layer, each submodule of each decoder has two inputs, one is the feature of the output of the previous layer, and the other is the output feature of the submodule of the corresponding level of the corresponding encoder through jump connection. The input of the KAN module of the fifth layer is the output of the KAN module of the sixth layer in the encoder.

[0063] In the U-Net convolution module of the decoder, the two inputs are matrix-added and then subjected to a 1×1 convolution for feature fusion to obtain the output of the U-Net convolution module; the U-Net convolution module of the first layer of the decoder is used as the output feature of the decoder and output to the Kolmogorov-Arnold transformer (KAN-Transformer) model structure.

[0064] The KAN-Transformer model structure includes four submodules, namely a bridge layer module at the first layer, a KAN-Feature Embedding module at the second layer, a KAN-Transformer module at the third layer, and a KAN-Feed-ForwardNet module at the fourth layer.

[0065] The output features of the decoder are upsampled and input into the bridge layer module. In the bridge layer module, the input is subjected to a self-similar convolution operation, flattened and output to the KAN-Feature Embedding module of the next layer. In the Kolmogorov-Arnold Feature Embedding module, the feature sequence is subjected to 3×3 convolution fusion and flattening before entering the KAN module, and then entering the KAN-Transformer module. In the Kolmogorov-Arnold Transformer module, the feature sequence is first standardized and then enters the multi-head attention layer. The output of the multi-head attention layer is concatenated with the standardized input and then enters a KAN module. The output of the KAN module is normalized for the second time and concatenated with the input of the KAN module to obtain the output of the KAN-Transformer module. After passing through the Kolmogorov-Arnold converter module, the feature sequence is output to the Kolmogorov-Arnold feed-forward network (KAN-Feed-ForwardNet) module. In the Kolmogorov-Arnold feed-forward network module, the feature sequence undergoes reshaping and upgrading operations, and then enters the KAN module and the CBR module to finally obtain the prediction graph.

[0066] Step 3: Use the dataset constructed in step 1 to train the UKAN-Transformer model based on the learnable activation function:

[0067] First, the hyperparameters of the UKAN-Transformer model (initial weights, learning rate, etc.) are randomly initialized, and the constructed UKAN-Transformer model is trained for 100 epochs using the training set. After each epoch of training, the validation set is used to verify the generalization ability of the deep learning model and prevent overfitting.

[0068] The learning rate determines the speed and direction in which the model updates the weight parameters during the training process. Specifically, in each iteration, the UKAN-Transformer model calculates the gradient of the loss function with respect to each parameter. This gradient indicates the direction in which the parameter should be adjusted to minimize the loss, and the learning rate is the size of the adjustment amplitude during the adjustment process, that is, the amount of parameter update. If the learning rate is set too large, then in each iteration process, the UKAN-Transformer model parameters may miss the optimal solution, resulting in oscillation or divergence; if the learning rate is set too small, the UKAN-Transformer model will converge to the optimal solution very slowly, and may fall into a local minimum point instead of a global optimal solution. The setting of the learning rate is as follows: The present invention adopts a method of first reducing the learning rate and then increasing the learning rate to control the size of the learning rate. At the beginning of training, first reduce the learning rate. If the loss function decreases slowly, the training enters a local minimum. At this time, you need to increase the learning rate to increase the step size and jump out of the minimum. If the loss function oscillates, the training enters a local maximum. You need to reduce the learning rate appropriately to move the training to a more precise position and gradually find the optimal solution. When the learning rate increases, it increases at a faster rate; when the learning rate decreases, it decreases at a slow rate. The former accelerates because of fast learning features, and the latter decelerates because the learning rate is too high. The update of model parameters may be very drastic, resulting in unstable training or even divergence. The purpose is to speed up the convergence of parameters to the appropriate area. The formula for the change of the i-th generation learning rate is as follows:

[0069] N1=a*T,

[0070]

[0071] Where T represents the number of training generations. In this embodiment, T=100, N1 represents the number of generations at which the learning rate increases, a represents the proportion of generations at which the learning rate increases, and b represents the maximum learning rate V max With the minimum learning rate V min The ratio of V i is the learning rate of the i-th round, β represents the coefficient of increase or decrease of the learning rate, and the calculation formula is shown above. Maximum learning rate V max The ratio a of the number of generations of learning rate increase is a fixed value, which is set at the beginning of training. max is a constant, so the learning rate V of the i-th generation i It is only related to the coefficient β by which the learning rate increases or decreases. max The value range of is between [0.005-0.01], the value of b is greater than 1, when i≤N1, that is, the learning rate needs to increase continuously, the value of β is It increases as i increases.

[0072] Secondly, the parameter model is adjusted using the loss function. The loss function of the model of the present invention is a weighted cross entropy loss function E, and the formula is as follows:

[0073]

[0074] Where Ω is the true label of each pixel, is the activation value of the class output given by the label corresponding to point x, and ω is the weight added to each pixel during training.

[0075] Set up a visualization chart to display the accuracy and loss curves in real time. During the training process, the current accuracy and time consumption will be output after each epoch. By observing the trend of the curve, analyze whether the convergence and accuracy of the UKAN-Transformer model meet the expected requirements. If the UKAN-Transformer model cannot converge or the accuracy is low after convergence, it needs to be adjusted by adjusting the hyperparameters. When the value of the loss function E is less than 0.001, it means that the UKAN-Transformer model has reached convergence and the training is stopped.

[0076] Step 4: Use the trained UKAN-Transformer model to segment the image of the cable tower construction progress to be segmented, and calculate the height of the cable tower during construction based on the segmentation results:

[0077] The cable tower construction progress image to be segmented is input into the trained UKAN-Transformer model for segmentation, and the cable tower construction progress image marked in red is obtained. That is, the cable tower construction progress area is represented by red pixels, and the area other than the cable tower construction progress area is segmented by black pixels, so that the pixel value of the cable tower construction progress area in the image can be obtained.

[0078] Then, the pixel counting method is used to calculate the number of pixels in the area and the number of pixels in the width of the tower construction progress area in the segmented image, represented by n and m. The number of pixels n in the entire tower construction progress area is multiplied by the square of the unit pixel size to obtain the area A of the tower construction progress area, in pixels squared. The number of pixels m at the bottom of the tower construction progress area is multiplied by the length of the unit pixel to obtain the pixel width B of the tower construction progress area, in pixels length.

[0079] The pixel area A and pixel width B are converted according to the proportional relationship with the actual distance to obtain the actual area S (in square meters) and width W (in meters) of the tower construction progress area, and then the formula is used:

[0080] L = S / W,

[0081] Get the actual height L (in meters) of the tower construction progress area.

[0082] By comparing the actual height L obtained with the preset expected height of the cable tower construction, the accurate progress of the cable tower construction can be obtained.

[0083] Example 2

[0084] The structures of the various parts of the UKAN-Transformer model in this embodiment are as follows:

[0085] In the encoder, the input image is first input to the preprocessing module of the first layer for processing, and an image with a width of W0, a height of H0, and a number of channels of C0 is obtained. W0 is the width of the image after the preprocessing module, which indicates the number of pixels in the horizontal dimension of the image. H0 is the height of the image after the preprocessing module, which indicates the number of pixels in the vertical dimension of the image (i.e., the image height). C0 is the number of channels, which indicates the number of channels in an image. The preprocessing module here includes 5×5 standardized convolutions, group normalization, ReLU activation functions, and pooling layers connected in sequence, which can solve the gradient explosion problem and prevent data overfitting. It has stable training dynamics, does not depend on the statistics between batches, does not depend on the batch size, is insensitive to the batch size, and is particularly suitable for situations where the batch size is limited. It improves the performance on small data sets, simplifies the deployment and maintenance of the model, and can accelerate the training efficiency of the model.

[0086] Then, in the U-Net convolution block of the second layer of the encoder, this layer will perform a 3×3 convolution operation on the input image, and then extract the multi-scale feature representation of the image. The U-Net convolution block of the second layer will compress the size of the extracted feature map to 1 / 4 of the original image. At this time, the width of the feature map is Height is The number of channels is C1, and it is used as the input of the U-Net convolution module in the third layer.

[0087] At this time, the U-Net convolution module of the third layer will perform 3×3 convolution and other operations on the feature map, further reducing the size of the image to 1 / 16 of the original image. The width of the feature map is Height is The number of channels is C2.

[0088] Then the output of the U-Net convolution module of the third layer is input to the U-Net convolution module of the fourth layer. The U-Net convolution module of the fourth layer will undergo 3×3 convolution and other operations to reduce the feature map to 1 / 64 of the original size. At this time, the width of the feature map is Height is The number of channels is C3. The number of channels and the size of the feature map change each time the processing is performed. The specific number of channels changes according to the number of input channels.

[0089] After receiving the feature map of size 1 / 64, the KAN module of the fifth layer first tokenizes it, that is, reshapes the features to obtain a series of flattened two-dimensional paths (Pacth). The purpose of tokenization here is to replace sensitive data with insensitive data. Next, linear projection is implemented through a convolutional layer with a kernel size of 3×3. After the convolutional layer obtains the Token, it passes them into the KAN structure. In the KAN structure, each layer will have a series of nodes, and there will be a set of specific functions in the nodes to process the input data and output it to the next layer; the function is usually parameterized by B-spline functions or other smooth functions, so that the KAN structure can learn the nonlinear transformation that best suits the data at each degree, and each input variable is embedded through a set of independent one-dimensional nonlinear learnable functions φ. p,q (x i ), where p represents the index of the input variable and q represents the index of the output dimension. After the KAN structure processing in the fifth layer of the KAN module, it will enter the efficient deep convolution layer (DwConv). After the convolution operation, the output of the deep convolution layer is connected to the feature sequence before tokenization using residual connection and input into the layer normalization. After layer normalization, the size of the feature map will be reduced to 1 / 256 of the original size, and the width of the feature map is Height is The channel number is D1.

[0090] After receiving the feature map from the KAN module of the fifth layer, the KAN module of the sixth layer first tokenizes it, and then implements linear projection through a convolution layer with a kernel size of 3×3. After the convolution layer obtains the token, it passes it into the KAN structure. After passing through the KAN structure, it will then enter the deep convolution layer (DwConv), and after the convolution operation, it will perform a residual connection with the feature sequence before tokenization and input it into the layer normalization. After layer normalization, the size of the feature map is reduced to 1 / 1024 of the original image. At this time, the width of the feature map is Height is The number of channels is D2.

[0091] The U-Net convolution module includes two convolutional neural networks (Conv), batch normalization function, and rectified linear unit ReLU activation function. In the convolution stage of the U-Net convolution module and the KAN module, the number of channels of each module is C1 to C3 and D1 to D2 respectively. The U-Net convolution module can be expressed by the formula:

[0092]

[0093] in and F i are the intermediate and final features of the i-th U-Net convolutional module, respectively, and F i-1 is the upper layer output, and i takes values ​​of 1, 2, and 3 in the encoder. in , C h , C o They are input, hidden and output layers respectively. The present invention introduces the KAN module in the fifth and sixth layers, which can achieve performance comparable to or even better than that of the traditional convolutional network with a smaller number of parameters, with a small demand for memory, and at the same time has a learnable nonlinear activation function, and can significantly improve the computational efficiency, and can be flexibly adapted to different application scenarios. The KAN structure is based on the Kolmogorov-Arnold representation theorem, that is, a multi-variable continuous function can be expressed as a finite combination of a single-variable continuous function, which can be expressed by the formula:

[0094]

[0095] Among them, φ q,p (x p ) is a one-dimensional function, p is the index of the input dimension, q is the index of the output dimension, and n is the number of referenced parameters. The activation function within the network node in the KAN structure avoids errors in sensitive data input, thereby reducing the model's convolution and pooling operations on sensitive data, so that the KAN structure further improves the accuracy of the output image and allows the KAN structure to achieve complex multi-dimensional input to multi-dimensional output mapping at each layer. In the decoder, the fifth-layer KAN module accepts the output features of the sixth-layer KAN module in the encoder using a jump connection. After passing through the fifth-layer KAN module of the decoder, the size of the feature map is 1 / 1024 of the original image, and the width is Height is The number of channels is D2. Next, the KAN module of the fourth layer of the decoder samples the features of the KAN module of the fifth layer of the decoder, and then fuses them with the features of the corresponding fifth layer in the encoder (Encoder) transmitted using a jump connection. After passing through the KAN module of the fourth layer of the decoder, the size of the feature map is enlarged to 1 / 256 of the original image, with a width of Height is The number of channels is D1. Then it is output to the U-Net convolution module of the third layer of the decoder. The U-Net convolution module of the third layer of the decoder upsamples the features of the KAN module of the fourth layer of the decoder, and then fuses the output features of the U-Net convolution module of the fourth layer of the encoder (Encoder) transmitted using the jump connection. After passing through the U-Net convolution module of the third layer of the decoder, the size of the feature map is enlarged to 1 / 64 of the original image, with a width of Height is The number of channels is C3. Next, the U-Net convolution module of the second layer of the decoder upsamples the size of the U-Net convolution module of the third layer of the decoder, and then fuses the features of the U-Net convolution module of the third layer of the encoder (Encoder) transmitted using the jump connection. After passing through the U-Net convolution module of the second layer of the decoder, the size of the feature map is enlarged to 1 / 16 of the input image, with a width of Height is The number of channels is C2. The U-Net convolution module of the first layer of the decoder upsamples the features of the U-Net convolution module of the second layer of the decoder, and then fuses them with the features of the U-Net convolution module of the second layer of the encoder (Encoder) transmitted using a skip connection. After passing through the U-Net convolution module of the first layer of the decoder, the feature map at this time is enlarged to 1 / 4 of the image processed by the preprocessing module, and the width is Height is The number of channels is C1. In the present invention, the KAN module is introduced into the decoder, which not only increases the timeliness of deconvolution, but also can further improve the accuracy of deconvolution output features, and the KAN module can reduce the waste of U-Net convolution module resources.

[0096] Finally, the U-Net convolution module of the first layer of the decoder will upsample the features to the same size as the image processed by the preprocessing module and pass it to the bridge layer module with a bridge channel. At this time, the width of the feature map is W0, the height is H0, and the number of channels is C. Y The output features of the first layer U-Net convolution block (Cnovolution Block) of the decoder are upsampled and passed to the bridge layer module, which processes the input F b After performing a self-similar convolution operation and then a flattening operation, the process of the bridge layer module can be expressed as:

[0097]

[0098] Among them, H0 and W0 are the height and width of the input feature map of the bridge layer module, Flatten() is the flattening operation, and C brepresents the convolutional layer, Represents a self-similar convolution operation, that is, the same convolution layer C b Convolve its own output again, H out and W out is the height and width of the output feature map of the bridge layer; F f Output features for the bridge layer.

[0099] The size of the convolution kernel F of the self-similar convolution operation is 3, the stride D is 4, the padding P is 1, and the output resolution of the bridge layer module is calculated as:

[0100]

[0101] Through the bridge layer module, F b Flatten into a sequence and pass it to the Kolmogorov-Arnold feature embedding layer (KAN-Feature-Embedding) module. The Kolmogorov-Arnold feature embedding layer (KAN-Feature-Embedding) module maps the feature space of the upper layer to a low-dimensional feature space, fuses and flattens it, and then the feature sequence passes through the KAN module for further convolution operations to reduce its size to 1 / 16 of the original size, with a width of Height is The number of channels is C Y . This helps reduce the dimensionality of the features.

[0102] Adding the KAN module to the feature embedding layer can combine traditional word embedding with domain-related knowledge graphs. The basic semantics can be captured through the embedded parameters, and then additional information about each parameter can be obtained by connecting the KAN module, which can improve the model's understanding ability and the overall performance of the model. The features are then passed to the Kolmogorov-Arnold Transformer (KAN-Transformer) module, which includes a multi-head attention (MHA) layer, a KAN module, normalization (Layer Norm), and a residual connection. The KAN-Transformer module can be expressed as:

[0103]

[0104] in, is the result of adding the output of the multi-head attention layer of the KAN-Transformer module to the output of the Kolmogorov-Arnold feature embedding module; MHA represents the multi-head attention layer operation; F jrepresents the output of the Kolmogorov-Arnold feature embedding module, i.e., the input of the KAN-Transformer module; f(x) is the output of the KAN structure in the KAN module. After the KAN structure passes through the deep convolution layer and layer normalization series operations in the KAN module, the KAN module output is obtained as Y(x), and LN() represents normalization; F * is the output of the KAN-Transformer module. In the multi-head attention layer (MHA), the dependencies between the sequence and other sequences are calculated by using cross attention. In this step, the computational complexity is C 2 , where C is the number of input sequences. Sequence reduction techniques are used, that is, data compression methods are used to make it suitable for high-resolution input images, and the complexity becomes C 2 / Z, where Z = 2 is the reduction rate. The output feature map of the KAN-Transformer module is also 1 / 16 of the original size, with a width of Height is The number of channels is C Y .

[0105] It is then output to the Kolmogorov-Arnold feedforward neural network (KAN-Feed-Forward Net) module of the next layer. After the Kolmogorov-Arnold feedforward neural network (KAN-Feature-Embedding) receives the embedding sequence from the KAN-Transformer module, the reshaping and upgrading module extracts the feature sequence, reshapes the feature sequence, and refines it. After that, it will pass through the KAN module and perform further convolution operations on the feature sequence, and then enter the CBR module. The CBR module contains 3 convolution blocks (Cnovolution Block), batch normalization (BatchNormalization) and ReLU activation function. The convolution kernel size of the first two convolution blocks is 3×3, and the convolution kernel size of the third convolution block is 1×1. The CBR module is expressed in mathematical formula as follows:

[0106]

[0107] Among them, F r8 is the output of the KAN-Transformer module, is the output of the first convolutional block in the CBR module, is the output of the second convolutional block in the CBR module, C h1 and C h2 The CBR module accepts the feature sequence from the KAN module and directly generates a prediction map with a width of W.out , height H out , the number of channels is 1.

[0108] The feedforward network is a basic deep learning network, where information is transmitted only in a single direction from the input layer to the hidden layer and then to the output layer, without loop connections. Adding the KAN module to the feedforward network can reduce the parameters that need to be trained in the module while preserving the model's expressiveness, reducing the risk of overfitting. At the same time, the module is more likely to discover the underlying feature combination, making the feature learning more hierarchical and regular. The nonlinear activation function of the KAN structure in the KAN module can enhance the module's understanding of complex patterns and relationships.

[0109] Here we use the KAN module. In data fitting and PDE solving, a smaller KAN can achieve comparable or better accuracy than a larger MLP. KAN can process faster neural scaling laws than MLP. For interpretability, KAN can be intuitively visualized and can easily interact with human users.

[0110] The encoder of the UKAN-Transformer model is used to extract the feature representation of the input image. The activation function φ(x) of the KAN structure in the KAN module is the sum of the basis function b(x) and the spline function, which can be expressed as:

[0111] φ x =w(b(x)+spline(x)),

[0112] Among them, w is used to control the overall size of the activation function.

[0113] Set the base function to:

[0114]

[0115] The spline function is parameterized as a linear combination of B-splines:

[0116]

[0117] Among them, c i are trainable parameters.

[0118] The KAN module changes the fixed nonlinear activation function and linear parameter learning to the direct learning of the parameterized nonlinear activation function. Although the learning difficulty of a single spline function is more difficult than that of a linear function, the KAN module requires a smaller network size to achieve the same effect. Using the KAN module in the encoder can not only avoid errors in sensitive data input, but also further improve the accuracy of the output image, and can also achieve complex multi-dimensional input to multi-dimensional output mapping on each layer. Adding the KAN module to the Kolmogorov-Arnold transformer KAN-Transformer module not only applies self-attention, which enables the model to better capture spatially related features, but also emphasizes local attention, which means that it will retain the pixel-level connection of the U-Net convolution module to a certain extent, rather than just global information interaction, which helps to keep the model's attention to details, dynamically switch between local and global information, increase the model's adaptability to input changes, and help improve the model's understanding of complex scenes.

[0119] The process of measuring and judging the progress of cable tower construction:

[0120] (1) The shooting camera is adjusted to a horizontal distance of 150 meters from the tower, and it is fixed at the bottom of the tower. The shooting angle is adjusted to capture the entire tower construction process, that is, the size of the tower construction site and the height range of the construction object after the tower construction is completed. The UKAN-Transformer model is trained using the training set of Example 1 to segment the tower construction image to obtain a tower construction progress image with red marking the tower construction progress area and black marking other areas (areas other than the tower construction progress area). Then, the pixel counting method is used to calculate the number of pixels in the entire area marked with the tower construction progress area and the number of pixels in the image width at the bottom of the area, which are represented by N and M.

[0121] Place a ruler in the image with a length of L1. The pixel length of the ruler in the image is measured to be Q. The pixel scale K can be calculated using the following formula:

[0122] K=L1 / Q,

[0123] Among them, K represents the conversion ratio between pixels and actual distances, that is, the pixel scale. The square of the unit pixel size is multiplied by N to obtain the pixel area S0 of the entire cable tower construction progress image, and the length of the unit pixel is multiplied by M to obtain the pixel width W1 of the cable tower construction progress image.

[0124] (2) Then, the pixel area S0 of the cable tower construction progress image is divided by the pixel width W1 of the cable tower construction progress image to obtain the pixel height H1 of the cable tower construction progress image, which can be expressed as:

[0125] H1=S0÷W1.

[0126] (3) Then convert H1 into actual distance to obtain the actual height H2 of the tower construction progress. It can be expressed as:

[0127] H2=H1×K

[0128] (4) By comparing the actual height H2 of the cable tower construction and the set target height H3 of the cable tower construction, the percentage of H2 to H3 is defined as the progress R0 of the cable tower construction. It can be expressed as:

[0129]

[0130] By comparing the progress of R0 with the expected target progress R, if R0 is less than R, the expected progress of the cable tower construction has not been completed. If R0 is greater than R, the expected progress of the cable tower construction has been exceeded and will be completed ahead of schedule.

[0131] During the training process, the k-fold cross-validation method can make full use of the limited data set, more accurately evaluate model performance, select the best model performance and parameters, and provide estimates of model variance and bias.

[0132] Table 1 below is a comparison of the effects of using the dataset of the present invention to train different types of models. The IoU and F1 indicators of the model of the present invention have been significantly improved. Among them, IoU is the intersection-over-union ratio, that is, the ratio of the area of ​​the intersection of the predicted image and the real image to the area of ​​the merged area. The larger the value, the more accurate the predicted image. F1 is used to comprehensively evaluate the precision and recall rate. The larger the value, the better the model performance.

[0133] Table 1

[0134]

[0135] The UKAN-Transformer model of the present invention aims to solve the problems of low monitoring accuracy, high computational complexity, insufficient intuitive feature fusion, strong data dependence, poor interpretability and high hardware requirements in the current UNet-Transformer model during large-scale cable tower construction. By adding a self-learning activation function, a more accurate output can be obtained, which can improve the efficiency of the model in calculating the input image, improve the adaptability of the model, accurately fit, improve the generalization ability and interpretability, and improve the accuracy of cable tower construction progress monitoring.

[0136] The present invention adopts the UKAN-Transformer model to greatly improve the accuracy of recognition and judgment, so that the system can monitor the construction progress of the cable tower in real time. The KAN module mechanism based on the learnable activation function is adopted to ensure the accuracy and timeliness of feature extraction, making the model more reliable and effective. In addition, the spatial relationship and semantic information between features can be effectively learned through the learnable activation function, thereby improving the accuracy of semantic segmentation.

[0137] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for monitoring the progress of cable tower construction, characterized in that: The method comprises the following steps: Step 1: Image data collection of cable tower construction height: Use the image capturing machine on the tower crane during the construction of the cable tower to take pictures. Fix the camera on the top of the cable tower. The image capturing machine can capture the construction object of the cable tower. During the data collection process, the horizontal distance between the image capturing machine and the side of the construction object remains unchanged. During the construction process, the image capturing machine can always capture pictures of the construction object at all heights. Collect all pictures during the entire construction process at regular intervals to build a data set. Step 2: Build a UKAN-Transformer model based on a learnable activation function: The UKAN-Transformer model adopts the framework of the UNet-Transformer model, including the encoder, decoder, and Kolmogorov-Arnold transformer model structure; The encoder includes a preprocessing module, three 3×3 U-Net convolution modules and two KAN modules which are sequentially connected from top to bottom; the decoder includes three 3×3 U-Net convolution modules and two KAN modules which are sequentially connected from top to bottom; each U-Net convolution module and KAN module of the encoder is connected to the corresponding U-Net convolution module and KAN module of the decoder using a skip connection at the same time; The output of the U-Net convolution module of the top layer of the decoder is used as the output feature of the decoder, and the output feature of the decoder is input into the Kolmogorov-Arnold converter model structure; the Kolmogorov-Arnold converter model structure includes a bridge layer module, a Kolmogorov-Arnold feature embedding layer KAN-Feature Embedding module, a Kolmogorov-Arnold converter KAN-Transformer module and a Kolmogorov-Arnold feedforward network KAN-Feed-Forward Net module connected in sequence from top to bottom; KAN modules are added to the Kolmogorov-Arnold feature embedding layer KAN-Feature Embedding module, the Kolmogorov-Arnold converter KAN-Transformer module and the Kolmogorov-Arnold feedforward network KAN-Feed-ForwardNet module; So far, the UKAN-Transformer model based on the learnable activation function is obtained; The KAN module includes a Tokenization layer, a 3×3 convolution layer, a KAN structure, a deep convolution layer and layer normalization. After the input feature sequence is processed by Tokenization, a series of flattened two-dimensional patches are obtained. The output structure of the 3×3 convolution layer, the KAN structure and the deep convolution layer is residually connected with the feature sequence before Tokenization. Finally, the output of the KAN module is obtained after layer normalization. The preprocessing module includes a 5×5 standardized convolution, a group normalization, a ReLU activation function and a pooling layer connected in sequence; Step 3: Use the data set constructed in step 1 to train the UKAN-Transformer model based on the learnable activation function, and use the trained UKAN-Transformer model based on the learnable activation function to segment the cable tower construction image to obtain the cable tower construction segmentation image that marks the cable tower construction progress area; Step 4: Count the pixel area of ​​the tower construction progress area in the tower construction segmentation image and the pixel width at the bottom of the area, and obtain the actual height of the tower construction based on the pixel area and pixel width; and judge the progress of the tower construction by comparing the actual height of the tower construction with the expected height of the tower construction.

2. The method according to claim 1, characterized in that In the Kolmogorov-Arnold feature embedding layer module, the feature sequence is fused and flattened by 3×3 convolution and then enters the KAN module, and then enters the Kolmogorov-Arnold transformer KAN-Transformer module; In the Kolmogorov-Arnold transformer KAN-Transformer module, the feature sequence is first standardized and then enters the multi-head attention layer. The output of the multi-head attention layer is spliced ​​with the standardized input and then enters a KAN module. The output of the KAN module is spliced ​​with the input of the KAN module after the second standardization process to obtain the output of the Kolmogorov-Arnold transformer KAN-Transformer module; Afterwards, in the Kolmogorov-Arnold feed-forward network (KAN-Feed-Forward Net) module, the feature sequence is reshaped and upgraded, and then processed by the KAN module and the CBR module to obtain the final prediction graph.

3. The method according to claim 1, characterized in that During the UKAN-Transformer model training process, The formula for the change of the learning rate of the i-th generation is: N1=a*T, Among them, T represents the number of training generations, N1 represents the number of generations with increased learning rate, a represents the proportion of generations with increased learning rate, and b represents the maximum learning rate V max With the minimum learning rate V min The ratio of V i is the learning rate of the i-th round, and β represents the coefficient of increase or decrease of the learning rate; The loss function is the weighted cross entropy loss function E, and the formula is: Where Ω is the true label of each pixel, is the activation value of the class output given by the label corresponding to point x, and ω is the weight added to each pixel during training.

4. The method according to claim 3, characterized in that T=100~150,V max The value range of is between [0.005-0.01], and the value of b is greater than 1.

5. The method according to claim 1, characterized in that The process of the bridge layer module is expressed as: Among them, H0 is the height of the feature map, Flatten() is the flattening operation, Represents a self-similar convolution operation, that is, the same convolution layer C b Convolve its own output again; F f is the output feature of the bridge layer module; F b is the input feature of the bridge layer module, that is, the output feature of the decoder after upsampling.

6. The method according to claim 5, characterized in that The size of the convolution kernel F of the self-similar convolution operation is 3, the stride D is 4, the padding P is 1, and the output resolution of the bridge layer module is calculated as: Among them, H0 is the height of the input feature map of the bridge layer module, H out and W out is the height and width of the output feature map of the bridge layer module.

7. A cable tower construction progress monitoring system, characterized in that: The system comprises: An image acquisition module is used to obtain images of the height direction of the construction object during the cable tower construction process; The data processing module is used to mark the tower construction progress area on the pictures collected by the image acquisition module; The cable tower construction progress area segmentation model is used to obtain a segmented image of the cable tower construction progress area according to the input image; A statistical calculation module is used to count the pixel area of ​​the cable tower construction progress area in the segmented image and the pixel width of the bottom of the area, and further calculate and convert to obtain the actual height of the cable tower construction progress area; The progress monitoring module is used to compare the actual height monitored by the statistical calculation module with the preset expected height of the cable tower construction to determine the real-time progress R0 of the cable tower construction; compare the real-time progress with the expected target progress R. If R0 is less than R, the expected progress of the cable tower construction has not been completed and acceleration measures should be taken; if R0 is greater than R, the expected progress of the cable tower construction has been exceeded and the construction is expected to be completed ahead of schedule.

8. The system according to claim 7, characterized in that The tower construction progress area segmentation model is implemented using the UKAN-Transformer model based on a learnable activation function as described in claim 1.

Images