An intelligent design method for beam bridge spans based on generative adversarial neural network

Abstract

The present invention proposes a method for intelligent design of beam bridge spans based on a generative adversarial neural network, the method comprising: (1) an automatic modeling and drawing device based on parametric modeling; (2) a bridge type selection network: the network receives a geological sketch and a condition vector composed of design information such as a predetermined span, and outputs a number of beam bridge type selection schemes suitable for the current geology; (3) a generative adversarial network composed of a generator and a discriminator: the generator includes a feature extraction network of the geological sketch, a construction network of the condition vector, a feature fusion network of the bridge and culvert design conditions, and a generative drawing network; the generator receives the geological sketch features and the condition vector composed of the designed bridge type, etc. as input, fuses the condition vector and the geological sketch features, and generates a span design drawing based on the fused image. The discriminator determines whether the span design drawing of the beam bridge is derived from an automatic design or a real design, and realizes an automatic span design adapted to local conditions in the adversarial training of the generator and the discriminator.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge and culvert design and machine learning application, and in particular relates to an intelligent design method for beam bridge spans based on a generative adversarial neural network. Background Art

[0002] In the preliminary design stage, bridge and culvert designers need to calculate the size and material limitations of various components according to the specifications on the one hand; on the other hand, they need to select appropriate component layouts based on the specific local geological and hydrological conditions according to their own accumulated structural design experience. In addition, the current design mainly relies on purely manual experience for comprehensive consideration, and the degree of automation and intelligence is insufficient. The design quality control is also greatly affected by the level of the designer himself. In terms of standard calculation, a number of mature commercial expert systems have appeared on the market for designers to quickly design calculation processes and structural modeling. However, for the part involving empirical design, there are almost no mature intelligent design methods that can integrate the existing design experience of human experts, so as to achieve rapid design based on local conditions.

[0003] Since its inception, machine learning methods have achieved great success in fields such as machine vision by virtue of their automatic learning characteristics, strong versatility, and good expansibility. In terms of empirical design integration in bridge and culvert design, machine learning methods have also been a strong candidate for competition and have received long-term attention from researchers. However, the use of machine learning methods in bridge and culvert design has long faced the problem of small and lack of training set samples. On the other hand, how to connect with mature expert systems, especially to distinguish which parts of the design are suitable for expert systems based on specification design and experience-based machine learning methods, and to consider how to organically combine the two, has always been an important issue that has led to the less application of machine learning methods in bridge and culvert design.

[0004] Therefore, the current application of machine learning methods in bridge and culvert design is rare, and is mainly limited to data analysis and causal inference of existing design factors, lacking practicality. The use of existing design drawings also relies on manual input, especially in the selection and span design stage of beam bridges. Although the technology and design process are relatively mature, it still relies on designers to manually estimate and draw design sketches, and make subjective comparisons and determinations. This brings advantages to the application of machine learning methods in beam bridges. There is an urgent need for a machine learning method with wide applicability, high automation, and high accuracy to assist in design. Summary of the invention

[0005] Purpose of the invention: The present invention aims to propose an intelligent design method for beam bridge spans based on generative adversarial neural networks. By proposing a set of standard procedures for automatically drawing sketches of design data, the design learning automation can be conveniently realized. It can smoothly connect to the existing commercial design expert system. By inputting a small amount of geological and hydrological data of the design bridge location, the bridge type and corresponding span design scheme that meet the design requirements and are adapted to local conditions can be quickly designed according to the real design cases entered into the database.

[0006] Technical solution: The present invention proposes a method for intelligent design of spans of beam bridges based on a generative adversarial neural network, which includes the following steps:

[0007] (1) Construct a training sample set and an automatic modeling drawer. Use the drawer to extract the terrain data of the beam bridge instance and generate a standardized geological sketch set T c ; Use the drawing tool to extract the design data of the beam bridge example and generate the span design atlas T s , extract the expected span, expected daily traffic volume and the number of available bridge types from the bridge instance to construct a preliminary condition vector set T v1 , extract the bridge type data selected by the bridge instance from the bridge instance to construct the bridge type selection set T of the training sample x , using the generated geological sketch set T c , Span Design Atlas T s , the initial condition vector set T v1 and bridge type selection set T x As training samples, the initial condition vector set T v1 and bridge type selection set T x The concatenation in the 0th dimension is called the complete conditional vector set T v2 ;

[0008] (2) Construct and train a bridge type selection network and transform the geological sketch set T c With the initial condition vector set T v1 Input the bridge type selection network, and obtain the network's bridge type selection set T through multimodal input processing, convolutional feature extraction and fully connected layer classification. xt , with T xt With T x The mean square error is the error function, and the network is trained by back propagation of the SGD algorithm;

[0009] (3) Constructing a span generator network, which is composed of a geological sketch set feature extraction network, a conditional vector construction network, a bridge and culvert design condition feature fusion network, and a generation and drawing network;

[0010] (4) Constructing a split-span discriminator network, which consists of a local discriminator and a global discriminator;

[0011] (5) Generate adversarial training and transform the geological sketch set T c Input the geological sketch set feature extraction network of the span generator and convert the complete condition vector set T v3 Input the conditional vector of the span generator to build a network, and train the generator and the discriminator alternately in several cycles, so that the difference between the image generated by the generator and the real span design is within a preset range, until the judgment accuracy of the discriminator on the generated scheme and the real scheme is stabilized at Nash equilibrium;

[0012] (6) Use the automatic modeling drawing device that has passed the performance evaluation, the bridge type selection network and the generator network, input the design data of the new design case, and construct a new preliminary condition vector set T v1x , using the automatic modeling sketcher to draw a new geological sketch set T cx , use the bridge type selection network to obtain the new bridge type selection set T for this design case xx , construct a new complete conditional vector set T v2x , and then use the generator network to complete the new cross-design atlas T sx Automatic design.

[0013] Furthermore, in step (1), constructing a training sample set comprises the following steps:

[0014] Step 1.1: Read the design geological information from the input data or drawings and draw a geological sketch set T c , Geological Sketches T c It is a longitudinal section of geological conditions, including the design water level, geographic line, bridge deck elevation, minimum vertical clearance above the bridge deck, beam bridge superstructure diagram, unfavorable geological section, navigation clearance area, cross-line lane clearance area, soil conditions, earthquake conditions, and design conditions of material use. Among them, the unfavorable geological section is defined as a geological section that is clearly determined to be unsuitable for piers in the design task. When input, it is expressed as a shape of [N b ,2], a two-dimensional matrix, N b The first column of the matrix indicates the starting point of each bad geological section, and the second column of the matrix indicates the end point of each bad geological section. The soil conditions are marked with different colors under the geographic line, and the material usage is marked with color blocks of each bridge and culvert structure. Different geological sketch information is added according to different bridge and culvert conditions:

[0015] The navigable bridge will read the input bridge deck elevation and expected span, and scale the larger of the expected span and the difference between the bridge deck elevation and the maximum water depth elevation to the pixel position predetermined by the design drawing, and scale the other to the pixel position while keeping the horizontal and vertical ratio unchanged. The upper structure building height is preliminarily estimated based on the optional minimum vertical clearance of the building height above the bridge deck and the expected height-span ratio, and the upper structure longitudinal section is drawn with different colors to mark different materials; if the database or input does not give the channel grade, the channel grade is calculated based on the width of the water body, and several polygonal navigation clearance blocks are drawn symmetrically in the center of the river, and then the water bodies at different water levels are colored according to the local flood control, drift ice, and bridge pier water defense water levels of the river;

[0016] The overpass will read the input bridge deck elevation and expected span, scale the larger of the expected span and the bridge deck elevation to the road ground elevation to the pixel position predetermined by the design drawing, and scale the other to the pixel position while keeping the horizontal and vertical ratio unchanged. Preliminary estimate the upper structure building height based on the optional minimum vertical clearance of the building height above the bridge deck and the expected height-span ratio, draw the upper structure longitudinal section with different colors to mark different materials, and draw the clearance area of ​​the roadway at the position of the roadway in imitation of the navigation bridge. For poor geology, if the poor geology comes from soil factors, the section below the geographic line will be marked in red; if it comes from obstacle factors, the detailed start and end points and clearance requirements will be marked under the water body and geographic line, and the poor geology will also be marked in red;

[0017] Step 1.2: Read the design geological information from the input beam bridge design data and draw the span design atlas T s , Span Design Atlas T s It is a longitudinal section of geological conditions, including the design water levels, geographic lines, bridge deck elevations, schematic diagrams of beam bridge superstructures, spans of beam bridges, positions and sizes of piers of beam bridge substructures, unfavorable geological sections, navigation clearance areas, clearance areas for cross-line carriageways, soil conditions, earthquake conditions, and material usage. The definition of unfavorable geological sections is the same as in step 1.1. Soil conditions are marked with different colors under the geographic lines. Material usage is marked with color blocks of various structures of bridges and culverts. Any selection of piers and material selections need to be marked with different colors. c Different, split across design atlas T s The detailed data of the bridge pier is added. When drawing, only the starting point of the bridge pier is accurately determined. The longitudinal section width of the bridge pier represents the nominal radius d of the plane section of the bridge pier. m :

[0018]

[0019] V: average cross-sectional area of ​​the pier

[0020] π: the circumference constant

[0021] Due to the different characteristics of navigation bridges and cross-line bridges, different additional annotations are required for the span design drawings: for navigation bridges, the colors of the piers are marked differently according to the water levels of the navigation bridges; for cross-line bridges, in addition to the annotation format of ordinary beam bridges, the span type, minimum left lateral clearance, minimum right lateral clearance and minimum vertical underground clearance elevation are additionally marked to calculate the specific restrictions of the cross-line on the pier design. Among them, the span type is divided into highway underground structure, railway underground structure and non-underground structure, and different clearance regulations and colors are used to identify different structure types. The minimum left lateral clearance of the passage refers to the distance from the right edge of the highway or the center line of the right track of the railway train to the nearest bridge structure or slope foot, and the minimum left lateral clearance refers to the distance from the left edge of the highway or the center line of the left track of the railway train to the nearest bridge structure or slope foot;

[0022] Step 1.3: Extract the estimated span, estimated daily traffic volume, and available bridge types directly from the bridge instance to construct the preliminary condition vector set T v1 , the initial condition vector set T v1 is a shape of [N s ,N v1 ], N s is the number of training samples, N v1 is the length of the initial condition vector set:

[0023] N v1 =2+n

[0024] n is the total number of selectable bridge types;

[0025] Preliminary condition vector set T v1 In [N si ,1] indicates the Nth si The estimated span of training samples, expressed as a single-precision number; [N si ,2] indicates the Nth si The estimated daily traffic volume of training samples is expressed as a single-precision value; [N si ,3] to [N si ,N v1 ] are all single-precision numbers 0;

[0026] Step 1.4: Extract the bridge type data selected by the bridge instance from the bridge instance to construct the bridge type selection set T of the training sample x , bridge type selection set T x is a shape of [N s ,n] two-dimensional tensor, N s is the number of training samples, n is the total number of selectable bridge types; T x Should be identified as one-hot encoding format:

[0027] T x =∑[Tx1 ,T x2 ,……T xn ]

[0028] T xi is the probability of using the i-th bridge type, which is a float value between 0 and 1. When marking, the unused bridge type is 0 and the used bridge type is 1;

[0029] Step 1.5: Set the initial condition vector set T v1 and bridge type selection set T x Concatenate in the 0th dimension to get the complete conditional vector set T v2 .

[0030] Furthermore, in step (2), the construction of the bridge selection network and training includes the following steps:

[0031] Step 2.1: Accept the geological sketch set T c And the initial condition vector set T v1 As input, T c With T v1 Each of the four-channel fusion bridge-type selection feature sets T is scaled to a unified shape through an embedding layer composed of multiple convolutional layers rq ;

[0032] Step 2.2: Use 3 fully connected layers as the middle layer to transform T rq Gradually map to N q dimensional multi-classification bridge selection output layer, N q is the number of available bridge types. The bridge type selection output layer uses the softmax layer as the activation layer to sort the priorities of each bridge type under the current geological conditions, and obtains the bridge type selection recommendation set T of the bridge selection network. xt :

[0033] T xt =∑[T xt1 ,T xt2 ,……T xtn ]

[0034] T xti The usage probability of the i-th bridge type obtained by the network is a float value between 0 and 1;

[0035] Step 2.3, define the loss function:

[0036]

[0037] T xi : The usage probability of the marked i-th bridge type, which is a float value between 0 and 1. When marking, the unused bridge type is 0 and the used bridge type is 1;

[0038] Using loss type Update the weights of the training samples and perform iterative training of the bridge selection network for N training times j .

[0039] Furthermore, in step (3), the construction of the generator includes the following parts and steps:

[0040] Step 3.1: The generator includes a feature extraction network of geological sketches, which extracts the geological sketches of the same batch into a set T c Input the network, extract and expand features in a sequential manner including but not limited to batch normalization, deconvolution layer, random dropout, and linear layer, analyze the characteristics of the geological sketch, and the output result is called the geological sketch feature vector X c , let the feature vector dimension be N c , the input sketch is an RGB three-channel color image, X c The size should be [3,N c ];

[0041] Step 3.2: The generator includes a network for constructing conditional vectors, which converts the complete conditional vector set T v2 Feature extraction is performed in a sequential manner including but not limited to batch normalization, deconvolution layer, random dropout, linear layer, and the conditional vector is expanded to the same feature vector dimension N as the geological sketch through multiple deconvolution layers. c , then the conditional vector eigenvector X t The scale is [1,N c ];

[0042] Step 3.3: The generator includes a feature fusion network for the bridge and culvert design conditions, and the geological sketch feature vector X c and the conditional feature vector X shown in step 3.2 t Input network, use multi-layer convolution for downsampling, sequentially including but not limited to batch normalization, random dropout, linear layer, use LeakyRelu activation function to reduce the influence of outliers, output fusion feature T of bridge and culvert design status r , let the dimension of fused feature vector be N r ,T r For [4,N r ];

[0043] Step 3.4, the generator includes generating a drawing network, and integrating the fusion features T of the bridge and culvert design conditions in step 3.3 r, perform multi-layer deconvolution to expand the size for upsampling, and automatically select the size of the deconvolution kernel that can be divided by the generated design image. Each layer is sequentially performed including but not limited to batch normalization, random dropout process, and several layers of positive convolution are added at the end to compress and reduce the checkerboard effect. The result is re-deformed and output as a design span map of a preset size.

[0044] Furthermore, in step (4), the construction of the span discriminator includes the following steps:

[0045] Step 4.1: Use the global discriminator of global features to input a span design diagram, use multi-layer convolutional layers for upsampling, followed by multi-layer linear fully connected layers, and the output layer is a judgment layer with only one parameter to determine whether it is generated by the system;

[0046] Step 4.2: Use the local discriminator of local geological features, input a span design drawing, and after upsampling using multiple convolutional layers, output a matrix in which each value represents the truth or falsehood of each area in the input design span draft. The average value of each area is taken as the final result of the local discriminator.

[0047] Furthermore, in step (5), performing generative adversarial training includes the following steps:

[0048] Step 5.1: Set the geological sketch set T c Input the geological sketch feature extraction network of the span generator, and estimate the complete condition vector set T of span and bridge type. v2 Input the conditional vector of the span generator to build the network. According to the generator structure mentioned above, the fusion feature set T of the bridge and culvert design status will be obtained. r , the fusion feature set T of bridge and culvert design conditions r Input the generated drawing network to obtain the generated span design atlas T g ;

[0049] Step 5.2: Generate the span design atlas T g Input the aforementioned span discriminator to obtain the span rationality vector X of the generated design g ;

[0050]

[0051] N s is the number of training samples;

[0052] X gi The rationality judged by the discriminator for the i-th generated design is a float value between 0 and 1;

[0053] The loss function of the generator is defined as:

[0054]

[0055] Through loss g Perform back-propagation optimization on the generator weights;

[0056] Step 5.3: Split the actual design atlas into s Input the discriminator to obtain the reasonableness vector X of the real design d :

[0057]

[0058] N s : Number of training samples

[0059] X di : The rationality judged by the discriminator of the i-th real design is a float value between 0 and 1;

[0060] The loss function of the discriminator for the true design is defined as:

[0061]

[0062] Through loss d_z Back-propagation optimization of the discriminator weights;

[0063] Step 5.4, define the loss function of the discriminator for generating the design as:

[0064]

[0065] N s : Number of training samples

[0066] X gi : The rationality judged by the discriminator of the i-th generated design is a float value between 0 and 1;

[0067] Through loss d_s Back-propagation optimization of the discriminator weights;

[0068] Step 5.5, train the generator and the discriminator alternately in several cycles, and repeat steps 5.2 to 5.4 so that the error between the image generated by the generator and the real span design drawing is within the preset range, and the accuracy of the discriminator in judging the generated case and the real design span drawing is gradually improved until the accuracy of the discriminator in judging the generated solution and the real solution is stabilized at the Nash equilibrium.

[0069] Furthermore, in step (6), the method comprises the following steps:

[0070] Step 6.1: Input the design data of the new design case and construct a new preliminary condition vector set Tv1x , using the automatic modeling sketcher to draw a new geological sketch set T cx ,

[0071] Step 6.2: Use the bridge type selection network to obtain the new bridge type selection set T for this design case xx , using the new set of preliminary condition vectors T v1x With the new bridge type selection set T xx Concatenate in the 0th dimension to get a new complete conditional vector set T v2s , using the generator network to complete the new cross-design atlas T sx Automatic design.

[0072] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0073] (1) The proposed method for intelligent design of beam bridges based on generative adversarial neural networks requires limited geological and hydrological data of the designed bridge locations as input data. It relies on a deconvolutional neural network to achieve upsampling and automatically analyze and extract high-dimensional features, thus ensuring applicability and simplicity.

[0074] (2) The method for intelligent design of beam bridges based on generative adversarial neural networks proposed in the present invention realizes the correspondence between the physical relationship between the data of the span design and the functional relationship between the high-dimensional parameters in the machine learning algorithm, so that the mutual influence between the design parameters in the beam bridge design can be mapped to the specific parameters in the machine learning network, thereby enhancing the physical and mechanical significance of the data-driven learning process.

[0075] (3) The method for intelligent design of spans of beam bridges based on generative adversarial neural networks proposed in the present invention uses a generative adversarial neural network discriminator that is a composite discriminator that is a secondary superposition of a local discriminator and a global discriminator. It takes into account the two tasks of avoiding local adverse geology and ensuring reasonable span distribution of waterways and the overall span distribution.

[0076] (4) The invention proposes an intelligent design method for beam bridges with multiple spans based on a generative adversarial neural network. The generative adversarial neural network used in the invention adopts an independent feature fusion network design. Since the design conditions are usually more complicated, the design scene data that can be collected may have drawings (pictures), data, texts and other formats with different contents and dimensions. The independent feature fusion network proposed in the invention can not only meet the automatic fusion and feature extraction of the aforementioned conditional vectors and geological sketches, but also realize the automatic fusion and feature extraction of design data in other formats through simple expansion.

[0077] (5) In the method for intelligent design of beam bridges with multiple spans based on generative adversarial neural networks proposed in the present invention, the material construction in the training process and the data entry in the use process all use the same automatic geological sketching program, so that the input of new design cases can be completed in a manner similar to the input of design data in the use process, which has good scalability and is simple to enter new design experience.

[0078] (6) The invention proposes an intelligent design method for beam bridges with multiple spans based on a generative adversarial neural network. The design of the generator of the generative adversarial neural network used in the invention can be expanded to consider more design criteria such as the moment balance between spans, thus leaving theoretical space for further professional and personalized development. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 A schematic diagram of a process of an intelligent design method for beam bridge spans based on a generative adversarial neural network proposed by the present invention;

[0080] Figure 2 This is an example diagram of a geological sketch of a cross-line bridge automatically drawn according to the bridge position data of the beam bridge design according to the present invention;

[0081] Figure 3 An example diagram of the span design drawing of the overpass drawn for the present invention;

[0082] Figure 4 This is an example diagram of a geological sketch of a navigation bridge automatically drawn according to the bridge position data of the beam bridge design according to the present invention;

[0083] Figure 5 This is an example diagram of the span design drawing of the navigation bridge drawn for the present invention. DETAILED DESCRIPTION

[0084] The following will introduce practical use cases of the present invention in conjunction with the accompanying drawings and embodiments.

[0085] The present invention proposes a method for intelligent design of beam bridge spans based on generative adversarial neural networks. The flow chart of the method is shown in the attached figure. Figure 1 The method specifically comprises the following steps:

[0086] (1) Construct a training sample set and an automatic modeling drawer. Use the drawer to extract the terrain data of the beam bridge instance and generate a standardized geological sketch set T c ; Use the drawing tool to extract the design data of the beam bridge example and generate the span design atlas T s , extract the expected span, expected daily traffic volume and the number of available bridge types from the bridge instance to construct a preliminary condition vector set T v1 , extract the bridge type data selected by the bridge instance from the bridge instance to construct the bridge type selection set T of the training sample x , using the generated geological sketch set T c, Span Design Atlas T s , the initial condition vector set T v1 and bridge type selection set T x As training samples, the initial condition vector set T v1 and bridge type selection set T x The concatenation in the 0th dimension is called the complete conditional vector set T v2 ;

[0087] (2) Construct and train a bridge type selection network and transform the geological sketch set T c With the initial condition vector set T v1 Input the bridge type selection network, and obtain the network's bridge type selection set T through multimodal input processing, convolutional feature extraction and fully connected layer classification. xt , with T xt With T x The mean square error is the error function, and the network is trained by back propagation of the SGD algorithm;

[0088] (3) Constructing a span generator network, which is composed of a geological sketch set feature extraction network, a conditional vector construction network, a bridge and culvert design condition feature fusion network, and a generation and drawing network;

[0089] (4) Constructing a split-span discriminator network, which consists of a local discriminator and a global discriminator;

[0090] (5) Generate adversarial training and transform the geological sketch set T c Input the geological sketch set feature extraction network of the span generator and convert the complete condition vector set T v2 Input the conditional vector of the span generator to build a network, and train the generator and the discriminator alternately in several cycles, so that the difference between the image generated by the generator and the real span design is within a preset range, until the judgment accuracy of the discriminator on the generated scheme and the real scheme is stabilized at Nash equilibrium;

[0091] (6) Use the automatic modeling drawing device that has passed the performance evaluation, the bridge type selection network and the generator network, input the design data of the new design case, and construct a new preliminary condition vector set T v1x , using the automatic modeling sketcher to draw a new geological sketch set T cx , use the bridge type selection network to obtain the new bridge type selection set T for this design case xx , construct a new complete conditional vector set T v2x , and then use the generator network to complete the new cross-design atlas T sx Automatic design.

[0092] Furthermore, in step (1), constructing a training sample set comprises the following steps:

[0093] Step 1.1: Read the design geological information from the input data or drawings and draw a geological sketch set T c , Geological Sketches T c It is a longitudinal section of geological conditions, including the design water level, geographic line, bridge deck elevation, minimum vertical clearance above the bridge deck, beam bridge superstructure diagram, unfavorable geological section, navigation clearance area, cross-line lane clearance area, soil conditions, earthquake conditions, and design conditions of material use. Among them, the unfavorable geological section is defined as a geological section that is clearly determined to be unsuitable for piers in the design task. When input, it is expressed as a shape of [N b ,2], a two-dimensional matrix, N b The first column of the matrix indicates the starting point of each bad geological section, and the second column of the matrix indicates the end point of each bad geological section. The soil conditions are marked with different colors under the geographic line, and the material usage is marked with color blocks of each bridge and culvert structure. Different geological sketch information is added according to different bridge and culvert conditions:

[0094] The drawing of the navigation bridge is as attached. Figure 2 As shown, the input bridge deck elevation and expected span are read, and the larger of the expected span and the difference between the bridge deck elevation and the maximum water depth elevation is scaled to the pixel position predetermined by the design drawing, and the other is scaled to the pixel position while keeping the horizontal and vertical ratio unchanged. The upper structure building height is preliminarily estimated according to the optional minimum vertical clearance of the building height above the bridge deck and the expected height-span ratio, and the upper structure 2 design bridge body is drawn with different colors to mark different materials; if the database or input does not give the channel grade, the channel grade is calculated according to the width of the water body, and several polygonal navigation clearance blocks are drawn symmetrically in the center of the river to represent 1 design channel, and then the water body at different water levels is colored according to the local flood control, drift ice, and bridge pier backwater defense water level of the river, as shown in the attached figure. Figure 2 The 3 minimum navigation clearances and 6 water body lines are included; and 4 geographical lines are drawn according to actual conditions.

[0095] The drawing of the overpass is shown in the attached Figure 4 As shown. The bridge deck elevation and the expected span are read in, and the larger of the expected span and the bridge deck elevation to the ground elevation of the road is scaled to the pixel position predetermined by the design drawing. The other is scaled to the pixel position while keeping the horizontal and vertical ratio unchanged. The upper structure building height is preliminarily estimated according to the optional minimum vertical clearance of the building height above the bridge deck and the expected height-span ratio. The 1 designed bridge body of the upper structure longitudinal section is drawn with different colors and different materials. The 3 lane clearance area is drawn at the lane position in imitation of the navigation bridge. For poor geology, if the poor geology comes from soil factors, the section below the geographic line is marked in red; if it comes from obstacle factors, the detailed start and end and clearance requirements are marked under the water body and geographic line, and the poor geology is also marked in red; for possible slope protection, the 4 span line slope needs to be marked.

[0096] Step 1.2: Read the design geological information from the input beam bridge design data and draw the span design atlas T s , Span Design Atlas T s It is a longitudinal section of geological conditions, including the design water levels, geographic lines, bridge deck elevations, schematic diagrams of beam bridge superstructures, spans of beam bridges, positions and sizes of piers of beam bridge substructures, unfavorable geological sections, navigation clearance areas, clearance areas for cross-line carriageways, soil conditions, earthquake conditions, and material usage. The definition of unfavorable geological sections is the same as in step 1.1. Soil conditions are marked with different colors under the geographic lines. Material usage is marked with color blocks of various structures of bridges and culverts. Any selection of piers and material selections need to be marked with different colors. c Different, split across design atlas T s The detailed data of the bridge pier is added. When drawing, only the starting point of the bridge pier is accurately determined. The longitudinal section width of the bridge pier represents the nominal radius d of the plane section of the bridge pier. m :

[0097]

[0098] V: average cross-sectional area of ​​the pier

[0099] π: the circumference constant

[0100] Due to the different characteristics of navigation bridges and overpasses, the span design drawings require different additional annotations: The annotations for navigation bridges are as follows: Figure 3 As shown in the figure, the colors of the designed bridge piers are marked according to the water levels of the navigation bridge; the markings for the overpass bridge are as shown in the attached figure. Figure 4 As shown in the figure, in addition to designing the bridge piers according to the annotation 5 of ordinary beam bridges, the span type is additionally marked, and the specific restrictions on the span pier design are calculated by 6 minimum lateral clearance and 2 minimum vertical underground clearance elevation. Among them, the span type is divided into highway underground structure, railway underground structure and non-underground structure. Different clearance regulations and colors are used to identify different structure types. The minimum lateral clearance of the passage refers to the distance from the edge of the highway or the center line of the railway train track to the nearest bridge structure or slope foot.

[0101] Step 1.3: Extract the estimated span, estimated daily traffic volume, and available bridge types directly from the bridge instance to construct the preliminary condition vector set T v1 , the initial condition vector set T v1 is a shape of [N s ,N v1 ], N s is the number of training samples, N v1 is the length of the initial condition vector set:

[0102] N v1 =2+n

[0103] n is the total number of selectable bridge types;

[0104] Preliminary condition vector set T v1 In [N si ,1] indicates the Nth si The estimated span of training samples, expressed as a single-precision number; [N si ,2] indicates the Nth si The estimated daily traffic volume of training samples is expressed as a single-precision value; [N si ,3] to [N si ,N v1 ] are all single-precision numbers 0;

[0105] Step 1.4: Extract the bridge type data selected by the bridge instance from the bridge instance to construct the bridge type selection set T of the training sample x , bridge type selection set T x is a shape of [N s ,n] two-dimensional tensor, N s is the number of training samples, n is the total number of selectable bridge types; T x Should be identified as one-hot encoding format:

[0106] T x =∑[T x1 ,T x2 ,……T xn ]

[0107] T xi is the probability of using the i-th bridge type, which is a float value between 0 and 1. When marking, the unused bridge type is 0 and the used bridge type is 1;

[0108] Step 1.5: Set the initial condition vector set T v1 and bridge type selection set T x Concatenate in the 0th dimension to get the complete conditional vector set T v2 .

[0109] Furthermore, in step (2), the construction of the bridge selection network and training includes the following steps:

[0110] Step 2.1: Accept the geological sketch set T c And the initial condition vector set T v1 As input, T c With T v1 Each of the four-channel fusion bridge-type selection feature sets T is scaled to a unified shape through an embedding layer composed of multiple convolutional layers rq ;

[0111] Step 2.2: Use 3 fully connected layers as the middle layer to transform T rqGradually map to N q dimensional multi-classification bridge selection output layer, N q is the number of available bridge types. The bridge type selection output layer uses the softmax layer as the activation layer to sort the priorities of each bridge type under the current geological conditions, and obtains the bridge type selection recommendation set T of the bridge selection network. xt :

[0112] T xt =∑[T xt1 ,T xt2 ,……T xtn ]

[0113] T xti The usage probability of the i-th bridge type obtained by the network is a float value between 0 and 1;

[0114] Step 2.3, define the loss function:

[0115]

[0116] T xi : The usage probability of the marked i-th bridge type, which is a float value between 0 and 1. When marking, the unused bridge type is 0 and the used bridge type is 1;

[0117] Using loss type Update the weights of the training samples and perform iterative training of the bridge selection network for N training times j .

[0118] Furthermore, in step (3), the construction of the generator includes the following parts and steps:

[0119] Step 3.1: The generator includes a feature extraction network of geological sketches, which extracts the geological sketches of the same batch into a set T c Input the network, extract and expand features in a sequential manner including but not limited to batch normalization, deconvolution layer, random dropout, and linear layer, analyze the characteristics of the geological sketch, and the output result is called the geological sketch feature vector X c , let the feature vector dimension be N c , the input sketch is an RGB three-channel color image, X c The size should be [3,N c ];

[0120] Step 3.2: The generator includes a network for constructing conditional vectors, which converts the complete conditional vector set T v2 Feature extraction is performed in a sequential manner including but not limited to batch normalization, deconvolution layer, random dropout, linear layer, and the conditional vector is expanded to the same feature vector dimension N as the geological sketch through multiple deconvolution layers.c , then the conditional vector eigenvector X t The scale is [1,N c ];

[0121] Step 3.3: The generator includes a feature fusion network for the bridge and culvert design conditions, and the geological sketch feature vector X c and the conditional feature vector X shown in step 3.2 t Input network, use multi-layer convolution for downsampling, sequentially including but not limited to batch normalization, random dropout, linear layer, use LeakyRelu activation function to reduce the influence of outliers, output fusion feature T of bridge and culvert design status r , let the dimension of fused feature vector be N r ,T r is [4,N r ];

[0122] Step 3.4, the generator includes generating a drawing network, and the fusion feature T of the bridge and culvert design conditions in step 3.3 r , perform multi-layer deconvolution to expand the size for upsampling, and automatically select the size of the deconvolution kernel that can be divided by the generated design image. Each layer is sequentially performed including but not limited to batch normalization, random dropout process, and several layers of positive convolution are added at the end to compress and reduce the checkerboard effect. The result is re-deformed and output as a design span map of a preset size.

[0123] Furthermore, in step (4), the construction of the span discriminator includes the following steps:

[0124] Step 4.1: Use the global discriminator of global features to input a span design diagram, use multi-layer convolutional layers for upsampling, followed by multi-layer linear fully connected layers, and the output layer is a judgment layer with only one parameter to determine whether it is generated by the system;

[0125] Step 4.2: Use the local discriminator of local geological features, input a span design drawing, and after upsampling using multiple convolutional layers, output a matrix in which each value represents the truth or falsehood of each area in the input design span draft. The average value of each area is taken as the final result of the local discriminator.

[0126] Furthermore, in step (5), performing generative adversarial training includes the following steps:

[0127] Step 5.1: Set the geological sketch set T c Input the geological sketch feature extraction network of the span generator, and estimate the complete condition vector set T of span and bridge type. v2Input the conditional vector of the span generator to build the network. According to the generator structure mentioned above, the fusion feature set T of the bridge and culvert design status will be obtained. r , the fusion feature set T of bridge and culvert design conditions r Input the generated drawing network to obtain the generated span design atlas T g ;

[0128] Step 5.2: Generate the span design atlas T g Input the aforementioned span discriminator to obtain the span rationality vector X of the generated design g ;

[0129]

[0130] N s is the number of training samples;

[0131] X gi The rationality judged by the discriminator for the i-th generated design is a float value between 0 and 1;

[0132] The loss function of the generator is defined as:

[0133]

[0134] Through loss g Perform back-propagation optimization on the generator weights;

[0135] Step 5.3: Split the actual design atlas into s Input the discriminator to obtain the reasonableness vector X of the real design d :

[0136]

[0137] N s : Number of training samples

[0138] X di : The rationality judged by the discriminator of the i-th real design is a float value between 0 and 1;

[0139] The loss function of the discriminator for the true design is defined as:

[0140]

[0141] Through loss d_z Back-propagation optimization of the discriminator weights;

[0142] Step 5.4, define the loss function of the discriminator for generating the design as:

[0143]

[0144] N s : Number of training samples

[0145] X gi : The rationality judged by the discriminator of the i-th generated design is a float value between 0 and 1;

[0146] Through loss d_s Back-propagation optimization of the discriminator weights;

[0147] Step 5.5, train the generator and the discriminator alternately in several cycles, and repeat steps 5.2 to 5.4 so that the error between the image generated by the generator and the real span design drawing is within the preset range, and the accuracy of the discriminator in judging the generated case and the real design span drawing is gradually improved until the accuracy of the discriminator in judging the generated solution and the real solution is stabilized at the Nash equilibrium.

[0148] Furthermore, in step (6), the method comprises the following steps:

[0149] Step 6.1: Input the design data of the new design case and construct a new preliminary condition vector set T v1x , using the automatic modeling sketcher to draw a new geological sketch set T cx ,

[0150] Step 6.2: Use the bridge type selection network to obtain the new bridge type selection set T for this design case xx , using the new set of preliminary condition vectors T v1x With the new bridge type selection set T xx Concatenate in the 0th dimension to get a new complete conditional vector set T v2x , using the generator network to complete the new cross-design atlas T sx Automatic design.

[0151] So far, the intelligent span design method of beam bridge based on generative adversarial neural network proposed by the present invention has been completed, and the intelligent span design of overpass and navigation bridge has been completed. However, the above embodiments are only for explaining the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for intelligent design of beam bridge spans based on generative adversarial neural networks, characterized in that: The method comprises the following steps: (1) Construct a training sample set and an automatic modeling drawer. Use the drawer to extract the terrain data of the beam bridge instance and generate a standardized geological sketch set T c ; Use the drawing tool to extract the design data of the beam bridge example and generate the span design atlas T s , extract the expected span from the bridge instance, and construct the preliminary condition vector set T according to the expected daily traffic volume and the number of available bridge types v1 , extract the bridge type data selected by the bridge instance from the bridge instance to construct the bridge type selection set T of the training sample x , using the generated geological sketch set T c , Span Design Atlas T s , the initial condition vector set T v1 and bridge type selection set T x As training samples, the initial condition vector set T v1 and bridge type selection set T x The concatenation in the 0th dimension is called the complete conditional vector set T v2 ; (2) Construct and train a bridge type selection network and transform the geological sketch set T c With the initial condition vector set T v1 Input the bridge type selection network, and obtain the network's bridge type selection set T through multimodal input processing, convolutional feature extraction and fully connected layer classification. xt , with T xt With T x The mean square error is the error function, and the network is trained by back propagation of the SGD algorithm; (3) Constructing a span generator network, which is composed of a geological sketch set feature extraction network, a conditional vector construction network, a bridge and culvert design condition feature fusion network, and a generation and drawing network; (4) Constructing a split-span discriminator network, which consists of a local discriminator and a global discriminator; (5) Generate adversarial training and transform the geological sketch set T c Input the geological sketch set feature extraction network of the span generator and convert the complete condition vector set T v2 Input the conditional vector of the span generator to build a network, and train the generator and the discriminator alternately in several cycles, so that the difference between the image generated by the generator and the real span design is within a preset range, until the judgment accuracy of the discriminator on the generated scheme and the real scheme is stabilized at Nash equilibrium; (6) Use the automatic modeling drawing device that has passed the performance evaluation, the bridge type selection network and the generator network, input the design data of the new design case, and construct a new preliminary condition vector set T v1x , using the automatic modeling sketcher to draw a new geological sketch set T cx , use the bridge type selection network to obtain the new bridge type selection set T for this design case xx , construct a new complete conditional vector set T v2x , and then use the generator network to complete the new cross-design atlas T sx Automatic design.

2. According to claim 1, a method for intelligent design of beam bridge spans based on generative adversarial neural networks is characterized in that: In step (1), constructing a training sample set comprises the following steps: Step 1.1: Read the design geological information from the input data or drawings and draw a geological sketch set T c , Geological Sketches T c It is a longitudinal section of geological conditions, including various design water levels, geographic lines, bridge deck elevations, minimum vertical clearance above the bridge deck, a simplified diagram of the upper structure of the beam bridge, unfavorable geological sections, navigation clearance areas, clearance areas for cross-line carriageways, soil conditions, earthquake conditions, and design conditions for material use. Among them, unfavorable geological sections are defined as geological sections that are clearly identified as unsuitable for piers in the design task. When input, they are expressed as a shape of [N b ,2], a two-dimensional matrix, N b The first column of the matrix indicates the starting point of each bad geological section, and the second column of the matrix indicates the end point of each bad geological section. The soil conditions are marked with different colors under the geographic line, and the material usage is marked with color blocks of each bridge and culvert structure. Different geological sketch information is added according to different bridge and culvert conditions: The navigable bridge will read the input bridge deck elevation and expected span, and scale the larger of the expected span and the difference between the bridge deck elevation and the maximum water depth elevation to the pixel position predetermined by the design drawing, and scale the other to the pixel position while keeping the horizontal and vertical ratio unchanged. The upper structure building height is preliminarily estimated based on the optional minimum vertical clearance of the building height above the bridge deck and the expected height-span ratio, and the upper structure longitudinal section is drawn with different colors to mark different materials; if the database or input does not give the channel grade, the channel grade is calculated based on the width of the water body, and several polygonal navigation clearance blocks are drawn symmetrically in the center of the river, and then the water bodies at different water levels are colored according to the local flood control, drift ice, and bridge pier water defense water levels of the river; The overpass will read the input bridge deck elevation and expected span, scale the larger of the expected span and the bridge deck elevation to the road ground elevation to the pixel position predetermined by the design drawing, and scale the other to the pixel position while keeping the horizontal and vertical ratio unchanged. Preliminary estimate the upper structure building height based on the optional minimum vertical clearance of the building height above the bridge deck and the expected height-span ratio, draw the upper structure longitudinal section with different colors to mark different materials, and draw the clearance area of ​​the roadway at the position of the roadway in imitation of the navigation bridge. For poor geology, if the poor geology comes from soil factors, the section below the geographic line will be marked in red; if it comes from obstacle factors, the detailed start and end points and clearance requirements will be marked under the water body and geographic line, and the poor geology will also be marked in red; Step 1.2: Read the design geological information from the input beam bridge design data and draw the span design atlas T s , Span Design Atlas T s It is a longitudinal section of geological conditions, including the design water levels, geographic lines, bridge deck elevations, schematic diagrams of beam bridge superstructures, spans of beam bridges, positions and sizes of piers of beam bridge substructures, unfavorable geological sections, navigation clearance areas, clearance areas for cross-line carriageways, soil conditions, earthquake conditions, and material usage conditions. The definition of unfavorable geological sections is the same as in step 1.

1. Soil conditions are marked with different colors under the geographic lines. Material usage conditions are marked with color blocks of various structures of bridges and culverts. Any selection of piers and material selections need to be marked with different colors. c Different, split across design atlas T s The detailed data of the bridge pier is added. When drawing, only the starting point of the bridge pier is accurately determined. The longitudinal section width of the bridge pier represents the nominal radius d of the plane section of the bridge pier. m : V: average cross-sectional area of ​​the pier π: the circumference constant Due to the different characteristics of navigation bridges and cross-line bridges, different additional annotations are required for the span design drawings: for navigation bridges, the colors of the piers are marked differently according to the water levels of the navigation bridges; for cross-line bridges, in addition to the annotation format of ordinary beam bridges, the span type, minimum left lateral clearance, minimum right lateral clearance and minimum vertical underground clearance elevation are additionally marked to calculate the restrictions of the cross-line on the pier design. Among them, the span type is divided into highway underground structure, railway underground structure and non-underground structure, and different clearance regulations and colors are used to identify different structure types. The minimum left lateral clearance of the passage refers to the distance from the right edge of the highway or the center line of the right track of the railway train to the nearest bridge structure or slope foot, and the minimum left lateral clearance refers to the distance from the left edge of the highway or the center line of the left track of the railway train to the nearest bridge structure or slope foot; Step 1.3: Extract the estimated span, estimated daily traffic volume and available bridge types directly from the bridge instance to construct the preliminary condition vector set T v1 , the initial condition vector set T v1 is a shape of [N s , N v1 ], N s is the number of training samples, N v1 is the length of the initial condition vector set: N v1 =2+n n is the total number of selectable bridge types; Preliminary condition vector set T v1 In [N si , 1] represents the Nth si The estimated span of training samples, expressed as a single-precision number; [N si , 2] represents the Nth si The estimated daily traffic volume of training samples is expressed as a single-precision value; [N si ,3] to [N si , N v1 ] are all single-precision numbers 0; Step 1.4: Extract the bridge type data selected by the bridge instance from the bridge instance to construct the bridge type selection set T of the training sample x , bridge type selection set T x is a shape of [N s ,n] two-dimensional tensor, N s is the number of training samples, n is the total number of selectable bridge types; T x Should be identified as one-hot encoding format: T x =∑[T x1 ,T x2 ,......T xn ] T xi is the probability of using the i-th bridge type, which is a float value between 0 and 1. When marking, the unused bridge type is 0 and the used bridge type is 1; Step 1.5: Set the initial condition vector set T v1 and bridge type selection set T x To concatenate: the initial condition vector set T v1 and bridge type selection set T x They are expressed in the form of two-dimensional tensors. v1 The data of each training sample is then concatenated with T x The data of the training sample in the training set is used to obtain the complete conditional vector set T v2 .

3. The method for intelligent design of beam bridge spans based on generative adversarial neural networks according to claim 1 is characterized in that: In step (2), the construction of the bridge selection network and training includes the following steps: Step 2.1: Accept the geological sketch set T c And the initial condition vector set T v1 As input, T c With T v1 Each of the four-channel fusion bridge-type selection feature sets T is scaled to a unified shape through an embedding layer composed of multiple convolutional layers rq ; Step 2.2: Use 3 fully connected layers as the middle layer to transform T rq Gradually map to N q dimensional multi-classification bridge selection output layer, N q is the number of available bridge types. The bridge type selection output layer uses the softmax layer as the activation layer to sort the priorities of each bridge type under the current geological conditions, and obtains the bridge type selection recommendation set T of the bridge selection network. xt : T xt =∑[T xt1 ,T xt2 ,......T xtn ] T xti The usage probability of the i-th bridge type obtained by the network is a float value between 0 and 1; Step 2.3, define the loss function: T xi : The usage probability of the marked i-th bridge type, which is a float value between 0 and 1. When marking, the unused bridge type is 0 and the used bridge type is 1; Using loss type Update the weights of the training samples and perform iterative training of the bridge selection network for N training times j .

4. The method for intelligent design of beam bridge spans based on generative adversarial neural networks according to claim 1 is characterized in that: In step (3), the construction of the generator includes the following parts and steps: Step 3.1: The generator includes a feature extraction network of geological sketches, which extracts the geological sketches of the same batch into a set T c Input the network, extract and expand features in a sequential manner including but not limited to batch normalization, deconvolution layer, random dropout, and linear layer, analyze the characteristics of the geological sketch, and the output result is called the geological sketch feature vector X c , let the dimension of the feature vector be N c , the input sketch is an RGB three-channel color image, X c Should be of size [3, N c ]; Step 3.2: The generator includes a network for constructing conditional vectors, which converts the complete conditional vector set T v2 Feature extraction is performed in a sequential manner including but not limited to batch normalization, deconvolution layer, random dropout, linear layer, and the conditional vector is expanded to the same feature vector dimension N as the geological sketch through multiple deconvolution layers. c , then the conditional vector eigenvector X t is of size [1, N c ]; Step 3.3: The generator includes a feature fusion network for the bridge and culvert design conditions, and the geological sketch feature vector X c and the conditional feature vector X shown in step 3.2 t Input network, use multi-layer convolution for downsampling, sequentially including but not limited to batch normalization, random dropout, linear layer, use LeakyRelu activation function to reduce the influence of outliers, output fusion feature T of bridge and culvert design status r , let the dimension of fused feature vector be N r , T r For [4, N r ]; Step 3.4, the generator includes generating a drawing network, and the fusion feature T of the bridge and culvert design conditions in step 3.3 r , perform multi-layer deconvolution to expand the size for upsampling, and automatically select the size of the deconvolution kernel that can be divided by the generated design image. Each layer is sequentially performed including but not limited to batch normalization, random dropout process, and several layers of positive convolution are added at the end to compress and reduce the checkerboard effect. The result is re-deformed and output as a design span map of a preset size.

5. The method for intelligent design of beam bridge spans based on generative adversarial neural networks according to claim 4 is characterized in that: In step (4), the construction of the span discriminator includes the following steps: Step 4.1: Use the global discriminator of global features to input a span design diagram, use multi-layer convolutional layers for upsampling, followed by multi-layer linear fully connected layers, and the output layer is a judgment layer with only one parameter to determine whether it is generated by the system; Step 4.2: Use the local discriminator of local geological features, input a span design drawing, and after upsampling using multiple convolutional layers, output a matrix in which each value represents the truth or falsehood of each area in the input design span draft. The average value of each area is taken as the final result of the local discriminator.

6. The method for intelligent design of beam bridge spans based on generative adversarial neural networks according to claim 5 is characterized in that: In step (5), performing generative adversarial training includes the following steps: Step 5.1: Set the geological sketch set T c Input the geological sketch feature extraction network of the span generator, and estimate the complete condition vector set T of span and bridge type. v2 Input the conditional vector of the span generator to build the network. According to the generator structure mentioned above, the fusion feature set T of the bridge and culvert design status will be obtained. r , the fusion feature set T of bridge and culvert design conditions r Input the generated drawing network to obtain the generated span design atlas T g ; Step 5.2: Generate the span design atlas T g Input the aforementioned span discriminator to obtain the span rationality vector X of the generated design g ; N s is the number of training samples; X gi The rationality judged by the discriminator for the i-th generated design is a float value between 0 and 1; The loss function of the generator is defined as: Through loss g Perform back-propagation optimization on the generator weights; Step 5.3: Split the actual design atlas into s Input the discriminator to obtain the reasonableness vector X of the real design d : N s : Number of training samples X di : The rationality judged by the discriminator of the i-th real design is a float value between 0 and 1; The loss function of the discriminator for the true design is defined as: Through loss d_z Back-propagation optimization of the discriminator weights; Step 5.4, define the loss function of the discriminator for generating the design as: N s : Number of training samples X gi : The rationality judged by the discriminator of the i-th generated design is a float value between 0 and 1; Through loss d_s Back-propagation optimization of the discriminator weights; Step 5.5, train the generator and the discriminator alternately in several cycles, and repeat steps 5.2 to 5.4 so that the error between the image generated by the generator and the real span design drawing is within the preset range, and the accuracy of the discriminator in judging the generated case and the real design span drawing is gradually improved until the accuracy of the discriminator in judging the generated solution and the real solution is stabilized at the Nash equilibrium.

7. The method for intelligent design of beam bridge spans based on generative adversarial neural networks according to claim 1 is characterized in that: In step (6), the method comprises the following steps: Step 6.1: Input the design data of the new design case and construct a new preliminary condition vector set T v1x , using the automatic modeling sketcher to draw a new set of geological sketches T cx , Step 6.2: Use the bridge type selection network to obtain the new bridge type selection set T for this design case xx , using the new set of preliminary condition vectors T v1x With the new bridge type selection set T xx Splice to get a new complete conditional vector set T v2x , using the generator network to complete the new cross-design atlas T sx Automatic design.

Images