JP7332068B1 - Automatic calculation system, automatic calculation method, machine learning model, and method for generating machine learning model - Google Patents

Abstract

An automatic calculation system capable of automatically calculating the length dimension of piping equipment from a design drawing and reducing the work load of picking up in integration. An automatic arithmetic system (1) of the present invention comprises a data acquisition unit for acquiring drawing image data (2) of a design drawing, an array data generation unit for generating array data from the drawing image data (2), The installation area of the piping equipment (5) is identified using a machine learning model generated by machine learning, in which the array data is input and the equipment binary mask indicating the installation area of the piping equipment (5) is output in the array data. It has an identification unit and a calculation unit that calculates the length dimension of the piping equipment (5) based on the array data representing the installation area of the piping equipment (5) identified by the identification unit. [Selection drawing] Fig. 1

Description

The present invention provides an automatic calculation system and an automatic calculation method for automatically calculating the length of piping equipment from design drawings describing piping equipment including pipes and ducts, and a machine learning model and machine learning used therefor. and related to model generation methods.

In the construction work of constructing a building, the work of calculating the costs required for the construction work is carried out by determining the materials, processes, construction period, etc. necessary for the construction work based on the design drawings and the like. In the cost estimation of building work, generally, a building surveyor with specialized knowledge picks up from the design drawing the type and quantity of piping equipment such as pipes and ducts used in the construction work. The construction cost is calculated based on the information obtained by picking it up.

Further, for example, Japanese Patent Application Laid-Open No. 2002-92393 (Patent Document 1) discloses an integration system for piping equipment used for integration work. In the accumulation system described in Patent Document 1, drawing data of a piping drawing in which piping and the like are described is displayed on the display of a personal computer. By specifying the position with the mouse and entering the height position information of the pipe etc. with the keyboard, the length of the pipe etc. is measured by the integration system based on the input data. Then, the measured length dimension is displayed on the display of the personal computer.

JP-A-2002-92393

Generally, in construction work, detailed estimates are often repeatedly made after a rough estimate is created in the initial basic planning stage until a final estimate is created. Also, when creating an estimate by a detailed estimate or when creating a final estimate, the above-mentioned building surveyor will perform the cost estimation each time. For this reason, when receiving an order for one construction work, the work of estimation is performed multiple times.

In addition, although an accumulation system such as that described in Patent Document 1, for example, is often used for picking work in an accumulation business, in such a conventional accumulation system, input using a mouse or keyboard is not possible. Manual work is often required. For this reason, the conventional pick-up work required a great deal of time and labor, and a heavy burden was placed on the estimating work.

Furthermore, at present, the number of building surveyors is insufficient for the required quantity surveying work, so there are cases where the surveying work that accompanies the work of picking out quantity surveys is outsourced. There is a concern that an increase in the number of outsourcing will lead to a deterioration in the quality of the estimation.

In addition, companies (so-called general contractors) that undertake construction work and equipment related work (so-called subcontractors), which are the main contractors for building and civil engineering work, face a large burden of survey work and a shortage of building surveyors. Due to this, it is difficult to increase the number of orders for construction work, and as a result, there is a problem that the options for construction work that can be ordered are limited.

The present invention has been made in view of the above-mentioned conventional problems, and its object is to automatically calculate the length of piping equipment including pipes and ducts from design drawings, thereby enabling An object of the present invention is to provide an automatic calculation system and an automatic calculation method that can reduce the burden, a machine learning model used in the automatic calculation system and the automatic calculation method, and a method for generating the machine learning model.

In order to achieve the above object, an automatic calculation system for automatically calculating the length dimension of the piping equipment from the design drawing describing the piping equipment including pipes and ducts provided by the present invention, a data acquisition unit that acquires drawing image data of the design drawing; an array data generation unit that generates array data including image information from the drawing image data acquired by the data acquisition unit; an identifying unit that identifies the installation area of the piping equipment using a machine learning model generated by machine learning that receives array data as an input and outputs a facility binary mask indicating the installation area of the piping equipment in the array data; and a computing unit that computes the length dimension of the piping equipment based on the installation area of the piping equipment specified by the specifying unit.

In the automatic arithmetic system of the present invention, it is preferable that the array data includes image width, image height, and pixel value information as the image information.
Moreover, it is preferable that the calculation unit acquires scale information of the design drawing and calculates the length dimension of the piping equipment.

The automatic arithmetic system of the present invention has a correct data acquisition unit for acquiring correct data indicating the correct area of the piping equipment when an error has occurred in the area of the piping equipment identified by the identifying unit, It is preferable that the specifying unit re-specify the installation area of the piping equipment based on the acquired correct data.
In this case, the automatic computing system preferably has a learning unit that further learns the machine learning model using the correct answer data.

Further, the automatic computing system of the present invention has a management server on which the machine learning model is deployed and data transmission/reception is performed with a web application, and the web application is acquired by the data acquisition unit. It is preferable to display the drawing image data of the design drawing, the installation area of the piping equipment specified by the specifying unit, and the calculated length dimension of the piping equipment.

The automatic calculation method provided by the present invention is an automatic calculation method for automatically calculating the length dimension of the piping equipment from the design drawing describing the piping equipment including pipes and ducts by an automatic calculation system. , a data acquisition step in which the data acquisition unit of the automatic calculation system acquires drawing image data of the design drawing; and an array data generation unit of the automatic calculation system, from the drawing image data acquired in the data acquisition step, an array data generation step of generating array data including image information; and an equipment binary indicating an installation area of the piping equipment in the array data, wherein the identification unit of the automatic arithmetic system receives the array data of the drawing image data as input. A specifying step of specifying an installation area of the piping equipment using a machine learning model generated by machine learning using a mask as an output; and a calculation step of calculating the length dimension of the piping equipment based on the installation area of the.
In the automatic calculation method of the present invention, it is preferable that the array data include image width, image height, and pixel value information as the image information.

Next, the machine learning model provided by the present invention is generated from the drawing image data of the design drawing in which piping facilities including pipes and ducts are described, and the width of the image, the height of the image, and the pixel value an output layer for outputting a facility binary mask indicating the installation area of the piping facility in the array data; the array data for the drawing image data; and the facility binary mask. and an intermediate layer whose parameters are machine-learned using a plurality of teacher data recorded in association with and, when the array data of the drawing image data is input to the input layer, the operation by the intermediate layer is performed machine learning model that causes a computer to perform and output the facility binary mask from the output layer.

A method for generating a machine learning model provided by the present invention includes pre-training processing in which a computer generates a pre-trained model, and transfer learning of the pre-trained model by the computer , whereby a piping installation including pipes and ducts is performed. A transfer learning process for generating a machine learning model that receives as input array data obtained from drawing image data of a design drawing in which is described, and outputs a facility binary mask that indicates the installation area of the piping facility in the array data and in the pre-training process, the computer provides provisional drawing image data in which a virtual line representing a virtual piping facility is added to first drawing image data of a first design drawing in which the piping facility is not described. a step of generating virtual array data including image width, image height, and pixel value information from the provisional drawing image data; and a plurality of virtual array data divided at a constant ratio a step of generating divided patch data, performing mask processing for masking a part of the obtained divided patch data to generate mask array data; and inputting the mask array data and outputting the virtual array data. and generating the pre-trained model by error backpropagation, and in the transfer learning process, the computer acquires second drawing image data of a second design drawing in which the piping equipment is described. generating second array data including image width, image height, and pixel value information from the second drawing image data; and generating by annotating the second array data obtaining a facility binary mask indicating an installation area of the piping facility; and using a plurality of teacher data, the second drawing image data as an input and the facility binary mask as an output, by an error backpropagation method. and transferring the training model to generate the machine learning model.

In the machine learning model generation method of the present invention, in the transfer learning process, the computer performs a step of selecting and designating a hint binary mask from the equipment binary mask generated by the annotation process, and performing the transfer learning Preferably, in the machine learning model generating step of processing, the computer uses the hint binary mask for the transfer learning of the pre-trained model.

The automatic calculation system and automatic calculation method of the present invention can automatically and accurately calculate the length dimensions of piping equipment including pipes and ducts from design drawings, so that the work load of picking up in integration can be reduced.

1 is a block diagram schematically showing the configuration of an automatic arithmetic system according to an embodiment of the present invention; FIG. FIG. 3 is a schematic diagram showing an example of displaying drawing image data of a design drawing in which a duct is described; 3 is a schematic diagram showing an example in which a part of the drawing image data shown in FIG. 2 is enlarged to display a duct installation area; FIG. FIG. 3 is a schematic diagram showing an example of displaying a duct installation area specified by a facility binary mask and a centerline extracted from the installation area; It is an explanatory view explaining pre-training processing in a generation method of a machine-learning model. It is a figure which shows the 1st design drawing in which the piping equipment is not described. 7 is a diagram showing provisional drawing image data generated from the first design drawing shown in FIG. 6; FIG. FIG. 4 is an explanatory diagram showing an algorithm of a software program for generating provisional drawing image data; FIG. 4 is an explanatory diagram for explaining transfer learning processing in the method of generating a machine learning model; 4 is a flow chart of an automatic calculation method according to an embodiment;

BEST MODE FOR CARRYING OUT THE INVENTION Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. It should be noted that the present invention is not limited to the embodiments described below, and various modifications are possible as long as they have substantially the same configuration as the present invention and have the same effects. is.

(Configuration of automatic computing system)
FIG. 1 is a block diagram schematically showing the configuration of the automatic arithmetic system of this embodiment.
The automatic arithmetic system 1 of the present embodiment uses the drawing image data 2 (see FIG. 2) of the design drawing in which piping equipment such as pipes and ducts 5 is described to identify the area where the piping equipment is installed. (See FIG. 3), and automatically calculates and displays the length dimension of the piping equipment whose area is specified.

In this embodiment, a case where a design drawing describes a duct 5 as piping equipment will be described. In the present invention, piping equipment includes pipes for circulating liquids, gases, etc., ducts for circulating air, pipes for housing and protecting electric wires and cables, and other pipes for circulating fluids.

As shown in FIG. 1, the automated computing system 1 of this embodiment includes a management server 10 and a plurality of user terminals (client terminals) 30 connected to the management server 10 via a network 20 so as to be able to communicate with each other. have. The network 20 is formed by the Internet, intranet, LAN, WAN, and the like.

The management server 10 of this embodiment is formed by one or more virtual servers (cloud servers). Note that the management server 10 may be formed by one or a plurality of physical servers. The management server 10 includes a web server 11 that transmits and receives data to and from the user terminal 30, an application server 12 that receives requests from the web server 11 and executes various processes, and a database that manages various data ( database server) 13. In addition, the management server 10 includes a serving server 14 to which a machine learning model (hereinafter abbreviated as a learning model) of this embodiment, which will be described later, is deployed, and a storage 15 that stores the learning model and various data related to the learning model. (e.g., cloud storage).

The web server 11 of the management server 10 receives the content requested from the web browser 31 of the user terminal 30 and performs processing according to the content. The application server 12 executes various processes in response to requests from the web server 11, and also accesses the database 13 to read, write, process, and search data.

The database 13 stores programs for executing each process and each function, and various data such as data necessary for the programs and data relating to users. The serving server 14 receives input from the application server 12 to execute the learning model, and outputs results obtained from the learning model to the application server 12 .

The management server 10 has a function of providing a web application to the web browser 31 operating on the user terminal 30 , and the web application is displayed on the web browser 31 of the user terminal 30 . Further, the web server 11 of the management server 10 and the web browser 31 of the user terminal 30 exchange data via an API (application programming interface).

Each user terminal 30 is formed in the form of a computer, tablet terminal, smartphone, or the like used by each user, and includes a web browser 31 that can use web applications. For example, when the user terminal 30 is a computer, the user terminal 30 includes, though not shown, a CPU that processes and controls the operation of the entire terminal, a communication interface (IF), and an input that accepts input operations from the user. A device, an output device for displaying information to a user, a memory for temporarily storing programs, data, and the like, and a storage unit for storing data are provided, and these are electrically connected.

In this case, the input device is formed by, for example, a mouse, keyboard, touch panel, touch pad, or the like. The output device is for example formed by a display. The memory is formed of volatile memory such as DRAM, for example. The storage unit is formed by, for example, an HDD, flash memory, or the like. In the present invention, the user terminal is not particularly limited as long as it has at least a web browser that displays web applications.

The web application used in each user terminal 30 includes a data acquisition unit that acquires the drawing image data 2 of the design drawing, and array data that generates array data including image information from the drawing image data 2 acquired by the data acquisition unit. It has a generation unit and a calculation unit that calculates the length dimension of the piping equipment based on the array data representing the installation area of the piping equipment received from the application server 12 or the equipment binary mask.

In this case, the computing unit of the web application can acquire the scale information of the design drawing and compute the length of the piping equipment. Furthermore, the web application has a correct data acquisition unit that acquires correct data indicating the correct area of the piping facility when the area of the piping facility specified by the specifying unit of the application server 12 is incorrect. The calculation unit of the application can recalculate the length of the piping installation based on the acquired correct data.

The application server 12 of the management server 10 includes an identification unit that identifies the installation area of the piping equipment in the design drawing based on the equipment binary mask 63 output from the learning model described later and generates array data. and a learning unit for further learning the learning model using the correct data indicating the correct area of the piping equipment when there is an error in the area of the piping equipment.

In this embodiment, the data acquisition unit, sequence data generation unit, calculation unit, and correct data acquisition unit of the web application, and the identification unit and learning unit of the application server 12 are each formed by a computer program. In the automatic arithmetic system 1 of the example, each function of the automatic arithmetic system 1 is realized by cooperation of software and hardware resources by executing each computer program described above.

Next, the learning model used in the automatic arithmetic system 1 of this embodiment will be described.
A learning model is a model using machine learning, and machine learning is sometimes called artificial intelligence. Machine learning itself can use various known methods, for example, a neural network may be used. Note that the learning model generated by machine learning in this embodiment is also called a fine-tuning model.

The learning model (fine-tuning model) of this embodiment includes an input layer to which array data generated from the drawing image data 2 of the design drawing describing the piping equipment is input, and an output layer to output the equipment binary mask 63. and an intermediate layer whose parameters are machine-learned using a plurality of teacher data recorded in association with the arrangement data of the drawing image data 2 and the equipment binary mask 63 . Here, the equipment binary mask 63 is binary data obtained by extracting the piping equipment from the array data, as will be described later.

In this embodiment, the design drawings include basic design drawings, working design drawings, construction drawings, and the like. Design drawings are also called architectural drawings. The drawing image data 2 is image data representing a design drawing, and includes, for example, image files such as JPEG, GIF, PNG, and BMP, and document files such as PDF.

The array data is data generated by converting the drawing image data 2 of the design drawing, and the array data includes image information. For example, in the case of this embodiment, the array data includes the image width W, image height H, and grayscale pixel value information in the drawing image data 2 . In the present invention, the array data may use RGB pixel values instead of grayscale pixel values. The equipment binary mask 63 is data extracted by separating the piping equipment from the background portion by masking the background portion other than the piping equipment such as the pipes and ducts 5 in the array data (for example, see FIG. 4). . This facility binary mask 63 is formed of two values "0" and "1".

The learning model (fine-tuning model) of this embodiment as described above undergoes a pre-training process for generating a pre-trained model 47 described below and a transfer learning process for performing transfer learning of the pre-trained model 47. Generated by
First, in the pre-training process, as shown in FIG. 5, a first training step of acquiring first drawing image data 41 of a plurality of first design drawings in which piping facilities such as pipes and ducts 5 are not described is performed. In this first training step, as the first drawing image data 41 of the first design drawing, public design drawing data that is open to the public, design drawing data at the rough estimate stage (basic planning stage), and the like are used. In this case, for example, a design drawing in which piping equipment is not described as shown in FIG. 6 is used as the first design drawing, and an image file such as JPEG or a document file such as PDF of the first design drawing is It is obtained as drawing image data 41a. Here, the first drawing image data 41a shown in FIG. 6 is the image data of the actual design drawing at the rough estimation stage.

Subsequently, a line 43a representing a virtual piping installation is automatically added to each of the obtained first drawing image data 41 to generate provisional drawing image data 42, thereby obtaining the provisional drawing image data 42. A training step 52 is performed (see FIG. 5). In this second training step 52, for example, using a software program having an algorithm as shown in FIG. By automatically adding a virtual solid line (hereinafter referred to as a virtual line) 43a representing, for example, piping equipment, provisional drawing image data 42a as shown in FIG. 7, for example, is generated. As a result, provisional drawing image data 42 is obtained by adding virtual piping equipment to the first drawing image data 41 .

The first drawing image data 41a shown in FIG. 6 and the provisional drawing image data 42a shown in FIG. 7 are merely examples, and the present invention is not limited to these. Further, in the second training step 52, if the provisional drawing image data 42 as described above can be acquired, the method and means for adding virtual piping equipment to the first drawing image data 41a of the first design drawing are particularly For example, the operator may manually add virtual piping equipment to the first drawing image data.

After generating a plurality of temporary drawing image data 42 as described above, image information (that is, image width W, image height H, and A third training step 53 is performed in which the transform generates virtual array data 44 containing grayscale pixel value information (see FIG. 5).

Next, a plurality of divided patch data obtained by dividing the virtual array data 44 generated in the third training step 53, and mask array data 45 are generated by masking a part of the obtained divided patch data. A fourth training step 54 is performed (see FIG. 5).
In the fourth training step 54, first, the virtual array data 44 containing image information is divided at a constant ratio using a dedicated software program to generate a plurality of divided patch data of the same size. . For example, in this embodiment, the virtual array data 44 is divided into a plurality of small divided patch data of 4 pixels×4 pixels.

Furthermore, in the fourth training step 54, a dedicated software program is used to randomly mask a part of the divided patch data from the virtual array data 44 to obtain mask array data. 45 is generated. The mask arrangement data 45 obtained by such masking is arrangement data corresponding to the design drawing in which a part of the duct 5 (piping equipment) is hidden, such as the "image diagram 46" schematically shown in FIG. becomes.

Furthermore, in a fourth training step 54, a large amount of data groups are created by pairing a plurality of small divided patch data obtained from one piece of virtual array data 44 with the divided patch data of the mask array data 45. FIG. Part of the process of masking part of the divided patch data in the fourth training step 54 may be performed manually by, for example, an operator without using a software program.

After performing the fourth training step 54, as shown in FIG. 5, the virtual array data 44 and the mask array data 45 generated in the fourth training step 54 are used to generate a pre-trained model 47 by error backpropagation. A fifth training step 55 is performed.

In the fifth training step 55, divided patch data of the masked array data 45 generated by masking the virtual array data 44 is used as input, and divided patch data of the virtual array data 44 (unmasked array data) is output. By using a plurality of pairs of data groups, the error back propagation method is used to learn the relationship between both data so that the virtual array data 44 can be restored from the mask array data 45, and the internal information of the pre-trained model 47 will be updated. Then, the pre-trained model 47 becomes capable of computing the expected output for any given input with a certain probability, thus completing the fifth training step 55 and generating the pre-trained model 47 .

Next, a transfer learning process is performed to transfer-learn the pre-trained model 47 generated in the pre-trained process (see FIG. 9). Here, the transfer learning is machine learning so that the pre-trained model 47 generated in the pre-training process is effectively used for the task of recognizing piping equipment. It means machine learning whether it works in the area or not.

In this transfer learning process, first, a first learning step of acquiring a large amount of second drawing image data 61 of a second design drawing in which plumbing equipment including piping etc. is already described in a computer that performs transfer learning processing work. I do. Also in this first learning step, the public data of the design drawing, the data of the design drawing at the detailed estimation stage or the final stage, etc. are used as the second drawing image data 61 of the second design drawing. As the second drawing image data 61, the provisional drawing image data 42 (for example, the provisional drawing image data 42a in FIG. 7) generated in the second training step 52 of the pre-training process described above may be used.

In this embodiment, as an example, a design drawing describing piping equipment as shown in FIG. 9 is used as the second design drawing. By reading an image file such as JPEG or a document file such as PDF of the second design drawing, the second drawing image data 61 can be obtained.

Subsequently, from the acquired second drawing image data 61, a dedicated software program is used to extract image information (that is, image width W, image height H, and grayscale pixel value information). A second learning step 72 is performed in which the second array data 62 are generated by converting the array data 62 (see FIG. 9). As image information, RGB pixel values may be used instead of grayscale pixel values.

After the second array data 62 is generated, a third learning step 73 is performed to acquire the facility binary mask 63 generated by annotating the second array data 62 .
In the third learning step 73, the second drawing image data 61 used in the second learning step 72 is displayed on a display by, for example, a computer GUI (graphical user interface) program, and an annotation process is performed by an operator (annotation operator). ) uses a GUI program to enclose the region of the piping facility described in the second drawing image data 61 displayed on the display with vertices, as shown in FIG. Perform operation 75 to mark as Such marking work 75 can be performed, for example, by an annotation operator operating a mouse of a computer.

By the annotation operator performing the operation 75 of adding the mark as described above to the piping equipment of the second drawing image data 61, a dedicated software program can be used to convert the pixels belonging to the original second array data 62. The data is assigned an appropriate numerical value (“0 (zero)” or “1 (one )”).

As a result, the above-described dedicated software program can mask the background portion other than the piping equipment in the second array data 62 and separate the piping equipment from the background portion. An extracted equipment binary mask 63 can be generated and obtained. By using such a facility binary mask 63, when specifying piping facilities such as pipes and ducts 5 in an actual design drawing, it is possible to determine whether the pixels of the array data obtained from the design drawing represent the piping facilities. It becomes possible to judge whether or not.

Further, for example, when the provisional drawing image data 42 (specifically, refer to the provisional drawing image data 42a in FIG. 7) as described above is used as the second drawing image data 61, extraction of piping equipment is performed. In the third learning step 73, based on the virtual line (virtual line 43a in FIG. 7) added by a dedicated software program (see FIG. 8), annotation processing is automatically performed to Since it is possible to assign regions other than the region, the equipment binary mask 63 can be generated without any work by the annotation operator. Even when the provisional drawing image data 42 (the provisional drawing image data 42a in FIG. 7) is used in this way, the equipment binary mask 63 may be generated based on the work of the annotation operator.

After acquiring the equipment binary mask 63 by the third learning step 73, as shown in FIG. A fourth learning step 74 is performed in which the model 47 is subjected to transfer learning to generate a learning model. In this fourth learning step 74, the second array data 62 of the second drawing image data 61 is used as an input, and the equipment binary mask 63 is used as an output. The pre-trained model 47 is transfer-learned. Through this transfer learning, a learning model is generated that can compute which part of the engineering drawing is the plumbing area. A learning model generated by performing the fourth learning step 74 is referred to as a fine-tuning model.

In this embodiment, in the transfer learning process (FIG. 9) for transferring the pre-trained model 47, after the equipment binary mask 63 is obtained by performing the annotation process as described above, the hint mask dedicated software program A hint step 76 is preferably performed in which some data is randomly selected from the facility binary mask 63 generated using , and designated (generated) as a hint binary mask 65 . The software program used here is, for example, a program for specifying hint masks by simulating that an annotation operator actually gives hints. Further, for example, when the above-described provisional drawing image data 42 (specifically, refer to the provisional drawing image data 42a in FIG. 7) is used as the second drawing image data 61, the provisional drawing image data 42 The hint binary mask 65 may be automatically specified by a software program based on the added virtual line (virtual line 43a in FIG. 7).

In the transfer learning process of this embodiment, when the pre-trained model 47 undergoes transfer learning in the fourth learning step 74 described above, the hint binary mask 65 specified in the hint step 76 as described above is transferred to the second drawing image. By using the data 61 together with the second array data 62 as an input, a fine tuning model (learning model) with a higher accuracy rate can be generated. Note that in the present invention, the hint binary mask 65 may be specified by an operator rather than by a software program. Also, according to the present invention, the fine-tuning model may be generated without performing such a hinting step 76 .

According to the fine-tuning model of the present embodiment generated by performing the above-described pre-training processing (Fig. 5) and transfer learning processing (Fig. 9), the input layer of the design drawing in which piping facilities are described is By inputting the array data of the drawing image data 2, the facility binary mask 63 indicating the installation area of the piping facility is calculated in the intermediate layer, and the calculated facility binary mask 63 is stabilized from the output layer with high accuracy. can be output as

Next, using the automatic computing system 1 of the present embodiment in which the fine tuning model (learning model) described above is deployed on the management server 10 shown in FIG. An automatic calculation method for automatically calculating the length dimension of the duct 5 from the drawing image data 2 (specifically, the partial region 6 selected on the drawing image data 2) will be described with reference to the drawings. do. Here, FIG. 10 is a flow chart of the automatic calculation method according to the embodiment.

When automatically calculating the length dimension of the duct 5 from the drawing image data 2 of FIG. 2 using the automatic calculation system 1 of this embodiment, first, as shown in FIG. The above-described data acquisition unit of the web application performs a data acquisition step 81 for acquiring the drawing image data 2 of the design drawing.

Specifically, in this data acquisition step 81, an image file (for example, a JPEG file) or a document file (for example, a PDF file) of the design drawing of FIG. 1 is loaded into the web application of the web browser 31 operating on the user terminal 30 shown in FIG. As a result, the web application of the user terminal 30 acquires the drawing image data 2 of the design drawing of FIG. 2 and displays the drawing image data 2 on the display (web application) of the user terminal 30 by the GUI of the web application. .

Next, inside the web application in the user terminal 30, the array data generation unit described above obtains the width W of the image, the height H of the image, and the gray scale from the drawing image data 2 of FIG. An array data generating step 82 is performed to convert and generate array data including pixel value information. Further, the array data of the generated drawing image data 2 is sent to the management server 10 via API, and is input to the application server 12 via the web server 11 of the management server 10 (see FIG. 1).

At this time, the user using the user terminal 30 uses, for example, a touch pen or a mouse to draw the drawing image data 2 displayed on the display of the user terminal 30, as shown in FIG. A region 6 whose length dimension is to be calculated can be selected and designated. When the user designates a partial area 6 of the drawing image data 2 on the display in this way, the web application displayed on the user terminal 30 displays the area 6 designated by the user from the drawing image data 2. , image width W, image height H, and array data including grayscale pixel value information is converted to generate array data of the region 6, and the generated array data is transmitted via API Send to the management server 10 . A case where the user designates a partial area 6 on the design drawing as shown in FIG. 2 will be described below. When the user selects and designates the entire design drawing, the array data of the entire drawing image data 2 is generated as described above.

At this time, the web application that acquired the drawing image data 2 shown in FIG. 2 uses the OCR function provided in the web application to read and acquire the scale information described in the drawing image data 2. Note that the scale information may be obtained by the web application, for example, by manually inputting the scale information into the web application using the input device of the user terminal 30 or selecting it on the web application. .

Next, in the application server 12 of the management server 10 to which the array data of the drawing image data 2 has been input, the above-described specifying unit of the application server 12 performs web application for the fine tuning model (learning model) of the management server 10. By inputting the arrangement data generated in 1., a specification step 83 is performed to specify the installation area of the piping equipment (duct 5) described within the range of area 6 of the design drawing of FIG.

In this identification step 83, in the application server 12, by inputting the array data generated by the web application into the input layer of the fine-tuning model (learning model) deployed on the serving server 14, An operation can be performed to output the facility binary mask within the specified region 6 of the drawing image data 2 from the output layer of the fine-tuning model.

After that, the application server 12 uses a post-processing software program to determine the duct 5 based on the array data input from the web application and the facility binary mask output from the fine-tuning model output layer. Generating sequence data in which the installation area is specified. Furthermore, the generated sequence data is transmitted to the web application of the user terminal 30 via the web server 11 together with the equipment binary mask. Note that the process of generating the arrangement data specifying the installation areas of the ducts 5 may be performed by the web application of the user terminal 30 instead of the application server 12 .

The web application of the user terminal 30 that has received the arrangement data specifying the installation area of the duct 5 from the application server 12 specifies the installation area of the duct 5 on the drawing image data 2 based on the received arrangement data, The specified installation area of the duct 5 is displayed on the display of the user terminal 30 by the GUI of the application, as shown in FIG. In addition, in FIG. 3 , the identified installation area of the duct 5 is indicated by diagonal hatching.

Furthermore, in the web application, in the above-described computing unit, the facility binary mask or array data received from the application server 12 (that is, the facility binary mask indicating the installation area of the duct 5 or array data specifying the installation area of the duct 5), Based on the reduced scale information acquired by the OCR function (or input or selection by the user), a calculation step 84 for calculating the length dimension of the piping equipment is performed.

In this calculation step 84, the center line (the solid line in the white part of FIG. 4) of the installation area (the white part of FIG. 4) of the duct 5 (the white part of FIG. 4) is calculated by dedicated software for the facility binary mask as shown in FIG. Extract and find the length of its center line. In the case of this embodiment, the center line of the installation area is formed by a straight line along the duct 5 and an intersection point where the two straight lines intersect. Furthermore, the actual length dimension of the duct 5 (piping equipment) can be determined by performing calculations using the length of the center line and scale information.

Further, in the calculation step 84, for example, by using the information of the arrangement data specifying the installation area of the duct 5 and the scale information, the ratio of the pixel interval in the drawing image data 2 of FIG. It is also possible to obtain the length dimension of the duct 5 of . In this embodiment, such calculation of length dimension is performed by the web application. may be sent to

Subsequently, the web application performs a display step 85 of displaying the numerical value of the actual length dimension of the duct 5 obtained by calculation on the display of the user terminal 30 . Thereby, the user can grasp the specified installation area of the duct 5 and the length dimension of the duct 5 by looking at the display on the display. In addition, in the calculation step 84 and the display step 85, for example, the length dimension of the piping equipment on the design drawing (the actual length dimension on the drawing) may be displayed without using the acquired scale information. is also possible.

Furthermore, in the automatic calculation system 1 of the present embodiment, as a result of calculation by the fine tuning model (learning model) in the application server 12, for example, the installation area of the duct 5 and/or the location of the duct 5 displayed on the display of the user terminal 30 If there is an error in the length dimension, the user who operates the user terminal 30 himself uses an input device such as a touch pen, mouse, keyboard, etc. of the user terminal 30 to check the installation area and / Or the length dimension of the duct 5 can be modified and entered into the web application.

At this time, in the web application used in the user terminal 30, the correct data acquisition unit described above acquires arrangement data (correct data) indicating the correct installation area of the ducts 5 corrected by the user, and the acquired correct arrangement is obtained. Send the data to the management server 10 .

Subsequently, the application server 12 that has received the correct sequence data generates a portion corrected by the user from the correct sequence data (correct data) as a hint binary mask, and inputs the generated hint binary mask to the fine tuning model. . As a result, an equipment binary mask based on the correct data can be output from the output layer of the fine tuning model, and array data in which the correct installation area of the duct 5 is respecified can be generated based on the equipment binary mask. After that, the arrangement data specifying the correct installation area of the duct 5 is transmitted to the web application of the user terminal 30, so that the display of the user terminal 30 can display the correct installation area of the duct 5 and the correct arrangement data based on the correct arrangement data. The recalculated length dimension of the duct 5 is displayed. The web application of the user terminal 30 may display the installation area of the duct 5 and/or the length dimension of the duct 5 corrected by the user on the web application as it is.

Furthermore, in the application server 12 that has received the correct sequence data (correct data), the above-described learning unit inputs the hint binary mask based on the above-described correct data into the fine tuning model, and further refines the fine tuning model (learning model). can be learned. As a result, the more correct sequence data (correct answer data) is acquired by the web application, the higher the probability of the fine tuning model (learning model) being correct, and the accuracy of the automatic arithmetic system 1 of the present embodiment can be improved. .

As described above, according to the automatic calculation system 1 of the present embodiment, the actual length dimension of the piping equipment is automatically determined from the drawing image data 2 of the design drawing in which the piping equipment such as the duct 5 is described. can be calculated and displayed on the display of the user terminal 30. This makes it possible to automate and simplify at least a part of the task of picking up the piping equipment such as the duct 5, which has conventionally been mainly performed manually, in the survey performed by the building surveyor. Therefore, it is possible to reduce the time and labor required for the pick-up work, improve the efficiency and rationalization of the work, shorten the work time, and greatly reduce the work load in the integration work. Furthermore, as a result, the quality of the estimation work can be improved, and the effect of reducing the need for outsourcing of the estimation work can be expected. In addition, it can be expected to contribute to resolving the problem of a shortage of building surveyors, which has been a concern for some time, and the problem of restrictions on the construction options available to companies.

1 automatic calculation system 2 drawing image data 5 duct 6 area 10 management server 11 web server 12 application server 13 database (database server)
14 serving server 15 storage 20 network 30 user terminal (client terminal)
31 web browser 41, 41a first drawing image data 42, 42a temporary drawing image data 43a solid line representing virtual piping equipment 44 virtual array data 45 mask array data 46 image diagram 47 pre-trained model 52 second training step 53 third training step 54 fourth training step 55 fifth training step 61 second drawing image data 62 second sequence data 63 facility binary mask 65 hint binary mask 72 second learning step 73 third learning step 74 fourth learning step 75 work 76 hint step 81 Data Acquisition Step 82 Data Generation Step 83 Identification Step 84 Calculation Step 85 Display Step

Claims (11)

An automatic calculation system that automatically calculates the length dimensions of piping equipment from design drawings describing piping equipment including pipes and ducts,
a data acquisition unit that acquires drawing image data of the design drawing;
an array data generation unit that generates array data including image information from the drawing image data acquired by the data acquisition unit;
Using a machine learning model generated by machine learning in which the array data of the drawing image data is input and an equipment binary mask indicating the installation area of the piping equipment is output in the array data, the installation area of the piping equipment is calculated. an identification part that identifies the
and a computing unit that computes the length dimension of the piping equipment based on the installation area of the piping equipment specified by the specifying unit. 2. The automatic arithmetic system according to claim 1, wherein said array data includes image width, image height, and pixel value information as said image information. 2. The automatic calculation system according to claim 1, wherein said calculation unit acquires scale information of said design drawing and calculates said length dimension of said piping equipment. a correct data acquisition unit that acquires correct data indicating the correct region of the piping equipment when an error has occurred in the region of the piping equipment identified by the identifying unit;
2. The automatic arithmetic system according to claim 1, wherein the identifying unit re-identifies the installation area of the piping equipment based on the acquired correct data. 5. The automatic arithmetic system according to claim 4, further comprising a learning unit for further learning said machine learning model using said correct data. Having a management server on which the machine learning model is deployed and on which data is sent and received from a web application;
The web application provides the drawing image data of the design drawing acquired by the data acquisition unit, the installation area of the piping equipment identified by the identification unit, and the calculated length dimension of the piping equipment. 2. The automatic computing system according to claim 1, which displays An automatic calculation method for automatically calculating the length dimensions of piping equipment from design drawings describing piping equipment including pipes and ducts by an automatic calculation system ,
a data acquisition step in which the data acquisition unit of the automatic calculation system acquires drawing image data of the design drawing;
an array data generation step in which the array data generation unit of the automatic arithmetic system generates array data including image information from the drawing image data acquired in the data acquisition step;
A machine learning model generated by machine learning in which the specifying unit of the automatic arithmetic system receives the array data of the drawing image data as an input and outputs an equipment binary mask indicating the installation area of the piping equipment in the array data. a identifying step of identifying the installation area of the piping equipment using
and a calculation step in which the calculation unit of the automatic calculation system calculates the length of the piping equipment based on the installation area of the piping equipment identified in the identifying step. . 8. The automatic calculation method according to claim 7, wherein said array data includes image width, image height, and pixel value information as said image information. an input layer that is generated from drawing image data of a design drawing that describes piping equipment including pipes and ducts, and that receives array data including image width, image height, and pixel value information;
an output layer for outputting a facility binary mask indicating the installation area of the piping facility in the array data;
An intermediate layer whose parameters are machine-learned using a plurality of teacher data recorded in association with the arrangement data of the drawing image data and the equipment binary mask,
A machine characterized in that, when the array data of the drawing image data is input to the input layer, the intermediate layer performs an operation and causes the computer to output the facility binary mask from the output layer. learning model. Pre-training processing in which a computer generates a pre-trained model, and the computer performs transfer learning of the pre-trained model. Obtained from drawing image data of design drawings in which piping equipment including pipes and ducts are described. a transfer learning process for generating a machine learning model having array data as an input and an equipment binary mask indicating an installation area of the piping equipment in the array data as an output,
In the pre-training process, the computer acquires provisional drawing image data in which a virtual line representing virtual piping equipment is added to first drawing image data of a first design drawing in which the piping equipment is not described. generating virtual array data including image width, image height, and pixel value information from the provisional drawing image data; and dividing the virtual array data into a plurality of divided patch data at a constant ratio. and generating mask array data by performing a masking process for masking a part of the divided patch data; and inputting the mask array data and outputting the virtual array data. generating a pre-trained model by error backpropagation;
In the transfer learning process, the computer acquires second drawing image data of a second design drawing in which the piping equipment is described, and image width and image height from the second drawing image data , and a step of generating second array data including pixel value information; obtaining a facility binary mask indicating an installation area of the piping facility generated by annotating the second array data; and generating the machine learning model by performing transfer learning on the pre-trained model by error backpropagation using a plurality of teacher data with the 2-drawing image data as input and the equipment binary mask as output.
A method of generating a machine learning model, characterized by: In the transfer learning process, the computer selects and designates a hint binary mask from the equipment binary masks generated by the annotation process,
11. The method of generating a machine learning model according to claim 10, wherein in the step of generating the machine learning model of the transfer learning process , the computer uses the hint binary mask for the transfer learning of the pre-trained model.

Images