Architectural design quality prediction method, device, electronic equipment and storage medium

Abstract

The present invention relates to the field of the intersection of construction industry and information technology, and in particular to a method, device, electronic device and storage medium for predicting architectural design quality, wherein the method comprises: obtaining log data and real-time record data of architectural design behaviors of multiple designers; generating a design behavior data set; extracting a design behavior feature vector of each design behavior data sample in the design behavior data set, generating test data according to the design behavior feature vector of each design behavior data sample and design quality quantitative evaluation data; constructing a design quality prediction model based on a machine learning algorithm, training the design quality prediction model using test data, and predicting the architectural design quality of a target prediction object using the trained design quality prediction model. Thus, the problems of incomplete design behavior data, lack of analysis methods for overall characteristics, and lack of methods for analyzing the correlation between design behavior and design quality are solved.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of construction industry and information technology, and in particular to a method, device, electronic equipment and storage medium for predicting building design quality. Background Art

[0002] Although the design phase of a construction project only takes up a small amount of time, money and manpower costs in the entire project, the quality of its results will play a decisive role in the quality and benefits of the entire project. Ensuring the quality of the design is of great significance to ensuring the smooth implementation of the entire project. In actual engineering practice, ensuring the quality of the design usually takes the form of reviewing the design results and discussing solutions to design problems, which is usually a process with high time and labor costs, and there are still omissions.

[0003] In the design process, the subject who performs the design behavior is the designer, and the final design result is formed by the accumulation of these design behaviors. Therefore, there is a close logical relationship between the designer's behavioral characteristics and the quality of the design results. Analyzing the data of design behavior and extracting information reflecting the designer's behavioral characteristics will help us understand the internal reasons for the change in design quality, help us reduce the occurrence of design problems from the root, and better control the quality of design. However, it is not clear how the two are related, and there is no mature method to explicitly and quantitatively describe this relationship.

[0004] At present, the technology that is relatively close to this method is mainly based on the design behavior feature analysis technology of BIM (Building Information Modeling). The data source of this type of technology is mainly based on the log files automatically generated by BIM software. Data mining technology is used to extract information related to certain local features of design behavior. Finally, various types of algorithms are applied to analyze the design behavior data, so as to obtain information such as design efficiency and design collaboration from the perspective of data.

[0005] For example, Related Technology 1 extracts design feature information from the log files automatically generated by BIM modeling software in the designer's work project, and uses the Generalized Suffix Trees algorithm to analyze the characteristics of the designer's execution of modeling commands, and finds that different designers have significantly different efficiencies in executing the same series of commands. Related Technology 2 uses a pattern recognition algorithm to obtain frequently occurring command sequences from log data, pointing out that different designers have significant differences in the efficiency of executing specific sequences. In the next year, they analyzed the log data of different designers working together in the same project to obtain the value played by different designers in the design team and the network of connections between designers. Related Technology 3 extracts command execution data from the log files automatically generated by Autodesk Revit software, uses LSTMNN (Long Short-Term Memory Neural Network) to learn command execution sequences, and predicts the next command that the designer may execute after completing a series of commands.

[0006] However, the methods proposed in these studies still have many limitations. First, it is impossible to obtain a complete description of design behavior by relying solely on log data. This is mainly because the logs automatically generated by the software are not designed to record all the operations of the designer. Secondly, the features and conclusions obtained in most studies are very local, such as the proficiency of certain operations, the relationship between several commands, etc., and there is a lack of analysis methods for overall features. Finally, there is a lack of research and methods to establish the relationship between design behavior and design quality. Summary of the invention

[0007] The present invention provides a method, device, electronic device and storage medium for predicting architectural design quality, so as to solve the problems of incomplete design behavior data, lack of analysis method for overall characteristics, lack of method for analyzing the correlation between design behavior and design quality, etc.

[0008] A first aspect of the present invention provides a method for predicting building design quality, comprising the following steps: obtaining log data and real-time record data about building design behaviors of multiple designers; generating a design behavior data set based on the log data and real-time record data of multiple designers, wherein the design behavior data set includes design behavior data samples of each designer; extracting a design behavior feature vector of each design behavior data sample in the design behavior data set, and generating test data based on the design behavior feature vector and design quality quantitative evaluation data of each design behavior data sample; constructing a design quality prediction model based on a machine learning algorithm, training the design quality prediction model using the test data, and predicting the building design quality of a target prediction object using the trained design quality prediction model.

[0009] Optionally, a design behavior data set is generated based on the log data and real-time record data of multiple designers, including: converting the log data and real-time record data into target format data; generating a design behavior data sample based on the target format data of each designer, and generating a design behavior data set based on the design behavior data sample of each designer.

[0010] Optionally, the target format includes timestamp, record type and record details, and a design behavior data sample is generated according to the target format data of each designer, including: time aligning all data in the target format data according to the timestamp; fusing all the time-aligned data to obtain the design behavior data sample.

[0011] Optionally, before converting the log data and the real-time record data into target format data, the method further includes: performing data cleaning and data screening on the log data and the real-time record data.

[0012] Optionally, extracting a design behavior feature vector for each design behavior data sample in a design behavior data set includes: identifying data corresponding to each statistical indicator in each design behavior data sample; calculating data density and time density of data corresponding to each statistical indicator in the corresponding design behavior data sample, wherein data density represents the ratio of the number of data corresponding to the statistical indicator to the total number of records in the sample, and time density represents the ratio of the total time occupied by the data corresponding to the statistical indicator to the total time span of the sample; and generating a design behavior feature vector based on the data density and time density.

[0013] Optionally, the design quality prediction model is trained using the test data, including: dividing the test data into a training set and a validation set; performing data augmentation on the training set, training the design quality prediction model based on the training set, and validating the design quality prediction model based on the validation set, until the training is stopped when a training stop condition is met.

[0014] Optionally, the real-time recorded data includes keyboard key operation data recorded in real time when the designer is working, data of creating, modifying and deleting components in the architectural design model, and execution command data.

[0015] The second aspect of the present invention provides a device for predicting architectural design quality, including: an acquisition module for acquiring log data and real-time record data about architectural design behaviors of multiple designers; a generation module for generating a design behavior data set based on the log data and real-time record data of multiple designers, wherein the design behavior data set includes design behavior data samples of each designer; an extraction module for extracting a design behavior feature vector of each design behavior data sample in the design behavior data set, and generating test data based on the design behavior feature vector and design quality quantitative evaluation data of each design behavior data sample; a training module for constructing a design quality prediction model based on a machine learning algorithm, training the design quality prediction model using test data, and predicting the architectural design quality of a target prediction object using the trained design quality prediction model.

[0016] The third aspect of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the architectural design quality prediction method of the first aspect.

[0017] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program or instruction stored thereon. When the computer program or instruction is executed, the method for predicting building design quality according to the first aspect is implemented.

[0018] Thus, the present invention includes the following beneficial effects:

[0019] The embodiment of the present invention generates a design behavior data set by acquiring log data and real-time record data about architectural design behaviors of multiple designers, extracting the design behavior feature vector of each design behavior data sample from the behavior data set, and generating test data based on the two, constructing a design quality prediction model based on a machine learning algorithm and using the test data to train the design quality prediction model, and using the trained design quality prediction model to predict the architectural design quality of the target prediction object, thereby obtaining more complete design behavior data, providing important overall features based on statistical analysis, and providing a method for analyzing design quality based on machine learning. Thus, the problems of incomplete design behavior data, lack of analysis methods for overall features, and lack of methods for analyzing the correlation between design behavior and design quality are solved.

[0020] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0022] Figure 1 A flowchart of a method for predicting building design quality according to an embodiment of the present invention;

[0023] Figure 2 An example diagram of a BIM design behavior feature recognition and design quality prediction method framework provided according to an embodiment of the present invention;

[0024] Figure 3 An example diagram of raw data format conversion and time alignment fusion provided according to an embodiment of the present invention;

[0025] Figure 4 A flowchart of behavior feature extraction according to an embodiment of the present invention;

[0026] Figure 5 A design quality prediction flow chart provided according to an embodiment of the present invention;

[0027] Figure 6 This is a diagram of an application architecture of a model quality prediction system provided according to an embodiment of the present invention;

[0028] Figure 7 A block diagram of a device for predicting building design quality according to an embodiment of the present invention;

[0029] Figure 8 It is a schematic diagram of the structure of an electronic device provided according to an embodiment of the present invention.

[0030] Explanation of the reference numerals: 10 - architectural design quality prediction device; 20 - electronic device; 201 - acquisition module; 202 - generation module; 203 - extraction module; 204 - training module; 301 - memory; 302 - processor; 303 - communication interface. DETAILED DESCRIPTION

[0031] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following describes the architectural design quality prediction method, device, electronic device and storage medium of the embodiment of the present invention with reference to the accompanying drawings. In view of the problems mentioned in the above background technology, such as incomplete design behavior data, lack of analysis methods for overall characteristics, lack of methods for analyzing the correlation between design behavior and design quality, the present invention provides a method for predicting architectural design quality. In the method, a design behavior data set is generated by acquiring log data and real-time record data about architectural design behavior of multiple designers; a design behavior feature vector of each design behavior data sample is extracted from the behavior data set, and test data is generated based on these design behavior feature vectors and design quality quantitative evaluation data, a design quality prediction model is constructed based on a machine learning algorithm, and the design quality prediction model is trained using the test data, and the architectural design quality of the target prediction object is predicted using the trained design quality prediction model, so as to obtain more complete design behavior data, provide important overall characteristics based on statistical analysis, and provide a method for analyzing design quality based on machine learning. Thus, the problems of incomplete design behavior data, lack of analysis methods for overall characteristics, lack of methods for analyzing the correlation between design behavior and design quality are solved.

[0033] Specifically, Figure 1 A schematic flow chart of a method for predicting building design quality provided by an embodiment of the present invention.

[0034] like Figure 1 As shown, the building design quality prediction method includes the following steps:

[0035] In step S101, log data and real-time record data of architectural design behaviors of multiple designers are obtained.

[0036] Among them, the acquisition method is based on BIM software, and the log data is directly extracted from the log data storage location set by the software; the real-time recording data is based on the open program interface of the BIM software. While the software is running, the designer's modification of the model, keyboard input and other behaviors are captured to obtain real-time recording data of the corresponding behaviors.

[0037] It can be understood that the embodiment of the present invention obtains log data and real-time record data of multiple designers' architectural design behaviors based on BIM software, and the log data can be directly obtained by extracting from the log data storage location set by the software; through the open program interface of the BIM software, while the software is running, the designer's modification of the model, keyboard input and other behaviors are captured, and the real-time record data of the corresponding behaviors can be obtained.

[0038] In the embodiment of the present invention, the real-time recorded data includes keyboard key operation data recorded in real time when the designer is working, data of creating, modifying and deleting components in the architectural design model, and execution command data.

[0039] It can be understood that the real-time recorded data in the embodiment of the present invention includes the keyboard key operation data recorded in real time when the designer is working, the data of creating, modifying and deleting components in the architectural design model, and the execution command data. The real-time collection method is based on the BIM software open program interface. While the software is running, it captures the designer's modification of the model, keyboard input and other behaviors by subscribing to specific events, obtains the data of the corresponding behaviors, and then outputs them in a structured manner to an external data file.

[0040] In step S102, a design behavior data set is generated according to the log data and real-time record data of a plurality of designers, wherein the design behavior data set includes a design behavior data sample of each designer.

[0041] It is understandable that the embodiment of the present invention can integrate the above-obtained log data and real-time record data into a data set, in which the design behavior data samples of each designer are included. The specific generation method will be described in detail below and will not be repeated here.

[0042] In an embodiment of the present invention, a design behavior data set is generated based on the log data and real-time record data of multiple designers, including: converting the log data and real-time record data into target format data; generating a design behavior data sample based on the target format data of each designer, and generating a design behavior data set based on the design behavior data sample of each designer.

[0043] The target format is a standard format set according to actual needs, which is not specifically limited here. The data is organized into target format data to facilitate data fusion.

[0044] It can be understood that the embodiment of the present invention first converts log data and real-time record data into target format data to facilitate subsequent data fusion, and then uses the obtained target format data to generate design behavior data samples, and generates a design behavior data set based on the design behavior data samples of each designer.

[0045] In an embodiment of the present invention, the target format includes a timestamp, a record type, and record details, and a design behavior data sample is generated based on the target format data of each designer, including: time aligning all data in the target format data according to the timestamp; and fusing all the time-aligned data to obtain the design behavior data sample.

[0046] The timestamp is the time when the data is recorded. Its original format may change with the user's region, system region, and language settings. It is converted into the time expressed in ticks for time alignment comparison.

[0047] It can be understood that in the process of converting log data and real-time record data into target format data in the embodiment of the present invention, since the original format of the timestamp may change with the user's region, the system's region and language settings, converting the timestamp in the data into time expressed in ticks can facilitate time alignment comparison.

[0048] In the embodiment of the present invention, before converting the log data and the real-time record data into target format data, it also includes: performing data cleaning and data screening on the log data and the real-time record data.

[0049] Among them, data cleaning refers to the integrity and consistency of the data. If the content of a real-time record data is incomplete or erroneous, the data will be discarded. The consistency of the execution command related records in the real-time record data and the related records in the log will be checked. If it is found that the time and command content of the real-time record are wrong and inconsistent with the log, these data will be removed; data screening refers to screening based on certain characteristics that log records related to designer behavior in the log usually have. For example, the data in the log is screened based on the characteristic that the beginning of the line of the design behavior record always contains the string "Jrn" or "jrn".

[0050] It is understandable that before converting log data and real-time record data into target format data, the embodiment of the present invention also needs to clean the data and the integrity and consistency of the data, and screen the data based on certain characteristics that log records related to designer behavior in the log usually have, so as to obtain complete, consistent and relevant data.

[0051] In step S103, a design behavior feature vector of each design behavior data sample in the design behavior data set is extracted, and test data is generated according to the design behavior feature vector of each design behavior data sample and the design quality quantitative evaluation data.

[0052] Among them, the design behavior feature vector is a feature vector that can be used for machine learning model training or prediction; the design quality quantitative evaluation data is based on the existing design quality evaluation method, and the data is obtained by quantitatively evaluating the BIM design quality corresponding to each design behavior fusion data file.

[0053] It can be understood that the embodiments of the present invention can extract the design behavior feature vector of each design behavior data sample in the design behavior data set, that is, convert a complete design behavior data sample into a feature vector that can be used for machine learning model training or prediction, and generate test data based on the design behavior feature vector and design quality quantitative evaluation data.

[0054] In an embodiment of the present invention, a design behavior feature vector of each design behavior data sample in a design behavior data set is extracted, including: identifying data corresponding to each statistical indicator in each design behavior data sample; calculating data density and time density of data corresponding to each statistical indicator in the corresponding design behavior data sample, wherein data density represents the ratio of the number of data corresponding to the statistical indicator to the total number of records in the sample, and time density represents the ratio of the total time occupied by the data corresponding to the statistical indicator to the total time span of the sample; and generating a design behavior feature vector according to the data density and time density.

[0055] It can be understood that the embodiments of the present invention obtain the data of the indicator by identifying the data corresponding to each statistical indicator in each design behavior data sample, such as the number of data records after the flag command is executed, the total number of graphics elements added in the real-time record, etc., and then divide the data by the total data length of each sample and the total time span of the sample to obtain the data density feature and the time density feature. The data density represents the ratio of the number of data corresponding to the statistical indicator to the total number of records in the sample, and the time density represents the ratio of the total time occupied by the data corresponding to the statistical indicator to the total time span of the sample. The design behavior feature vector is generated based on the two.

[0056] In step S104, a design quality prediction model is constructed based on a machine learning algorithm, the design quality prediction model is trained using test data, and the trained design quality prediction model is used to predict the architectural design quality of the target prediction object.

[0057] Among them, the design quality prediction model based on machine learning algorithm adopts Extremely RandomTrees, an extreme random tree algorithm; the test data consists of two parts, one is the design behavior feature data obtained in the previous step, and the other part needs to be based on the existing design quality evaluation method.

[0058] It can be understood that the embodiment of the present invention constructs a design quality prediction model based on a machine learning model of an extreme random tree algorithm, uses the design behavior feature vector obtained in the previous step and the design quality quantitative evaluation data obtained based on an existing design quality evaluation method, trains the design quality prediction model with the design behavior feature vector, and uses the trained design quality prediction model to predict the architectural design quality of the target prediction object.

[0059] In an embodiment of the present invention, the design quality prediction model is trained using test data, including: dividing the test data into a training set and a validation set; performing data augmentation on the training set, training the design quality prediction model based on the training set, and validating the design quality prediction model based on the validation set, until the training stop condition is met.

[0060] The data augmentation adopts the slicing method. By setting a sliding window, each fused data file is divided into several parts with N data records, and the step length between the previous and the next part is set to s, so that the two parts will overlap partially, and the length of the overlapping part is n=Ns. The corresponding quality score of each newly generated data sample is consistent with the original sample. This method introduces two hyperparameters, and it is necessary to use the control variable method to conduct a model effect comparison experiment. The data reliability and model effect are used as indicators to determine the optimal set of hyperparameter values, which are not specifically limited here.

[0061] It can be understood that the embodiment of the present invention divides the test data into a training set and a validation set, and uses the slicing method to perform data augmentation on the training set so that the model can obtain better results. After the training set is augmented, the design quality prediction model is trained according to the training set, and the design quality prediction model is verified according to the validation set until the training stop condition is met.

[0062] According to the architectural design quality prediction method proposed in the embodiment of the present invention, log data and real-time record data about architectural design behaviors of multiple designers are obtained to generate a design behavior data set, a design behavior feature vector of each design behavior data sample is extracted from the behavior data set, and test data is generated based on the two. A design quality prediction model is constructed based on a machine learning algorithm and the test data is used to train the design quality prediction model. The trained design quality prediction model is used to predict the architectural design quality of the target prediction object, so that more complete design behavior data is obtained, important overall features based on statistical analysis are given, and a method for analyzing design quality based on machine learning is provided.

[0063] The following is a further description of the architectural design quality prediction method through a specific embodiment. The BIM design behavior feature analysis and quality prediction method proposed in this embodiment extracts design behavior features, builds a machine learning model for predicting design quality, and finally realizes the prediction of design quality. Its framework is as follows: Figure 2 As shown, the details are as follows:

[0064] Design behavior data collection and preprocessing This study adopts a design behavior data collection method that combines logs and real-time collection. On the one hand, it makes full use of the data in the log that is closely related to the design behavior. On the other hand, based on the secondary development technology of BIM software, it records the data of keyboard key presses and pop-ups in real time when the designer is working, and tracks and records the relevant data of creating, modifying, and deleting components. The collection of log files adopts the existing technical solutions and directly extracts them from the log file storage location set by the software. The real-time collection method is based on the open program interface of BIM software. While the software is running, it captures the designer's modification of the model, keyboard input and other behaviors by subscribing to specific events, obtains the corresponding behavior data, and then outputs it to an external data file in a structured manner. The real-time collected data includes the following: the time when each model component is created, modified, or deleted, the id, quantity, type, and other information of the components involved; the time when each key on the keyboard is pressed and popped up, and the name of the key involved; the relevant records of executing commands, including command names and command function descriptions.

[0065] The log data and real-time record data will be stored in two data files respectively, and paired one by one according to the designer and the design time. Next, the real-time record data needs to be cleaned, and the basis for cleaning is the integrity and consistency of the data. If the content of a real-time record data is incomplete or erroneous, then this data must be discarded; the consistency of the execution command related records in the real-time record data and the related records in the log is checked. If the time and command content of the real-time record are found to be wrong and inconsistent with the log, these data are removed. After the data cleaning is completed, according to the time information of each operation record in the two paired files, each data record from the two data files is aligned and integrated in chronological order and merged into one data file.

[0066] Data fusion mainly includes three steps: relevance screening, format conversion and time alignment fusion of each specific data record from log files and real-time record files.

[0067] 1) Relevance screening

[0068] Since the log files automatically generated by BIM software contain a lot of information that is not related to design behavior, such as computer system configuration, software configuration information, etc., the log data needs to be filtered. The log records related to the designer's behavior usually have certain characteristics, and filtering is based on this feature. For example, the data in the log is filtered based on the characteristic that the beginning of the design behavior record always contains the string "Jrn" or "jrn".

[0069] 2) Format conversion

[0070] In this step, each piece of data in the log and real-time record is converted into a standard format to facilitate data fusion. A piece of data in a standard format includes a timestamp, a record type, and a record details. The timestamp is the time when the data is recorded. Its original format may change with the user's region, the system's region, and the language settings. Therefore, timestamps in different formats need to be uniformly converted into time expressed in ticks for time alignment comparison. The time expressed in ticks is the number of milliseconds from 0:0:0:0:0:0 on January 1, 0001 to this time. The record type uses a word as a label to describe the type of the data. For example, "Command" indicates that the data record is a command information, and "AccelKey" indicates that the data record is a shortcut key command. After the record type, use the record details to record other important information related to the data, including but not limited to the specific name and function of the executed command, the specific key combination and function of the shortcut key command, etc.

[0071] Figure 3 A specific example of format conversion and time alignment fusion is given. The figure shows the information related to the standard format in the log file and the real-time record file. The detailed information of the command recorded in the log shows that the command of this data record is "cancel the current operation", while the record type of the real-time record data refers to "components are created and modified", and the details list the number information of the relevant components.

[0072] 3) Time alignment fusion

[0073] By comparing the tick time of each record in the paired data files, sorting them according to the chronological order, the data from the two sources are merged.

[0074] Behavior feature extraction The behavior feature extraction method proposed in this embodiment can convert a complete design behavior data sample into a feature vector that can be used for machine learning model training or prediction. The features used are obtained by statistical methods to obtain the overall density of various types of behavior record data in the sample, which is mainly divided into two categories: data density and time density. The extraction process is as follows: Figure 4 Data density refers to the ratio of the number of records of a certain type to the total number of records in the sample, while time density refers to the ratio of the total time occupied by such records to the total time span of the sample.

[0075] To calculate the density data, it is necessary to first count the total data length (L) of each sample and the total time span (T) of the sample. This embodiment provides the following 15 indicators directly obtained from the sample statistics, as shown in Table 1. Table 1 is a corresponding table of the 15 indicators obtained by statistics. After obtaining the data of these indicators, they are divided by L and T respectively to obtain the data density features and time density features as the extracted behavioral data features. Among them, the "effective time" statistical indicator does not calculate the data density, so there are a total of 29 features. When generating feature vectors, these 29 features are arranged first in data density and then in time density, and the features calculated by each statistical indicator are arranged in sequence in each type of density.

[0076] Table 1

[0077]

[0078] The meaning of these statistical indicators and the specific classification of statistical methods are explained below.

[0079] Indicator 1: In BIM design behavior, there are multiple consecutive commands that are executed before the model is substantially modified. If only the number of executed commands is counted, it is not enough to describe the modeling results brought about by the design behavior. Therefore, this statistical indicator is set. In practical applications, it is necessary to combine the log formats of different software, analyze the characteristics of data records that can mark the completion of command execution, and then count the number of these records. For example, for Autodesk Revit software, the record data with "Transaction Successful" in the log record indicates that a modification to the BIM model has been successfully executed, and such data can mark the completion of command execution.

[0080] Indicator 2-6: Directly count the corresponding recorded data, which comes from real-time records.

[0081] Indicators 7-11: These five statistical indicators are related to the execution of commands. According to the different ways of executing commands, the "Number of menu bar command operations" and "Number of shortcut key commands" are set. When counting the "Number of times a command is canceled or withdrawn", various different methods need to be considered. For example, the software provides two ways to cancel the command, namely, through shortcut key commands and menu bar commands. Therefore, the cancellation behaviors using these two methods need to be counted. The "Number of button presses" mainly counts the number of times users click buttons in menu bars, dialog boxes, etc. when executing commands.

[0082] Indicators 12-15: These four statistical indicators are related to time. The time interval refers to the time difference between the timestamps of two adjacent data records. Time intervals of different lengths need to be counted separately, mainly because time intervals of different lengths represent different behavioral characteristics. For example, a short interval may mean that the designer is engaged in thinking activities, while a long time interval may mean an inefficient working state, interruptions by other things, etc. The effective working time refers to the total time span of the sample minus the sum of the duration of time intervals of more than 5 minutes. Dividing the effective time by the total time span of the sample gives the proportion of the effective time in the total time, which may reflect whether the designer's working time is complete, whether he can stay focused for a long time, and other characteristics. The data density calculated using the effective time statistic does not have obvious practical significance, so it is not used as a data feature.

[0083] The test data used to construct the machine learning model for design quality prediction model construction consists of two parts: one is the design behavior feature data obtained in the previous step, and the other part needs to be based on the existing design quality assessment method to quantitatively evaluate the BIM design quality corresponding to each design behavior fusion data file to obtain the design quality quantitative assessment data. The machine learning model uses the Extremely Random Trees algorithm, which divides the test data into a training set and a validation set, takes the design behavior features as input, and takes the quality quantitative assessment data as the expected output.

[0084] Under the condition of sufficient data, it is only necessary to use the existing machine learning model training and testing methods to obtain a model that can evaluate and predict the design quality based on the design behavior characteristics. However, considering that it is not easy to obtain quantitative evaluation data for BIM design quality, there are small samples in practical applications. Usually, the small sample situation refers to the number of original data samples being less than 10 times the number of features, which is 290 in this embodiment. Even if this quantity requirement is met, the use of data augmentation strategies can also enable the model to achieve better results.

[0085] Data augmentation uses the slicing method (Window cropping or slicing). By setting a sliding window, each fused data file is divided into several parts with N data records, and the step length between the previous and next parts is set to s, so that the previous and next parts will overlap partially, and the length of the overlapping part is The quality score of each newly generated data sample is consistent with that of the original sample. The data augmentation method introduces two hyperparameters, and it is necessary to use the control variable method to conduct a model effect comparison experiment, using data reliability and model effect as indicators to determine the optimal set of hyperparameter values.

[0086] Design quality prediction method The process of design quality prediction method is as follows Figure 5 As shown, it mainly includes three steps: design behavior data collection, data preprocessing and feature extraction, and model prediction.

[0087] Before designers who need to conduct quality prediction work, they need to install data collection software in the BIM system they use. According to the needs of regular quality prediction work, data files are regularly collected from the computers used by designers and stored in a dedicated database.

[0088] 2) Data preprocessing and feature extraction

[0089] In order to reduce the randomness of the prediction results of a single sample, the collected data files need to be divided into multiple samples, each of which contains at least one complete session, and then the above methods are used to preprocess the data and extract feature data to form feature vectors. If the amount of collected data is insufficient, the above methods need to be used to augment the data to obtain more samples before subsequent operations.

[0090] 3) Model prediction

[0091] Input the feature values ​​of all samples into the trained model, and the model will give a predicted quality score. Averaging the scores of all samples will give a comprehensive prediction value of the design quality based on the design behavior data at this stage.

[0092] Model quality prediction system application architecture The application architecture of the entire system is as follows Figure 6 As shown in the figure, it includes two major subsystems: model acquisition and quality prediction. The purpose of the model acquisition subsystem is to obtain a trained machine learning model, and export and store the model so that it can be loaded and used in actual application scenarios; the purpose of the quality prediction subsystem is to realize model quality prediction for specific evaluation objects in specific scenarios. The system mainly includes the following four functional modules: data collection module, data processing module, model training module, model saving and loading module, and quality prediction module.

[0093] Data collection module The data collection module is used to collect the required design behavior data from the data source of model training and the object of quality prediction, using the aforementioned design behavior data collection method. According to the different BIM software used by designers, different corresponding software or plug-in systems need to be developed to implement this module. The module also needs to have the function of temporarily storing the collected data on the computer used by the designer, and can store the data files of different designers for different projects separately according to a certain file hierarchy management method.

[0094] Data processing module The data processing module mainly implements the data preprocessing and feature extraction functions for the original data. The main functions of the module are: a) access the original data storage medium and obtain the specified original data; b) apply the above method to preprocess the original data, determine whether data augmentation is needed according to user needs and data volume, and if necessary, execute the data augmentation strategy and output the hyperparameter value introduced by augmentation; c) apply the above method to realize feature extraction and output feature vector.

[0095] The model training module data training module applies the above method to train the machine learning model based on the feature vector data and the design quality quantitative evaluation data. This module can randomly divide the training set and the validation set according to the proportion, and output the root mean square error and correlation coefficient of the model on the validation set to evaluate the effect of the machine learning model.

[0096] Model saving and loading module The model saving and loading module needs to export and save the trained machine learning model, and provide the function of reloading the specified model. For example, based on the pickle package in Python, the model can be exported and saved as a .pkl file, and then loaded when needed.

[0097] Quality Prediction Module The quality prediction module can receive feature vector input, call the machine learning model, get the predicted result and return it. In the case of multiple feature vectors input at a time, the module will get the result corresponding to each feature vector and then return their average as the comprehensive result.

[0098] The operation of this embodiment is further described below by substituting the data source:

[0099] This embodiment uses the process of students completing modeling assignments in these courses as the data source: The Civil Engineering Department of a certain university has opened several courses related to the application of computer technology in the AEC field from 2022 to 2024. Four of these courses focused on the principles of BIM and BIM modeling methods based on Autodesk Revit software, and set up Revit modeling assignments to evaluate students' learning and practice effects of related content.

[0100] The design behavior data collection and preprocessing of this embodiment are as follows: the real-time data recording uses a plug-in developed based on the secondary development interface provided by Autodesk Revit, which records data and exports data files in txt format while students are modeling. The specific development idea is to monitor the changes of the model and the execution of commands by subscribing to the events provided by the software. Once the model is changed or a command is executed, the relevant information is further obtained to realize the recording of component creation, modification, deletion data and command data; based on the existing keyboard hook technology, the monitoring and recording of keyboard behavior is realized.

[0101] The raw data obtained from the homework files uploaded by the students were cleaned and fused to obtain a fused data set. The fused data set contains the design behavior data of 68 students. The data of each student contains several data files, and each data file corresponds to a complete Revit session process. The total amount of data of different students varies greatly. The average total length of a single fused data file is 58045 lines, and the maximum value is close to 350000. In view of this data characteristic, the present embodiment adopts the slicing method for data augmentation. On the one hand, this can reduce the errors that may be caused by excessive differences in data length, and on the other hand, it can further expand the amount of data to ensure the effect of machine learning model training. However, this method will introduce two new hyperparameters, the length N of the data obtained by slicing and the slicing step size s. The optimal values ​​of these two hyperparameters need to be determined through subsequent experiments.

[0102] This embodiment conducts comparative experiments on the effects of different machine learning algorithms, conducts machine learning model experiments and optimizations, proves the advantages of using the extreme random tree algorithm proposed in the embodiment of the present invention, and determines the optimal values ​​of the slicing method hyperparameters through experiments.

[0103] The data set used in the experiment to compare the effects of different machine learning algorithms is a behavioral feature data set obtained through data augmentation and feature extraction steps, and is randomly divided into a training set and a test set at 80% / 20%. Considering that the data augmentation step introduces two hyperparameters, all algorithms in the model selection experiment phase are trained with the same set of hyperparameters. Combined with the analysis of the fusion data set, a set of hyperparameter values ​​needs to be fixed in the experiment. The experiment uses Python's Sklearn software package to traverse the available candidate supervised machine learning regression algorithms for modeling training. The expected output of the model is the score corresponding to the sample, and the mean square error of the output score is used as the loss function. The trained model uses the root mean square error and The model is evaluated on the test set and the best model is selected. The best model will be used in subsequent hyperparameter optimization experiments to determine the appropriate hyperparameter values.

[0104] The Sklearn package provides a large number of different available machine learning algorithms. This study uses the following categories of algorithms as candidates: linear algorithms (including ordinary least squares, ridge regression, least absolute shrinkage and selection operators, elastic networks, etc.), support vector machines, stochastic gradient descent, nearest neighbor algorithms, Gaussian processes, decision trees, and ensemble methods (including adaptive boosting, bootstrap aggregation, random forests, etc.).

[0105] The evaluation model effect is mainly based on RMSE (Root Mean Square Error) and correlation coefficient ( ) two indicators. In the specific context of this embodiment, the root mean square error is used to describe the average error between the score value output by the model and the actual true score value, and The value can indicate the proportion of the variance that the model can explain to the total variance of the sample data. The smaller the RMSE value, The closer the value is to 1, the better the model effect is.

[0106] Among the 42 algorithms included in the experiment, the 6 algorithms with the best results are shown in Table 1, which is a comparison table of machine learning algorithm results. According to the experimental results, the extreme random tree algorithm has the best effect.

[0107] Table 1

[0108]

[0109] The experiment of slicing method hyperparameter optimization adopts the control variable method, that is, fix one parameter, change another parameter, and observe the change of the effect of the same model. In addition, in order to ensure the effect of the model, the value of the hyperparameter cannot make the total number of samples too small.

[0110] fixed The experimental results of changing s are shown in Table 2. Table 2 is the experimental results of changing s with fixed N. The experimental results show that increasing the step size will significantly reduce the number of samples and have a significant impact on the effect of model training. When s is 0, the number of samples is only 190, and it is difficult to achieve good results. On the other hand, reducing s can improve the effect of the model, but when the value of s is too small (for example, 500), the number of samples becomes relatively large. On the one hand, the improvement of the model effect becomes relatively limited. On the other hand, the number of samples increases by nearly 70 times, resulting in a high similarity between a large number of samples, which may cause overfitting and lead to low reliability of the results.

[0111] Table 2

[0112]

[0113] The experimental results of controlling the value of step length s and changing the value of N are shown in Table 3. Table 3 is the experimental results of changing N with fixed s. Controlling the value of step length s to be fixed can maintain the number of samples to a certain extent without causing a large change. The number of samples was not too low in the experiment. When N decreases, although the number of samples increases, the effect of the algorithm decreases. This may be because each sample is too short, resulting in insufficient information, and it is more susceptible to extreme situations, resulting in increased differences between data. When , the model is already good enough and further improvement is no longer obvious.

[0114] Table 3

[0115]

[0116] Based on the above experimental results, for the same N value, the smaller the step size s, the better, but too small a value may lead to a decrease in the reliability of the model; and for the same s value, the larger the N, the better the model effect, but too large a value of N will not bring about a significant improvement in effect. A hyperparameter N of 30,000 is a more appropriate value, and s of 2,000 can ensure a good model effect without causing a significant decrease in model reliability.

[0117] The best model effect is achieved by using the extreme random tree algorithm to build a machine learning model under the optimal hyperparameter value. The RMSE value on the test set is 2.29, which is significantly smaller than the standard deviation of the fusion data set score of 7.61, and the R-square value reaches 0.88. Considering that the range of the score data is 38 points, the relative error is only 6%, which shows that the model has a good ability to explain the data and has the ability to accurately predict the design quality based on behavioral characteristics.

[0118] Next, the architectural design quality prediction device proposed in the embodiment of the present invention will be described with reference to the accompanying drawings.

[0119] Figure 7 It is a block diagram of a device for predicting building design quality according to an embodiment of the present invention.

[0120] like Figure 7 As shown, the building design quality prediction device 10 includes: an acquisition module 201, a generation module 202, an extraction module 203 and a training module 204.

[0121] Among them, the acquisition module 201 is used to obtain the log data and real-time record data of multiple designers about the architectural design behavior; the generation module 202 is used to generate a design behavior data set based on the log data and real-time record data of multiple designers, wherein the design behavior data set includes a design behavior data sample of each designer; the extraction module 203 is used to extract the design behavior feature vector of each design behavior data sample in the design behavior data set, and generate test data according to the design behavior feature vector and design quality quantitative evaluation data of each design behavior data sample; the training module 204 is used to construct a design quality prediction model based on a machine learning algorithm, train the design quality prediction model with test data, and use the trained design quality prediction model to predict the architectural design quality of the target prediction object.

[0122] In an embodiment of the present invention, the generation module 202 is further used to: convert log data and real-time record data into target format data; generate design behavior data samples according to the target format data of each designer, and generate design behavior data sets according to the design behavior data samples of each designer.

[0123] In the embodiment of the present invention, the target format includes a timestamp, a record type and record details, and the generation module 202 is further used to: time-align all data in the target format data according to the timestamp; merge all the time-aligned data to obtain the design behavior data sample.

[0124] In the embodiment of the present invention, a processing module is further included, wherein the processing module is further used to: perform data cleaning and data screening on the log data and the real-time record data before converting the log data and the real-time record data into target format data.

[0125] In an embodiment of the present invention, the extraction module 203 is further used to: identify the data corresponding to each statistical indicator in each design behavior data sample; calculate the data density and time density of the data corresponding to each statistical indicator in the corresponding design behavior data sample, wherein the data density represents the ratio of the number of data corresponding to the statistical indicator to the total number of records in the sample, and the time density represents the ratio of the total time occupied by the data corresponding to the statistical indicator to the total time span of the sample; generate a design behavior feature vector according to the data density and the time density.

[0126] In an embodiment of the present invention, the training module 204 is further used to: divide the test data into a training set and a validation set; perform data augmentation on the training set, train the design quality prediction model based on the training set, and validate the design quality prediction model based on the validation set, until the training stop condition is met.

[0127] In the embodiment of the present invention, the real-time recorded data includes keyboard key operation data recorded in real time when the designer is working, data of creating, modifying and deleting components in the architectural design model, and execution command data.

[0128] It should be noted that the aforementioned explanation of the embodiment of the building design quality prediction method is also applicable to the building design quality prediction device of this embodiment, and will not be repeated here.

[0129] According to the architectural design quality prediction device proposed in the embodiment of the present invention, through the synergistic effect of the acquisition module, the generation module, the extraction module and the training module, it is achieved by acquiring the log data and real-time record data of multiple designers on architectural design behavior, generating a design behavior data set, extracting the design behavior feature vector of each design behavior data sample from the behavior data set, and generating test data based on the two, building a design quality prediction model based on a machine learning algorithm and using the test data to train the design quality prediction model, using the trained design quality prediction model to predict the architectural design quality of the target prediction object, obtaining more complete design behavior data, providing important overall characteristics based on statistical analysis, and providing a method for analyzing design quality based on machine learning.

[0130] Figure 8 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device may include:

[0131] A memory 301 , a processor 302 , and a computer program stored in the memory 301 and executable on the processor 302 .

[0132] When the processor 302 executes the program, the architectural design quality prediction method provided in the above embodiment is implemented.

[0133] Furthermore, the electronic device further comprises:

[0134] The communication interface 303 is used for communication between the memory 301 and the processor 302 .

[0135] The memory 301 is used to store computer programs that can be run on the processor 302 .

[0136] The memory 301 may include a high-speed RAM (Random Access Memory) memory, and may also include a non-volatile memory, such as at least one disk memory.

[0137] If the memory 301, the processor 302 and the communication interface 303 are implemented independently, the communication interface 303, the memory 301 and the processor 302 can be connected to each other through a bus and communicate with each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0138] Optionally, in a specific implementation, if the memory 301, the processor 302 and the communication interface 303 are integrated on a chip, the memory 301, the processor 302 and the communication interface 303 can communicate with each other through an internal interface.

[0139] The processor 302 may be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention.

[0140] An embodiment of the present invention further provides a computer-readable storage medium having a computer program or instruction stored thereon. When the computer program or instruction is executed, the above-mentioned architectural design quality prediction method is implemented.

[0141] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0142] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0143] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

[0144] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, the steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0145] A person of ordinary skill in the art may understand that all or part of the steps carried by the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the above-mentioned program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0146] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for predicting architectural design quality, characterized in that: The following steps are involved: Obtain log data and real-time record data of architectural design behaviors of multiple designers, wherein the real-time record data includes keyboard key operation data recorded in real time when the designer is working, data of creating, modifying and deleting components in the architectural design model, and execution command data; generating a design behavior data set according to the log data and the real-time record data of a plurality of designers, wherein the design behavior data set includes a design behavior data sample of each designer; Extracting the design behavior feature vector of each design behavior data sample in the design behavior data set, and generating test data according to the design behavior feature vector of each design behavior data sample and the design quality quantitative evaluation data, specifically including identifying the data corresponding to each statistical indicator in each design behavior data sample, calculating the data density and time density of the data corresponding to each statistical indicator in the corresponding design behavior data sample, and generating the design behavior feature vector according to the data density and the time density, wherein the data density represents the ratio of the number of data corresponding to the statistical indicator to the total number of records of the sample, and the time density represents the ratio of the total time occupied by the data corresponding to the statistical indicator to the total time span of the sample; A design quality prediction model is constructed based on a machine learning algorithm, the design quality prediction model is trained using the test data, and the trained design quality prediction model is used to predict the architectural design quality of a target prediction object.

2. The architectural design quality prediction method according to claim 1, characterized in that: The generating of the design behavior data set according to the log data and the real-time record data of the multiple designers includes: Converting the log data and the real-time record data into target format data; The design behavior data sample is generated according to the target format data of each designer, and the design behavior data set is generated according to the design behavior data sample of each designer.

3. The architectural design quality prediction method according to claim 2, characterized in that: The target format includes a timestamp, a record type, and record details, and the generating the design behavior data sample according to the data in the target format of each designer includes: Time-aligning all data in the target format data according to the timestamp; All the time-aligned data are fused to obtain the designed behavior data sample.

4. The architectural design quality prediction method according to claim 2, characterized in that: Before converting the log data and the real-time record data into target format data, the method further includes: The log data and the real-time recorded data are cleaned and screened.

5. The architectural design quality prediction method according to claim 1, characterized in that: The using the test data to train the design quality prediction model includes: Dividing the test data into a training set and a validation set; Data augmentation is performed on the training set, the design quality prediction model is trained according to the training set, and the design quality prediction model is verified according to the verification set, until the training stop condition is met, the training is stopped.

6. A device for predicting building design quality, characterized in that: include: An acquisition module is used to acquire log data and real-time record data of architectural design behaviors of multiple designers, wherein the real-time record data includes keyboard key operation data recorded in real time when the designer is working, data of creating, modifying and deleting components in the architectural design model, and execution command data; A generating module, configured to generate a design behavior data set according to the log data and the real-time recording data of a plurality of designers, wherein the design behavior data set includes a design behavior data sample of each designer; An extraction module is used to extract the design behavior feature vector of each design behavior data sample in the design behavior data set, and generate test data according to the design behavior feature vector of each design behavior data sample and the design quality quantitative evaluation data, specifically including identifying the data corresponding to each statistical indicator in each design behavior data sample, calculating the data density and time density of the data corresponding to each statistical indicator in the corresponding design behavior data sample, and generating the design behavior feature vector according to the data density and the time density, wherein the data density represents the ratio of the number of data corresponding to the statistical indicator to the total number of records of the sample, and the time density represents the ratio of the total time occupied by the data corresponding to the statistical indicator to the total time span of the sample; The training module is used to construct a design quality prediction model based on a machine learning algorithm, train the design quality prediction model using the test data, and predict the architectural design quality of a target prediction object using the trained design quality prediction model.

7. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the architectural design quality prediction method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed, the architectural design quality prediction method according to any one of claims 1 to 5 is implemented.

Images