CN117540174B

CN117540174B - Building structure multi-source heterogeneous data intelligent analysis system and method based on neural network

Info

Publication number: CN117540174B
Application number: CN202410026224.3A
Authority: CN
Inventors: 钮慧娟; 汤东婴; 魏晓斌; 陆斌; 汪晟; 王若晨
Original assignee: Jiangsu Testing Center For Quality Of Construction Engineering Co ltd
Current assignee: Jiangsu Testing Center For Quality Of Construction Engineering Co ltd
Priority date: 2024-01-09
Filing date: 2024-01-09
Publication date: 2024-03-26
Anticipated expiration: 2044-01-09
Also published as: CN117540174A

Abstract

The invention discloses a building structure multi-source heterogeneous data intelligent analysis system and method based on a neural network, which are used for determining a building to be detected and acquiring a super matrix of the building to be detected and monitoring data of at least two different modes; preprocessing the monitoring data one by one; aiming at multi-mode monitoring data, a multi-channel data processing module is constructed, and each mode at least comprises two channels; outputting the trained multichannel data processing module as a building multi-mode data analysis unit; collecting monitoring data of a building to be detected within a preset time, processing the monitoring data by adopting a building multi-mode data analysis unit, obtaining building risk information and a building risk fusion result output by each channel, and updating a super matrix of the building to be detected; and giving a building risk analysis result based on the updated super matrix. According to the invention, through the combination of the physical model and the neural network, a more accurate analysis result is given, and the calculation efficiency and accuracy are improved.

Description

Building structure multi-source heterogeneous data intelligent analysis system and method based on neural network

Technical Field

The invention relates to a building quality monitoring technology, in particular to a system and a method for intelligent analysis of multi-source heterogeneous data of a building structure based on big data analysis and a neural network.

Background

Building structures are important places for human life and work, and the safety and reliability of the building structures are directly related to the life and property safety of people. Therefore, the monitoring and evaluation of the building structure is an important engineering task, and aims to discover and prevent diseases and risks of the building structure in time and ensure the normal operation and use of the building structure.

However, conventional building structure monitoring and assessment methods have a number of problems and challenges, such as: the building structure has complex and various forms and sizes, various sensors and devices are needed to collect data, such as images, sounds, vibration, temperature, humidity and the like, and the acquisition of the data requires a great deal of manpower and material resources and time cost, and is easy to be interfered by environment and devices and error. Because various data sources are different and the format difference is very large, the data is cleaned, fused, analyzed, mined and the like, and the data is characterized by large quantity, high dimension, poor quality, dynamic, unstructured and the like, so that the processing of the data requires a large amount of calculation resources and algorithm support, and the validity and the accuracy of the data are difficult to ensure. The purpose of monitoring and evaluating building structures is to provide valuable information and knowledge, such as states, properties, diseases, risks, etc. of building structures, which information and knowledge extraction and expression requires the combination of professional domain knowledge and experience, such as structural mechanics, material science, normative standards, etc., which knowledge and experience acquisition and application requires a great deal of expert involvement and human intervention, which is time consuming and laborious.

To solve the above problems, the skilled person has developed methods such as the invention patent CN116596321a filed by the applicant. However, due to the characteristics of large building monitoring data volume and complex coupling relation, the problems of poor interpretability of the neural network prediction result, huge computing resources required by a physical model and the like are caused, and the efficiency and quality of building monitoring are affected.

Disclosure of Invention

The invention aims to provide a system and a method for intelligently analyzing multi-source heterogeneous data of a building structure based on a neural network, which are used for solving at least one technical problem existing in the prior art.

According to one aspect of the application, the intelligent analysis method for the multi-source heterogeneous data of the building structure based on the neural network comprises the following steps:

s1, determining a building to be detected, and acquiring a supermatrix of the building to be detected and monitoring data of at least two different modes; preprocessing the monitoring data one by one to enable the monitoring data to reach a preset input standard, and forming a training set and a testing set;

step S2, constructing a multi-channel data processing module aiming at multi-mode monitoring data, wherein each mode at least comprises two channels; training the multi-channel data processing module by adopting a training set, testing by adopting a testing set, and outputting the trained multi-channel data processing module as a building multi-mode data analysis unit;

S3, collecting monitoring data of the building to be detected within a preset time, processing the monitoring data by adopting a building multi-mode data analysis unit, obtaining building risk information and a building risk fusion result output by each channel, and updating a super matrix of the building to be detected;

and S4, giving out a construction risk analysis result based on the updated super matrix, and giving out a final risk prompt and suggestion together with a construction risk fusion result.

According to one aspect of the application, the step S1 is further:

s11, determining a building to be detected and a peripheral building affecting the building to be detected according to a monitoring target, and constructing a super matrix of the building to be detected and the peripheral building according to a pre-configured super matrix parameter; the supermatrix is obtained by carrying out Hadamard multiplication operation on a superelliptic matrix, a pose transformation matrix, a materialized parameter matrix and a risk coefficient matrix;

step S12, collecting monitoring data of a monitoring instrument, and classifying according to the type of the monitoring data to form monitoring data of at least two modes, wherein the monitoring data comprise vibration, stress, deformation, corrosion and temperature; the monitoring data at least come from a mechanical sensor and a camera;

step S13, aiming at the monitoring data of each type of mode, preprocessing the monitoring data one by one to enable the monitoring data to reach a preset input standard, and forming a training set and a testing set, wherein the preprocessing comprises denoising, time-frequency conversion and feature extraction;

Step S14, the monitoring data serving as a training set and a testing set are configured in a memory.

According to one aspect of the present application, the step S2 is further:

step S21, aiming at the monitoring data of each mode, constructing a data processing channel, wherein the monitoring data of each mode comprises at least two channels to form a multi-mode multi-channel data processing module; the method comprises the steps that a crack region extraction channel and a corrosion texture extraction channel are arranged for image data from a camera; setting a stress extraction channel and a vibration extraction channel aiming at time sequence data from a mechanical sensor;

s22, constructing a plurality of cross-modal building attention modules according to the number of channels to form a multi-channel data processing module;

and S23, calling a training set and a testing set in a memory to train and test the multi-channel data processing module, and outputting the multi-channel data processing module meeting the expected performance standard as a building multi-mode data analysis unit.

According to one aspect of the present application, the step S3 is further:

step S31, determining a monitoring time interval and a duration according to the risk level and the monitoring requirement of the building to be detected, and regularly collecting multi-mode monitoring data of the building;

S32, inputting the collected monitoring data into a building multi-mode data analysis unit to obtain the output of a risk identification channel of each mode, namely obtaining building risk information of each mode;

step S33, combining the cross-modal attention module to fuse the building risk information of each mode to obtain a building risk fusion result, and reflecting the overall risk condition of the building;

and step S34, updating a materialized parameter matrix and a risk coefficient matrix in the super matrix of the building to be detected according to the building risk information of each mode, and adjusting the structural characteristics, environmental factors and risk grades of the building.

According to one aspect of the present application, the step S4 is further:

s41, based on the updated super matrix, calling a preconfigured finite element simulation calculation module to calculate the building risk, and giving a building risk analysis result;

step S42, comparing the building risk fusion result with the risk analysis result, checking the consistency of the building risk fusion result and the risk analysis result, judging whether the error is in a preset range, and if not, carrying out data analysis and checking;

step S43, according to the risk analysis result and the risk fusion result, giving a final risk prompt and corresponding advice, wherein the risk prompt comprises no risk, low risk, medium risk and high risk; the advice includes continuing monitoring, reinforcing inspection, and emergency repair.

According to an aspect of the application, the step S13 further includes:

step S131, reading monitoring data of each type of mode, and performing time alignment on data of the mechanical sensor and the camera;

step S132, acquiring monitoring data of at least two mechanical sensors, extracting mutation points, and extracting at least two sections of monitoring data with preset lengths according to the mutation points to form risk period monitoring data;

and step S133, preprocessing the risk period monitoring data and the data of the corresponding cameras to form a training set and a testing set.

According to an aspect of the present application, in the step S21, the processing procedure of extracting the channel through the crack region and the etched texture specifically includes:

step S211, obtaining feature graphs of at least two scales after feature extraction;

step S212, constructing sub-channels with different scales of crack region extraction channels and corrosion texture extraction channels according to the preset scale number;

step S213, a fusion unit for constructing a crack region extraction channel and a corrosion texture extraction channel comprises an up-sampling unit, a fusion unit and a classification unit, wherein the up-sampling unit up-samples a low-resolution feature map to the size of a high-resolution feature map by using a bilinear interpolation method; the fusion unit carries out weighted fusion on the feature images with different scales by using a weighted average method to obtain a comprehensive feature image; the classification unit uses a 1 multiplied by 1 convolution layer and a softmax layer to carry out pixel-level classification on the comprehensive feature map, and a binary fracture segmentation map and a binary corrosion texture map are respectively obtained;

And step S214, performing optimization operation on the crack segmentation map by using a post-processing technology to obtain final crack segmentation and corrosion texture segmentation results, wherein the optimization operation comprises removing small areas, filling holes and smoothing edges.

According to an aspect of the application, the step S34 further includes:

s341, constructing a homotypic materialized parameter updating matrix and a risk coefficient updating matrix according to the characteristics of the materialized parameter matrix and the risk coefficient matrix in the super matrix;

step S342, building risk information of each mode is obtained, and a materialized parameter updating matrix and a risk coefficient updating matrix are generated;

and S343, replacing the original matrixes by adopting the materialized parameter updating matrix and the risk coefficient updating matrix to finish the updating process.

According to an aspect of the application, the step S41 is further:

step S411, acquiring an updated supermatrix, comparing the updated supermatrix with the supermatrix before updating, extracting columns of a superellipse matrix with physical and chemical parameters and risk coefficient changes larger than a threshold value in the supermatrix, and constructing a risk building component set;

step S412, clustering the risk building component set based on the pose transformation matrix to form at least one risk area building component graph structure;

Step S413, aiming at the building component graph structure of each risk area, arranging the building component graph structures in descending order according to the change percentage of the risk coefficient to obtain a building component risk attenuation chain;

and step S414, extracting building components based on the building component risk attenuation chain, forming a regional building component super matrix, calling a preconfigured finite element simulation calculation module to calculate building risks, and giving a building risk analysis result.

According to another aspect of the application, a system for intelligent analysis of heterogeneous data of a building structure based on a neural network, is characterized by comprising:

at least one processor; and

a memory communicatively coupled to at least one of the processors; wherein,

the memory stores instructions executable by the processor for execution by the processor to implement the neural network-based intelligent analysis method for building structure multi-source heterogeneous data according to any one of the above technical schemes.

The method has the beneficial effects that the accuracy and the robustness of building risk identification are improved through monitoring data of different modes; by updating the super matrix, the structural state and risk change of the building are dynamically reflected, and the timeliness and sensitivity of building risk analysis are improved; the risk analysis result is given out through the neural network, and the risk analysis result is given out by combining the neural network with the physical simulation model, so that the problem of poor interpretability of the neural network can be solved. Other technical advantages will be described below in connection with the detailed description.

Drawings

Fig. 1 is a flow chart of the present invention.

Fig. 2 is a flowchart of step S1 of the present invention.

Fig. 3 is a flow chart of step S2 of the present invention.

Fig. 4 is a flowchart of step S3 of the present invention.

Fig. 5 is a flowchart of step S4 of the present invention.

Detailed Description

Since the solutions are relatively complex, in order to highlight the content of the present application, the prior knowledge known or able to be known to the person skilled in the art is not described in the present application, reference being made to the applicant or to the co-ordinator's prior application patents, such as CN116596321a. In order to solve the problems of large data volume and high processing consumption resources in the prior art. The data are analyzed and processed manually, time and labor are consumed, a physical model, such as a simulation calculation unit based on a digital model, is used for calculation, and for a building model with a large volume, a large calculation resource is needed, so that the simulation calculation speed is low. The neural network is adopted for processing, and because the neural network belongs to a black box model, the interpretability is relatively poor, and the traceability and theoretical analysis cannot be performed. Based on the above, the technical scheme of the application is provided. The problems of large data processing capacity and poor analysis and calculation interpretability are solved through the novel technical scheme.

As shown in fig. 1, according to an aspect of the present application, a method for intelligently analyzing multi-source heterogeneous data of a building structure based on a neural network includes the following steps:

Aiming at the limitation that the building risk analysis method in the prior art relies on single-mode data and manual analysis, the monitoring data of different modes are utilized, and the efficiency and reliability of building risk analysis are improved; the real-time and dynamic problems of building risk analysis are solved, and real-time monitoring and dynamic assessment of building risks are realized through quick update of the super matrix; in other words, through the supermatrix based on the superellipse model, the local part can be updated without reading all information of the supermatrix, the structural state and risk change of the building are dynamically reflected, and the timeliness and the sensitivity of building risk analysis are improved. The calculation and updating efficiency of the super matrix is greatly improved, and a new solution is provided for more efficient and accurate risk analysis.

In this embodiment, by means of the patent technology previously applied by the applicant, the model and the data are decoupled based on the superellipse model, and then an automatic extraction module of risk and physicochemical data is established, that is, the physicochemical parameters and risk parameters of the building are extracted through the neural network, and then the partial matrix in the supermatrix is updated without adjusting the whole model information. Therefore, the speed of information processing is greatly increased. On the basis of the calculation, parallel calculation can be performed, one data processing path is through the calculation of multiple channels and multiple modes of a neural network, and on the basis of physical and chemical parameters and risk information, overall risk information of the building is given. And secondly, the super matrix model of the building is corrected based on the intermediate calculation result, and then the risk information of the building is given based on the physical model in a finite element calculation mode, so that the interpretation is better, the possible problems of intermediate links can be found, and then the links such as input data, the intermediate calculation result, part of building component modules and the like are analyzed and traced. Therefore, the application provides a new thought of model optimization and acceleration calculation on the basis of the original patent, and provides a new method for intelligent analysis of multi-source heterogeneous data of a building structure. By analyzing the multi-source heterogeneous data, not only the risk analysis result can be given, but also the digital model of the building can be corrected. Based on the prior patent, the updated building super-ellipse model can be converted into common digital models such as cad and sw, so that simulation calculation is conveniently performed through the existing finite element calculation module, two intelligent analysis paths are provided for multi-source heterogeneous data, and the two paths can be mutually verified and compared, so that the reasons for the occurrence of problems are analyzed.

As shown in fig. 2, according to an aspect of the present application, the step S1 is further:

s11, determining a building to be detected and a peripheral building affecting the building to be detected according to a monitoring target, and constructing a super matrix of the building to be detected and the peripheral building according to a pre-configured super matrix parameter; the supermatrix is obtained by carrying out Hadamard multiplication operation on a superelliptic matrix, a pose transformation matrix, a materialized parameter matrix and a risk coefficient matrix; the super matrix is used for comprehensively reflecting structural features and risk states of the building and provides a basis for subsequent data analysis and risk assessment.

Step S12, collecting monitoring data of a monitoring instrument, and classifying according to the type of the monitoring data to form monitoring data of at least two modes, wherein the monitoring data comprise vibration, stress, deformation, corrosion and temperature; the monitoring data at least come from a mechanical sensor and a camera; from different angles and dimensions, the health condition and risk factors of the building are reflected, and data support is provided for subsequent data analysis and risk assessment.

Step S13, aiming at the monitoring data of each type of mode, preprocessing the monitoring data one by one to enable the monitoring data to reach a preset input standard, and forming a training set and a testing set, wherein the preprocessing comprises denoising, time-frequency conversion and feature extraction; the automatic feature extraction and fusion of the monitoring data are realized, and effective input is provided for subsequent data analysis and risk assessment.

Step S14, the monitoring data serving as a training set and a testing set are configured in a memory. And the monitoring data is conveniently called and updated, and a data source is provided for subsequent data analysis and risk assessment.

In summary, in this embodiment, monitoring data of different modes, such as sound waves, images, videos, stress, infrared, electromagnetism, and the like, can be utilized to improve accuracy and robustness of building risk identification; through construction of the super matrix, unified representation and calculation of the building are realized, and structural features and risk states of the building are reflected; the monitoring data is preprocessed to reach a preset input standard, so that the quality and usability of the data are improved; and finally, storing and managing the preprocessed monitoring data, and improving the safety and accessibility of the data.

In this embodiment, the above analysis is merely an example, and in different building monitoring, by arranging different types of sensors, expected data of various modes can be obtained, so as to provide input data with more dimensions, perform mutual verification, and extract deformation and corrosion information of a building through image and video data of a building captured by an unmanned aerial vehicle or the like, which are described in the above patent and are not described in detail herein. In other embodiments, there may be three, four, or more modalities of monitoring data.

As shown in fig. 3, according to an aspect of the present application, the step S2 is further:

step S21, aiming at the monitoring data of each mode, constructing a data processing channel, wherein the monitoring data of each mode comprises at least two channels to form a multi-mode multi-channel data processing module; the method comprises the steps that a crack region extraction channel and a corrosion texture extraction channel are arranged for image data from a camera; setting a stress extraction channel and a vibration extraction channel aiming at time sequence data from a mechanical sensor; the data processing channel is a module for carrying out specific processing and conversion on the monitoring data of each mode, can extract and enhance the effective information in the monitoring data, and inhibit and eliminate the ineffective information in the monitoring data, thereby providing optimized input for subsequent data fusion and analysis. The crack region and the corrosion texture in the image can be detected and positioned, the characteristics of the length, width, depth, direction, distribution and the like of the crack are extracted, and the characteristics of the color, shape, density, range and the like of the corrosion are extracted. And setting a stress extraction channel and a vibration extraction channel aiming at time sequence data from the mechanical sensor, wherein the stress extraction channel and the vibration extraction channel are respectively used for calculating and analyzing stress values and vibration frequencies in the time sequence data, extracting the characteristics of stress, change, extremum, trend and the like, and extracting the characteristics of vibration such as amplitude, period, waveform, energy and the like.

S22, constructing a plurality of cross-modal building attention modules according to the number of channels to form a multi-channel data processing module; the cross-modal attention module is used for realizing the neural network module for mutual learning and fusion among different modal data, can automatically distribute the weight of the different modal data by using an attention mechanism, highlights important information, suppresses redundant information and improves the expression capability and semantic consistency of the data. Each cross-modal attention module can receive data of two or more modes as input and output data of one or more modes as output, so that fusion and conversion of the data are realized. Such as crack-stress cross-modal attention module: and receiving data of the crack region extraction channel and the stress extraction channel as input, and outputting crack-stress fusion data as output, so that mutual learning and fusion between cracks and stress are realized, and the crack detection precision and the stress analysis effect are improved. Corrosion-vibration cross-modal attention module: and receiving data of the corrosion texture extraction channel and the vibration extraction channel as input, and outputting corrosion-vibration fusion data as output, so that mutual learning and fusion between corrosion and vibration are realized, and the identification accuracy of corrosion and the evaluation effect of vibration are improved.

And S23, calling a training set and a testing set in a memory to train and test the multi-channel data processing module, and outputting the multi-channel data processing module meeting the expected performance standard as a building multi-mode data analysis unit. The training set and the testing set in the memory are used for adjusting and verifying parameters such as weight, bias, activation function and the like of the multi-channel data processing module respectively, so that the multi-channel data processing module can effectively extract and fuse characteristics of different mode data, output high-quality data and meet expected performance standards such as accuracy, recall rate, F1 value and the like. The training and testing result can output a multichannel data processing module meeting the expected performance standard as a building multi-mode data analysis unit, and can be used as a core module for subsequent data analysis and risk assessment to realize multi-dimensional, multi-level and multi-angle analysis of building risks

In summary, in this embodiment, different data processing channels are set for monitoring data of different modes, so as to effectively extract building risk indexes such as cracks, stress, corrosion, vibration and the like. The attention mechanism is utilized to automatically allocate the weight of the data in different modes, highlight important information, inhibit redundant information and improve the expression capacity and semantic consistency of the data.

It should be noted that, the cross-mode attention module may be configured independently, that is, after the intermediate result is obtained, the cross-mode attention module is input into a parallel module to extract the coupling relationship between the modes, thereby obtaining the cross-mode data analysis result. The system architecture of the neural network has various design modes, so that various topology architectures can be formed, and the system architecture is not particularly limited, so long as the functions can be realized, and a person skilled in the art can combine and configure the system architecture according to actual situations, thereby realizing the functions.

As shown in fig. 4, according to an aspect of the present application, the step S3 is further:

the monitoring time interval and the duration refer to the frequency and the range of multi-mode monitoring of the building to be detected, and can be determined according to the risk level and the monitoring requirement of the building to be detected so as to ensure the validity and the sufficiency of monitoring data. The risk level refers to the risk level of a building to be detected, and can be evaluated according to a risk coefficient matrix in a super matrix, wherein the higher the risk level is, the shorter and longer the monitoring time interval and duration are, and the risks are discovered and prevented in time. The monitoring requirement refers to the monitoring purpose and requirement of the building to be detected, and can be determined according to the type, function, position, environment and other factors of the building, and the more complex and strict the monitoring requirement is, the shorter and longer the monitoring time interval and duration are, so as to comprehensively and accurately monitor the state and risk of the building. The regular collection of the multi-mode monitoring data of the building means that different types of monitoring instruments are used for multi-mode monitoring of the building to be detected according to the determined time interval and duration, namely, the monitoring data of different modes such as stress, vibration, images, sound waves, infrared waves, electromagnetism and the like are collected simultaneously or sequentially to form a multi-mode monitoring data set, and a data source is provided for subsequent data processing and analysis.

For example, according to the risk coefficient matrix in the super matrix, the risk level of the building to be detected is evaluated, and if the risk level is moderate, the monitoring time interval can be set to be once a day, and the duration is one hour, so as to ensure the validity and sufficiency of the monitoring data. According to factors such as type, function, position, environment of building, confirm the monitoring demand, assume that the monitoring demand is for detecting structural integrity and stability of building, then can set up the time interval of monitoring to once a day, duration is an hour to monitor state and risk of building comprehensively and accurately. According to the determined time interval and duration, different types of monitoring instruments are used for carrying out multi-mode monitoring on the building to be detected, namely, monitoring data of different modes such as stress, images, sound waves, infrared rays, electromagnetism and the like are collected simultaneously or sequentially to form a multi-mode monitoring data set, for example, the following monitoring instruments and methods can be used: the acoustic wave sensor is used for transmitting and receiving acoustic wave signals, and monitoring data of acoustic wave modes are collected and reflect internal defects of a building, such as holes, cracks, delamination and the like. Electromagnetic sensors are used for measuring the electromagnetic field of a building, and monitoring data of electromagnetic modes are collected to reflect metal corrosion of the building, such as corrosion, potential, resistance and the like. And extracting corrosion, deformation and other conditions by adopting image data.

the building multi-mode data analysis unit is a neural network module formed by a multi-channel data processing module, can automatically extract and fuse multi-mode monitoring data, outputs high-quality data, and realizes multi-dimensional, multi-level and multi-angle analysis of building risks. Various risk indicators for a building include cracks, stress, corrosion, vibration, and the like.

Step S33, combining the cross-modal attention module to fuse the building risk information of each mode to obtain a building risk fusion result, and reflecting the overall risk condition of the building; the attention mechanism is utilized to automatically allocate the weight of the data in different modes, highlight important information, inhibit redundant information and improve the expression capacity and semantic consistency of the data. And the building risk fusion result reflects the overall risk condition of the building, such as a crack-stress fusion result and a corrosion-vibration fusion result.

And step S34, updating a materialized parameter matrix and a risk coefficient matrix in the super matrix of the building to be detected according to the building risk information of each mode, and adjusting the structural characteristics, environmental factors and risk grades of the building. The physical and chemical parameter matrix refers to a matrix reflecting physical properties of a building and environmental conditions, such as density, elastic modulus, temperature, humidity, and the like. The risk factor matrix refers to a matrix reflecting the risk level and risk factors of the building, such as cracks, stress, corrosion, vibration, etc. According to the building risk information of each mode, the materialized parameter matrix and the risk coefficient matrix are modified and adjusted, so that the structural state and risk change of the building can be dynamically reflected, and the structural characteristics, environmental factors and risk level of the building are adjusted.

After training is completed, the training data is configured in a server or an edge server, and processing operation is carried out according to the newly acquired data, so that a specific analysis result is given.

As shown in fig. 5, according to an aspect of the present application, the step S4 is further:

the finite element simulation calculation module can carry out numerical simulation and calculation on structural performance and risk level of a building according to a physical and chemical parameter matrix and a risk coefficient matrix in the super matrix, the building can be divided into a plurality of finite elements by utilizing a finite element method, a finite element equation set is established, the equation set is solved, states of stress, deformation, damage and the like of the building are obtained, and the risk degree of the building is estimated. And according to the physicochemical parameter matrix and the risk coefficient matrix in the updated super matrix, the updated super matrix is used as the input of a finite element simulation calculation module, a preconfigured finite element simulation calculation module is called, finite element simulation calculation is carried out, a building risk analysis result is obtained, and the structural performance and risk level of the building such as stress, deformation, damage and the like are reflected.

the construction risk fusion result is data output by the construction multi-mode data analysis unit and reflects overall risk conditions of the construction, such as a crack-stress fusion result and a corrosion-vibration fusion result. The risk analysis result refers to data output by the finite element simulation calculation module, and reflects structural performance and risk level of the building, such as stress, deformation, damage and the like. And then checking and analyzing the source, processing, output and other processes of the data, finding out the abnormality and error of the data, and correcting and optimizing the abnormality and error.

And comparing the construction risk fusion result with the risk analysis result, and analyzing whether the data of the construction risk fusion result and the risk analysis result are consistent, such as the position, the size, the shape, the depth, the size, the change, the extreme value, the trend and the like of the crack, the color, the shape, the density, the range, the vibration amplitude, the period, the waveform, the energy and the like of corrosion.

And (3) checking the consistency of the two, judging whether the error is within a preset range, for example, setting an error threshold value, such as 5%, calculating the difference and proportion of the data of the two, such as the length of a crack, the size of stress and the like, judging whether the error is within the error threshold value, if so, judging that the data of the two are consistent, and if not, judging that the data of the two are inconsistent, and carrying out data analysis and check.

Data analysis and verification are performed to find out anomalies and errors of the data, correction and optimization are performed, for example, checking sources of the data, such as accuracy, stability, calibration conditions and the like of a monitoring instrument, checking data processing, such as parameters, algorithms, logic and the like of a data processing channel, checking data output, such as format, range, units and the like of the data, finding out anomalies and errors of the data, such as noise, missing, outliers, error values and the like of the data, and correcting and optimizing, such as filtering, interpolation, rejection, conversion and the like of the data.

According to the analysis result of the neural network and the simulation result of the neural network combined with the physical model, the more accurate analysis result is given, and the analysis accuracy is improved.

According to an aspect of the application, the step S13 further includes:

step S131, reading monitoring data of each type of mode, and performing time alignment on data of the mechanical sensor and the camera; the monitoring data of each type of mode is data of different modes, such as sound waves, infrared rays, electromagnetism and the like, collected by different types of monitoring instruments. The data of the mechanical sensor and the camera refer to the data of mechanical modes and visual modes, such as stress, deformation, images and the like, collected by the mechanical sensor and the camera. After synchronization, the data of each pair of mechanical sensors and cameras have the same or similar time stamps to facilitate subsequent data processing and analysis.

Step S132, acquiring monitoring data of at least two mechanical sensors, extracting mutation points, and extracting at least two sections of monitoring data with preset lengths according to the mutation points to form risk period monitoring data; the monitoring data of the mechanical sensor are data of mechanical modes collected by the mechanical sensor, such as stress, deformation and the like. The abrupt points are points where abrupt changes occur in the monitored data, such as peaks, valleys, jumps, etc. The mutation point extraction means that mutation points are detected and extracted from the monitoring data by using a certain algorithm and a threshold value, such as wavelet transformation, kalman filtering, threshold value judgment and the like. According to the position of the mutation point, at least two pieces of monitoring data with preset length are extracted, namely, at least two pieces of data with preset length, such as 1 second, 10 hours and the like, are cut out from the monitoring data, and are used as risk period monitoring data to reflect the risk state of a building, such as cracks, damages and the like.

And step S133, preprocessing the risk period monitoring data and the data of the corresponding cameras to form a training set and a testing set. The dangerous period monitoring data refers to at least two pieces of monitoring data of a predetermined length extracted in step S132, reflecting the risk status of the building, such as cracks, damages, etc. The data corresponding to the camera is the data of the visual mode aligned in time in step S131, such as image and video, and the preprocessed data is randomly divided into a training set and a testing set according to a certain proportion, such as 8:2, wherein the training set is used for training the multi-channel data processing module, and the testing set is used for testing the performance of the multi-channel data processing module, such as accuracy, recall rate, F1 value and the like.

features include edges, textures, colors, etc. The different scales may be feature maps of different resolutions, such as a high resolution feature map and a low resolution feature map. For example, using certain methods and parameters, feature maps of different resolutions are obtained from the raw monitored data, such as using the output of different layers of a convolutional neural network as feature maps of different scales.

Step S212, constructing sub-channels with different scales of crack region extraction channels and corrosion texture extraction channels according to the preset scale number; the characteristic diagrams with different scales are used as inputs of the sub-channels with different scales by using a certain method and parameters, such as a convolution layer, a pooling layer, an activation layer and the like. For example, using a convolution layer, pooling layer, activation layer, etc., a high resolution feature map is taken as an input to a high resolution sub-channel, and a low resolution feature map is taken as an input to a low resolution sub-channel, to further extract and compress the features of the crack and corrosion textures, respectively.

The image information is extracted through multiple channels, and different characteristics are respectively obtained, so that the calculation speed is improved, and the accuracy of crack and corrosion areas is improved.

According to an aspect of the application, the step S34 further includes:

s341, constructing a homotypic materialized parameter updating matrix and a risk coefficient updating matrix according to the characteristics of the materialized parameter matrix and the risk coefficient matrix in the super matrix; and constructing a homotypic materialized parameter updating matrix and a risk coefficient updating matrix, namely, matrixes with the same dimension, type, range and other characteristics, and storing the updated materialized parameters and risk coefficients. The parameters of the physical and chemical parameter matrix comprise material elastic modulus, cross-sectional area, length and the like; parameters of the risk coefficient array include crack width, corrosion degree, load size and the like.

the building risk information includes crack segmentation maps, corrosion texture maps, stress distribution maps, and the like. The corresponding construction risk information may be obtained from each type of monitoring data using certain methods and parameters.

And S343, replacing the original matrixes by adopting the materialized parameter updating matrix and the risk coefficient updating matrix to finish the updating process. By the updating mode, the matrix can be updated rapidly, and the updating efficiency is improved greatly.

According to an aspect of the application, the step S41 is further:

By extracting the building structure of the local area, the whole building structure does not need to be input and simulated, and the calculation efficiency is greatly improved. In other words, the embodiment not only realizes the decoupling of the building digital model and the physicochemical parameters and risk information, but also can rapidly extract the areas with risks based on the characteristics of the hyper-elliptic model, and then calculate the areas, thereby reducing the calculated amount and improving the calculation efficiency. In other words, because the building is huge, when a risk occurs in one place, the influence on the area far away from the building is smaller, and in this case, important simulation can be performed on the risk area, especially on information such as vibration and stress. So that the calculation efficiency can be greatly improved. This is also an advantage of the model decoupling of the present application.

A simple case is given to illustrate the implementation of this example.

The pre-update and post-update supermatrices, denoted M0 and M1, respectively, are read from the database, assuming that their dimensions are mxn, i.e., there are n superelliptic matrices, each with M elements.

And calculating a difference matrix D=M1-M0 of the two supermatrixes, and taking an absolute value of each element to obtain |D|.

Summing each column of |d| results in an n-dimensional vector S representing the degree of variation of each super-elliptic matrix.

Setting a threshold tau, screening out indexes of super elliptic matrixes with variation greater than tau according to the value of S, and marking the indexes as a set I, namely I= { i|S _i >τ,i=1,2,…,n}。

From set I, the columns of the corresponding super-elliptic matrix are extracted from M1 to form a matrix R of mX I, where I represents the number of elements of set I. R is a set of risk building elements, and each column represents the characteristics and risk factors of a building element with higher risk.

Reading a set of risk building elements R from a database, and a pose transformation matrix T for each building element _i Where i is the index of the building element, T _i Is a 4 x 4 homogeneous matrix representing the position and attitude of the building element relative to a reference coordinate system.

For each building element, its pose transformation matrix T is utilized _i Calculate the center point coordinate Ci, i.e. C, in the reference coordinate system _i =T _i [0、0、0、1] ^T Wherein C _i Is a 4 x 1 vector, the first threeThe elements represent the x, y, z coordinates of the center point, with the last element being 1.

For all center point coordinates C _i A clustering algorithm, such as K-means, is used to divide them into categories, each category representing a risk area, containing several adjacent or nearby building elements.

For each risk area, a risk area building element graph structure is constructed, namely an undirected graph g= (V, E), wherein V is a vertex set, each vertex corresponds to one building element, E is an edge set, each edge represents a connection relationship between two building elements, and the weight of the edge can be defined by the distance or similarity between the two building elements.

Reading from the database the building element graph structure g= (V, E) for each risk area, and the percentage change P of the risk factor for each building element _i Where i is the index of the building element, P _i Is a real number indicating how much the risk factor of the building element has been increased after the update compared to before the update.

For each risk zone, the building element with the greatest percentage change in risk coefficient is selected from the graph structure G and denoted as v ₀ It is taken as the starting point of the risk attenuation chain.

From v ₀ Initially, building elements with a large percentage of risk factor variation are selected in sequence along the edges of graph structure G until no more building elements are available to be selected, or the length of the preset chain is reached, forming an ordered sequence of building elements, denoted as l= { v ₀ ,v ₁ ,…,v _k Where k is the length of the chain, v _i Is the ith building element, P _vi >P _vi+1 I=0, 1, …, k-1.L is the building element risk decay chain, representing the path of gradual decay of risk from v 0.

Reading building element risk decay chain l= { v for each risk area from the database ₀ ,v ₁ ,…,v _k -and super elliptic matrix E of each building element _i Where i is the index of the building element, E _i Is an m x 1 vector representing the physicochemical parameters and risk coefficients of the building element.

For each risk zone, according to its building element risk attenuation chain L, from the supermatrix M ₁ Extracting columns of the corresponding super elliptic matrix to form a matrix Q of m× (k+1), wherein the ith column of Q is E _vi-1 I=1, 2, …, k+1.Q is the regional building element supermatrix and represents the characteristics and risk level of the building elements within the risk region.

And calling a preconfigured finite element simulation calculation module, taking the regional building component super matrix Q as input, carrying out finite element analysis, calculating building risks, and giving a building risk analysis result. Finite element analysis is a numerical method that can be used to solve complex structural mechanical problems such as stress, strain, displacement, vibration, etc. The construction risk analysis results may include the following: overall risk assessment of risk areas, such as risk level, risk index, risk probability, etc. Local risk assessment of individual building elements within the risk area, such as maximum stress, maximum strain, maximum displacement, failure mode, etc. Risk impact analysis of risk areas, such as risk propagation paths, risk impact ranges, risk impact levels, and the like. Risk control suggestions of the risk areas, such as risk early warning, risk precaution, risk elimination, risk reduction and the like.

It should be noted that the above-mentioned embodiments are only examples, and those skilled in the art can design other implementations according to the disclosure of the present application.

at least one processor; and

a memory communicatively coupled to at least one of the processors; wherein,

The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to the specific details of the above embodiments, and various equivalent changes can be made to the technical solution of the present invention within the scope of the technical concept of the present invention, and all the equivalent changes belong to the protection scope of the present invention.

Claims

1. The intelligent analysis method for the multi-source heterogeneous data of the building structure based on the neural network is characterized by comprising the following steps of:

step S4, based on the updated super matrix, a construction risk analysis result is given, and a final risk prompt and suggestion are given together with a construction risk fusion result;

the step S1 is further:

step S14, configuring monitoring data serving as a training set and a testing set in a memory;

the step S3 is further:

Step S34, updating a physical and chemical parameter matrix and a risk coefficient matrix in a super matrix of the building to be detected according to building risk information of each mode, and adjusting structural features, environmental factors and risk grades of the building;

the step S34 further includes:

2. The intelligent analysis method for multi-source heterogeneous data of a building structure based on a neural network according to claim 1, wherein the step S2 is further:

3. The intelligent analysis method for multi-source heterogeneous data of a building structure based on a neural network according to claim 2, wherein the step S4 is further:

4. The intelligent analysis method for multi-source heterogeneous data of a building structure based on a neural network according to claim 3, wherein the step S13 further comprises:

5. The intelligent analysis method for multi-source heterogeneous data of a building structure based on a neural network according to claim 4, wherein in the step S21, the processing procedure for extracting the channel through the crack region and the corrosive texture extraction channel specifically comprises:

6. The intelligent analysis method for multi-source heterogeneous data of a building structure based on a neural network according to claim 4, wherein the step S41 is further:

7. The utility model provides a building structure multisource heterogeneous data intelligent analysis system based on neural network which characterized in that includes:

at least one processor; and

a memory communicatively coupled to at least one of the processors; wherein,

the memory stores instructions executable by the processor for execution by the processor to implement the neural network-based architectural structure multi-source heterogeneous data intelligent analysis method of any one of claims 1 to 6.