CN-122017018-A - Deep learning-based building crack scanning monitoring method

CN122017018ACN 122017018 ACN122017018 ACN 122017018ACN-122017018-A

Abstract

The invention discloses a deep learning-based building crack scanning monitoring method which comprises the following steps of collecting sound vibration response time sequence data of an outer facade of a building, constructing a sound vibration response time sequence data set, executing frequency spectrum analysis processing on the sound vibration response time sequence data set to generate a time-frequency feature set, constructing a multi-level space grid structure, calculating and generating the space grid feature set, mapping the time-frequency feature set, the space grid feature set and a scanning parameter set into a multi-domain input field set, constructing an improved ANO model, executing multi-domain attention operator processing to generate a coarse grid structure response feature field and a fine grid crack local feature field, establishing a cross-scale attention mechanism, modulating the fine grid crack local feature field to generate a crack risk field, executing threshold judgment and connectivity analysis to generate a crack region marking result. The invention realizes the multi-scale structure response fusion and improves the crack identification capability.

Inventors

MENG QIFENG
Fang Dao
DAI HONGDA
LIAO JINGHAO
ZHENG TIANLUN
REN YIWEI

Assignees

华东交通大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (8)

1. The building crack scanning monitoring method based on deep learning is characterized by comprising the following steps of: radiating sweep frequency sound vibration excitation signals to the outer facade of the building, collecting sound vibration response time sequence data of the outer facade of the building, and constructing a sound vibration response time sequence data set; performing spectrum analysis processing on the sound vibration response time sequence data set, constructing a time-frequency characteristic tensor based on a spectrum analysis result, and generating a time-frequency characteristic set; Constructing a multi-level space grid structure based on the measuring point position records of the acoustic vibration response time sequence data set, attributing each measuring point to grid cells in the multi-level space grid structure, calculating the vibration energy characteristic and the statistical characteristic of each grid cell, and generating a space grid characteristic set; Mapping the time-frequency characteristic set into a time-frequency characteristic field input field, mapping the space grid characteristic set into a space heat map field input field, and combining the space grid characteristic set into a multi-field input field set; constructing an improved ANO model, taking a multi-level space grid structure as an operator acting domain, executing multi-domain attention operator processing, defining a coarse grid operator and a fine grid operator, and generating a coarse grid structure response characteristic field and a fine grid crack local characteristic field; Establishing a cross-scale attention mechanism based on a multi-level space grid structure, modulating a fine grid crack local feature field by using a coarse grid structure response feature field, and generating a crack risk field; And based on the crack risk field, performing threshold judgment and connectivity analysis on the crack risk value corresponding to each grid unit in the crack risk field, and generating a building facade crack region marking result.
2. The deep learning-based building crack scanning monitoring method according to claim 1, wherein the construction of the acoustic vibration response time series data set comprises: Setting a sound source parameter subset, a radar parameter subset and a scanning control parameter subset, and respectively performing numerical range setting, step resolution setting and unit unified processing on the three subsets to generate a scanning parameter set; Based on geometric outline information and a scanning mode of the outer facade of the building, acquiring a scanning angle and a scanning speed from a scanning control parameter subset, generating a preset scanning path covering a detection area under a coordinate system of the outer facade of the building, and determining a scanning track point sequence and a corresponding time sequence of a laser vibration detection radar to form preset scanning path data; Based on the scanning parameter set and preset scanning path data, controlling a directional sound source to radiate sweep-frequency sound vibration excitation signals to the outer facade of the building, and controlling a laser vibration measuring radar to sequentially acquire sound vibration response time sequence data corresponding to a time sequence, so as to generate original sound vibration response time sequence data; And carrying out time synchronization arrangement and measuring point number arrangement on the original sound vibration response time sequence data to construct a sound vibration response time sequence data set.
3. The deep learning-based building crack scanning monitoring method according to claim 1, wherein the generating of the time-frequency feature set comprises: Performing spectral analysis processing on the acoustic vibration response time sequence data set, wherein the spectral analysis processing comprises Fourier transformation, time-frequency transformation, noise estimation, effectiveness judgment and power spectrum calculation; Recombining and stacking the spectrum analysis result set according to the measuring point dimension, the time dimension and the frequency dimension to construct a time-frequency characteristic tensor; And extracting the data segments corresponding to each measuring point in the time-frequency characteristic tensor as measuring point level characteristic units, and organizing to form a time-frequency characteristic set.
4. The deep learning based building crack scanning monitoring method of claim 1, wherein the generating of the spatial grid feature set comprises: Based on the measuring point position record in the acoustic vibration response time sequence data set, the polar coordinate measuring point coordinates output by the laser vibration measuring radar are paired with the image rectangular coordinates of the outer facade of the building one by one, coordinate conversion is carried out on the polar coordinate measuring point coordinates, measuring point space coordinates are obtained, and the measuring point space coordinates are arranged to generate a measuring point space coordinate set; dividing coarse grids and fine grids according to preset grid sizes under a building outer elevation coordinate system corresponding to the measuring point space coordinate set, respectively setting coarse grid indexes and fine grid indexes, and constructing a multi-level space grid structure covering the building outer elevation; and attributing the time-frequency characteristics corresponding to each measuring point in the time-frequency characteristic set to the grid cells corresponding to the coarse grid and the fine grid according to the position of the measuring point space coordinate in the multi-level space grid structure, calculating the vibration energy characteristics and the statistical characteristics of the grid cells based on the time-frequency characteristics corresponding to the measuring points attributing to the grid cells, and organizing to generate a space grid characteristic set.
5. The deep learning based building crack scanning monitoring method of claim 1, wherein the generating of the multi-domain input field set comprises: Based on a time-frequency characteristic set, a measuring point space coordinate set and a multi-level space grid structure, determining measuring points belonging to each coarse grid and each fine grid in the multi-level space grid structure, performing weighted summation and normalization processing on time-frequency characteristics of the belonging measuring points in the time-frequency characteristic set according to a preset aggregation rule, writing processing results into corresponding grid units, and forming a time-frequency characteristic domain input field after all grid units are written; The method comprises the steps of reading vibration energy characteristics and statistical characteristics corresponding to grid cells from a space grid characteristic set in grid cells in a multi-level space grid structure, encoding the vibration energy characteristics and the statistical characteristics into space characteristic vectors of the grid cells, writing the space characteristic vectors into the corresponding grid cells, and forming a space heat map input field after writing all the grid cells; performing numerical normalization processing and encoding processing on the scanning parameter set, arranging each encoded scanning parameter into a scanning parameter vector, copying the scanning parameter vector into an operator core physical modulation vector corresponding to the grid cell in a multi-level space grid structure, and arranging to generate a scanning parameter modulation field; Based on the time-frequency characteristic domain input field, the space heat map domain input field and the scanning parameter modulation field, the three input fields are spliced and organized into a multi-domain input field set.
6. The deep learning based building crack scanning monitoring method of claim 1, wherein the construction and use of the improved ANO model comprises: based on the multi-domain input field set and the multi-level space grid structure, respectively setting input channels for the multi-domain input field set on grid nodes aligned with the multi-level space grid structure, carrying out channel mapping and alignment on the three input channels on the multi-level space grid structure, constructing a multi-domain operator input structure of the improved ANO model, setting a multi-domain attention operator structure, and taking the multi-level space grid structure as an operator acting domain; In the improved ANO model, a multi-domain operator input structure is taken as input, a time-frequency operator branch is constructed on an input channel corresponding to a time-frequency characteristic domain input field, a space operator branch is constructed on an input channel corresponding to a space heat map domain input field, an attention weight calculation unit is arranged, and attention weights of grid nodes in the time-frequency operator branch and the space operator branch are modulated in a combined mode based on an operator core physical modulation vector; on coarse grids and fine grids of the multi-level space grid structure, configuring coarse grid operators and fine grid operators for modulated time-frequency operator branches and space operator branches respectively based on coarse grid indexes and fine grid indexes; and performing coarse grid operator operation to generate a coarse grid structure response characteristic field, and performing fine grid operator operation to generate a fine grid crack local characteristic field.
7. The deep learning based building crack scanning monitoring method of claim 1, wherein the generation of the crack risk field comprises: Designating a coarse grid index for each coarse grid node in the coarse grid structure response characteristic field, designating a fine grid index for each fine grid node in the fine grid crack local characteristic field, determining the coarse grid node to which each fine grid node belongs according to the space inclusion relation between the coarse grid index and the fine grid index, and generating a coarse and fine grid mapping relation set; based on the coarse and fine grid mapping relation sets, reading coarse grid structure response characteristics of each coarse grid node and fine grid crack local characteristics of fine grid nodes corresponding to the coarse grid nodes, calculating cross-scale correlation metrics, executing normalization processing, and generating a cross-scale attention weight set; And constructing a cross-scale attention mechanism on the multi-level space grid structure, performing weighted combination on the coarse grid structure response characteristic of the affiliated coarse grid node and the fine grid crack local characteristic of the fine grid node by using the cross-scale attention weight on each fine grid node, generating a crack risk value of the fine grid node, and arranging to generate a crack risk field.
8. The method for scanning and monitoring the cracks of the building based on the deep learning according to claim 1, wherein the generation of the marking result of the crack area of the outer facade of the building comprises the following steps: Based on the distribution of the crack risk fields on the multi-level space grid structure, reading the crack risk value of each grid unit from the crack risk fields, executing threshold judgment, marking the crack candidate grid units and the non-crack grid units, organizing marking results of the crack candidate grid units and the non-crack grid units, and generating a crack candidate marking matrix; Based on the crack candidate marking matrix and the multi-level space grid structure, determining the space adjacent relation among grid cells, performing connectivity analysis on the crack candidate grid cells, identifying a crack connected region set formed by the crack candidate grid cells which are adjacent in space, calculating the grid cell number and the crack risk value accumulation sum in the crack connected region, eliminating the grid cell number and the crack risk value accumulation sum which are lower than a threshold value, and generating an effective crack connected region set; Based on the effective crack communication region set, grid cells belonging to the same effective crack communication region are aggregated into a single crack region, and an outer contour extraction operation is carried out on the fine grid cell set in each crack region to generate a corresponding crack region contour; Calculating a crack area estimation result of each crack area, generating a crack area estimation result set, correspondingly storing a crack area contour and a crack area estimation result set and an effective crack communication area set, and generating a building facade crack area marking result; And carrying out color coding on the crack risk values in the crack risk field to generate a building facade crack risk heat map, superposing and displaying a crack region contour and a crack area estimation result set, and outputting a building crack scanning monitoring report containing the crack risk heat map, the crack region contour and the crack area estimation result.

Description

Deep learning-based building crack scanning monitoring method Technical Field The invention relates to the field of building outer elevation detection, in particular to a building crack scanning monitoring method based on deep learning. Background Crack detection of building facades usually depends on manual inspection, close-range photographing evidence obtaining or a crack identification method based on two-dimensional images. The manual inspection is limited by experience of inspection staff, continuous and systematic scanning identification of the whole condition of a large-scale outer elevation is difficult, and the problems of missed inspection risk and subjectivity exist. The traditional image recognition method only carries out deduction according to the appearance texture information, cannot reflect response characteristics of cracks on a structural layer, is sensitive to illumination change, surface pollution and material difference, and has high misjudgment rate. In addition, the detection method based on the single image mode lacks the acquisition capability of structural vibration behaviors, and early cracks or hidden cracks are difficult to accurately identify. Although the existing acoustic detection and vibration detection technology can be used for identifying structural diseases, the existing acoustic detection and vibration detection technology generally relies on discrete sampling point measurement, the acquisition mode is scattered, and the continuous scanning capability of the whole area of the outer facade of the building is lacking. Meanwhile, the traditional vibration signal analysis method is more in a feature extraction level of a time domain, a frequency domain or a time-frequency domain, so that the multi-source response features and the spatial distribution information are difficult to deeply fuse, and unified characterization of the structural response and the local abnormal features with different scales cannot be established. In addition, the existing multi-scale analysis method generally adopts a fixed resolution or single-scale network structure, and the association relationship between the coarse-scale structure change and the fine-scale local crack in the complex outer vertical surface is underutilized. In the fusion processing of multisource vibration response data and space image coordinates, the prior art generally lacks an overall modeling method for structural states of all areas of a building facade, and particularly lacks a trans-scale analysis mechanism capable of synchronously deducing crack risks based on thick and thin scale structural response. When the existing deep learning model processes the task of fusion of the vibration response of the outer facade and the spatial characteristics, the problems of insufficient sensitivity to physical parameters, weak coupling of the spatial structure, inaccurate identification of a local high-risk area and the like exist, and the application of the scanning type crack monitoring technology in engineering practice is limited. Therefore, how to provide a building crack scanning monitoring method based on deep learning is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a building crack scanning monitoring method based on deep learning. The invention constructs a multi-level space grid structure, introduces a multi-domain input field, uniformly codes time-frequency characteristics, space grid characteristics and scanning physical parameters of sound vibration response, realizes synchronous modeling of coarse grid structure response and fine grid crack local characteristics through an improved ANO model, and generates a crack risk field based on a cross-scale attention mechanism, thereby realizing continuous scanning identification of building facades. The invention fully utilizes the structural response information of the acoustic vibration excitation and the laser vibration detection radar, introduces innovative structures such as physical modulation, multi-domain operator and cross-scale fusion, and the like, can realize the fine detection of the risk of the complex facade crack, and has the advantages of high identification accuracy, large scanning range and strong physical consistency of the model. According to the embodiment of the invention, the method for scanning and monitoring the building cracks based on deep learning comprises the following steps: radiating sweep frequency sound vibration excitation signals to the outer facade of the building, collecting sound vibration response time sequence data of the outer facade of the building, and constructing a sound vibration response time sequence data set; performing spectrum analysis processing on the sound vibration response time sequence data set, constructing a time-frequency characteristic tensor based on a spectrum analysis result, and generating a time-frequency characteristic set; Constructing