Bridge structure erosion degree monitoring method based on deep learning

Abstract

The invention provides a deep learning-based bridge structure erosion degree monitoring method, which comprises the following steps of: for bridge steel samples of different types and parts, characterizing grain boundary structures and grain orientation microstructure characteristics by utilizing an electron back scattering diffraction method, and determining typical microstructure types of crystals; performing preprocessing and feature engineering processing on the obtained large sample data set, extracting chloride ion concentration and microstructure features, and constructing training data of an intercrystalline corrosion prediction model; according to the change rule of the chloride ion concentration and the typical microstructure type of the crystal, a bridge steel intergranular corrosion prediction model is constructed, and training and verification are carried out on a large sample data set; a wireless sensor network is arranged in a bridge steel service environment, chloride ion concentration and temperature and humidity environment parameters are collected in real time, and data needed by model input are obtained through data preprocessing and feature extraction.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for monitoring the degree of corrosion of a bridge structure based on deep learning. Background Art

[0002] When bridge steel is in service for a long time in a salt spray environment, chloride ions will be enriched on the surface of the steel and gradually penetrate into the intergranular space, resulting in aggravated intergranular corrosion. At present, conventional corrosion monitoring methods are difficult to reflect the development of intergranular corrosion in steel in real time and dynamically. How to establish a quantitative relationship between chloride ion concentration and the degree of intergranular corrosion, and to achieve continuous monitoring and early warning of intergranular corrosion based on deep learning algorithms, is a technical problem that needs to be solved urgently. The enrichment and diffusion process of chloride ions on the surface of steel is also difficult to observe directly, and the occurrence and development of intergranular corrosion is affected by grain boundary structure and grain orientation. Therefore, obtaining large sample data of intergranular corrosion under different chloride ion concentration conditions and establishing a reasonable mathematical model are the basis for applying deep learning algorithms for prediction. In order to overcome these technical difficulties, it is necessary to study a method that can monitor chloride ion concentration and its effect on intergranular corrosion in real time. Summary of the invention

[0003] The present invention provides a bridge structure corrosion degree monitoring method based on deep learning, which mainly includes:

[0004] According to the enrichment of chloride ions on the bridge surface, the electrochemical impedance spectroscopy method was used to obtain the distribution data of chloride ion concentration under different environmental conditions, and the change law of chloride ion concentration was obtained through Fourier transform analysis.

[0005] For bridge steel samples of different types and locations, the electron backscatter diffraction method is used to characterize the grain boundary structure and grain orientation microstructure characteristics, and determine the typical microstructure type of the crystal;

[0006] Conduct large-scale bridge steel exposure tests in a salt spray environment, regularly test the chloride ion concentration and intergranular corrosion on the surface and inside of bridge steel, and obtain a large sample data set with different chloride ion concentrations, environmental conditions and intergranular corrosion levels;

[0007] Preprocess and feature engineer the acquired large sample data set to extract chloride ion concentration and microstructure characteristics, and construct training data for the intergranular corrosion prediction model;

[0008] According to the variation of chloride ion concentration and typical crystal microstructure types, a prediction model for intergranular corrosion of bridge steel was constructed, and trained and verified on a large sample data set;

[0009] A wireless sensor network is deployed in the service environment of bridge steel to collect chloride ion concentration, temperature and humidity environmental parameters in real time, and obtain the data required for model input through data preprocessing and feature extraction;

[0010] The real-time collected data is input into the trained intergranular corrosion prediction model to dynamically predict the development trend and degree of intergranular corrosion inside the bridge steel. When the prediction result exceeds the preset threshold, an early warning signal is issued to provide a decision-making basis for the maintenance and replacement of bridge steel.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a method for monitoring the corrosion degree of bridge structures based on deep learning. The method analyzes the distribution of chloride ions and the microstructure of steel by electrochemical impedance spectroscopy and electron backscatter diffraction, obtains corrosion data by conducting large-scale exposure tests in a salt spray environment, and constructs a convolutional neural network prediction model. The present invention uses a wireless sensor network to collect environmental parameters in real time, inputs the data into a trained model, and dynamically predicts the development trend and degree of intergranular corrosion inside bridge steel. When the prediction result exceeds a preset threshold, an early warning signal is issued in time to provide a decision-making basis for the maintenance and replacement of bridge steel. The present invention realizes the accurate prediction and early warning of intergranular corrosion of bridge steel, effectively improves the safety and service life of the bridge structure, and is of great significance to the long-term service performance and maintenance management of bridge projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a method for monitoring the degree of corrosion of a bridge structure based on deep learning of the present invention.

[0014] Figure 2 It is a schematic diagram of a bridge structure corrosion degree monitoring method based on deep learning according to the present invention. DETAILED DESCRIPTION

[0015] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] like Figure 1-2 In this embodiment, a bridge structure corrosion degree monitoring method based on deep learning may specifically include:

[0017] S101. According to the enrichment of chloride ions on the bridge surface, the electrochemical impedance spectroscopy method was used to obtain the distribution data of chloride ion concentration under different environmental conditions, and the change pattern of chloride ion concentration was obtained through Fourier transform analysis.

[0018] The impedance data of different areas on the bridge surface are obtained according to preset measurement parameters, wherein the measurement parameters include frequency range, sampling interval and electrode type; the chloride ion concentration value of each area is calculated by the impedance data, and the degree of chloride ion enrichment is quantitatively evaluated; the chloride ion concentration value is subjected to time-frequency analysis to obtain the variation characteristics of the chloride ion concentration in time and space dimensions; the main frequency components are extracted from the variation characteristics, and a relationship function between the chloride ion concentration and environmental factors is established, wherein the environmental factors include temperature, humidity and pH value; the model parameters of the relationship function are estimated by the least squares method, and the model fit is evaluated by the R-square value; the relationship function is used to combine with meteorological forecast data to predict the distribution of chloride ion concentration in the future; if the predicted concentration exceeds a preset warning threshold, the potential corrosion risk area is marked.

[0019] Specifically, according to the preset measurement parameters, electrochemical impedance spectroscopy was used to obtain impedance data of different areas on the bridge surface. Platinum electrodes were used, the frequency range was 0.1Hz to 100kHz, and the sampling interval was 1 hour. The chloride ion concentration values ​​of each area were calculated through impedance data, and the degree of chloride ion enrichment was quantitatively evaluated. For the obtained chloride ion concentration values, a time-frequency analysis was performed by combining fast Fourier transform and wavelet transform to obtain the variation characteristics of chloride ion concentration in time and space dimensions. The main frequency components were extracted from the variation characteristics, and the relationship function between chloride ion concentration and environmental factors, including temperature, humidity, and pH value, was established by multivariate regression. The model parameters were estimated by the least squares method, and the model fit was evaluated by the R square value. Using the established multivariate regression function, combined with meteorological forecast data, ARIMA was used to predict the chloride ion concentration distribution in the next week, and the warning threshold was set to 0.4% (weight percentage). When the predicted concentration exceeded the threshold, the potential corrosion risk area was marked. Multiple measurement points were set on the bridge surface, and electrochemical impedance spectroscopy was measured using platinum electrodes. The frequency range was set to 0.1Hz to 100kHz, the sampling interval was 1 hour, and the measurement was continuous for 72 hours. The chloride ion concentration values ​​of each region were calculated using the equivalent circuit model through the measured impedance data. For example, at a certain measurement point, when the impedance value was 500Ω, the corresponding chloride ion concentration was 0.2% (weight percentage). The degree of chloride ion enrichment was quantitatively evaluated, and the areas with concentrations greater than 0.3% were marked as highly enriched areas. Next, the acquired chloride ion concentration time series data were subjected to time-frequency analysis. The data were converted to the frequency domain using fast Fourier transform, and the main frequency components were obtained as 12 hours and 24 hours, reflecting the diurnal variation of chloride ion concentration. Then, wavelet transform was applied, Morlet wavelet was selected, and the decomposition scale was set to 1-8 to obtain the concentration variation characteristics under different time scales. Based on the results of time-frequency analysis, a multivariate regression model was established, with independent variables including temperature, humidity and pH value, and the dependent variable being chloride ion concentration. The least squares method is used to estimate the model parameters, and the regression equation is C = 0.1 + 0.005T + 0.002H-0.02pH, where C is the chloride ion concentration, T is the temperature, H is the humidity, and pH is the acidity. The R square value of the model is 0.85, indicating that the fitting effect is good. Finally, combined with the weather forecast data (temperature, humidity, pH value) for the next week, these data are substituted into the regression equation to obtain preliminary prediction results. The ARIMA model is then used to adjust the prediction results in time series, and the model parameters are ARIMA (1,1,1). The warning threshold is set to 0.4%. When the predicted concentration exceeds this value, it is automatically marked as a potential corrosion risk area. For example, the predicted concentration of a certain area on the 5th day is 0.45%, which exceeds the threshold, and the system automatically generates a risk alert.

[0020] S102. For bridge steel samples of different types and locations, the electron backscatter diffraction method is used to characterize the grain boundary structure and grain orientation microstructure characteristics to determine the typical microstructure type of the crystal.

[0021] The diffraction patterns of samples at key parts of the bridge are obtained, and the diffraction patterns are collected by an electron backscatter diffractometer; the grain boundary structure data and the grain orientation data are extracted according to the diffraction patterns; the grain boundary angle distribution and the grain boundary length distribution are calculated for the grain boundary structure data, and if the grain boundary angle is less than a preset threshold, it is determined to be a small-angle grain boundary, otherwise it is determined to be a large-angle grain boundary; the small-angle grain boundary energy is calculated, and the large-angle grain boundary energy takes a fixed value; the grain orientation data is statistically analyzed to obtain the average grain size, shape factor and texture index; the grain boundary characteristics and grain characteristics are subjected to dimensionality reduction processing, and the grain boundary characteristics include angle distribution and energy distribution, and the grain characteristics include size, shape factor and texture index; the K-means clustering algorithm is applied to the feature data after dimensionality reduction, and the number of clusters of the K-means clustering algorithm is determined by a comprehensive evaluation of the silhouette coefficient and the Calinski-Harabasz index; the typical microstructure type of the steel sample is obtained according to the clustering results.

[0022] Specifically, according to the bridge structure diagram, 3 samples were taken from key locations such as the main beam, cross beam, and piers. The sample size was 10mm×10mm×5mm. The diffraction pattern of the sample was obtained using an electron backscatter diffractometer. The diffraction pattern was processed using TSLOIM software. The step size was set to 0.1μm and the scanning area was 100μm×100μm. The grain boundary structure and grain orientation data were extracted from the diffraction pattern. Based on the obtained grain boundary structure data, the grain boundary angle distribution and grain boundary length distribution were calculated to determine the ratio of large-angle grain boundaries to small-angle grain boundaries. At the same time, the grain boundary energy distribution was calculated. The Read-Shockley formula was used to estimate the small-angle grain boundary energy. The large-angle grain boundary energy was set to a fixed value of 0.5J / m 2. The grain orientation data were statistically analyzed, the average size and shape factor of the grains were calculated, the pole figures and anti-pole figures were drawn to characterize the grain orientation distribution, the directional distribution function was used to quantitatively describe the grain orientation distribution, and the texture index was calculated to characterize the orientation concentration. The grain boundary characteristics, including the angle distribution, energy distribution and grain characteristics, including the size, shape factor, and texture index, were reduced by principal component analysis, and then the K-means clustering algorithm was applied. The number of clusters was set to 3-5, and the silhouette coefficient and Calinski-Harabasz index were used to comprehensively evaluate the optimal number of clusters to obtain the typical microstructure type of the steel sample. Three 10mm×10mm×5mm steel samples were taken from the main beam, cross beam and pier of the bridge, and the diffraction patterns were obtained using an electron backscatter diffractometer with an accelerating voltage of 20kV. The TSLOIM software was set to a step size of 0.1μm, and a 100μm×100μm area was scanned to extract the grain boundary structure and grain orientation data. The calculation of the grain boundary angle distribution shows that the high-angle grain boundaries around 60° account for 65%, and the small-angle grain boundaries less than 15° account for 35%. In the grain boundary energy distribution, the small-angle grain boundary energy E is calculated by the Read-Shockley formula E=E0θ(A-lnθ), where the high-angle grain boundary energy, that is, the grain boundary energy when the tilt angle θ is close to 90°, is E0=0.5J / m 2 , A = 1-1 / 4ln(2π). A is a constant related to the material, which depends on the elastic modulus, stacking fault energy and other properties of the material. θ is the inclination angle of the grain boundary. The high-angle grain boundary energy is 0.5 J / m 2 . The average grain size is 5μm and the shape factor is 0.75. The pole figures and inverse pole figures are drawn using equal-area projection, and the directional distribution function is calculated using spherical harmonics expanded to the 16th order. The texture index J=7.2, indicating a high orientation concentration. After dimensionality reduction by principal component analysis, the first three principal components were selected as K-means clustering inputs, explaining 85% of the variance. The number of clusters was set to 3-5, and the silhouette coefficient and Calinski-Harabasz index were calculated. The results showed that both indicators were optimal when there were 4 categories, which were 0.68 and 152.3, respectively. Finally, the four typical microstructure types were determined to be fine-grained strong texture type, coarse-grained weak texture type, mixed grain type and abnormal grain type, with proportions of 40%, 30%, 20% and 10%, respectively.

[0023] Continuous sectioning and electron backscattering diffraction characterization were used in turn to obtain the three-dimensional microstructure information of bridge steel, reconstruct the grain morphology and grain boundary network, and analyze the influence of three-dimensional microstructure characteristics on the mechanical properties of steel.

[0024] The method comprises the following steps: obtaining a slice request information carrying a unique identification of a bridge steel sample, wherein the slice request information is issued by a continuous slice device; controlling an electron backscatter diffractometer to collect grain orientation and grain boundary distribution data of each layer according to the slice request information; receiving a registration instruction carrying an adjacent slice identification, wherein the registration instruction is generated by an image registration algorithm based on mutual information; supplementing missing information between slices using a cubic spline interpolation method according to the registration instruction to obtain a complete three-dimensional data set; segmenting the three-dimensional grains and reconstructing the grain boundary network for the three-dimensional data set; reconstructing the grain boundary network includes calculating the volume, surface area, and shape factor of the three-dimensional grains, and the area, curvature, and mismatch of the grain boundary network; receiving a comparison instruction carrying an X-ray tomography result; calculating voxel consistency according to the comparison instruction, and determining that the reconstruction accuracy is qualified if the voxel consistency reaches a preset threshold; if the reconstruction accuracy is qualified, performing mechanical property simulation on the three-dimensional microstructure model; and simultaneously calculating the corrosion current density distribution and corrosion depth using an electrochemical corrosion model.

[0025] Specifically, according to the preset slice thickness of 0.5μm and scanning interval of 0.1μm, a continuous slicing device is used to cut the bridge steel samples layer by layer, and the grain orientation and grain boundary distribution data of each layer are obtained by electron backscatter diffractometer. For the acquired two-dimensional slice data, the image registration algorithm based on mutual information is used to align adjacent slices, and the missing information between slices is supplemented by cubic spline interpolation to obtain a complete three-dimensional data set. The grain and grain boundary features are extracted from the three-dimensional data set, the three-dimensional grains are segmented using the watershed algorithm, and the grain boundary network is reconstructed using the March cube algorithm. The volume, surface area, shape factor of the grains, the area, curvature, and mismatch of the grain boundaries are calculated, and the three-dimensional grain morphology and grain boundary network model are constructed. The reconstructed three-dimensional model is compared with the X-ray tomography results, and the voxel consistency is calculated to ensure that the reconstruction accuracy reaches more than 95%. The mechanical properties of the verified three-dimensional microstructure model were simulated using ABAQUS finite element software, and the boundary conditions and loading methods were set. At the same time, the electrochemical corrosion model was used to simulate the corrosion process, calculate the corrosion current density distribution and corrosion depth, and determine the relationship between microstructure characteristics and macroscopic performance. On the bridge steel samples, the FEI Helios NanoLab 600i dual-beam focused ion beam system was used for continuous sectioning, and the section thickness was set to 0.5 μm, and a total of 200 layers were cut. After each layer of sectioning, the Tescan MIRA 3GMH field emission scanning electron microscope was used for electron backscatter diffraction scanning with a scanning interval of 0.1 μm to obtain 200×200 μm 2The grain orientation and grain boundary distribution data of the region. The 200 acquired two-dimensional slice data were aligned using an image registration algorithm based on mutual information, and the registration accuracy reached 0.05μm. The cubic spline interpolation method was used to insert 4 data points in the z-axis direction to increase the data resolution to 0.1μm×0.1μm×0.1μm. The watershed algorithm was used to segment the three-dimensional data and 3267 complete grains were identified. The March cube algorithm was used to reconstruct the grain boundary network, and a total of 12853 crystal interfaces were reconstructed. The average grain volume was calculated to be 125μm 3 , with an average surface area of ​​78 μm 2 , the average shape factor is 0.73. The average grain boundary area is 15μm 2 The average radius of curvature is 2.5 μm and the average mismatch is 32°. The reconstructed 3D model was compared with the Zeiss Xradia520 Versa X-ray tomography results, and the voxel consistency reached 97.2%. The mechanical properties of the verified 3D model were simulated using ABAQUS 2021 finite element software, and the stretching rate was set to 0.001 s -1 The yield strength is calculated to be 420MPa and the tensile strength is 565MPa. At the same time, the Butler-Volmer equation is used to establish an electrochemical corrosion model to simulate the corrosion process in 0.1mol / LNaCl solution, and the average corrosion current density is calculated to be 2.3μA / cm 2 , the maximum corrosion depth is 15μm. Through comparative analysis, it is found that the stress concentration coefficient and corrosion depth of the high-angle grain boundary area are higher than those of the low-angle grain boundary area, revealing the correlation between microstructural characteristics and macroscopic properties.

[0026] Based on the microstructure information of bridge steel obtained by electron backscatter diffraction characterization, a multi-scale mechanical properties testing method is used to evaluate material properties at different scales, including nanoindentation, micro-tensile and small-scale bending. By establishing a multi-scale correlation model between microstructure characteristics and macroscopic mechanical properties, the macroscopic properties of materials can be predicted from microstructure characteristics.

[0027] According to the microstructure information of bridge steel obtained by electron backscatter diffraction, a nanoindenter is used to perform indentation tests at different grains and grain boundaries of the steel to obtain the spatial distribution data of the hardness and elastic modulus of the steel. According to the indentation data, combined with the microstructure characteristics, micro-tensile samples are prepared, tensile tests are performed on the samples, and the stress-strain curves of the samples are recorded to obtain the yield strength, tensile strength and elongation of the steel. Small-sized specimens are cut from the samples of the micro-tensile test, and the bending strength and bending modulus of the specimens are determined by a three-point bending test, and the acoustic emission signals of the specimens are recorded using an AE signal acquisition system. The nanoindentation, micro-tensile and bending test data are statistically analyzed, the average and standard deviation of each index of the steel are calculated, and a scatter plot of the correlation between the microstructure characteristics and mechanical properties is drawn. A correlation model between the microstructure characteristics and multi-scale mechanical properties is established using a support vector machine regression algorithm with a radial basis function kernel. If the input characteristics of the model include parameters such as average grain size, grain boundary mismatch angle distribution, and nanoindentation hardness distribution, the output of the model includes macroscopic mechanical performance indicators such as yield strength, tensile strength, and elongation.

[0028] Specifically, according to the microstructure information of bridge steel obtained by electron backscatter diffraction, a nanoindenter was used to perform indentation tests at different grain and grain boundary positions, with the maximum indentation depth set to 200nm and the loading rate set to 10mN / min to obtain the spatial distribution data of hardness and elastic modulus. According to the obtained indentation data, combined with the microstructure characteristics, a micro-tensile sample with a gauge length of 2mm and a width of 0.5mm was prepared, and the strain rate was set to 10^-3s^-1 for tensile testing, and the stress-strain curve was recorded to calculate the yield strength, tensile strength and elongation. Small-sized specimens were cut from the micro-tensile test samples, and the bending strength and bending modulus were determined by a three-point bending test. At the same time, the AE signal acquisition system was used to record the acoustic emission signal, with the sampling frequency set to 5MHz and the threshold set to 40dB. The signal was denoised and feature extracted by wavelet transform to analyze the damage evolution process of the material. The nanoindentation, micro-tensile and bending test data were statistically analyzed, the average and standard deviation of each index were calculated, and a scatter plot of the correlation between microstructure characteristics and mechanical properties was drawn. The support vector machine regression algorithm with radial basis function kernel was used to establish the correlation model between microstructure characteristics and multi-scale mechanical properties. The input features included 10 parameters such as average grain size, grain boundary mismatch angle distribution, and nanoindentation hardness distribution. The output included 5 macroscopic mechanical properties such as yield strength, tensile strength, and elongation. The model performance was evaluated by 5-fold cross validation, and the root mean square error and determination coefficient were calculated. Electron backscatter diffraction analysis was performed on bridge steel samples using a JEOL JSM-7800F field emission scanning electron microscope. The scanning area was 1 mm × 1 mm and the step size was 1 μm. According to the obtained grain orientation map, 50 grains with different orientations and 30 typical grain boundary positions were selected and tested using a Hysitron TI950 Tribo Indenter nanoindenter. The maximum indentation depth was set to 200 nm and the loading rate was set to 10 mN / min to obtain the spatial distribution data of hardness and elastic modulus. The average grain hardness was 3.2±0.3 GPa and the grain boundary hardness was 3.8±0.4 GPa. Micro-tensile samples with a gauge length of 2 mm and a width of 0.5 mm were processed by wire cutting, and tensile tests were performed on the Instron 5948 micromechanical testing system, with a strain rate set to 10^-3s^-1. The stress-strain curve was recorded, and the yield strength was calculated to be 420±15MPa, the tensile strength was 560±20MPa, and the elongation was 18±2%. Small-sized specimens of 10mm×2mm×0.5mm were cut from the micro-tensile samples, and a three-point bending test was performed on the MTS Criterion Model 43 bending tester with a span of 8mm. At the same time, the MISTRASPCI-2 acoustic emission system was used to record the acoustic emission signal, with a sampling frequency of 5MHz and a threshold of 40dB. The signal was decomposed and denoised in 5 layers using the db6 wavelet to extract characteristic parameters such as energy, count, and amplitude.The bending strength is 980±30MPa, and the bending modulus is 205±8GPa. All test data were statistically analyzed, the mean and standard deviation were calculated, and MATLAB was used to draw a scatter plot of the correlation between microstructure characteristics and mechanical properties. The multi-scale correlation model was established using the support vector machine regression algorithm with radial basis function kernel. The input features included 10 parameters such as average grain size, grain boundary mismatch angle distribution, and nanoindentation hardness distribution, and 5 macroscopic mechanical performance indicators were output. The model performance was evaluated using 5-fold cross validation, with a root mean square error of 15MPa and a determination coefficient R. 2 It is 0.92, indicating that the model has good predictive ability.

[0029] S103. Conduct large-scale bridge steel exposure tests in a salt spray environment, regularly detect the chloride ion concentration and intergranular corrosion on the surface and inside of the bridge steel, and obtain a large sample data set of different chloride ion concentrations, environmental conditions and intergranular corrosion degrees.

[0030] An exposure test was conducted on bridge steel samples, and environmental data was monitored and recorded in real time. Outlier detection and filtering were performed based on the environmental data to obtain an environmental parameter time series data set. For bridge steel samples, the surface chloride ion concentration was measured, and the internal chloride ion concentration distribution was determined at the same time. Samples were randomly selected from the bridge steel samples for metallographic analysis, and the intergranular corrosion depth and area ratio were analyzed based on the intergranular corrosion morphology. According to the surface chloride ion concentration, internal chloride ion concentration distribution, environmental parameter time series data set, and intergranular corrosion depth and area ratio, the stepwise regression method was used to screen significant variables and establish a multivariate linear regression model. If the goodness of fit of the multivariate linear regression model meets the preset conditions, the standardized regression coefficients of the respective variables are calculated to determine the degree of influence of the respective variables on intergranular corrosion.

[0031] Specifically, according to the preset environmental parameters such as salt spray concentration of 5% NaCl solution, temperature range of 20-40℃, relative humidity of 70-95%, pH value of 6.5-7.2, an automatic controlled salt spray test chamber was used to conduct exposure tests on a large number of bridge steel samples. The environmental data was monitored and recorded in real time through a network composed of temperature and humidity sensors, pH electrodes and conductivity sensors. The real-time monitored environmental data was detected and filtered for outliers, and the daily average, maximum and minimum values ​​were calculated to form an environmental parameter time series data set. For the bridge steel samples in the exposure test, a portable X-ray fluorescence spectrometer was used to measure the surface chloride ion concentration of 100 samples every week. Five measuring points were selected for each sample and the position information was recorded. At the same time, an electrochemical impedance spectrometer was used to determine the internal chloride ion concentration distribution. Ten samples were randomly selected from the samples of the exposure test every month for metallographic analysis. Three cross sections were prepared for each sample. The intergranular corrosion morphology was observed by a metallographic microscope with a magnification of 500 times. The intergranular corrosion depth and area ratio were analyzed using image processing software to obtain quantitative data on the degree of intergranular corrosion. The significant variables were screened by stepwise regression method, and a multivariate linear regression model was established. The input parameters included surface and internal chloride ion concentration, temperature, humidity, pH value, etc., and the intergranular corrosion depth and area ratio were output. 2 , adjusted R 2 The goodness of fit of the model was evaluated by the F test, and the standardized regression coefficients of each variable were calculated to analyze their influence on intergranular corrosion.

[0032] In the Q-FOG CCT-1100 salt spray test chamber, a 5% NaCl solution was set, the temperature cycle was 30±2℃ (16 hours) and 40±2℃ (8 hours), the relative humidity was maintained at 85±5%, and the pH value was controlled at 7.0±0.2. 500 bridge steel samples of 100mm×50mm×5mm were placed, and a network consisting of DHT22 temperature and humidity sensor, E-201-C pH electrode and DJS-1C conductivity sensor was used to collect data every 5 minutes. The 3σ rule was used for outlier detection, and the Savitzky-Golay filter was used to smooth the data. The daily average, maximum and minimum values ​​of the environmental parameters were calculated. The surface chloride ion concentration of 100 samples was measured using Olympus Vanta portable XRF every week. Five measurement points were selected at the center and four corners of each sample, with an accuracy of ±0.01%. At the same time, a Gamry Reference 600+ electrochemical workstation was used to measure the internal chloride ion concentration distribution, with a scanning frequency range of 10^-2Hz to 10^5Hz. Ten samples were randomly selected each month, and metallographic specimens were prepared using a Struers CitoPress-20 hot press mounting machine. After grinding and polishing, the intergranular corrosion morphology was observed under an Olympus GX71 metallographic microscope at 500 times magnification. ImageJ software was used to analyze the depth and area ratio of intergranular corrosion, and 100 grain boundaries were measured to obtain the average value. SPSS software was used for stepwise regression analysis, with 0.05 as the significance level for entering the model and 0.10 as the significance level for removing from the model. A multivariate linear regression model was established, and 7 independent variables were entered, including surface and internal chloride ion concentrations, temperature, humidity, pH value, exposure time, and number of cycles. The R of the model 2 Reaching 0.89, the adjusted R 2 The p value of the F test is 0.87, and the p value of the F test is less than 0.001. The calculation of the standardized regression coefficient shows that the internal chloride ion concentration (0.45) and the exposure time (0.38) have the greatest impact on intergranular corrosion.

[0033] S104. Preprocess and feature engineer the acquired large sample data set, extract chloride ion concentration and microstructure characteristics, and construct training data for the intergranular corrosion prediction model.

[0034] A large sample data set is obtained, and the missing values ​​in the data set are processed by the median filling method. According to the filled data set, outliers are identified and eliminated by the box plot method. For the data set after eliminating outliers, the continuous variables are z-score standardized to obtain standardized data with a mean of 0 and a standard deviation of 1. If the variable is categorical, the categorical variable is converted to a single hot encoding. Chloride ion concentration-related features are extracted from the standardized data, including surface concentration, depth distribution, and concentration gradient. Grain boundaries are extracted from microstructure images, and the area and perimeter of the grain boundaries are calculated. The Pearson correlation coefficient method is used for the extracted features to screen features that are highly correlated with the degree of intergranular corrosion. The dimension of the screened features is reduced by the principal component analysis method, and the minimum number of principal components that meets the cumulative variance contribution rate threshold is selected. The minority class samples in the reduced dimension data set are oversampled, and the nearest neighbor number is set to a preset value, and the oversampling ratio is set to a preset ratio. The oversampled data set is divided into a training set, a validation set, and a test set by a stratified sampling method to form training data for an intergranular corrosion prediction model.

[0035] Specifically, based on the large sample data set obtained, the median filling method is used to handle missing values, the box plot method is used to identify and eliminate outliers, the overall mean and standard deviation are used to perform z-score standardization on continuous variables, the data is converted to a distribution with a mean of 0 and a standard deviation of 1, and the categorical variables are converted to one-hot encoding. For the preprocessed data, the chloride ion concentration-related features are extracted, including surface concentration, depth distribution, concentration gradient, etc. At the same time, the image segmentation algorithm is used to extract the grain boundary from the microstructure image, the grain area and perimeter are calculated, and the grain boundary angle distribution is analyzed by Fourier transform. From the extracted features, the Pearson correlation coefficient method is used to screen the features that are highly correlated with the degree of intergranular corrosion, the principal component analysis method is used to reduce the feature dimension, the cumulative variance contribution rate threshold is set to 95%, and the minimum number of principal components that meet the threshold is selected. The variance inflation factor is used to detect multicollinearity, and the features with a variance inflation factor greater than 10 are eliminated. The recursive feature elimination method is used to rank the feature importance, and the top 80% of the features are selected. The SMOTE algorithm was used to oversample the minority class samples, the nearest neighbor number was set to 5, the oversampling ratio was set to 200%, the distribution of samples of each category in the data set was balanced, and the data set was divided into training set, validation set and test set by stratified sampling method, with a ratio of 6:2:2, to ensure that the proportion of each category in each subset was consistent with the original data set. When processing 10,000 bridge steel sample data, the fillna() function of the pandas library was first used to fill the missing values ​​with the median, and a total of 342 missing values ​​were processed. The statistical functions in the SciPy library were used to calculate the quartiles (Q1 and Q3) and interquartile range (IQR) of each continuous variable, and 127 data points exceeding Q3+1.5IQR or below Q1-1.5IQR were marked as outliers and removed. For the remaining 9,873 data, the StandardScaler() function in the Scikit-learn library was used for z-score standardization, and 15 continuous variables were converted to a distribution with a mean of 0 and a standard deviation of 1. The three categorical variables were encoded using the get_dummies() function of pandas to generate 12 new features. From the preprocessed data, the chloride ion concentration gradient was calculated using the sobel() function of scipy's ndimage to obtain five features such as surface concentration and depth distribution. For the microstructure image, the watershed() algorithm in the Scikit-image library was used to extract the grain boundaries, and the average grain area was calculated to be 25μm 2, with a circumference of 18μm. The fft function in the NumPy library was used for Fourier transform to analyze the distribution of grain boundary angles, which were mainly concentrated around 30° and 60°. The DataFrame.corr() function in the Pandas library was used to calculate the Pearson correlation coefficient, and 12 features with a correlation coefficient greater than 0.6 with the degree of intergranular corrosion were screened out. The PCA() of decomposition in the sklearn library was used for principal component analysis, and n_components=0.95 was set to obtain 7 principal components. The outliers_influence.variance_inflation_factor() function in the statsmodels library was used to calculate the variance inflation factor, and 2 features with variance inflation factors (VIF)>10 were eliminated. The RFE() in the sklearn library was used for recursive feature elimination, and the top 80% of the 15 features were selected. Finally, the SMOTE() algorithm in the Imbalanced-Learn library is used to oversample the minority class samples, with k_neighbors=5 and sampling_strategy=2.0, generating 2376 new samples. Setting k_neighbors=5 means considering the 5 nearest neighbors, and setting sampling_strategy=2.0 means increasing the number of minority class samples to twice that of the majority class. StratifiedShuffleSplit() in the Scikit-learn library is used for stratified sampling, dividing the 12249 data into 7349 training sets, 2450 validation sets, and 2450 test sets, with the proportion of each category kept consistent.

[0036] S105. Based on the variation pattern of chloride ion concentration and typical crystal microstructure types, a prediction model for intergranular corrosion of bridge steel is constructed, and trained and verified on a large sample data set.

[0037] A convolutional neural network structure suitable for intergranular corrosion prediction of bridge steel is established, wherein the structure comprises a preset number of convolutional layers, a preset number of maximum pooling layers, and a preset number of fully connected layers, wherein the convolution kernel size is a preset size, and the pooling size is a preset size. According to the convolutional neural network structure, the structural parameters are adjusted to adapt to the intergranular corrosion prediction task. A large sample data set is obtained, and the chloride ion concentration and microstructure characteristics are extracted from the large sample data set as input, and the intergranular corrosion degree is output, wherein the input and output are set with an initial learning rate of a preset value, an Adam optimizer is used, the batch size is a preset size, and the number of training rounds is a preset number of rounds. The prediction accuracy and generalization ability of the convolutional neural network are evaluated by a cross-validation method. If the prediction accuracy is higher than a preset range and the mean absolute error is less than a preset range, it is determined that the convolutional neural network meets the performance requirements. The robustness of the convolutional neural network is verified by a test set under different temperature, humidity, and pH conditions. If the performance of the convolutional neural network drops by more than a threshold, the convolutional neural network is subjected to data enhancement and regularization processing until the convolutional neural network reaches a preset performance indicator.

[0038] Specifically, according to the obtained chloride ion concentration variation law and typical crystal microstructure type, a convolutional neural network structure suitable for intergranular corrosion prediction of bridge steel is designed, which includes 4 convolutional layers, 4 maximum pooling layers, and 2 fully connected layers. The convolution kernel size is 3x3, the pooling size is 2x2, the input layer receives chloride ion concentration and microstructure characteristics, and the output layer predicts the degree of intergranular corrosion. For the designed network structure, the ResNet50 model pre-trained on the MaterialsProject dataset is used, the parameters of the first 40 layers are frozen, and the last fully connected layer is replaced to adapt to the intergranular corrosion prediction task. The chloride ion concentration and microstructure characteristics are extracted from the large sample dataset as input, and the degree of intergranular corrosion is used as output. The initial learning rate is set to 0.001, the Adam optimizer is used, the batch size is 64, the number of training rounds is 100, and the learning rate decay strategy is used, which decays to 0.1 times the original every 20 rounds. The prediction accuracy and generalization ability of the model are evaluated by the 5-fold cross-validation method, and the prediction accuracy threshold is set to 90%, and the average absolute error is less than 5%. The robustness of the model is verified by testing sets under different temperature, humidity, and pH conditions. If the performance drops by more than 10%, data enhancement and regularization are added to fine-tune and optimize the model until the preset performance indicators are reached. When constructing the intergranular corrosion prediction model for bridge steel, a convolutional neural network structure consisting of 4 convolutional layers, 4 maximum pooling layers, and 2 fully connected layers is first designed. The input layer receives a 128x128 pixel chloride ion concentration distribution map and a microstructure feature map. The first convolutional layer uses 32 3x3 convolution kernels with a step size of 1 and a padding of 1, followed by a ReLU activation function and a 2x2 maximum pooling. Subsequent convolutional layers use 64, 128, and 256 convolution kernels, respectively, maintaining the same configuration. The number of neurons in the two fully connected layers is 1024 and 512, respectively, and the last layer outputs the degree of intergranular corrosion from 0 to 100%. The ResNet50 model pre-trained on the MaterialsProject dataset, which contains the structure and property data of 1 million materials, is used for transfer learning. The parameters of the first 40 layers are frozen and the last fully connected layer is replaced with the above custom structure. The bridge steel dataset containing 10,000 images is used for training, of which 70% are used for training, 15% are used for validation, and 15% are used for testing. The initial learning rate is set to 0.001, and the Adam optimizer is used with a batch size of 64 for 100 rounds of training. The learning rate is decayed to 0.1 times the original every 20 rounds. The model performance is evaluated using 5-fold cross validation, with an average prediction accuracy of 92.5% and an average absolute error of 3.8%. The model robustness is tested under different environmental conditions, such as temperature 5-40℃, relative humidity 30-90%, and pH 5-9. It is found that under extreme conditions, such as 40℃, 90% humidity, and pH=5, the performance drops by 12%.By adding data enhancement, including random rotation, scaling and flipping and L2 regularization, where λ = 0.001, the performance degradation under extreme conditions is controlled within 8%, meeting the preset threshold requirement.

[0039] S106. Deploy a wireless sensor network in the service environment of bridge steel to collect chloride ion concentration, temperature and humidity environmental parameters in real time, and obtain the data required for model input through data preprocessing and feature extraction.

[0040] The layout positions of chloride ion concentration sensors and temperature and humidity sensors are determined according to the structural characteristics of the bridge, and the topological structure of the wireless sensor network is obtained; the short-distance and low-power wireless communication data of the sensor is processed in real time, and the communication data is stored in the InfluxDB time series database; the original communication data is obtained from the time series database, the median filtering algorithm is used to remove outliers, the db4 wavelet is used to perform n-layer decomposition, and the high-frequency coefficients are processed using the soft threshold method to obtain the denoised data; feature extraction is performed on the denoised data, the statistical characteristics of the chloride ion concentration are calculated, the time series characteristics of the temperature and humidity data are extracted, and the m-dimensional feature vector is constructed in combination with the environmental parameters. The feature vector is used as the input of the intergranular corrosion prediction model.

[0041] Specifically, according to the structural characteristics and service environment of the bridge, the topology of the wireless sensor network is established, the location and number of chloride ion concentration sensors and temperature and humidity sensors are determined, the ZigBee protocol is selected for short-distance low-power wireless communication, and a hybrid power supply method of solar panels combined with lithium batteries is adopted. For the deployed wireless sensor network, the chloride ion concentration sensor is configured to collect data every 10 minutes, and the temperature and humidity sensor is configured to collect data every 5 minutes. The data is transmitted to the cloud server through the MQTT protocol for centralized processing. ApacheKafka is used for real-time data stream processing, and the sensor data is stored in the InfluxDB time series database to facilitate subsequent data preprocessing and analysis. From the raw data collected by the wireless sensor network, the median filtering algorithm is used to remove outliers, the db4 wavelet is used for 5-layer decomposition, the soft threshold method is used to process the high-frequency coefficients to achieve data noise reduction, and the interpolation algorithm is used to supplement the missing data to obtain continuous and smooth time series data. The preprocessed data was subjected to feature extraction, and 10 statistical features such as the mean, standard deviation, skewness, and kurtosis of chloride ion concentration were calculated. Eight time series features such as the daily variation and weekly variation trend of temperature and humidity data were extracted. Combined with four environmental parameters such as pH value and wind speed, a 22-dimensional feature vector was constructed as the input of the intergranular corrosion prediction model. 100 wireless sensor nodes were deployed on the steel structure components of a cross-sea bridge, including 50 chloride ion concentration sensors and 50 temperature and humidity sensors, which were connected to 10 ZigBee coordinators using a star topology. Each node was equipped with 100cm2 The solar panel and 3000mAh lithium battery ensure all-weather operation. The chloride ion concentration sensor has a range of 0-2000ppm and an accuracy of ±1%. It collects data every 10 minutes. The temperature and humidity sensor has a temperature accuracy of ±0.3℃ and a humidity accuracy of ±2%RH. It collects data every 5 minutes. The data is transmitted to the AWS cloud server via the MQTT protocol, and Kafka is used for real-time stream processing, processing about 1000 messages per second. The raw data is stored in InfluxDB, and a data retention policy of 1 hour is set. The raw data is first filtered by a 5x5 window median filter to remove outliers, and then the db4 wavelet is used for 5-layer decomposition. The high-frequency coefficients are processed using a soft threshold (λ=3σ) to achieve data noise reduction. Missing data points are filled with cubic spline interpolation. In the feature extraction stage, 10 statistical features of chloride ion concentration are calculated, including mean, standard deviation, skewness, kurtosis, maximum, minimum, median, interquartile range, coefficient of variation, and trend coefficient. Eight time series features of temperature and humidity data are extracted, including daily average, daily variation, weekly variation trend, autocorrelation coefficient, partial autocorrelation coefficient, Fourier transform main frequency, wavelet energy, and entropy. Combined with four environmental parameters, pH value, wind speed, atmospheric pressure, and rainfall, a 22-dimensional feature vector is finally constructed. These features are updated every hour as real-time input to the intergranular corrosion prediction model.

[0042] S107. Input the real-time collected data into the trained intergranular corrosion prediction model to dynamically predict the development trend and degree of intergranular corrosion inside the bridge steel. When the prediction result exceeds the preset threshold, an early warning signal is issued to provide a decision-making basis for the maintenance and replacement of the bridge steel.

[0043] According to the real-time collected environmental parameter data such as chloride ion concentration, temperature and humidity, after denoising and normalization, the time-frequency domain features and statistical features are extracted to obtain a feature vector that conforms to the input format of the intergranular corrosion prediction model. For the feature vector, the trained intergranular corrosion prediction model is input, a sliding time window is set, and the prediction results are updated every preset time to obtain the development trend of intergranular corrosion in the future. From the predicted time series data, the key characteristic parameters of corrosion rate, acceleration and cumulative corrosion amount are extracted to determine the severity of the current corrosion state. A comprehensive evaluation is performed on the key characteristic parameters. If the preset threshold is exceeded, an early warning information containing corrosion level, location, degree and expected deterioration time is generated. Based on the early warning information, maintenance recommendations for adjusting the detection frequency, anti-corrosion treatment and replacement time estimation are generated to determine the maintenance and replacement decision basis of bridge steel.

[0044] Specifically, based on the real-time collected data of environmental parameters such as chloride ion concentration, temperature and humidity, wavelet denoising and normalization are used to extract time-frequency domain features and statistical features, and construct a feature vector that conforms to the input format of the intergranular corrosion prediction model. For the constructed feature vector, the trained intergranular corrosion prediction model is input, and the long short-term memory network model is used to set a 24-hour sliding time window, update the prediction results every hour, and dynamically predict the development trend of intergranular corrosion in the future period of time to obtain the time series data of corrosion depth and area ratio. From the predicted time series data, key characteristic parameters are extracted, including corrosion rate, acceleration and cumulative corrosion amount, and four warning thresholds are set, including mild 0-20%, moderate 20-50%, severe 50-80% and dangerous >80%. A comprehensive assessment is conducted based on the corrosion depth and area ratio to determine the severity of the current corrosion state. A rule engine based on a decision tree is used to comprehensively evaluate the comparison results. Combined with the corrosion rate, acceleration and cumulative corrosion amount, when the preset threshold is exceeded, detailed warning information including corrosion level, location, degree and expected deterioration time is automatically generated, and a warning signal is sent to relevant personnel through a multi-channel alarm system. According to the warning information, maintenance suggestions are automatically generated, including adjustment of detection frequency, anti-corrosion treatment and replacement time estimation, providing a decision-making basis for the maintenance and replacement of bridge steel. In the steel structure monitoring of a cross-sea bridge, data of chloride ion concentration 0-2000ppm, temperature -20~50℃ and relative humidity 0-100% are collected every 5 minutes. The db4 wavelet is used to perform 5-layer decomposition and denoising on the original data, and then Min-Max normalization is performed. 13 time-frequency domain features such as wavelet energy, main frequency and 9 statistical features such as mean and standard deviation are extracted to construct a 22-dimensional feature vector. The feature vector is input into the pre-trained LSTM model (2 layers, 128 neurons in each layer), and the prediction results are updated every hour using a 24-hour sliding time window to predict the development trend of intergranular corrosion in the next 7 days. The corrosion rate (μm / year), acceleration (μm / year2) and cumulative corrosion amount (μm) are extracted from the prediction results. The four-level warning thresholds are set as mild (0-20μm), moderate (20-50μm), severe (50-80μm) and dangerous (>80μm). A rule engine based on a decision tree with a depth of 5 and a minimum number of leaf node samples of 10 comprehensively evaluates the corrosion status. When the corrosion depth in a certain area exceeds 50μm and the area ratio exceeds 30%, a warning information is automatically generated, including corrosion level and GPS positioning, with an accuracy of

[0045] ±1m, depth, area ratio and expected deterioration after 7 days. Early warnings are sent to managers via SMS, email and mobile application push. At the same time, maintenance recommendations are automatically generated according to the early warning level: mild - inspection within 3 months, moderate - anti-corrosion treatment within 1 month, severe - partial replacement within 2 weeks, dangerous - immediate full replacement. These recommendations serve as an important basis for bridge steel maintenance decisions and are stored in the maintenance management database for subsequent analysis and decision-making.

[0046] The above only lists some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and changes can be made. As long as the improvements and changes are made on the basis of the basic principles of the present invention, they should be regarded as falling within the protection scope of the present invention.

Claims

1. A bridge structure erosion degree monitoring method based on deep learning, characterized in that: The method comprises: According to the enrichment of chloride ions on the bridge surface, the electrochemical impedance spectroscopy method was used to obtain the distribution data of chloride ion concentration under different environmental conditions, and the change law of chloride ion concentration was obtained through Fourier transform analysis. For bridge steel samples of different types and locations, the electron backscatter diffraction method is used to characterize the grain boundary structure and grain orientation microstructure characteristics, and determine the typical microstructure type of the crystal; Conduct large-scale bridge steel exposure tests in a salt spray environment, regularly test the chloride ion concentration and intergranular corrosion on the surface and inside of bridge steel, and obtain a large sample data set with different chloride ion concentrations, environmental conditions and intergranular corrosion levels; Preprocess and feature engineer the acquired large sample data set to extract chloride ion concentration and microstructure characteristics, and construct training data for the intergranular corrosion prediction model; According to the variation of chloride ion concentration and typical crystal microstructure types, a prediction model for intergranular corrosion of bridge steel was constructed, and trained and verified on a large sample data set; A wireless sensor network is deployed in the service environment of bridge steel to collect chloride ion concentration, temperature and humidity environmental parameters in real time, and obtain the data required for model input through data preprocessing and feature extraction; The real-time collected data is input into the trained intergranular corrosion prediction model to dynamically predict the development trend and degree of intergranular corrosion inside the bridge steel. When the prediction result exceeds the preset threshold, an early warning signal is issued to provide a decision-making basis for the maintenance and replacement of bridge steel.

2. The method according to claim 1, wherein: According to the enrichment of chloride ions on the bridge surface, the electrochemical impedance spectroscopy method is used to obtain the distribution data of chloride ion concentration under different environmental conditions, and the change law of chloride ion concentration is obtained by Fourier transform analysis, including: Acquiring impedance data of different areas on the bridge surface according to preset measurement parameters, wherein the measurement parameters include frequency range, sampling interval and electrode type; The chloride ion concentration value of each area is calculated by the impedance data, and the degree of chloride ion enrichment is quantitatively evaluated; Conduct time-frequency analysis on chloride ion concentration values ​​to obtain the variation characteristics of chloride ion concentration in time and space dimensions; Extracting main frequency components from the variation characteristics and establishing a relationship function between chloride ion concentration and environmental factors, wherein the environmental factors include temperature, humidity and pH value; The least squares method is used to estimate the model parameters of the relationship function, and the model fit is evaluated by the R square value; Using the relationship function and combining with weather forecast data, the distribution of chloride ion concentration in the future is predicted; If the predicted concentration exceeds a preset warning threshold, the potential corrosion risk area is marked.

3. The method according to claim 1, wherein: The electron backscatter diffraction method is used to characterize the grain boundary structure and grain orientation microstructure characteristics of bridge steel samples of different types and locations, and determine the typical microstructure type of the crystal, including: Obtaining a diffraction pattern of a sample at a key part of a bridge, wherein the diffraction pattern is collected by an electron backscatter diffractometer; extracting grain boundary structure data and grain orientation data according to the diffraction pattern; Calculating the grain boundary angle distribution and the grain boundary length distribution according to the grain boundary structure data, if the grain boundary angle is less than a preset threshold, it is determined to be a small-angle grain boundary, otherwise it is determined to be a large-angle grain boundary; The energy of small-angle grain boundaries is calculated, and the energy of large-angle grain boundaries is fixed; Performing statistical analysis on the grain orientation data to obtain average grain size, shape factor and texture index; Performing dimensionality reduction processing on the grain boundary characteristics and grain characteristics, the grain boundary characteristics include angle distribution and energy distribution, and the grain characteristics include size, shape factor and texture index; Applying a K-means clustering algorithm to the feature data after dimensionality reduction, wherein the number of clusters of the K-means clustering algorithm is determined by a comprehensive evaluation of the silhouette coefficient and the Calinski-Harabasz index; The typical microstructure types of steel samples were obtained based on the clustering results; It also includes: using continuous slicing and electron backscattering diffraction characterization in sequence to obtain three-dimensional microstructural information of bridge steel, reconstruct the grain morphology and grain boundary network, and analyze the influence of three-dimensional microstructural characteristics on the mechanical properties of steel; based on the microstructural information of bridge steel obtained by electron backscattering diffraction characterization, a multi-scale mechanical property testing method is used to evaluate material properties at different scales, including nanoindentation, micro-stretching and small-size bending, and by establishing a multi-scale correlation model between microstructural characteristics and macroscopic mechanical properties, the macroscopic properties of materials can be predicted by microstructural characteristics.

4. The method according to claim 3, wherein: The method uses continuous sectioning and electron backscattering diffraction characterization in sequence to obtain three-dimensional microstructure information of bridge steel, reconstruct grain morphology and grain boundary network, and analyze the influence of three-dimensional microstructure characteristics on the mechanical properties of steel, including: Obtaining slicing request information carrying a unique identifier of a bridge steel sample, wherein the slicing request information is issued by a continuous slicing device; Controlling an electron backscatter diffractometer to collect grain orientation and grain boundary distribution data of each layer according to the slice request information; receiving a registration instruction carrying adjacent slice identifiers, wherein the registration instruction is generated by an image registration algorithm based on mutual information; Using cubic spline interpolation method to supplement missing information between slices according to the registration instruction to obtain a complete three-dimensional data set; For the three-dimensional data set, three-dimensional grains are segmented and then a grain boundary network is reconstructed; Reconstructing the grain boundary network includes calculating the volume, surface area, and shape factor of three-dimensional grains, and the area, curvature, and mismatch of the grain boundary network; receiving a comparison instruction carrying an X-ray tomography scan result; Calculating voxel consistency according to the comparison instruction, and if the voxel consistency reaches a preset threshold, determining that the reconstruction accuracy is qualified; If the reconstruction accuracy is qualified, performing mechanical property simulation on the three-dimensional microstructure model; At the same time, the electrochemical corrosion model is used to calculate the corrosion current density distribution and corrosion depth.

5. The method according to claim 3, wherein: Based on the microstructure information of bridge steel obtained by electron backscatter diffraction characterization, a multi-scale mechanical property testing method is used to evaluate material properties at different scales, including nanoindentation, micro-stretching and small-size bending. By establishing a multi-scale correlation model between microstructure characteristics and macroscopic mechanical properties, the macroscopic properties of the material can be predicted from the microstructure characteristics, including: Based on the microstructure information of the bridge steel obtained by electron backscatter diffraction, a nanoindenter is used to perform indentation tests at different grains and grain boundaries of the steel to obtain the spatial distribution data of the hardness and elastic modulus of the steel; According to the indentation data, combined with the microstructure characteristics, a micro-tensile sample is prepared, a tensile test is performed on the sample, a stress-strain curve of the sample is recorded, and the yield strength, tensile strength and elongation of the steel are obtained; Cutting a small-sized specimen from the sample of the micro-tensile test, measuring the bending strength and bending modulus of the specimen by a three-point bending test, and recording the acoustic emission signal of the specimen by using an AE signal acquisition system; Performing statistical analysis on the nanoindentation, microtensile and bending test data, calculating the average value and standard deviation of various indicators of the steel, and drawing a scatter plot of the correlation between the microstructure characteristics and mechanical properties; A correlation model between the microstructure characteristics and multi-scale mechanical properties is established using a support vector machine regression algorithm with a radial basis function kernel. If the input characteristics of the model include parameters such as average grain size, grain boundary mismatch angle distribution, and nanoindentation hardness distribution, the output of the model includes macroscopic mechanical performance indicators such as yield strength, tensile strength, and elongation.

6. The method according to claim 1, wherein: The large-scale bridge steel exposure test is conducted in a salt spray environment, and the chloride ion concentration and intergranular corrosion on the surface and inside of the bridge steel are regularly detected to obtain a large sample data set of different chloride ion concentrations, environmental conditions and intergranular corrosion degrees, including: Conduct exposure tests on bridge steel samples, monitor and record environmental data in real time; Perform outlier detection and filtering processing on the environmental data to obtain an environmental parameter time series data set; For bridge steel samples, the surface chloride ion concentration is measured, and the internal chloride ion concentration distribution is determined at the same time; Randomly select samples from bridge steel samples for metallographic analysis, and analyze the depth and area ratio of intergranular corrosion based on the intergranular corrosion morphology; According to the surface chloride ion concentration, internal chloride ion concentration distribution, environmental parameter time series data set, and intergranular corrosion depth and area ratio, the stepwise regression method was used to screen significant variables and establish a multivariate linear regression model. If the goodness of fit of the multivariate linear regression model meets the preset conditions, the standardized regression coefficients of the respective variables are calculated to determine the degree of influence of the respective variables on intergranular corrosion.

7. The method according to claim 1, wherein: The method of preprocessing and feature engineering the acquired large sample data set, extracting chloride ion concentration and microstructure characteristics, and constructing training data for the intergranular corrosion prediction model includes: Obtain a large sample data set, and use a median filling method to handle missing values ​​in the data set; According to the filled data set, outliers are identified and removed through the box plot method; For the data set after removing outliers, z-score standardization was performed on continuous variables to obtain standardized data with a mean of 0 and a standard deviation of 1; If the variable is categorical, the categorical variable is converted to one-hot encoding; Extract chloride ion concentration-related features from the normalized data, including surface concentration, depth distribution, and concentration gradient; extracting grain boundaries from the microstructure image, and calculating the area and perimeter of the grain boundaries; The Pearson correlation coefficient method was used for the extracted features to screen out the features that were highly correlated with the degree of intergranular corrosion; The dimension of the filtered features is reduced by principal component analysis, and the minimum number of principal components that meets the cumulative variance contribution rate threshold is selected; Oversample the minority class samples in the reduced-dimensional data set, set the nearest neighbor number to a preset value, and set the oversampling ratio to a preset ratio; The oversampled data set is divided into a training set, a validation set and a test set by using a stratified sampling method to form training data for an intergranular corrosion prediction model.

8. The method according to claim 1, wherein: According to the variation law of chloride ion concentration and typical microstructure type of crystal, a prediction model for intergranular corrosion of bridge steel is constructed, and trained and verified on a large sample data set, including: Establishing a convolutional neural network structure suitable for predicting intergranular corrosion of bridge steel, the structure comprising a preset number of convolutional layers, a preset number of maximum pooling layers, and a preset number of fully connected layers, wherein the convolution kernel size is a preset size, and the pooling size is a preset size; According to the convolutional neural network structure, adjusting the structural parameters to adapt to the intergranular corrosion prediction task; Obtain a large sample data set, extract chloride ion concentration and microstructure characteristics from the large sample data set as input, and extract the degree of intergranular corrosion as output, wherein the input and output are set with initial learning rates as preset values, an Adam optimizer is used, a batch size is a preset size, and the number of training rounds is a preset number of rounds; The prediction accuracy and generalization ability of the convolutional neural network are evaluated by a cross-validation method. If the prediction accuracy is higher than a preset range and the mean absolute error is less than a preset range, the convolutional neural network is determined to meet the performance requirements. The robustness of the convolutional neural network is verified by testing sets under different temperature, humidity and pH conditions. If the performance of the convolutional neural network drops beyond a threshold, the convolutional neural network is subjected to data augmentation and regularization until the convolutional neural network reaches the preset performance indicators.

9. The method according to claim 1, wherein: The wireless sensor network is deployed in the service environment of the bridge steel to collect chloride ion concentration, temperature and humidity environmental parameters in real time, and the data required for model input is obtained through data preprocessing and feature extraction, including: According to the structural characteristics of the bridge, the locations of the chloride ion concentration sensor and the temperature and humidity sensor are determined, and the topological structure of the wireless sensor network is obtained; Perform real-time data stream processing on the short-distance and low-power wireless communication data of the sensor, and store the communication data in the InfluxDB time series database; Obtaining original communication data from the time series database, removing outliers using a median filter algorithm, performing n-layer decomposition using a db4 wavelet, and processing high-frequency coefficients using a soft threshold method to obtain denoised data; Feature extraction is performed on the denoised data, statistical features of chloride ion concentration are calculated, time series features of temperature and humidity data are extracted, and an m-dimensional feature vector is constructed in combination with environmental parameters. The feature vector is used as an input for an intergranular corrosion prediction model.

10. The method according to claim 1, wherein: The real-time collected data is input into the trained intergranular corrosion prediction model to dynamically predict the development trend and degree of intergranular corrosion inside the bridge steel. When the prediction result exceeds the preset threshold, an early warning signal is issued to provide a decision-making basis for the maintenance and replacement of the bridge steel, including: According to the real-time collected environmental parameter data such as chloride ion concentration, temperature and humidity, after denoising and normalization, the time-frequency domain features and statistical features are extracted to obtain the feature vector that meets the input format of the intergranular corrosion prediction model; For the feature vector, a trained intergranular corrosion prediction model is input, a sliding time window is set, and the prediction result is updated every preset time to obtain the development trend of intergranular corrosion in the future; Extract the key characteristic parameters of corrosion rate, acceleration and cumulative corrosion amount from the predicted time series data to determine the severity of the current corrosion state; Comprehensively evaluate the key characteristic parameters, and if they exceed a preset threshold, generate early warning information including corrosion level, location, degree and expected deterioration time; Based on the warning information, maintenance recommendations for adjusting the detection frequency, anti-corrosion treatment and replacement time estimation are generated to determine the basis for maintenance and replacement decisions of bridge steel.

Images