Concrete structure steel bar remaining life prediction method based on steel bar corrosion rate

Abstract

The invention relates to the technical field of concrete structure steel bar corrosion monitoring and life prediction, and discloses a concrete structure steel bar residual life prediction method based on a steel bar corrosion rate. According to the method, reinforced concrete surface multi-mode electrochemical data are collected in real time through a distributed intelligent sensor array, and a three-dimensional corrosion feature tensor is generated through wavelet packet decomposition and feature extraction. A prediction model is constructed by using cross-domain feature transfer learning and an attention mechanism, and feature weights of different corrosion stages are dynamically allocated. On-line reasoning and dynamic correction are carried out in combination with the edge computing node and the cloud platform, a steel bar section corrosion rate distribution cloud picture and a residual life prediction value are output and pushed to the terminal operation and maintenance platform to realize early warning information dynamic visual display of a multi-level user interface, and the efficiency and level of concrete structure maintenance and management are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete structure steel bar corrosion monitoring and life prediction, and in particular to a method for predicting the remaining life of concrete structure steel bars based on the steel bar corrosion rate. Background Art

[0002] In concrete structures, steel corrosion is one of the main causes of reduced structural durability. Traditional monitoring methods rely on periodic manual inspections or single electrochemical parameter measurements, which are unable to fully reflect the spatiotemporal evolution of steel corrosion. Furthermore, due to the complex and variable environmental factors (such as temperature and humidity), existing prediction models often struggle to accurately capture the dynamic changes in steel corrosion rates, resulting in significant uncertainty in remaining life predictions.

[0003] There are two major technical bottlenecks in current projects that need to be addressed urgently: first, there is a lack of methods that can accurately locate the location of steel bar rust pits in real time and quantitatively evaluate the geometric characteristics (depth, area) and spatial distribution of rust pits; second, the safety assessment of reinforced concrete structures urgently requires the establishment of a quantitative correlation model between the degree of corrosion, the development rate of rust pits and the degradation of structural performance.

[0004] Therefore, developing a method that can monitor the corrosion status of steel bars in real time and accurately and dynamically predict their remaining life is of great significance for ensuring the safety and durability of concrete structures. Summary of the Invention

[0005] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a method for predicting the remaining life of concrete structure steel bars based on the steel bar corrosion rate, so as to solve the problems of inaccurate steel bar corrosion monitoring and inaccurate remaining life prediction in the prior art.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate, the method comprising: Real-time acquisition of multimodal electrochemical data on the reinforced concrete surface using a distributed intelligent sensor array, wherein the multimodal electrochemical data includes potential distribution patterns, current density waveforms, and impedance spectrum signals; Wavelet packet decomposition and feature extraction are performed on multimodal electrochemical data. A signal quality evaluation function is established to perform adaptive weighted fusion on heterogeneous data to generate a three-dimensional corrosion feature tensor with spatiotemporal correlation. The 3D corrosion feature tensor is input into a pre-built steel bar remaining life prediction model. The steel bar remaining life prediction model uses a cloud platform to perform cross-domain feature transfer learning on historical corrosion samples and uses an attention mechanism to dynamically assign feature weights to different corrosion stages. Based on the edge computing node, the trained remaining life model of the steel bar is called for online inference. Combining the environmental temperature and humidity monitoring data, a dynamic correction coefficient of the corrosion rate is constructed, and a cloud map of the corrosion rate distribution of the steel bar cross-section and the predicted value of the remaining life are output; The cloud map of the corrosion rate distribution of the steel bar cross-section and the predicted value of the remaining life are pushed to the terminal operation and maintenance platform to realize the dynamic visualization display of the warning information on the multi-level user interface.

[0007] Preferably, in a possible implementation manner of the first aspect, the db4 wavelet basis is used to perform 5-layer wavelet packet decomposition on the multi-modal electrochemical data, and the energy entropy of each node is extracted as the frequency domain feature; Perform time-domain statistical analysis on the original signal, and extract the peak-to-peak value, kurtosis factor, and waveform factor as the time-domain features; Construct a signal quality evaluation function based on information entropy, calculate the confidence weight of each sensor data, and perform weighted fusion on the potential, current, and impedance features in the time-space dimension to generate a three-dimensional corrosion feature tensor containing time series, spatial position, and modal dimensions.

[0008] Preferably, in a possible implementation manner of the first aspect, the signal quality evaluation function satisfies:

[0009] Among them, is the fusion weight of the th sensor, is the data information entropy of the th sensor, is the th sensor's signal-to-noise ratio, is the data information entropy of the th sensor, is the th sensor's signal-to-noise ratio, is the total number of sensors, 、 are the contribution coefficients of entropy and signal-to-noise ratio, is the normalization adjustment factor, and satisfies .

[0010] Preferably, in a possible implementation manner of the first aspect, the formula for constructing the three-dimensional corrosion feature tensor is:

[0011] Among them, represents the tensor element, is the time-domain feature vector of the th sensor, is the spatial position encoding of the th sensor, is the modal feature matrix of the th sensor, is the dynamic weight of the th sensor, represents the tensor product operation.

[0012] Preferably, in a possible implementation manner of the first aspect, the remaining life prediction model of the steel bar is a multi-channel convolutional neural network. After receiving the three-dimensional rust feature tensor at the input layer, the distribution difference between the historical samples and the real-time data is aligned through the cross-domain feature transfer learning module; The channel attention mechanism is embedded in the convolutional layer to dynamically adjust the weight ratio of the feature maps at each rust stage; The output layer uses a bidirectional LSTM network to predict the rust development trajectory, and the model is trained based on the joint loss function of the mean square error and the KL divergence.

[0013] Preferably, in a possible implementation manner of the first aspect, the cross-domain feature transfer learning divides the historical rust samples into a source domain and a target domain, and generates domain-invariant features through adversarial training; The maximum mean difference constraint is introduced during the feature alignment process, and the gradient reversal layer is used to optimize the parameter update direction of the feature discriminator.

[0014] Preferably, in a possible implementation manner of the first aspect, the weight allocation formula of the attention mechanism is:

[0015] where is the attention score of the th channel, is the Sigmoid function, , are the learnable parameters of the pooling operation, is the input feature map, is the average pooling operation, is the maximum pooling operation, and is satisfied; The final feature map is updated through the gating mechanism , is the updated feature map, is the scaling factor.

[0016] Preferably, in a possible implementation manner of the first aspect, the edge computing node deploys a lightweight inference engine to perform dimensionality reduction processing on the real-time input three-dimensional rust feature tensor; Call the environmental temperature and humidity monitoring data to construct a dynamic correction coefficient for the rust rate, and use the fuzzy logic method to quantify the non-linear impact of environmental factors on the rust process; Input the corrected corrosion rate into the life prediction model to generate the cloud map of the corrosion rate distribution of the steel bar cross-section and the predicted value of the remaining life interval.

[0017] Preferably, in a possible implementation manner of the first aspect, the construction process of the corrosion rate dynamic correction coefficient includes: Collect the environmental temperature and humidity time series data and calculate its covariance matrix with the historical corrosion samples; Extract the main component features of the covariance matrix as the environmental impact factors, detect environmental mutation events through a sliding time window, and adjust the update frequency of the correction coefficient based on the event trigger mechanism.

[0018] Preferably, in a possible implementation manner of the first aspect, the calculation formula of the corrected corrosion rate is:

[0019] Where is the reference corrosion rate, is the th main component eigenvalue, is the environmental sensitivity coefficient, is the number of main components, is the contribution weight of the th main component, and satisfies .

[0020] The beneficial effects of the present invention are as follows: The present invention collects multi-modal electrochemical data in real time through a distributed intelligent sensor array, combines wavelet packet decomposition and feature extraction technologies to generate a three-dimensional corrosion feature tensor with spatio-temporal correlation, effectively improving the accuracy of steel bar corrosion monitoring.

[0021] At the same time, using cross-domain feature transfer learning and attention mechanism, the constructed remaining life prediction model of the steel bar can dynamically adapt to different corrosion stages and environmental conditions, significantly improving the prediction accuracy.

[0022] In addition, through the collaborative work of the edge computing node and the cloud platform, online inference and dynamic correction are realized, further enhancing the real-time performance and reliability of the system. Finally, through the dynamic visualization display of the multi-level user interface of the terminal operation and maintenance platform, intuitive and convenient decision-making support is provided for the operation and maintenance personnel, effectively improving the efficiency and level of the maintenance management of the concrete structure. Description of the Drawings

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0024] Figure 1 This application provides a flowchart of a method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate. Detailed implementation manners

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0026] Embodiment 1: As Figure 1 shown, the present invention provides a method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate, including: Real-time collecting multimodal electrochemical data on the surface of reinforced concrete through a distributed intelligent sensor array, where the multimodal electrochemical data includes potential distribution maps, current density waveforms, and impedance spectrum signals.

[0027] In this embodiment, the distributed intelligent sensor array is arranged on the surface of the reinforced concrete of the concrete structure to be measured in a three-dimensional grid topology using micro electrochemical sensing units. The spacing of the sensing units is set to 20mm × 20mm according to the steel bar diameter and the corrosion-sensitive area, and isopotential contact with the surface of the reinforced concrete is achieved through a flexible conductive colloid.

[0028] Each sensing unit is integrated with a potential detection module, a potentiostat module, and a frequency response analysis module. Among them, the potential detection module is equipped with an Ag / AgCl reference electrode and a Pt counter electrode to real-time collect the potential distribution in the micro area on the surface of the reinforced concrete with a resolution of 0.5mV; the potentiostat module uses a three-electrode system to perform a dynamic potential scan within a 200mV polarization range and record the current density waveform at a sampling frequency of 100Hz; the frequency response analysis module applies a 10mV sinusoidal excitation signal in the frequency band of 10mHz - 100kHz and synchronously obtains the real and imaginary part data of the impedance spectrum.

[0029] The sensor array establishes a communication connection with the edge computing gateway through the industrial Ethernet protocol. The built-in signal conditioning circuit performs second-order Butterworth low-pass filtering and 24-bit analog-to-digital conversion on the original signal, effectively suppressing 50Hz power frequency interference and electromagnetic noise. To ensure the spatio-temporal consistency of multi-source data, the system uses a GPS timing module to trigger synchronous acquisition of each sensor node, and aligns the time stamps of potential, current, and impedance signals through a sliding time window mechanism, forming a multi-modal data cube indexed by spatial coordinates.

[0030] Among them, the potential distribution map is converted into a local corrosion tendency index according to the Nernst equation, the current density waveform is analyzed for the instantaneous corrosion rate through the Tafel extrapolation method, and the impedance spectrum signal is used to obtain the charge transfer resistance and double-layer capacitance parameters through equivalent circuit fitting. During the data acquisition process, the system dynamically monitors the contact impedance and baseline drift of each sensor node. When the detected contact impedance exceeds 50kΩ or the baseline drift is greater than 10% of the full scale, the self-calibration program is automatically triggered and the abnormal data channel is marked.

[0031] Perform wavelet packet decomposition and feature extraction on the multi-modal electrochemical data, establish a signal quality evaluation function for adaptive weighted fusion of heterogeneous data, and generate a three-dimensional corrosion feature tensor with spatio-temporal correlation.

[0032] In this embodiment, the wavelet packet decomposition and feature extraction process of multi-modal electrochemical data is realized by using the discrete wavelet transform technology. Specifically, for the three types of heterogeneous data of potential distribution map, current density waveform, and impedance spectrum signal, the db4 wavelet basis function is selected for 5-layer wavelet packet decomposition respectively, and a decomposition tree structure containing 32 frequency sub-signals is constructed. For the time-domain signal collected by each sensor node , it is decomposed to the 5th layer through recursive downsampling and filtering operations, and the wavelet packet coefficient matrix corresponding to each decomposition node in the frequency band is extracted , where represents the decomposition level, is the node index. On this basis, the energy entropy of each frequency band signal is calculated as the frequency-domain feature, and its expression is:

[0033] In the formula, is the probability distribution of the wavelet packet coefficients of the th frequency band within the time window, is the number of coefficient bins. This energy entropy feature can effectively characterize the variation of the frequency-domain energy distribution of the electrochemical signal caused by the corrosion process.

[0034] In terms of time-domain feature extraction, the system performs sliding time window analysis on the original signal, with the window length set to 10 seconds and the overlap rate of 50%. Three types of time-domain statistics are synchronously calculated within each time window: the peak-to-peak value is the difference between the maximum and minimum values of the signal, the kurtosis factor characterizes the pulse characteristics of the signal, and is defined as the ratio of the fourth-order central moment to the fourth power of the standard deviation:

[0035] where is the original signal sequence within the time window, is the mean value of the signal within the time window, is the standard deviation of the signal within the time window, is the expected value operation, is the result of the kurtosis factor.

[0036] The waveform factor S is calculated by the ratio of the effective value of the signal to the absolute mean value:

[0037] is the effective value of the signal, is the mean value of the signal within the time window, is the result of the waveform factor. The above-mentioned time-domain feature set can comprehensively reflect the electrochemical activity changes caused by steel bar corrosion. Among them, the kurtosis factor has a sensitive capture ability for the transient current fluctuations caused by local pitting corrosion.

[0038] To achieve the effective fusion of multi-source heterogeneous data, this embodiment constructs a signal quality evaluation function based on information entropy. First, the original data collected by each sensor node is preprocessed, and its information entropy and signal-to-noise ratio are calculated. Among them, the information entropy is calculated using the probability distribution histogram of the signal amplitude within the sliding time window, and the window length is consistent with the aforementioned time-domain analysis. The signal-to-noise ratio is estimated by the improved singular value decomposition method, and the ratio of the main component energy of the signal to the remaining component energy is used as . The signal quality evaluation function is defined as:

[0039] In the formula, is the fusion weight of the th sensor, is the data information entropy of the th sensor, is the signal-to-noise ratio of the th sensor, is the data information entropy of the th sensor, is the signal-to-noise ratio of the th sensor, is the total number of sensors, 、 are the contribution coefficients of entropy and signal-to-noise ratio, is a normalization adjustment factor and satisfies . In this embodiment, and are taken. This function provides a quantitative weight for subsequent feature fusion by dynamically evaluating the confidence of each sensor's data. In particular, when baseline drift or abnormal noise is detected in a certain sensor's data, its signal-to-noise ratio will be significantly reduced, thereby automatically reducing the weight ratio of this node in the fusion process.

[0040] The construction of the three-dimensional rust feature tensor adopts a spatio-temporal-modal multi-dimensional fusion strategy. For each sensor node , its time-domain feature vector consists of the peak-to-peak value, kurtosis factor, and waveform factor, with a dimension of 3×T (T is the length of the time series). The spatial position encoding adopts a normalized position vector in a three-dimensional coordinate system and is generated through a pre-calibrated sensor topology relationship. The modal feature matrix integrates the frequency-domain energy entropy features of three modalities: potential, current, and impedance, forming a 32×3 feature plane. The above features are fused across dimensions through a tensor product operation:

[0041] In the formula, represents a tensor element, is the time-domain feature vector of the th sensor, [[ID=3,2]] is the spatial position encoding of the th sensor, is the modal feature matrix of the th sensor, represents the tensor product operation, is the dynamic weight calibrated by the signal quality evaluation function. The finally generated three-dimensional rust feature tensor retains the temporal correlation of the sampling sequence in the time dimension, encodes the rust differences at different positions on the surface of reinforced concrete in the spatial dimension, and integrates the complementary information of multi-source electrochemical signals in the modal dimension. The tensor structure is a three-dimensional data block of 128×128×96, where 128×128 corresponds to the spatial resolution of the sensor grid on the surface of reinforced concrete, and 96 is the depth of the time-modal joint feature.

[0042] In this embodiment, through the spatial topology structure and spatio-temporal fusion tensor of the distributed sensor array, the precise positioning and morphology reconstruction of the rust pits on the steel bar surface are realized. The multi-modal data collected by each sensor node are encoded in space to obtain a three-dimensional rust feature tensor, solving the problem that traditional methods cannot quantitatively characterize local rust concentration areas.

[0043] Input the three-dimensional rust feature tensor into the pre-built prediction model for the remaining life of steel bars. The prediction model for the remaining life of steel bars performs cross-domain feature transfer learning on historical rust samples through a cloud platform, and uses an attention mechanism to dynamically allocate feature weights for different rust stages.

[0044] In this embodiment, the prediction model for the remaining life of steel bars adopts a multi-channel convolutional neural network architecture, and its input layer is designed as a three-dimensional convolutional kernel adapter to receive a rust feature tensor with a dimension of 128×128×96.

[0045] The model integrates the historical rust sample library through a cloud training platform to perform cross-domain feature transfer learning, and constructs a deep network structure containing 12 convolutional layers. The cross-domain feature transfer learning module at the front end of the network adopts an adversarial training mechanism, divides the historical rust samples into a source domain (laboratory accelerated rust data) and a target domain (field long-term monitoring data), and generates domain-invariant features through the dynamic game between the domain discriminator and the feature generator. Specifically, during implementation, the feature generator consists of 3 layers of dilated convolutions, the domain discriminator adopts a fully connected network structure, and the two implement backpropagation of the parameter update direction through a gradient reversal layer, enabling the generator to extract common features while obscuring the distribution differences between domains. During the training process, the maximum mean discrepancy is introduced as a regularization constraint, and the similarity of the feature distributions of the source domain and the target domain is calculated in the reproducing kernel Hilbert space. The expression is:

[0046] where is the maximum mean discrepancy, is the norm in the reproducing kernel Hilbert space, represents the kernel function mapping, is the feature of the -th sample in the source domain (laboratory accelerated rust data), is the feature of the -th sample in the target domain (field long-term monitoring data), is the number of source domain samples, is the number of target domain samples. In this embodiment, the Gaussian kernel function is selected, and the kernel width parameter is optimized to 0.5 through grid search. This constraint term and the adversarial loss together constitute the multi-objective optimization function of transfer learning, effectively alleviating the problem of the degradation of the model's generalization performance caused by the distribution drift of on-site data.

[0047] The model embeds a channel attention mechanism during the feature extraction stage to dynamically allocate feature weights for different rust stages. Specifically, the feature map output by each convolutional layer undergoes parallel global average pooling and max pooling operations to generate a spatially compressed feature vector. The pooling results pass through learnable parameters , Perform linear weighted fusion, where is initialized to 0.7, is initialized to 0.3, and through the constraint layer, ensure that . The weighted feature vector is activated by the Sigmoid function to generate the channel attention score:

[0048] where, is the attention score of the th channel, is the Sigmoid function, , are the learnable parameters of the pooling operation, is the input feature map, is the average pooling operation, is the max pooling operation. This score matrix controls the feature enhancement intensity through the scaling factor (the initial value is 1.2 in this embodiment and is trainable), and finally updates the feature map through the gating mechanism: . This design enables the model to strengthen the response of the potential distribution features in the early rust stage, focus on the weight allocation of the current density features during the accelerated rust period, and achieve adaptive feature focusing of the corrosion process.

[0049] The backend of the network is connected to a bidirectional LSTM module for predicting the rust development trajectory. Its hidden layer dimension is set to 256, and the time step is aligned with the time dimension of the input tensor. This module captures the long-range dependence relationship of the rust rate through forward and backward time series processing, and the output layer uses a fully connected network to map to the remaining life prediction value. The model training adopts a joint loss function of mean squared error (MSE) and KL divergence:

[0050] where is used to constrain the similarity between the predicted value distribution and the true life distribution , is the true remaining life value of the th sample, is the model-predicted remaining life value of the th sample. In this embodiment, the weight coefficients , . The optimization process uses the adaptive moment estimation algorithm, the initial learning rate is set to , and the cosine annealing strategy is applied to dynamically adjust the learning rate. The cloud training platform deploys a distributed computing framework, divides 200,000 groups of historical rust samples into a training set, a validation set, and a test set according to the ratio of 8:1:1. Each training batch contains 32 three-dimensional feature tensors, the training period is set to 100 rounds, and an early stopping mechanism is adopted to prevent overfitting.

[0051] During the model inference stage, the edge computing node loads the trained network parameters to generate a lightweight inference engine. The real-time input three-dimensional rust feature tensor first undergoes layer normalization to eliminate the feature scale differences caused by environmental interference. The cross-domain feature transfer learning module performs domain adaptation operations on the real-time data and embeds it into the latent space consistent with the historical sample library through feature projection. The channel attention mechanism dynamically adjusts the response intensity of the convolution kernel according to the current rust stage. When an abnormally high potential fluctuation is detected in a local area, it automatically enhances the current density feature weight at the corresponding spatial position. After the bidirectional LSTM module integrates the temporal features, the output layer generates the rust rate prediction curve for the next 30 days and the remaining life confidence interval. The final prediction result is transmitted to the correction module in the form of a three-dimensional probability cloud map for environmental factor compensation.

[0052] Based on the edge computing node, the trained remaining life model of the steel bar is called for online inference. Combining the environmental temperature and humidity monitoring data, a dynamic correction coefficient for the rust rate is constructed, and the rust rate distribution cloud map of the steel bar cross-section and the predicted remaining life value are output.

[0053] In this embodiment, the edge computing node deploys a lightweight inference engine based on the TensorRT framework, and uses channel pruning and quantization-aware training techniques to optimize and compress the pre-constructed remaining life prediction model of the steel bar. For the real-time input three-dimensional rust feature tensor, first perform feature dimensionality reduction processing in the space-time dimension: apply principal component analysis in the spatial dimension to extract the first 16 principal component feature vectors, retaining more than 95% of the original information; in the time dimension, use the dynamic time warping algorithm to align the temporal sampling deviation of the sensor array, and compress the 96-dimensional time feature to 32-dimensional latent representation. The size of the optimized three-dimensional feature tensor is reduced to 64×64×32, and the inference latency is controlled within 200 ms, meeting the requirements of on-site real-time monitoring.

[0054] The environmental temperature and humidity monitoring data are synchronously collected through the LoRa wireless sensor network at a frequency of 1 Hz. The edge node is built-in with a Kalman filter to denoise and smooth the original data, and constructs a time series database to store the historical records of environmental parameters for the past 30 days. The construction of the dynamic correction coefficient for the rust rate adopts a hybrid method combining covariance analysis and fuzzy logic: first calculate the covariance matrix of the current temperature and humidity sequence and the historical rust sample library, and extract the first 3 principal component features (cumulative contribution rate ≥ 85%) as environmental impact factors through singular value decomposition; then use the sliding time window mechanism (window length 30 minutes, step length 5 minutes) to detect the mutation events of environmental parameters. When the standard deviation of temperature and humidity within the window exceeds the threshold (temperature , humidity ), it triggers an emergency update of the correction coefficient, and the coefficient calibration is performed every 60 minutes under normal working conditions.

[0055] The input variables of the fuzzy logic system are the normalized principal component scores of temperature and humidity (PC1, PC2, PC3), and the output variable is the corrosion rate correction coefficient. Three fuzzy subsets are defined: low temperature and low humidity (LL), normal temperature and normal humidity (NN), and high temperature and high humidity (HH). Their membership functions adopt Gaussian distributions, and the standard deviations are dynamically adjusted according to the statistical values of historical data. The defuzzification process uses the centroid method to calculate the exact output value, and the rule weights are optimized by the genetic algorithm to make the correction coefficient within the adjustment range of 0.8 to 1.5. The calculation formula for the finally corrected corrosion rate is:

[0056] In the formula, the reference corrosion rate is obtained by decoding the feature tensor output by the prediction model through a fully connected layer. is the th principal component eigenvalue, is the principal component contribution weight (the preset values are 0.55, 0.35, and 0.10 respectively), is the standardized principal component score. The environmental sensitivity coefficient is dynamically generated by the fuzzy inference system, and its value is updated in real time with the occurrence of temperature and humidity mutation events, effectively quantifying the non-linear coupling effect of concrete carbonation depth and chloride ion penetration rate on steel bar corrosion.

[0057] The generation of the corrosion rate distribution cloud map uses three-dimensional interpolation rendering technology: mapping the 64×64 spatial feature matrix to the actual geometric coordinate system of the steel bar cross-section, filling the data gaps between sensor nodes through the thin plate spline interpolation algorithm; superimposing the corrected corrosion rate prediction values in the time dimension, and using the Level Set method to simulate the expansion trajectory of the corrosion front. The remaining life prediction value is obtained through integral operation, and the specific formula is:

[0058] where is the critical corrosion rate threshold (set to 15% of the steel bar cross-sectional area loss rate). The integral process is numerically solved using the fourth-order Runge-Kutta method, and Monte Carlo simulation is introduced to evaluate the prediction uncertainty caused by environmental parameter fluctuations. Finally, the life estimation range under an 80% confidence interval is output. The result data is transmitted to the terminal operation and maintenance platform through the OPC UA protocol, and the dynamic heat map rendering of the corrosion cloud map is realized using WebGL technology, and AR augmented reality labels are superimposed to display the real-time corrosion parameters at key positions.

[0059] During the project implementation, the system was configured with dual redundant edge computing nodes and a heartbeat detection mechanism to ensure the continuity of online inference services. If the primary node fails, the backup node can take over data processing tasks within 500ms, and an incremental synchronization mechanism is used to maintain the consistency of the environmental database. Historical correction coefficient records and prediction results are stored in a blockchain format. Each data block includes a timestamp, environmental parameter hash value, and digital signature, ensuring the immutability and auditability of the monitoring data.

[0060] This example identifies rust pit development hotspots through corrosion rate analysis. A "corrosion-structural performance" mapping model, based on the time-varying law of corrosion rates, quantitatively predicts the degradation of bearing capacity due to steel bar cross-sectional loss. This allows for a coordinated assessment of remaining service life and structural safety level under non-destructive conditions, meeting the need for quantitative prediction of overall structural performance degradation.

[0061] The steel bar section corrosion rate distribution cloud map and remaining life prediction value are pushed to the terminal operation and maintenance platform to realize the dynamic visualization display of early warning information on the multi-level user interface.

[0062] In this embodiment, the terminal operation and maintenance platform adopts a multi-level visualization system based on B / S architecture, and receives the corrosion rate distribution cloud map and remaining life prediction data pushed by the edge computing node through the industrial Internet of Things protocol. The system deploys a distributed message middleware to achieve high-concurrency data access, and uses the MQTT protocol to establish a transmission channel with a service quality level of QoS2 to ensure the reliable delivery and temporal integrity of early warning information. The corrosion rate distribution cloud map is encapsulated in a three-dimensional point cloud data format. Each spatial coordinate point contains the corrosion rate value, confidence interval and material attribute label. Dynamic heat map rendering is achieved on the front-end interface through WebGL technology, supporting perspective zooming, section cutting and isosurface extraction operations. The remaining life prediction value is stored in a time series database, combined with the ECharts visualization library to generate a corrosion process trend curve, and superimposed with an 80% confidence interval shadow band generated by Monte Carlo simulation.

[0063] The multi-level user interface design adopts a responsive layout and is compatible with various devices such as desktop workstations, mobile terminals and AR glasses. Operation and maintenance personnel access different levels of visualization content through role-based permission management: the primary view shows the overall corrosion level and remaining life overview of the steel bar section, using a red-yellow-green three-color warning coding system; the intermediate view can drill down to the local corrosion cloud map of the specified steel bar number, and display the environmental temperature and humidity time series curve and the correction coefficient change trend in a linked manner; the advanced view provides a three-dimensional BIM model overlay display function, and uses AR space anchoring technology to map the prediction results to the digital twin of the actual engineering structure, supporting multi-person collaborative labeling and abnormal area annotation. The warning information push adopts a hierarchical trigger mechanism. When it is detected that the local corrosion rate exceeds the threshold (section loss rate ) or when the predicted remaining life is lower than the design life, an early warning work order containing positioning coordinates, risk level and disposal suggestions will be automatically generated, and multi-channel alarms will be issued through sound and light prompts, text messages and emails.

[0064] A dual-mode storage architecture of relational database and graph database is established in the background of the system. The corrosion prediction data is compressed and encrypted and then written into the time series database partition. A global data snapshot is generated every 12 hours for historical tracing. User operation logs and warning event records are stored using blockchain technology. Each transaction contains the hash value of the previous block, a timestamp and a digital signature, forming an audit traceability chain that cannot be tampered with. The visual interface has a built-in intelligent diagnosis module, which can call historical data of the same period to compare and analyze the corrosion development rate, and recommend anti-corrosion maintenance strategies based on the expert knowledge base. At the same time, the API interface is open to connect with third-party operation and maintenance systems, supporting online verification and feedback optimization of prediction model parameters. Users can adjust the corrosion rate threshold and confidence parameters through the drag-and-drop interactive interface, and observe the spatial distribution changes of the prediction results in real time.

[0065] In this embodiment, the terminal operation and maintenance platform's multi-level visualization interface deeply integrates rust pit location information with structural safety warnings. The primary warning view highlights rust pit concentrations in the form of a heat map; the intermediate analysis interface displays the coupled relationship between corrosion rate and remaining life; and the advanced decision-making view simulates the impact of rust pit expansion on the mechanical properties of key locations such as beam-column joints through a structural digital twin. This integrated display of "corrosion morphology, development rate, and structural response" provides decision support for both localized corrosion control and overall structural safety assessment.

[0066] Embodiment 2: The present invention provides a method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate, the method further comprising: During the deployment phase of the distributed sensor array, programmable magnetic sensor nodes are used to dynamically reinforce blind spots in complex structural components (such as beam-column joints and shear wall edges). When the system detects persistently low data confidence in a local area, it automatically generates an optimal replenishment plan. Using a drone-mounted magnetic device, a spare node is precisely positioned to form an adaptive monitoring network. Redundant communication links are established between nodes using an ad hoc networking protocol to ensure robust data transmission even in the presence of localized damage to the concrete overburden.

[0067] To achieve the continuous optimization of the prediction model during long-term service, the system is built with an online incremental learning mechanism. The edge computing nodes regularly encrypt and upload the on-site environmental parameters and rust feature tensors to the cloud knowledge base. When the detected difference in the distribution of new and old data exceeds the threshold, the model fine-tuning process is automatically triggered. This process uses the elastic weight consolidation technology to directionally update the channel weight parameters in the attention mechanism with new on-site samples while retaining historical knowledge, effectively adapting to the differences in rust evolution patterns caused by different regional climate conditions.

[0068] Furthermore, the terminal operation and maintenance platform integrates a decision support engine to convert the remaining life prediction value into an executable maintenance strategy. Based on the spatial gradient features of the rust rate distribution cloud map of the steel bar cross-section, the hot spots of rust development are automatically identified, and the critical bearing margin is calculated by combining structural mechanics simulations. When the predicted life approaches the safety threshold, the system dynamically generates a hierarchical disposal plan: recommends the injection parameters of the rust inhibitor and the optimal construction window period for low-risk areas; outputs the local reinforcement design drawings and material lists for high-risk areas, and links with the engineering resource scheduling system to reserve maintenance resources. All decision-making schemes are virtually deduced through the digital twin model and pushed to the mobile terminal for execution after verifying their effectiveness.

[0069] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A prediction method for the remaining life of steel bars in concrete structures based on the corrosion rate of steel bars, characterized in that, The method includes: Collecting multimodal electrochemical data on the surface of reinforced concrete in real time through a distributed intelligent sensor array, where the multimodal electrochemical data includes potential distribution maps, current density waveforms, and impedance spectrum signals; Performing wavelet packet decomposition and feature extraction on the multimodal electrochemical data, establishing a signal quality evaluation function to adaptively weight and fuse heterogeneous data, and generating a three-dimensional corrosion feature tensor with spatio-temporal correlation; Inputting the three-dimensional corrosion feature tensor into a pre-constructed prediction model for the remaining life of the reinforcement. The prediction model for the remaining life of the reinforcement performs cross-domain feature transfer learning on historical corrosion samples through a cloud platform, and uses an attention mechanism to dynamically allocate feature weights for different corrosion stages; Based on an edge computing node, calling the trained prediction model for the remaining life of the reinforcement for online inference, combining environmental temperature and humidity monitoring data to construct a dynamic correction coefficient for the corrosion rate, and outputting a cloud map of the corrosion rate distribution of the reinforcement cross-section and a predicted value of the remaining life; Pushing the cloud map of the corrosion rate distribution of the reinforcement cross-section and the predicted value of the remaining life to the terminal operation and maintenance platform to realize the dynamic visualization display of early warning information on multiple-level user interfaces.

2. The method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate as described in claim 1, wherein, Performing 5-layer wavelet packet decomposition on the multimodal electrochemical data using the db4 wavelet basis, and extracting the energy entropy of each node as a frequency-domain feature; Performing time-domain statistical analysis on the original signal, and extracting the peak-to-peak value, kurtosis factor, and waveform factor as time-domain features; Constructing a signal quality evaluation function based on information entropy, calculating the confidence weights of each sensor data, and performing weighted fusion on potential, current, and impedance features in the spatio-temporal dimension to generate a three-dimensional corrosion feature tensor including time series, spatial position, and modal dimensions.

3. The method for predicting the remaining service life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 2, wherein The signal quality evaluation function satisfies: Among them, is the fusion weight of the th sensor, is the data information entropy of the th sensor, is the signal-to-noise ratio of the th sensor, is the data information entropy of the th sensor, is the signal-to-noise ratio of the th sensor, is the total number of sensors, , are the contribution coefficients of entropy and signal-to-noise ratio, is the normalization adjustment factor, and it satisfies .

4. The method for predicting the remaining service life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 3, characterized in that, The construction formula of the three-dimensional corrosion feature tensor is: Among them, represents a tensor element, is the time-domain feature vector of the th sensor, is the spatial position encoding of the th sensor, is the modal feature matrix of the th sensor, is the dynamic weight of the th sensor, represents the tensor product operation.

5. The method for predicting the remaining service life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 4, characterized in that, The prediction model for the remaining life of the reinforcement is a multi-channel convolutional neural network. After its input layer receives the three-dimensional corrosion feature tensor, it aligns the distribution differences between historical samples and real-time data through a cross-domain feature transfer learning module; Embedding a channel attention mechanism in the convolutional layer to dynamically adjust the weight ratio of feature maps at each corrosion stage; The output layer uses a bidirectional LSTM network to predict the corrosion development trajectory, and performs model training based on a combined loss function of mean square error and KL divergence.

6. The method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 5, wherein The cross-domain feature transfer learning divides historical corrosion samples into a source domain and a target domain, and generates domain-invariant features through adversarial training; Introducing a maximum mean discrepancy constraint during the feature alignment process, and using a gradient reversal layer to optimize the parameter update direction of the feature discriminator.

7. The method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 6, wherein, The weight allocation formula of the attention mechanism is: Among them, is the attention score of the th channel, is the Sigmoid function, , are the learnable parameters of the pooling operation, is the input feature map, is the average pooling operation, is the max pooling operation, and satisfies ; The final feature map is updated through a gating mechanism updated, is the updated feature map, is the scaling factor.

8. The method for predicting the remaining service life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 7, characterized in that, The edge computing node deploys a lightweight inference engine to perform dimensionality reduction processing on the real-time input three-dimensional corrosion feature tensor; Calling environmental temperature and humidity monitoring data to construct a dynamic correction coefficient for the corrosion rate, and using a fuzzy logic method to quantify the non-linear impact of environmental factors on the corrosion process; Inputting the corrected corrosion rate into the life prediction model to generate a cloud map of the corrosion rate distribution of the reinforcement cross-section and a predicted value of the remaining life interval.

9. The method for predicting the remaining service life of steel bars in a concrete structure based on the steel bar corrosion rate as claimed in claim 8, wherein, The construction process of the dynamic correction coefficient for the corrosion rate includes: Collecting environmental temperature and humidity time series data, and calculating its covariance matrix with historical corrosion samples; Extract the principal component features of the covariance matrix as environmental impact factors, detect environmental mutation events through a sliding time window, and adjust the update frequency of the correction coefficient based on an event-triggered mechanism.

10. The method for predicting the remaining life of steel bars in a concrete structure based on the steel bar corrosion rate according to claim 9, characterized in that, The calculation formula for the corrected corrosion rate is as follows: Among them, is the reference corrosion rate, is the th principal component eigenvalue, is the environmental sensitivity coefficient, is the number of principal components, is the contribution weight of the th principal component, and satisfies .

Images