A method for identifying concrete cracks based on building foundation pillars

Abstract

The application relates to the technical field of structure detection and visual recognition, and discloses a concrete crack identification method based on a building base column. The method comprises the following steps: acquiring a multi-angle image sequence of a base column surface and internal multi-type sensor readings, performing time scale alignment and feature decoupling, and forming a fusion perception feature set; using the set to construct a crack knowledge graph with space-time correlation, and training a base column structure behavior model based on historical normal data; inputting the real-time acquired fusion features into the knowledge graph and the structure model at the same time, performing bidirectional feature comparison and abnormal scoring, and outputting a preliminary abnormal event with a confidence score; and according to the event, triggering the collaborative iterative optimization of the knowledge graph and the structure model, and generating a deep analysis report containing a specific crack mode label and a traceability direction. The method can improve the accuracy and system adaptive ability of crack identification, and enhance the explainability of analysis results.

Description

Technical Field

[0001] This invention relates to the field of structural inspection and visual recognition technology, specifically a method for identifying concrete cracks in building foundation columns. Background Technology

[0002] In the field of building structural health monitoring, crack identification in concrete columns mainly relies on manual inspection and single-mode sensing technology. Manual inspection is highly subjective, inefficient, and struggles to detect early-stage hidden damage. Automated methods based on computer vision typically only acquire surface images, are susceptible to environmental interference, and cannot perceive internal conditions. While sensor-based monitoring technologies can acquire internal parameters, the data and visual information are analyzed separately, making it difficult to form a comprehensive judgment.

[0003] Existing technical solutions suffer from two key drawbacks. At the multimodal data fusion level, visual images and sensor readings suffer from asynchronous acquisition times and heterogeneous data formats. Conventional simple stitching or early fusion methods lack precise time-scale alignment and deep feature decoupling, leading to information mixing and feature redundancy, severely hindering the accurate extraction of subsequent crack features. At the intelligent analysis and decision-making level, existing methods are mostly unidirectional and static recognition models. These models lack explicit modeling of crack semantic relationships and spatiotemporal evolution patterns; their knowledge bases and parameters are fixed, making continuous optimization with new data impossible. This results in poor system adaptability, insufficient interpretability of output results, and difficulty in providing source analysis of crack development.

[0004] The technical challenges that need to be addressed are: how to achieve high-quality fusion and pure feature extraction of multi-source heterogeneous data, and how to build a crack recognition system that can dynamically evolve and has self-optimization and deep analysis capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying concrete cracks based on building foundation columns, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for identifying concrete cracks based on building foundation columns, the method comprising:

[0007] Acquire multi-angle image sequences of the base column surface and multi-type sensor readings inside the base column, perform time-scale alignment and feature decoupling on the multi-angle image sequences and the multi-type sensor readings to form a fused sensing feature set;

[0008] Multi-scale crack feature extraction and semantic association are performed on the fused perception feature set to construct a crack knowledge graph with spatiotemporal correlation.

[0009] Based on the historical normal state patterns in the crack knowledge graph, a column structure behavior model is trained and generated. The column structure behavior model outputs the behavior confidence and state reconstruction error.

[0010] The real-time generated fusion perception feature set is synchronously input into the column structure behavior model and crack knowledge graph to perform bidirectional feature comparison and anomaly score calculation, and output the preliminary anomaly event and the confidence score of the preliminary anomaly event.

[0011] Upon receiving the preliminary abnormal event and confidence score, a collaborative iterative optimization of the crack knowledge graph and the column structure behavior model is triggered, and a deep analysis report containing crack mode labels and source tracing points is generated.

[0012] Preferably, a multi-angle image sequence of the base column surface and multi-type sensor readings inside the base column are acquired. Time-scale alignment and feature decoupling are performed on the multi-angle image sequence and the multi-type sensor readings to form a fused sensing feature set. This is specifically achieved through the following steps:

[0013] Simultaneously trigger the image acquisition equipment arranged on the surface of the building's foundation column and the sensing device embedded inside the foundation column to acquire image frame sequences and physical quantity reading sequences with a unified timestamp, respectively;

[0014] The image frame sequence is subjected to texture enhancement and region segmentation processing to extract the visual feature vector of the base column surface, which includes grayscale distribution, edge gradient and local binary pattern;

[0015] The physical quantity reading sequence is subjected to noise reduction and normalization processing to extract the internal state feature vector of the column representing strain, vibration and humidity changes;

[0016] The visual feature vector on the surface of the base column and the state feature vector inside the base column are modally mapped and aligned based on their unified timestamps to generate the fused perception feature set.

[0017] Preferably, the fused perception feature set is subjected to multi-scale crack feature extraction and semantic association to construct a crack knowledge graph with spatiotemporal correlation, which is specifically achieved through the following steps:

[0018] From the fused sensing feature set, sub-feature sets belonging to different time segments and different spatial regions are separated;

[0019] For each of the sub-feature sets, feature extraction operators at the micro and macro scales are applied respectively to identify potential crack initiation features, crack propagation features, and crack stabilization features;

[0020] All identified crack features are used as nodes, and directed connection edges are established between nodes based on the co-occurrence relationship, temporal relationship and spatial adjacency relationship between features.

[0021] The nodes and the directed edges are stored in a graph database to form the crack knowledge graph, and time and confidence attributes are attached to each node and edge.

[0022] Preferably, a behavior model of the base column structure is trained and generated based on the historical normal state patterns in the crack knowledge graph, specifically through the following steps:

[0023] From the crack knowledge graph, nodes and connecting edges marked as normal historical states are selected to form a normal state sample set;

[0024] Using the normal state sample set as training data, a deep autoencoder network is constructed. The deep autoencoder network learns to compress and reconstruct the normal state sample set.

[0025] The state reconstruction error is obtained by calculating the difference between the reconstructed data and the original input data;

[0026] Gaussian process regression is used to model the distribution of the state reconstruction error in the time dimension, and the behavior confidence score, which characterizes the stability of the structural behavior, is obtained. Together, they constitute the behavior model of the pillar structure.

[0027] Preferably, the step of obtaining the state reconstruction error by calculating the difference between the reconstructed data and the original input data is specifically achieved through the following steps:

[0028] Obtain the reconstructed data of a sample in the normal state sample set output by the deep autoencoder network, and the original input data corresponding to the sample. Both the original input data and the reconstructed data are multi-dimensional feature vectors.

[0029] The original input data vector and the reconstructed data vector are subtracted element by element to obtain the absolute error vector for each feature dimension;

[0030] Calculate the statistical features of the original input data vector and the reconstructed data vector, including mean, variance and higher-order moments, and calculate the relative differences between the statistical features;

[0031] The error values ​​of each dimension in the absolute error vector are weighted and summed to obtain a preliminary error scalar. At the same time, the relative differences between the statistical features are integrated to obtain another difference scalar.

[0032] The initial error scalar and the difference scalar are linearly combined to generate the final state reconstruction error value used to characterize the accuracy of a single reconstruction.

[0033] Preferably, the real-time generated fusion perception feature set is synchronously input into the base column structure behavior model and crack knowledge graph for bidirectional feature comparison and anomaly scoring calculation, which is specifically achieved through the following steps:

[0034] The fused sensing feature set generated at the current moment is input into the deep autoencoder network to obtain the real-time state reconstruction error at the current moment;

[0035] The real-time state reconstruction error at the current moment is input into the Gaussian process regression model to calculate the real-time behavior confidence at the current moment.

[0036] Simultaneously, using the fusion perception feature set at the current moment as the query condition, a subgraph matching search is performed in the crack knowledge graph to find the historical state pattern with the highest similarity.

[0037] Calculate the feature distance between the fused perception feature set and the historical state pattern with the highest similarity, and combine it with the real-time behavior confidence to generate the anomaly score through a weighted fusion algorithm;

[0038] When the anomaly score exceeds a preset static threshold, or the real-time behavior confidence level is lower than a dynamically adjusted confidence threshold, it is determined to be a preliminary anomaly event.

[0039] Preferably, receiving the preliminary abnormal event and confidence score triggers a collaborative iterative optimization of the crack knowledge graph and the column structure behavior model, specifically achieved through the following steps:

[0040] The set of fused perception features corresponding to the preliminary abnormal event is taken as a new node to be analyzed and attempted to be fused into the crack knowledge graph;

[0041] In the crack knowledge graph, starting from the new node to be analyzed, multiple paths are searched to identify existing historical nodes that may have a causal or related relationship with the abnormal event.

[0042] The found associated paths and nodes, along with the confidence scores of the preliminary abnormal events, are submitted to the column structure behavior model.

[0043] The column structure behavior model uses the submitted associated paths and node information to locally recalibrate the parameters of the Gaussian process regression model and update the calculation benchmark of the behavior confidence.

[0044] The updated column structure behavior model is then used to re-evaluate the initial abnormal event, forming a collaborative iterative optimization loop between the crack knowledge graph and the column structure behavior model.

[0045] Preferably, the step of using Gaussian process regression to model the distribution of the state reconstruction error over the time dimension to obtain the behavioral confidence score characterizing the stability of the structural behavior is specifically achieved through the following steps:

[0046] Using timestamps as input variables and the state reconstruction error sequence as observation variables, a training dataset for Gaussian process regression is constructed.

[0047] Choose the squared exponential kernel function as the covariance function, and initialize the length scaling parameter and variance parameter of the kernel function;

[0048] The hyperparameters of the Gaussian process regression model are optimized by maximizing the marginal likelihood function to obtain a trained Gaussian process regression model.

[0049] The trained Gaussian process regression model is used to predict the state reconstruction error at the current time and in future time windows, and the prediction mean and prediction variance are obtained.

[0050] The inverse of the prediction variance is normalized and mapped to the range of zero to one, which is used as the confidence level of the behavior, where a higher confidence level indicates smaller fluctuations in the state reconstruction error.

[0051] Preferably, the method further includes:

[0052] Continuously collect the state reconstruction error sequence generated by the behavior model of the base column structure within a continuous monitoring period;

[0053] The state reconstruction error sequence is segmented, and the statistical and trend features of each segment are extracted.

[0054] A multi-scale time series clustering algorithm is applied to perform unsupervised clustering analysis on all extracted features to identify recurring error change patterns with specific shapes.

[0055] Each identified error change pattern is defined as a typical structural behavior pattern, and a feature template and trigger condition library are established for each of these structural behavior patterns.

[0056] Preferably, the method further includes:

[0057] Obtain the associated paths found in the crack knowledge graph;

[0058] Analyze the attribute change gradients of each node on the associated path, especially the evolution direction and rate of crack feature vectors;

[0059] Based on the gradient of the attribute change, the direction of the most dramatic feature evolution is traced in reverse to locate the most likely origin node of the initial abnormal event in the spatiotemporal dimension.

[0060] By combining sensor readings and image features of the origin node, the type of initial cause leading to crack initiation is inferred, and the origin node and the type of initial cause are used as the source tracing conclusion and output to the in-depth analysis report.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By performing time-scale alignment and feature decoupling operations, the time base from two heterogeneous data streams—visual and sensor—was unified, and independent feature components characterizing the essence of cracks were separated. This process eliminates information distortion caused by asynchronous acquisition devices and reduces interference from non-correlated information between modalities, resulting in a more pure and representative fusion sensing feature set. Crack feature extraction based on this feature set significantly improves the quality of the underlying data, enhances the reliability of the initial input signal for subsequent crack detection and analysis, and reduces the probability of false alarms and false negatives caused by data defects.

[0063] A crack knowledge graph was constructed to achieve a structured expression of multi-scale crack features and their spatiotemporal semantic relationships, while the structural behavior model of the foundation column learned the "healthy" behavior patterns of the structure from historical normal data. Real-time data was synchronously input into both for bidirectional feature comparison, combining a dual verification mechanism of symbolic knowledge reasoning and numerical model simulation. This comparison not only detects anomalies by calculating deviations but also finds potential correlation patterns of anomalous features from the knowledge graph. When an initial anomaly is identified, the system triggers a collaborative iterative optimization mechanism, using newly discovered anomalous samples to simultaneously update the parameters of the semantic network of the knowledge graph and the structural behavior model. This allows the entire system to continuously learn from actual operation, and its crack recognition and pattern cognition capabilities continuously evolve over monitoring time. The output in-depth analysis report combines accurate pattern classification with clear causal traceability clues. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the concrete crack identification method based on building foundation columns described in this invention.

[0065] Figure 2 A flowchart for constructing a fusion-sensory feature set;

[0066] Figure 3 A flowchart for training the behavior model of the column base structure;

[0067] Figure 4A line graph showing the time distribution of errors in the reconstruction of the structural state of the foundation column;

[0068] Figure 5 This is a time-series monitoring graph of the knowledge graph node scores for column cracks. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides a method for identifying concrete cracks in building foundation columns. The method includes: acquiring multi-angle image sequences of the column surface and readings from multiple types of sensors inside the column; performing time-scale alignment and feature decoupling on the multi-angle image sequences and multi-type sensor readings to form a fused sensing feature set; extracting multi-scale crack features and semantically associating them with the fused sensing feature set to construct a crack knowledge graph with spatiotemporal correlation; training a column structural behavior model based on historical normal state patterns in the crack knowledge graph, which outputs behavior confidence and state reconstruction error; synchronously inputting the real-time generated fused sensing feature set into the column structural behavior model and the crack knowledge graph for bidirectional feature comparison and anomaly scoring calculation, outputting preliminary anomaly events and their confidence scores; receiving the preliminary anomaly events and confidence scores to trigger collaborative iterative optimization of the crack knowledge graph and the column structural behavior model, and generating a deep analysis report containing crack pattern labels and source tracing directions.

[0071] Example 1: See Figure 2Simultaneously triggering image acquisition devices deployed on the surface of building foundation columns and sensing devices embedded inside the columns, the system acquires image frame sequences and physical quantity reading sequences with unified timestamps. Texture enhancement and region segmentation are performed on the image frame sequences to extract visual feature vectors of the column surface, including grayscale distribution, edge gradients, and local binary patterns. Noise reduction and normalization are performed on the physical quantity reading sequences to extract internal state feature vectors representing strain, vibration, and humidity changes within the column. The visual feature vectors of the column surface and the internal state feature vectors are modally mapped and aligned based on their unified timestamps to generate a fused sensing feature set. From this fused sensing feature set, sub-feature sets belonging to different time segments and spatial regions are separated. For each sub-feature set, micro-scale and macro-scale feature extraction operators are applied to identify potential crack initiation features, crack propagation features, and crack stabilization features. All identified crack features are used as nodes, and directed connections are established between nodes based on co-occurrence relationships, temporal relationships, and spatial adjacency relationships between features. The nodes and directed edges are stored in a graph database to form a crack knowledge graph, and time and confidence attributes are attached to each node and edge.

[0072] In the specific implementation, a high-resolution industrial camera deployed at a specific location on the surface of the building's foundation column and a fiber Bragg grating sensor array embedded inside the concrete of the foundation column are simultaneously triggered. The high-resolution industrial camera acquires images of the foundation column surface from three different angles at a fixed frequency of one frame per second, while the fiber Bragg grating sensor array acquires strain and temperature readings at a frequency of one hundred sampling points per second. The data acquisition module adds a millisecond-accurate timestamp to each frame of image and each set of sensor readings from the same timing source, forming an image frame sequence and a physical quantity reading sequence with a unified timestamp. In the specific implementation, texture enhancement processing based on contrast-limited adaptive histogram equalization is performed on the acquired image frame sequence. Then, the watershed algorithm is used to segment the enhanced image into regions, separating the foundation column surface from the background environment. Within the segmented foundation column surface region, the gray-level histogram distribution of the image patch, the edge gradient magnitude calculated by the Sobel operator, and the local binary pattern features extracted using a neighborhood with a radius of two pixels are calculated. These features are then concatenated to form a 1000-dimensional visual feature vector of the foundation column surface. In some embodiments, wavelet threshold-based noise reduction is performed on the physical quantity reading sequence to filter out high-frequency noise. Then, the strain, vibration acceleration, and humidity readings from sensors with different ranges are mapped to a numerical range of zero to one, and normalization is performed. The mean, variance, peak value, and valley value of each sensor reading in the past time window are extracted and combined into a 200-dimensional internal state feature vector of the base column that characterizes the internal state change of the base column.

[0073] In practical implementation, based on a unified timestamp, a 1000-dimensional visual feature vector of the base column surface and a 200-dimensional internal state feature vector of the base column generated at the same time are modally mapped. The mapping process uses a fully connected neural network layer to project the two vectors onto the same feature space. Then, the projected vectors are concatenated end to end to generate a 1200-dimensional fused feature vector. All fused feature vectors in the continuous time series are arranged in chronological order to form a fused sensing feature set. It can be understood that when separating sub-feature sets from the fused sensing feature set, the feature set is divided into multiple time segment subsets based on the timestamp, with each hour as a segment. Simultaneously, based on the spatial encoding corresponding to the sensor and camera numbers of the data source, the feature set is divided into spatial region subsets corresponding to the top, middle, and bottom of the base column. In practice, for each sub-feature set formed by the interweaving of time segments and spatial regions, feature extraction operators at both the micro-scale and macro-scale are applied. The micro-scale operators focus on minute abrupt changes in local feature values ​​to identify potential crack initiation features, such as abnormal increases in local strain gradients or sudden disruptions in the consistency of local image texture. The macro-scale operators analyze the statistical regularities and evolution trends of features over a larger area to identify crack propagation and crack stability features, such as the changing trends of crack length and width, as well as the shift in vibration mode frequencies.

[0074] In practical implementation, each identified crack initiation feature, crack propagation feature, and crack stabilization feature is instantiated as a node. Node attributes include feature type, feature vector, spatial location, and timestamp. Directed connections between nodes are established based on the co-occurrence, temporal, and spatial adjacency relationships between features. Co-occurrence refers to two feature nodes being observed simultaneously within the same time segment; temporal relationship refers to one feature node appearing before another in a time sequence; and spatial adjacency refers to the spatial regions corresponding to two feature nodes being adjacent in three-dimensional space. It can be understood that in constructing the crack knowledge graph, the aforementioned nodes and directed connections are stored in the Neo4j graph database, and each node is assigned a "timestamp" and "confidence" attribute, while each directed connection is assigned a "relationship type" and "transition probability" attribute. In some embodiments, the distance between feature vectors is used to calculate the similarity between nodes. A specific distance calculation formula is as follows:

[0075]

[0076] in: This represents the distance between node A and node B. and These are the feature vectors of node A and node B, respectively. It is the Euclidean distance between the two. It is the normalized time difference between the timestamps of node A and node B. It is the normalized spatial distance between the spatial coordinates of node A and node B. , , These are preset weighting coefficients used to balance the contributions of features, time, and space dimensions in distance calculation. Optionally, in the early stages of knowledge graph construction, the confidence attributes of nodes and edges can be initialized based on the probability scores output by the model during the feature extraction stage, and updated during subsequent collaborative iterative optimization.

[0077] Example 2: See Figure 3 From the crack knowledge graph, nodes and edges marked as normal historical states are selected to form a normal state sample set. Using this normal state sample set as training data, a deep autoencoder network is constructed. The deep autoencoder network learns to compress and reconstruct the normal state sample set. The state reconstruction error is obtained by calculating the difference between the reconstructed data and the original input data. Gaussian process regression is used to model the distribution of the state reconstruction error over time, obtaining a behavioral confidence score representing the stability of the structure. The reconstructed data corresponding to a sample in the normal state sample set, output by the deep autoencoder network, and the corresponding original input data are obtained. Both the original input data and the reconstructed data are multi-dimensional feature vectors. Element-wise subtraction of the original input data vector and the reconstructed data vector yields the absolute error vector for each feature dimension. Statistical features of the original input data vector and the reconstructed data vector, including mean, variance, and higher-order moments, are calculated for each, and the relative differences between these statistical features are also calculated. The error values ​​of each dimension in the absolute error vector are weighted and summed to obtain a preliminary error scalar. Simultaneously, the relative differences between the statistical features are integrated to obtain another difference scalar. The initial error scalar and the difference scalar are linearly combined to generate the final state reconstruction error value used to characterize the accuracy of a single reconstruction.

[0078] In practical implementation, all nodes and connections marked as "normal historical state" are selected from the crack knowledge graph. These nodes and connections originate from periods when the pillar did not experience any crack alarms or significant performance degradation during historical monitoring cycles, forming a normal state sample set. Each sample in the normal state sample set corresponds to a composite data structure consisting of a historical fusion sensing feature vector and its contextual relationship in the graph. Using the normal state sample set as training data, a deep autoencoder network with a symmetrical structure is constructed. The encoder part of the deep autoencoder network consists of three fully connected layers, progressively compressing the 1200-dimensional fusion feature vector into a 50-dimensional latent space vector. The decoder part consists of another three fully connected layers, responsible for reconstructing the 1200-dimensional output vector from the latent space vector. The deep autoencoder network learns to efficiently compress and reconstruct the normal state sample set with high fidelity by minimizing the reconstruction loss function.

[0079] In practice, the state reconstruction error is obtained by calculating the difference between the reconstructed data and the original input data. The reconstructed data vector corresponding to a sample in the normal state sample set, output by the deep autoencoder network, and the corresponding original input data vector are obtained. Both the original input data vector and the reconstructed data vector are 1200-dimensional feature vectors. Element-wise subtraction is performed between the original input data vector and the reconstructed data vector to obtain the absolute error vector for each feature dimension, which is also 1200-dimensional. The statistical features of the original input data vector and the reconstructed data vector are calculated separately. These statistical features include the four higher-order moments: mean, variance, skewness, and kurtosis. The relative difference between the statistical features of the original input data vector and the reconstructed data vector is calculated by dividing the absolute difference of the corresponding statistical feature values ​​by the statistical feature values ​​of the original input data vector.

[0080] In some embodiments, a preliminary error scalar is obtained by weighted summation of the error values ​​of each dimension in the absolute error vector. During weighted summation, different weight coefficients are assigned to the dimensions of visual features from the base column surface and the dimensions of state features within the base column. Simultaneously, the relative differences between statistical features are integrated to obtain another difference scalar. This integration is achieved by taking the arithmetic mean of the four relative differences: mean, variance, skewness, and kurtosis. The preliminary error scalar and the difference scalar are then linearly combined to generate the final state reconstruction error value, which characterizes the accuracy of a single reconstruction. The coefficients of this linear combination are pre-calibrated based on the reconstruction error distribution of the normal state sample set. The specific calculation formula for the state reconstruction error value is as follows:

[0081]

[0082] in: This represents the state reconstruction error value. The total dimension of the feature vectors is 1200. Represents the original input data vector of the th eigenvalues ​​of dimension Represents the reconstructed data vector. eigenvalues ​​of dimension It is the first Preset weight coefficients corresponding to the dimensional features and These represent the first and second halves of the original input data vector and the reconstructed data vector, respectively. One statistical characteristic value, The values ​​from 1 to 4 correspond to the mean, variance, skewness, and kurtosis, respectively. and These are preset combination coefficients used to balance the contributions of absolute error and statistical difference. In some embodiments, Gaussian process regression is used to model the distribution of state reconstruction error in the time dimension. The Gaussian process regression model is trained based on the state reconstruction error value sequence and its timestamps corresponding to historical normal state samples. The Gaussian process regression model outputs a behavioral confidence score that represents the uncertainty of future state reconstruction error prediction. The behavioral confidence score and the state reconstruction error together constitute the pillar structure behavioral model. Optionally, the training process of the deep autoencoder network and the training process of the Gaussian process regression model can be performed alternately. The state reconstruction error sequence is recalculated using the updated deep autoencoder network, and the Gaussian process regression model is retrained using the new state reconstruction error sequence.

[0083] Example 3: A training dataset for Gaussian process regression is constructed using timestamps as input variables and the state reconstruction error sequence as observed variables. The squared exponential kernel function is chosen as the covariance function, and its length scaling and variance parameters are initialized. The hyperparameters of the Gaussian process regression model are optimized by maximizing the marginal likelihood function, resulting in a trained Gaussian process regression model. The trained Gaussian process regression model is used to predict the state reconstruction error at the current time and in future time windows, yielding the predicted mean and predicted variance. The reciprocal of the predicted variance is normalized and mapped to the range of zero to one, serving as the behavioral confidence score, where a higher confidence score indicates smaller fluctuations in the state reconstruction error.

[0084] In practice, a training dataset for Gaussian process regression is constructed using timestamps as input variables and state reconstruction error sequences as observed variables. The timestamps are continuous time point sequences, and the state reconstruction error sequences are a series of state reconstruction error values ​​arranged chronologically after being reconstructed from historical normal state sample sets by a deep autoencoder network. The squared exponential kernel function is chosen as the covariance function of the Gaussian process regression model. The squared exponential kernel function can characterize the smooth changes in state reconstruction errors over time. The length scale parameter and variance parameter of the squared exponential kernel function are initialized; the length scale parameter controls the smoothness of the function's changes, and the variance parameter controls the range of the function's output amplitude. The hyperparameters of the Gaussian process regression model are optimized by maximizing the marginal likelihood function. The marginal likelihood function represents the probability of model parameters given the input timestamps and observed state reconstruction error sequences. The gradient ascent algorithm is used to iteratively adjust the length scale parameter and variance parameter until the marginal likelihood function value converges to its maximum, resulting in a trained Gaussian process regression model. The Gaussian process regression model is used to predict the state reconstruction error at the current time and in future time windows. The input is the timestamps of the current and future time points. The Gaussian process regression model outputs the predicted mean and predicted variance of the state reconstruction error at each time point. The predicted mean represents the most likely state reconstruction error value, and the predicted variance represents the degree of uncertainty in the prediction.

[0085] In some embodiments, the reciprocal of the prediction variance is normalized to a range of zero to one and used as the behavioral confidence score. Specifically, for a future point in time, the prediction variance is... Calculate its reciprocal Then, the min-max normalization method is used to calculate the reciprocals of all time points to be evaluated. The values ​​are scaled to the interval [0,1], and the scaled values ​​are defined as the behavioral confidence level at that time point. It can be understood that a higher confidence level indicates that the Gaussian process regression model has less uncertainty in predicting the state reconstruction error at that time point, meaning the fluctuation of the state reconstruction error is smaller and the structural behavior is more stable; conversely, a lower confidence level indicates greater prediction uncertainty, and the state reconstruction error may fluctuate significantly. Optionally, the specific form of the squared exponential kernel function is as follows:

[0086]

[0087] in: Indicates a point in time With time point Covariance between It is the variance parameter, which controls the output amplitude of the covariance function. It is a length scale parameter that controls the rate at which the covariance decays with time intervals. and These are any two input time points. In some embodiments, the state reconstruction error sequence corresponding to the historical normal state sample set is divided into a training set and a validation set. The training set is used to optimize the hyperparameters of the Gaussian process regression model, and the validation set is used to evaluate the predictive performance of the optimized Gaussian process regression model to prevent overfitting. Optionally, the length of the future time window can be set according to actual monitoring needs, such as the next 24 hours or the next 7 days.

[0088] See Figure 4 This is a line graph showing the time distribution of errors in the reconstruction of the structural state of a foundation column, corresponding to the "training dataset construction stage" in the concrete crack identification process for building foundation columns. Its core purpose is to verify the rationality of the Gaussian process regression training data. The overall error fluctuates smoothly within the range of 0.02-0.08, without abrupt changes or outliers, consistent with the stability characteristics of the structural behavior of foundation columns under normal conditions. The uniform width of the error fluctuation range (light blue area) indicates strong consistency in structural behavior under normal conditions. This dataset satisfies the training assumptions of Gaussian process regression (smooth data, stable fluctuations) and can be used to learn the normal behavior patterns of foundation columns. These types of charts are pre-training validation tools for foundation column structural behavior models: by displaying the time distribution of state reconstruction errors, the rationality of the training data can be confirmed, providing a reliable input foundation for subsequent hyperparameter optimization and behavior confidence calculation of the Gaussian process regression model, ensuring that the model can accurately characterize the normal structural behavior of the foundation column.

[0089] Example 4: The fused sensing feature set generated at the current moment is input into a deep autoencoder network to obtain the real-time state reconstruction error at the current moment. The real-time state reconstruction error at the current moment is input into a Gaussian process regression model to calculate the real-time behavior confidence score at the current moment. Simultaneously, using the fused sensing feature set at the current moment as a query condition, a subgraph matching search is performed in the crack knowledge graph to find the historical state pattern with the highest similarity. The feature distance between the fused sensing feature set and the historical state pattern with the highest similarity is calculated, and combined with the real-time behavior confidence score, an anomaly score is generated through a weighted fusion algorithm. When the anomaly score exceeds a preset static threshold, or the real-time behavior confidence score is lower than a dynamically adjusted confidence threshold, it is determined as a preliminary anomaly event. The fused sensing feature set corresponding to the preliminary anomaly event is treated as a new node to be analyzed and attempted to be fused into the crack knowledge graph. In the crack knowledge graph, starting from the new node to be analyzed, multiple path searches are performed to identify existing historical nodes that may have a causal or related relationship with the anomaly event. The identified association paths and nodes, along with the confidence scores of the initial anomaly events, are submitted to the column structure behavior model. The column structure behavior model uses the submitted association path and node information to locally recalibrate the parameters of the Gaussian process regression model and update the calculation benchmark for the behavior confidence. The updated column structure behavior model is then re-used to evaluate the initial anomaly events, forming a collaborative iterative optimization loop between the crack knowledge graph and the column structure behavior model.

[0090] In practice, the fused sensing feature set generated at the current moment is input into a deep autoencoder network. The deep autoencoder network encodes and reconstructs the input fused sensing feature vector, outputting a real-time reconstructed data vector. The real-time state reconstruction error at the current moment is obtained by calculating the difference between the real-time reconstructed data vector and the original input fused sensing feature vector. The real-time state reconstruction error at the current moment is input into a pre-trained Gaussian process regression model. Based on the temporal correlation of the historical state reconstruction error sequence, the Gaussian process regression model calculates the real-time behavior confidence score, which characterizes the stability of the structural behavior prediction at the current moment. Simultaneously, using the fused sensing feature set at the current moment as the query condition, a subgraph matching search is performed in the crack knowledge graph. The subgraph matching search uses the crack feature vector extracted at the current moment and its associated spatiotemporal attributes as the query subgraph. The graph database is searched for the existing historical state pattern subgraph with the most similar topology and node attributes, and the historical state pattern with the highest similarity and its feature vector set are returned.

[0091] The system calculates the feature distance between the fused sensing feature set and the historical state pattern with the highest similarity. This feature distance calculation integrates the Euclidean distance between feature vectors of nodes, timestamp differences, and spatial location differences, and combines this with real-time behavior confidence obtained from a Gaussian process regression model. An anomaly score is generated through a weighted fusion algorithm. The weighted fusion algorithm linearly combines the feature distance and real-time behavior confidence; the larger the feature distance and the lower the real-time behavior confidence, the higher the generated anomaly score. When the anomaly score exceeds a preset static threshold, or the real-time behavior confidence falls below a dynamically adjusted confidence threshold, the system determines that a preliminary anomaly event has occurred. The preliminary anomaly event includes the time, location, anomaly score, and the specific condition type that triggered the determination. Table 1 shows the weight coefficient configuration used in one type of weighted fusion algorithm for anomaly score calculation.

[0092] Table 1: Weighting Coefficient Configuration Table for Anomaly Scoring Calculation

[0093]

[0094] The formula for calculating the anomaly score is as follows:

[0095]

[0096] in: This represents the calculated anomaly score. Represents the normalized feature distance. Indicates the confidence level of real-time behavior. and These are preset weighting coefficients. It is a positive number. A negative value indicates that an increase in feature distance will increase anomaly scores, while a decrease in behavioral confidence will also increase anomaly scores.

[0097] The logical condition for preliminary abnormal event determination is: if or If it is established, a preliminary abnormal event is triggered, in which It is a static threshold. It is a dynamically adjusted confidence threshold.

[0098] In practice, the system receives preliminary anomaly events and confidence scores, triggering collaborative iterative optimization of the crack knowledge graph and the column structure behavior model. The fused perceptual feature set corresponding to the preliminary anomaly event is treated as a new node to be analyzed and attempted to be integrated into the crack knowledge graph. Within the crack knowledge graph, starting from the new node to be analyzed, multiple path searches are performed. These searches are conducted using a breadth-first traversal along the directed edges between nodes, identifying existing historical nodes within a preset hop count that may have causal or related connections to the anomaly event node. These connections include temporal sequence associations, spatial proximity associations, or feature similarity associations. The found associated paths and nodes, along with the confidence score of the preliminary anomaly event, are submitted to the column structure behavior model.

[0099] The column structure behavior model uses the submitted association paths and node information to locally recalibrate the parameters of the Gaussian process regression model. This local recalibration process involves adding the state reconstruction error data of historical nodes associated with the new nodes to the training set of the Gaussian process regression model with higher weights, re-optimizing the hyperparameters of the model, and updating the calculation benchmark for behavioral confidence based on the optimized model. Essentially, the updated column structure behavior model is reused to evaluate initial anomalies, the updated model is used to recalculate the real-time behavioral confidence at the corresponding moment of the anomaly, and the anomaly score is recalculated in conjunction with the updated association information in the knowledge graph, forming a collaborative iterative optimization loop between the crack knowledge graph and the column structure behavior model. Optionally, after one collaborative iterative optimization loop, if the judgment conditions for the initial anomaly event are no longer met, the label of the anomaly event node in the crack knowledge graph may be corrected; if the judgment conditions are still met, the node and its association information will be retained and used for subsequent analysis. In some embodiments, a dynamic confidence threshold is used. It is not a fixed value, but rather dynamically adjusted based on the historical average of the recent prediction variance of the Gaussian process regression model. The dynamic confidence threshold is adjusted when the overall uncertainty of the model's recent predictions increases. This will be reduced accordingly to accommodate a wider range of normal fluctuations. Optionally, the upper limit for the number of hops in association path finding is typically set to three to five hops to balance computational complexity with the comprehensiveness of association discovery.

[0100] Example 5: Continuously collect the state reconstruction error sequence generated by the structural behavior model of the column base over a continuous monitoring period. The state reconstruction error sequence is segmented, and the statistical and trend features of each segment are extracted. A multi-scale time series clustering algorithm is applied to perform unsupervised clustering analysis on all extracted features to identify recurring error change patterns with specific morphologies. Each identified error change pattern is defined as a typical structural behavior pattern, and a feature template and trigger condition library are established for each pattern. The associated paths found in the crack knowledge graph are obtained. The attribute change gradients of each node on the associated paths are analyzed, especially the evolution direction and rate of the crack feature vector. Based on the attribute change gradients, the direction of the most drastic feature evolution is traced backward to locate the most likely origin node of the initial abnormal event in the spatiotemporal dimension. Combining the sensor readings and image features of the origin node, the initial cause type leading to crack initiation is inferred, and the origin node and initial cause type are used as the source tracing conclusion and output to the deep analysis report.

[0101] In practice, the state reconstruction error sequence generated by the column structure behavior model is continuously collected over a continuous monitoring period of thirty consecutive calendar days. The column structure behavior model outputs one state reconstruction error value per hour, forming a sequence containing 720 state reconstruction error values ​​arranged in chronological order. The state reconstruction error sequence is segmented, with each segment consisting of 24 error values ​​(one day's worth of data). The entire sequence is divided into 30 segments. Statistical and trend characteristics are extracted from each segment. The statistical characteristics include the segment's mean, standard deviation, skewness, and kurtosis. The trend characteristics are obtained by linearly fitting the error values ​​within the segment to obtain the slope and intercept of the fitted line, and by calculating the root mean square of the difference between adjacent error values ​​within the segment. A multi-scale time series clustering algorithm is applied to perform unsupervised clustering analysis on all extracted features. The multi-scale aspect is reflected in the simultaneous consideration of feature aggregation at the hour, day, and week levels. First, the feature vectors extracted from each segment are standardized. Then, the dynamic time warping algorithm is used to calculate the similarity distance of the time series morphology of different segments. Finally, the k-means clustering algorithm is combined to perform comprehensive clustering of segment feature vectors and morphological distances, identifying recurring error change patterns with specific morphologies.

[0102] In practical implementation, the dynamic time warping algorithm used in multi-scale time series clustering algorithms is employed to measure the morphological similarity between two time series segments. The dynamic time warping algorithm does not require the two sequences to be exactly the same length; it uses a non-linear warping path to slightly stretch one sequence along the time axis, thereby finding the optimal morphological correspondence with the other sequence. In the application scenario described in this embodiment, the sequence segments to be compared are 24-hour state reconstruction error value sequence segments, each containing 24 error values ​​arranged in chronological order. The input to the dynamic time warping algorithm is two such sequence segments, denoted as sequence A, sequence B, and sequence C. and sequence In specific examples, and Both are twenty-four. The algorithm first constructs a matrix of size twenty-four. The distance matrix, in which the nth Line 1 Column elements It is calculated as the absolute difference between the values ​​at corresponding points in the two sequences, i.e. Subsequently, the dynamic time warping algorithm searches for a path in the distance matrix starting from the lower left corner of the matrix. Go to the top right corner Regular path Each element in the path Represents a sequence The Points and Sequences The Points must match. This regular path needs to satisfy boundary conditions, continuity conditions, and monotonicity conditions: the path must start from the starting point and end at the ending point; adjacent steps on the path must be adjacent matrix elements (including diagonally adjacent); and the path must have matching points. and The index must monotonically remain constant over time. The algorithm aims to find a regular path that minimizes the cumulative distance, defined as the distance between all matching points on the path. The sum of all. The final dynamic time-warped distance. This is the minimum cumulative distance. The smaller this distance value, the more similar the two time series segments are in terms of morphology. It is used as the core metric for measuring the similarity between sequences in subsequent cluster analysis.

[0103] The process involves retrieving association paths from the crack knowledge graph. These paths are sequences of nodes connected by directed edges, obtained through multi-hop searches starting from the anomaly node in the crack knowledge graph after initial anomaly event detection. The analysis focuses on the attribute change gradients of each node along these paths, particularly the evolution direction and rate of crack feature vectors. The evolution direction is determined by comparing the signs of the differences in feature vectors between adjacent nodes along the path, while the evolution rate is obtained by dividing the feature vector difference by the time difference between nodes. For a given sequence of nodes... The associated path, nodes arrive In feature dimension gradient of change Defined as:

[0104]

[0105] in: and Representing nodes respectively and The eigenvectors in the th Values ​​in the dimension and These are nodes and The timestamps are used to compute the gradient of change across all key feature dimensions for all node pairs along the path. By taking the absolute value and averaging, we can obtain the average change intensity of the entire path.

[0106] By tracing the gradient of attribute changes backwards along the direction of the most dramatic feature evolution, the most likely origin node of the initial anomaly in the spatiotemporal dimension is located. Specifically, this involves tracing back along the associated path from the initial anomaly node to earlier nodes, comparing the average change intensity between each adjacent node segment, and identifying the interval where the feature mutation occurred along the path between nodes with the highest average change intensity. The starting node of this interval is marked as the most likely origin node. The initial cause type leading to crack initiation is inferred by combining the sensor readings and image features of the origin node. For example, if the sensor readings of the origin node show a sudden drop in temperature and a sudden increase in humidity, while the image features do not show obvious cracks, the initial cause type is inferred to be "freeze-thaw cycle stress"; if the vibration sensor readings of the origin node show a sudden increase in energy at a specific frequency, the initial cause type is inferred to be "external impact load". In some embodiments, the origin node and the initial cause type are used as the source tracing conclusion and output to a deep analysis report. The deep analysis report records the source tracing time, spatial location, associated path summary, and inferred initial cause in a structured format. Optionally, for smooth evolution paths with average change intensity below a preset threshold, the origin node may not be located, but the initial cause type may be marked as "long-term cumulative effect".

[0107] See Figure 5 This is a time-series monitoring chart of the knowledge graph node scores for column cracks, corresponding to the "two-way feature comparison and anomaly scoring" stage in the concrete crack identification process for building column foundations. Its core purpose is to monitor the crack risk status of column foundation nodes in real time. The crack scores generally fluctuate between -2.0 and 1.5, mostly remaining within the safe range of near 0. Only green dots (normal nodes) appear in the chart, with no suspicious / abnormal nodes, indicating that the crack risk of column foundation nodes is at a safe level during the monitoring period. Significant fluctuations in scores occurred at some times, but did not exceed the risk threshold; these may be temporary fluctuations caused by environmental interference (such as vibration). This type of chart is a visualization tool for real-time monitoring of column cracks—by displaying the crack scores of knowledge graph nodes in a time sequence, it can quickly identify the crack risk status (normal / suspicious / abnormal) of the column foundation, providing real-time risk information for subsequent "abnormal event triggering" and "co-optimization of knowledge graph and behavioral models," and assisting in the timely detection of potential crack hazards in the column foundation.

[0108] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying a concrete crack based on a building foundation pillar, characterized by, The method comprises: Obtaining a multi-angle image sequence of the base column surface and a multi-type sensor reading of the inside of the base column, performing time alignment and feature decoupling on the multi-angle image sequence and the multi-type sensor reading to form a fusion perception feature set; Performing multi-scale crack feature extraction and semantic association on the fusion perception feature set to construct a crack knowledge graph with spatiotemporal correlation; Based on the historical normal state mode in the crack knowledge graph, a base column structure behavior model is trained and generated, which outputs a behavior confidence and a state reconstruction error; Based on the historical normal state mode in the crack knowledge graph, a base column structure behavior model is trained and generated, which outputs a behavior confidence and a state reconstruction error; From the crack knowledge graph, nodes and connecting edges marked as normal historical states are selected to form a normal state sample set; A deep autoencoder network is constructed using the normal state sample set as training data, which learns to compress and reconstruct the normal state sample set; The state reconstruction error is obtained by calculating the difference between the reconstructed data and the original input data; The state reconstruction error is modeled in the time dimension using Gaussian process regression to obtain the behavior confidence representing the stability of the structure behavior, which together constitutes the base column structure behavior model; The fusion perception feature set generated in real time is input into the base column structure behavior model and the crack knowledge graph for bidirectional feature comparison and abnormal score calculation, and the preliminary abnormal event and the confidence score of the preliminary abnormal event are output; The fusion perception feature set generated in real time is input into the base column structure behavior model and the crack knowledge graph for bidirectional feature comparison and abnormal score calculation, which is implemented by the following steps: The fusion perception feature set generated at the current time is input into the deep autoencoder network to obtain the real-time state reconstruction error at the current time; The real-time state reconstruction error at the current time is input into the Gaussian process regression model to calculate the real-time behavior confidence at the current time; At the same time, the fusion perception feature set at the current time is used as a query condition to search for the historical state mode with the highest similarity in the crack knowledge graph; The feature distance between the fusion perception feature set and the historical state mode with the highest similarity is calculated, and the real-time behavior confidence is combined to generate the abnormal score through a weighted fusion algorithm; When the abnormal score exceeds the preset static threshold or the real-time behavior confidence is lower than the dynamically adjusted confidence threshold, a preliminary abnormal event is determined; The preliminary abnormal event and the confidence score are received to trigger the cooperative iterative optimization of the crack knowledge graph and the base column structure behavior model, and a deep analysis report containing a crack mode label and a traceability direction is generated.

2. The method of claim 1, wherein the method is based on a building foundation column. Obtaining a multi-angle image sequence of a base column surface and a multi-type sensor reading inside the base column, performing time alignment and feature decoupling on the multi-angle image sequence and the multi-type sensor reading to form a fusion perception feature set, specifically by the following steps: Synchronizing the image acquisition device on the surface of the building base column and the sensing device buried in the base column, respectively obtaining image frame sequences and physical quantity reading sequences with uniform timestamps; Performing texture enhancement and region segmentation processing on the image frame sequence to extract base column surface visual feature vectors containing gray distribution, edge gradient and local binary pattern; Performing noise reduction and normalization processing on the physical quantity reading sequence to extract base column internal state feature vectors representing strain, vibration and humidity change; Mapping and aligning the base column surface visual feature vectors and the base column internal state feature vectors according to their uniform timestamps to generate the fusion perception feature set.

3. The method of claim 2, wherein the method is based on a building foundation column. Performing multi-scale crack feature extraction and semantic association on the fusion perception feature set to construct a crack knowledge graph with spatiotemporal correlation, specifically by the following steps: Separate the sub-feature sets belonging to different time segments and different spatial regions from the fusion perception feature set; Apply micro-scale and macro-scale feature extraction operators to each sub-feature set to identify potential crack initiation features, crack propagation features and crack stability features; Take all the identified crack features as nodes and establish directed connection edges between the nodes according to the co-occurrence relationship, temporal relationship and spatial adjacency relationship between the features; Store the nodes and directed connection edges in a graph database to form the crack knowledge graph, and attach time attributes and confidence attributes to each node and connection edge.

4. The method of claim 1, wherein the method is based on a building foundation column. The state reconstruction error is obtained by calculating the difference between the reconstructed data and the original input data, specifically by the following steps: Obtain the reconstructed data output by the deep autoencoder network corresponding to one sample in the normal state sample set, and the original input data corresponding to the sample, both of which are multi-dimensional feature vectors; Subtract the original input data vector and the reconstructed data vector element by element to obtain an absolute error vector in each feature dimension; Calculate the statistical features of the original input data vector and the reconstructed data vector, including mean, variance and higher-order moments, and calculate the relative difference between the statistical features; Weighted sum the error values in each dimension of the absolute error vector to obtain a preliminary error scalar, and integrate the relative difference between the statistical features to obtain another difference scalar; Linearly combine the preliminary error scalar and the difference scalar to generate the state reconstruction error value finally used to represent the single reconstruction accuracy.

5. The method of claim 1, wherein the method is based on a building foundation column. Receive the preliminary abnormal event and confidence score to trigger the collaborative iterative optimization of the crack knowledge graph and the base column structure behavior model, specifically by the following steps: The fusion perception feature set corresponding to the preliminary abnormal event is fused into the crack knowledge graph as a new node to be analyzed; In the crack knowledge graph, a plurality of path searches are performed starting from the new node to be analyzed to identify existing historical nodes that may have a cause-effect or correlation with the abnormal event; The identified associated paths and nodes are submitted to the base column structure behavior model along with the confidence score of the preliminary abnormal event; The base column structure behavior model recalibrates the parameters of the Gaussian process regression model based on the submitted associated path and node information and updates the calculation basis for the behavior confidence; The recalibrated base column structure behavior model is used to reevaluate the preliminary abnormal event, forming a collaborative iterative optimization cycle of the crack knowledge graph and the base column structure behavior model.

6. The method of claim 1, wherein the method is based on a building foundation column. The Gaussian process regression is used to model the distribution of the state reconstruction error in the time dimension to obtain the behavior confidence representing the stability of the structure behavior, which is achieved by the following steps: A training data set for the Gaussian process regression is constructed by taking the timestamp as the input variable and the state reconstruction error sequence as the observed variable; A square exponential kernel function is selected as the covariance function, and the length scale parameter and variance parameter of the kernel function are initialized; The hyperparameters of the Gaussian process regression model are optimized by maximizing the marginal likelihood function to obtain a trained Gaussian process regression model; The trained Gaussian process regression model is used to predict the state reconstruction error at the current time and in the future time window to obtain the predicted mean and predicted variance; The reciprocal of the predicted variance is normalized to a range of zero to one and used as the behavior confidence, where a higher confidence indicates smaller fluctuations in the state reconstruction error.

7. The method of claim 4, wherein the method is based on a building foundation column. The method further includes: The state reconstruction error sequence generated by the base column structure behavior model over a continuous monitoring period is continuously collected; The state reconstruction error sequence is segmented and the statistical features and trend features of each error sequence are extracted; All extracted features are subjected to unsupervised clustering analysis using a multi-scale time series clustering algorithm to identify error change patterns that repeatedly occur and have specific morphologies; Each identified error change pattern is defined as a typical structure behavior pattern, and a feature template and trigger condition library are established for each structure behavior pattern.

8. The method of claim 1, wherein the method is based on a building foundation column. The method further includes: The associated path found in the crack knowledge graph is obtained; The attribute change gradient of each node in the associated path is analyzed, particularly the evolution direction and rate of the crack feature vector; The most likely origin node of the preliminary abnormal event in the spatio-temporal dimension is located by tracing the direction of the most dramatic feature evolution based on the attribute change gradient; The initial cause type leading to crack initiation is inferred based on the sensor readings and image features of the origin node, and the origin node and initial cause type are output as the traceability conclusion to the deep analysis report.

Images