Concrete bridge crack abnormity intelligent monitoring and early warning method based on GAF image classification and GRU prediction

Abstract

The invention relates to the technical field of civil engineering structure health monitoring, in particular to a concrete bridge crack abnormity intelligent monitoring and early warning method based on GAF image classification and GRU prediction.The method comprises the steps that an intelligent monitoring framework integrating GAF image coding, CNN classification and recognition and GRU time sequence predication is constructed, high-frequency noise is removed through wavelet packet decomposition, and then the GRU time sequence predication is carried out; smoothing the crack-temperature coupling time sequence data; encoding the image into a two-dimensional image through a GAF method, and enabling the CNN to recognize an abnormal mode; and training a GRU model based on the high-quality data set after abnormity elimination, and realizing accurate modeling of crack width evolution under temperature driving. According to the method, residual error approximate normal distribution is predicted, the crack width early warning decision coefficient (R) is stabilized to be more than 0.93, a + / -3 sigma dynamic residual error threshold early warning mechanism is combined, structural damage trends under different disturbance scenes can be identified in a graded mode, and the method is suitable for online monitoring and maintenance decision support of bridge crack diseases in actual engineering.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural health monitoring technology, specifically to an intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction. Background Technology

[0002] As urban infrastructure ages, cracks in concrete bridges, a typical indicator of structural degradation, directly impact the bridge's load-bearing capacity and service life. Current crack monitoring relies primarily on manual inspections and traditional sensor data collection, methods that are time-consuming, have slow response times, and are ill-suited for real-time health assessments under complex operating conditions during bridge service. Crack width is influenced by a combination of physical factors, such as ambient temperature, humidity, and traffic loads, exhibiting significant nonlinearity and strong nonstationarity in its response characteristics. Traditional statistical models have clear limitations in feature extraction, anomaly removal, and multi-source coupling modeling. Furthermore, monitoring is often accompanied by data drift, missing data, noise disturbances, and sensor malfunctions, leading to frequent false alarms and missed alarms, severely affecting the stability and reliability of structural health monitoring systems.

[0003] In recent years, deep learning technology has gradually demonstrated superior performance in the field of structural health monitoring. Convolutional Neural Networks (CNNs) have significant advantages in extracting complex spatial features, while Gated Recurrent Units (GRUs) have shown good accuracy and generalization ability in temporal modeling. However, there is currently a lack of a complete solution that organically combines anomaly identification, data cleaning, temporal modeling, and early warning mechanisms. For the dynamic mapping modeling problem of temperature-cracks, existing research mostly builds prediction models based on global data, ignoring the destructive impact of outliers on model performance. Therefore, this invention proposes an integrated framework that combines GAF image transformation, CNN anomaly classification, GRU temporal modeling, and residual dynamic threshold early warning. This framework systematically improves crack prediction accuracy and monitoring robustness, and is suitable for intelligent condition assessment and crack early warning tasks of multi-condition bridge structures in practical engineering. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent monitoring and early warning method for crack anomalies in concrete bridges based on GAF image classification and GRU prediction, in order to solve the problems mentioned in the background art, such as insufficient crack prediction accuracy, limited anomaly identification capability, and lack of data cleaning and early warning linkage mechanism in the existing technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for intelligent monitoring and early warning of crack anomalies in concrete bridges based on GAF image classification and GRU prediction includes the following steps:

[0007] S1 Data Acquisition and Preprocessing: Time-series data of crack width and ambient temperature at key parts of the bridge structure were collected, with a sampling frequency of once every 10 minutes; wavelet packet decomposition was performed on the raw data to extract the main components of temperature-induced crack response, high-frequency noise interference was removed, and all feature sequences were normalized to unify the dimensions and scale, thereby improving the stability of the model.

[0008] S2 Time Series Image Conversion: The normalized crack time series is converted into a two-dimensional image using Gramian Angular Field (GAF) technology, forming a GASF or GADF image that retains the time autocorrelation features, providing visual input data for subsequent anomaly identification models;

[0009] S3CNN Image Classification and Anomaly Detection: A convolutional neural network model based on GAF image input is constructed, trained, and used to identify crack development patterns. Classification labels include NSWP (normal cycle pattern), ASWP (abnormal cycle pattern), Drift (drift pattern), Square (step mutation), and Missing (missing data). The classification accuracy can reach over 98%. Identified abnormal data channels will be removed from subsequent modeling data to ensure the quality of model training data.

[0010] S4GRU Prediction Model Construction: Based on high-quality crack-temperature coupled data after outlier removal, a GRU neural network model is constructed. The input features include: raw temperature, moving average temperature (MA3), and normalized time index. The output is the predicted crack width value at the corresponding time. The GRU model consists of two stacked layers, each containing 128 neurons. The dropout rate is set to 0.2, and the optimizer is trained using the Adam algorithm.

[0011] S5 Residual Dynamic Threshold Early Warning Mechanism: Construct a dynamic threshold early warning mechanism based on the predicted residual sequence; update the mean and standard deviation of the prediction error through a sliding statistical window (24 hours), and set an adaptive anomaly threshold of ±3σ accordingly; when the predicted residuals of multiple consecutive data points exceed this range, different levels of risk warning signals are triggered, providing a basis for bridge structure safety assessment and maintenance decisions.

[0012] Preferably, the data collected in step S1 covers multiple measurement points and multiple seasonal operating conditions. The established dataset is processed by wavelet packet to form a high-fidelity temperature crack coupling feature sequence, which has unified dimensions and strong cross-regional adaptability. The ambient temperature data is simultaneously extracted with moving average features. The data is reconstructed according to the time series sliding window principle, with a window length of 24 hours, i.e., 144 time steps, to generate input-output sample pairs in the form of a three-dimensional array. The first dimension of the three-dimensional array is the number of samples, the second dimension is the time step length, and the third dimension is the input / output feature dimension.

[0013] Preferably, the normalization process in step S1 uses the Min-Max scaling method, mapping all features to the interval between -1 and 1, satisfying the formula:

[0014] ;

[0015] in, These are the original eigenvalues. This is the minimum value of the feature. This is the maximum value of this feature.

[0016] Preferably, step S2 introduces a sliding time window and a periodic reconstruction mechanism to segment the crack data, so that the GAF image can better reflect the influence of daily cycles and seasonal temperature changes on crack evolution and improve the accuracy of subsequent classification. When generating the GAF image, the crack time series is first normalized to the interval of -1 to 1 and converted into polar coordinates. Then, the angle matrix is ​​constructed using the cosine function, and the angle inner product relationship between any pair of time points is calculated to generate a symmetrical two-dimensional image matrix. The image size is uniformly 144×144 pixels. During the generation process, interpolation smoothing and noise filtering are used, and the GAF image is re-encoded according to the daily cycle in combination with the periodic reconstruction strategy.

[0017] As a preferred approach, the crack response patterns in step S3 are divided into five categories. Among them, the Drift and Square categories expand the number of samples through data augmentation and image synthesis methods to alleviate the sample imbalance problem. The CNN model used adopts a dual convolutional layer structure combined with Softmax classification output. The convolutional neural network classifier in step S3 includes two convolutional layers, two max pooling layers, one fully connected layer, and a Softmax output layer. The convolutional kernel size is 3×3, the activation function is ReLU, and the dropout rate is 0.2. The loss function during training is cross-entropy loss, the optimizer is Adam, the initial learning rate is 0.001, the batch size is 64, and the number of training epochs is 100. For the few sample categories, GAF graphs with noise are used for data augmentation.

[0018] Preferably, in step S3, accuracy, precision, recall, and F1 score are used as performance evaluation metrics for model classification. Here, let TP be the number of correctly predicted samples in a given class, FP be the number of samples that are incorrectly predicted as belonging to that class, and FN be the number of samples that are incorrectly predicted as belonging to another class. Then:

[0019] The formula for calculating accuracy is: ;

[0020] The formula for calculating accuracy is: ;

[0021] The formula for calculating recall rate is: ;

[0022] The formula for calculating the F1 score is: .

[0023] Preferably, in step S4, the gated recurrent unit neural network model adopts a two-layer stacked structure, with each layer containing 128 neurons. Dropout layers are set between layers with a dropout rate of 0.2. The training uses the Adam optimizer with an initial learning rate of 0.0001 and a loss function of root mean square error. An EarlyStopping strategy is also introduced. The training data is constructed using a sliding window strategy with a window length of 144 time steps and a step size of 1.

[0024] Preferably, the model performance evaluation metrics in step S4 include the coefficient of determination R², root mean square error RMSE, mean absolute error MAE, and the interval prediction evaluation metric Winkler score SWS; let... This represents the true value of the crack width. This is the predicted crack width. The average of the true values ​​is n, where n is the sample size. For confidence level, and Confidence levels The upper and lower limits of the prediction interval If the true value is:

[0025] The formula for calculating the coefficient of determination is:

[0026] ;

[0027] The formula for calculating the root mean square error is:

[0028] ;

[0029] The formula for calculating the mean absolute error is:

[0030] ;

[0031] The Winkler score is calculated using the following formula:

[0032] .

[0033] As a preferred embodiment, in step S5, the predicted residual is defined as the difference between the predicted crack width and the measured value. The mean μ and standard deviation σ of the residual are updated in real time through a 24-hour sliding window, and the anomaly discrimination interval is set to [μ-3σ, μ+3σa]. The early warning mechanism supports multi-point deployment and multi-channel fusion. After each monitoring channel is independently evaluated, a structural overall status assessment report is generated through weighted fusion. The early warning information can be pushed to the maintenance personnel terminal and automatically record the abnormal time, severity, and trend indicator metadata.

[0034] As a preferred option, the early warning mechanism constructed in step S5 dynamically updates the early warning threshold range based on the residual distribution, and sets the risk classification standard as follows: not exceeding the limit is normal, a single sudden change is an early warning, and continuous abnormality is severe, which is used to support bridge health assessment and operation and maintenance response.

[0035] This invention adopts the above technical solution, based on the engineering needs of monitoring and identifying crack evolution in concrete bridges, aiming to provide an intelligent solution for structural condition perception, defect identification, and safety early warning. Firstly, based on long-term crack and temperature monitoring data from actual bridges, wavelet packet decomposition is used to extract the dominant temperature-influencing component, and Gramian Angular Field (GAF) technology is employed to convert the time series into a two-dimensional image to enhance the pattern recognition capability of the crack data. Combined with a constructed convolutional neural network (CNN) classifier, high-precision identification of typical crack response patterns such as NSWP, ASWP, and Drift is achieved, eliminating abnormal data channels and improving the purity and effectiveness of subsequent modeling data. Subsequently, a GRU neural network structure is designed to construct a crack width time-series prediction model based on temperature and time features, forming a high-dimensional nonlinear mapping mechanism of temperature-driven – structural response. Experimental results show that the proposed method, on a typical bridge crack dataset, can stably maintain a prediction sample determination coefficient R² above 0.93 and a root mean square error below 0.0025 mm, significantly outperforming traditional machine learning models; among them, the GRU-based model maintains good fitting performance even under complex temperature changes and data perturbation conditions. By combining residual statistical characteristics, a ±3σ dynamic threshold early warning mechanism is further constructed to realize real-time anomaly identification and risk-level response in the crack evolution process.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] This invention breaks through the bottlenecks of traditional models in handling temporal anomalies, modeling crack evolution, and responding to early warnings by introducing a combined framework of GAF image conversion, CNN anomaly detection, and GRU temporal modeling. It provides high-precision and robust technical support for crack monitoring of concrete bridges under complex working conditions, and also provides theoretical basis and methodological reserves for the intelligent upgrading and widespread engineering promotion of structural health monitoring systems. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0039] Figure 1 This is a flowchart illustrating the overall process of an intelligent monitoring and early warning method for crack anomalies in concrete bridges based on GAF image classification and GRU prediction, according to the present invention.

[0040] Figure 2 This invention relies on a monitoring platform based on actual monitoring data and on-site images;

[0041] Figure 3 This is a schematic diagram of the GAF encoding method for crack time-series data in this invention;

[0042] Figure 4 This is a schematic diagram of the basic CNN classification model structure of this invention;

[0043] Figure 5 This is a diagram showing the classification results of the basic CNN model used in this invention.

[0044] Figure 6 This is a schematic diagram of the neural network model based on the gated recurrent unit (GRU) of this invention;

[0045] Figure 7 This is a graph showing the prediction results of the GRU prediction model of this invention;

[0046] Figure 8 This is a typical data set for the present invention, along with the model prediction residuals and probability distribution. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0048] The following is a detailed description of the embodiments in conjunction with the accompanying drawings:

[0049] In embodiments of the present invention, such as Figure 1 As shown, the steps of the intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction include:

[0050] [1] Step 1: Collect long-term monitoring data of crack width and corresponding ambient temperature of key parts of concrete bridge, with a sampling interval of 10 minutes. Use wavelet packet decomposition technology to extract low-frequency trend components from the raw data, remove high-frequency interference, and normalize the crack and temperature sequences to improve the stability and robustness of subsequent modeling.

[0051] [2] Step 2: Convert the processed crack time series into a two-dimensional image using the Gramian Angular Field (GAF) encoding method. Choose GASF or GADF encoding form according to the requirements to preserve the autocorrelation structure of the time series and provide clear and discriminative input feature maps for subsequent CNN image classification models.

[0052] [3] Step 3: Construct a convolutional neural network classifier based on GAF images to identify different patterns of crack response data. The model training adopts a supervised learning framework, and the classification labels include five response patterns: NSWP, ASWP, Drift, Square, and Missing. The model achieves an accuracy of over 98% on the test set. Abnormal samples identified will be removed to improve the data quality of subsequent prediction models.

[0053] [4] Step 4: Use the high-quality crack-temperature dataset after removing outlier data to train a gated recurrent unit (GRU) neural network model. The model input includes the original temperature, the moving average temperature and the normalized time index, and the output is the predicted crack width at the corresponding time. The GRU model adopts a two-layer structure with 128 units in each layer, and the dropout rate is set to 0.2 to avoid overfitting.

[0054] [5] Step 5: Construct a dynamic early warning method based on the predicted residual sequence. Use a 24-hour sliding window to update the statistical characteristics of the residual in real time, and set an adaptive anomaly discrimination interval according to the ±3σ principle. When the predicted residuals of multiple consecutive sampling points exceed the interval, different levels of early warning signals are automatically triggered to assist bridge structure operation and maintenance personnel in risk classification management and maintenance strategy formulation.

[0055] The working principle of this invention will be explained in detail below:

[0056] Regarding step one, such as Figure 2 As shown, in an embodiment of the present invention, the crack anomaly monitoring system relies on a long-term health monitoring platform for bridge structures to acquire time-series monitoring data of crack width and corresponding ambient temperature in key parts, which serves as the data basis for subsequent modeling.

[0057] Furthermore, the sampling frequency for crack width and ambient temperature is set to once every 10 minutes, covering at least a six-month period. This can encompass the combined effects of diurnal temperature differences, seasonal changes, and operational load fluctuations on structural cracks, ensuring that the dataset has long-term validity and representativeness.

[0058] Furthermore, in order to reduce the impact of high-frequency measurement noise and occasional interference in the monitoring data, wavelet packet decomposition (WPD) was used to decompose the original crack width sequence, retaining the low-frequency temperature-driven components and filtering out the high-frequency components, so as to highlight the slow-changing trend of crack evolution.

[0059] Furthermore, the ambient temperature data is also processed synchronously to extract its moving average features (such as MA3 and MA5), which are used to construct a combination of input features that are synchronous with the crack response, thereby improving the neural network's ability to fit the temperature-driven crack trend model.

[0060] Furthermore, all crack and temperature feature data must be normalized before being fed into the model. The normalization uses the Min-Max scaling method to map all features to the [-1,1] interval to eliminate the influence of different feature dimensions and scales, and to avoid gradient explosion or convergence difficulties during model training.

[0061] Furthermore, the data is reconstructed according to the time series sliding window principle to form a standardized training sample structure. The window length is set to 24 hours (144 time steps), and input / output sample pairs are generated using the sliding window method. These pairs are then formatted as a three-dimensional array and input into the GRU model, where the first dimension is the number of samples, the second dimension is the time step, and the third dimension is the input / output feature dimension.

[0062] Regarding step two, such as Figure 3 As shown, in an embodiment of the present invention, the preprocessed crack time series data is visualized and converted into a two-dimensional image using the Gramian Angular Field (GAF) encoding method, which is then used as the image classification input for subsequent deep learning models.

[0063] Furthermore, the GAF image generation process includes two common forms: Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF), which are used to highlight the additive correlation and differential structure between time points in the sequence, respectively. Users can switch or combine them according to the actual classification performance.

[0064] Furthermore, before encoding, the crack time series needs to be normalized to the [-1,1] interval and converted to polar coordinates. Then, the angle matrix is ​​constructed using the cosine function, and the inner product relationship between any two time points is calculated to generate a symmetrical two-dimensional image matrix, which fully preserves the autocorrelation and trend information of the original time series.

[0065] Furthermore, to improve the model's ability to recognize periodic patterns, this invention introduces a sliding time window mechanism to segment the crack time series. Each GAF image corresponds to a fixed time window (24 hours), and a sliding window strategy with a step size of 1 is used to expand the image data, thereby enhancing the model's temporal resolution.

[0066] Furthermore, this invention introduces a periodic reconstruction strategy. For crack responses with obvious typical daily periodic characteristics, the GAF image is re-encoded according to the daily cycle by combining the daily cycle curve of ambient temperature. This helps the classification model capture recurring features and potential abnormal patterns in the time domain.

[0067] Furthermore, to reduce the impact of image boundary effects and sampling errors, interpolation smoothing and noise filtering are used in the GAF image generation process, making the generated image structurally smoother and with clearer edges, which helps the subsequent CNN model extract deep spatial features.

[0068] Furthermore, the generated GAF ​​image is input into the CNN classification network in the form of a two-dimensional grayscale image, with a uniform image size of 144×144 pixels (the pixel size can be appropriately reduced if speed and accuracy need to be considered later). This image retains the overall trend of change, abrupt change nodes, and periodic information in the time series, providing a clear and identifiable feature basis for subsequent crack response pattern classification.

[0069] Regarding step three, such as Figure 4 As shown, in an embodiment of the present invention, a convolutional neural network (CNN) classification model is constructed, using GAF images as input, to achieve automatic identification and anomaly detection of crack response patterns in concrete bridges.

[0070] Furthermore, the crack response modes are classified into five categories in this invention: NSWP (Normal Seasonal Width Pattern), ASWP (Abnormal Seasonal Width Pattern), Drift, Square, and Missing. These classification criteria are defined based on the time-varying trend of crack response coupled with temperature, and have clear engineering physical implications.

[0071] Furthermore, the CNN model structure is as follows: Figure 4As shown, the network consists of two convolutional layers (Conv), two max-pooling layers (MaxPool), one fully connected layer (FC), and a softmax output layer. The kernel size is set to 3×3, the activation function is ReLU, and the dropout rate is 0.2 to enhance the network's generalization ability and reduce the risk of overfitting.

[0072] Furthermore, to address the class imbalance problem during training for few-sample classes (such as Drift and Square), this invention introduces a data augmentation mechanism, using methods such as adding noise to GAF graphs to increase the number of images for few-sample classes, thereby improving the model's recognition accuracy and robustness.

[0073] Furthermore, the training process adopts a supervised learning approach, with the loss function being cross-entropy loss, the optimizer being Adam, the initial learning rate being set to 0.001, the batch size being 64, and the number of training epochs being set to 100.

[0074] Furthermore, to evaluate the model's classification performance, accuracy, precision, recall, and F1 score are used as evaluation metrics. The specific calculation formulas are as follows:

[0075]

[0076]

[0077]

[0078] In the formula, for each category, TP represents the number of samples in that category that are correctly predicted to belong to that category. FP represents the number of samples that do not belong to that category but are incorrectly predicted to belong to that category. TN represents the number of samples that do not belong to that category but are correctly predicted to belong to another category. FN represents the number of samples that belong to that category but are incorrectly predicted to belong to another category.

[0079] In addition, the F1 score, used as a comprehensive classification assessment indicator, is the harmonic mean of Precision and Recall, and can be calculated using the following formula:

[0080]

[0081] Furthermore, the classification results of the basic CNN model can be found in... Figure 5 In multiple test sets, the CNN model maintained an overall classification accuracy of over 98% for the five types of crack response patterns. The model performed particularly well in ASWP and Drift category recognition, significantly outperforming traditional statistical methods.

[0082] Furthermore, after classification, the system marks the detected abnormal category data (including ASWP, Drift, Square, and Missing) as unstable sample channels and removes them in the subsequent modeling process, retaining only the normal channel data of the NSWP category for the GRU prediction modeling stage, thereby improving the data quality and stability of the prediction model.

[0083] Regarding step four, such as Figure 6 As shown, in an embodiment of the present invention, a gated recurrent unit (GRU) neural network model is constructed based on high-quality crack-temperature coupling data after anomaly removal processing to achieve time-series prediction of future crack width trends.

[0084] Furthermore, to improve the representativeness of the model's input features, the input vector of the GRU model includes the following three dimensions:

[0085] (1) Original ambient temperature value (Temp)

[0086] (2) Temperature moving average (MA3) is used to smooth temperature change trends and weaken short-term disturbances.

[0087] (3) Normalized Time Index (TimeIndex) is used to introduce time periodic factors.

[0088] Furthermore, the GRU network structure is as follows: Figure 6 As shown, the network consists of two stacked GRU units, each containing 128 neurons. A dropout layer with a dropout rate of 0.2 is placed between each layer to prevent overfitting. The network is finally mapped to the output space through a fully connected layer, outputting the predicted crack width for the corresponding time step.

[0089] Furthermore, the Adam optimizer is used to update network parameters during training, with an initial learning rate set to 0.0001. The loss function is the root mean square error (RMSE), and an EarlyStopping strategy is introduced: when the validation set loss does not improve within a certain number of consecutive epochs, the training process is automatically stopped, improving training efficiency and model generalization performance.

[0090] Furthermore, the training data employs a sliding window strategy to construct input-output sequences, with a window length of 144 time steps (corresponding to 24 hours) and a step size of 1, ensuring that each predicted sample contains sufficient historical information. All sequences are formatted as three-dimensional tensors before being input into the model, with dimensions representing the number of samples, the time step size, and the number of input features, respectively.

[0091] Furthermore, the model performance evaluation metrics include: coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). The specific formulas are as follows:

[0092]

[0093]

[0094]

[0095] In the formula, , and These represent the true value, the predicted value, and the average of the true values, respectively. R² is the maximum size metric, while smaller MAE and RMSE values ​​indicate better model performance.

[0096] Furthermore, in addition to evaluation metrics for point prediction results such as R2, MAE, and RMSE, interval prediction evaluation metrics should also be established to verify the model's impact on residual results from multiple perspectives. The comprehensive metric, Winkler score (WS) (denoted as SWS), which considers both the reliability and width of the prediction interval coverage, can be expressed as:

[0097]

[0098] In the formula, and They are respectively Upper and lower limits of the confidence level interval. This is the actual value.

[0099] Furthermore, Figure 7 The prediction results of the GRU prediction model are presented. In actual tests, the GRU model achieved an average R² value of over 0.93 on typical bridge monitoring data, with RMSE controlled within 0.0025 mm, verifying the high accuracy and stability of the method of this invention in crack trend prediction.

[0100] Furthermore, comparative experimental analysis revealed that, compared with the traditional LSTM model, the simplified gating structure of the GRU model outperforms the traditional LSTM model in terms of training efficiency and prediction accuracy. It has a significant advantage in processing long-sequence low-frequency trend modeling and is more suitable for modeling high-frequency small-amplitude fluctuation time series data in bridge crack monitoring scenarios.

[0101] Regarding step five, such as Figure 8 As shown in the embodiments of the present invention, data with a specific number is presented. Based on the residual between the output of the GRU prediction model and the actual monitoring data, a dynamic threshold early warning mechanism is constructed to achieve real-time identification and intelligent early warning of abnormal evolution of cracks in concrete bridges.

[0102] Furthermore, the prediction residual is defined as the sequence of differences between the model prediction value and the measured crack width value. The residual sequence is dynamically statistically analyzed using a sliding time window method, with the window length set to 24 hours (corresponding to 144 sampling points), and its mean (μ) and standard deviation (σ) are updated in real time.

[0103] Furthermore, the warning threshold is set based on the ±3σ statistical principle, that is, the residual within the range of [μ−3σ,μ+3σ] is considered to be within the normal fluctuation range. If the residual exceeds this range for multiple consecutive time points, it is considered to be an abnormal crack evolution trend and triggers the warning mechanism. At the same time, it has been verified for noise resistance.

[0104] Furthermore, to enhance the system's adaptability in actual engineering environments, this early warning mechanism supports the formation of a global risk map of bridge cracks based on multi-point deployment and multi-channel fusion. Each monitoring channel independently evaluates and generates local early warning results, and a weighted fusion is used to generate an overall structural status assessment report.

[0105] Furthermore, the early warning results can be integrated with the bridge maintenance system or IoT platform, supporting the real-time push of alarm information to maintenance personnel terminals via edge computing nodes, and automatically recording metadata such as abnormal time, severity, and trend indicators, facilitating subsequent data traceability and engineering handling decisions.

[0106] Furthermore, the residual threshold early warning mechanism and the GRU prediction model together form a closed-loop feedback system that can be continuously updated online and dynamically adjusted according to sensor deployment density, environmental changes, and historical performance, thereby improving the robustness and practicality of the system.

[0107] Comparative analysis shows that the proposed method for monitoring and early warning of crack anomalies in concrete bridges based on GAF image classification and GRU prediction exhibits significant advantages in handling complex nonlinear crack evolution data. First, wavelet packet decomposition and moving average strategies effectively extract the dominant temperature component, and combined with GAF image encoding, enhance the ability to express periodicity, abrupt changes, and trends in time series, overcoming the insufficient representational ability of traditional methods in high-dimensional time series modeling. Second, the CNN classification model accurately identifies and removes anomalous pattern data, significantly improving the input data quality of the crack prediction model. Furthermore, the constructed GRU neural network architecture possesses strong long-term dependency modeling capabilities, achieving high-precision prediction of crack width temporal changes. Experimental results show that the proposed method achieves a prediction sample determination coefficient (R²) exceeding 0.93 on a typical bridge monitoring dataset, with overall prediction errors controlled below the millimeter level, significantly outperforming traditional machine learning and statistical methods. Comparative analysis reveals that the simplified gated GRU model converges faster and has smaller errors in predicting crack development sections, while its performance is similar to LSTM in capturing non-periodic trend changes. In summary, this invention can effectively adapt to various crack evolution modes, has good adaptability and promotion value in practical engineering, and provides an efficient, stable and portable technical solution for bridge structural anomaly identification and intelligent early warning.

[0108] The five steps described above constitute the implementation process of this prediction method.

[0109] This invention adopts the above technical solution, based on the engineering needs of monitoring and identifying crack evolution in concrete bridges, aiming to provide an intelligent solution for structural condition perception, defect identification, and safety early warning. Firstly, based on long-term crack and temperature monitoring data from actual bridges, wavelet packet decomposition is used to extract the dominant temperature-influencing component, and Gramian Angular Field (GAF) technology is employed to convert the time series into a two-dimensional image to enhance the pattern recognition capability of the crack data. Combined with a constructed convolutional neural network (CNN) classifier, high-precision identification of typical crack response patterns such as NSWP, ASWP, and Drift is achieved, eliminating abnormal data channels and improving the purity and effectiveness of subsequent modeling data. Subsequently, a GRU neural network structure is designed to construct a crack width time-series prediction model based on temperature and time features, forming a high-dimensional nonlinear mapping mechanism of temperature-driven – structural response. Experimental results show that the proposed method, on a typical bridge crack dataset, can stably maintain a prediction sample determination coefficient R² above 0.93 and a root mean square error below 0.0025 mm, significantly outperforming traditional machine learning models; among them, the GRU-based model maintains good fitting performance even under complex temperature changes and data perturbation conditions. By combining residual statistical characteristics, a ±3σ dynamic threshold early warning mechanism is further constructed to realize real-time anomaly identification and risk-level response in the crack evolution process.

[0110] This invention breaks through the bottlenecks of traditional models in handling temporal anomalies, modeling crack evolution, and responding to early warnings by introducing a combined framework of GAF image conversion, CNN anomaly detection, and GRU temporal modeling. It provides high-precision and robust technical support for crack monitoring of concrete bridges under complex working conditions, and also provides theoretical basis and methodological reserves for the intelligent upgrading and widespread engineering promotion of structural health monitoring systems.

[0111] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring and early warning of crack anomalies in concrete bridges based on GAF image classification and GRU prediction, characterized in that, Includes the following steps: S1 Data Acquisition and Preprocessing: Time-series data of crack width and ambient temperature at key parts of the bridge structure were collected. The sampling frequency was set to once every 10 minutes. Wavelet packet decomposition was performed on the raw data to extract the main control component of temperature-induced crack response. After filtering out high-frequency noise, normalization was performed. S2 time series image conversion: The normalized crack time series is encoded into a two-dimensional image using Gramian Angular Field (GAF). The encoding method includes GASF or GADF to enhance the expressive power of time series features. S3CNN Image Classification and Anomaly Detection: Convolutional neural network model based on GAF image input is constructed to identify crack response pattern categories, including NSWP, ASWP, Drift, Square and Missing, and abnormal data channels are removed to improve the training quality of the model. S4GRU prediction model construction: Based on high-quality crack-temperature coupled data after cleaning, the input features include the original temperature, the moving average temperature and the normalized time index. A GRU model is constructed to predict crack width. A two-layer structure is adopted and the training is completed by combining Dropout and Adam optimization strategies. S5 Residual Dynamic Threshold Early Warning Mechanism: Based on the residual sequence of predicted and actual values, a ±3σ dynamic early warning threshold is constructed, which is dynamically updated using a 24-hour sliding statistical window, and different early warning responses are triggered by setting risk level standards.

2. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction as described in claim 1, characterized in that, The data collected in step S1 covers multiple measuring points and multiple seasonal working conditions. The dataset is processed by wavelet packets to form a high-fidelity temperature crack coupling feature sequence with unified dimensions and strong cross-regional adaptability. The ambient temperature data is synchronously extracted with moving average features. The data is reconstructed according to the time series sliding window principle, with a window length of 24 hours, or 144 time steps, to generate input-output sample pairs in the form of a three-dimensional array. The first dimension of the three-dimensional array is the number of samples, the second dimension is the time step length, and the third dimension is the input / output feature dimension.

3. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction as described in claim 1, characterized in that, The normalization process in step S1 uses the Min-Max scaling method to map all features to the interval between -1 and 1, satisfying the formula: ； in, These are the original eigenvalues. This is the minimum value of the feature. This is the maximum value of this feature.

4. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, In step S2, a sliding time window and periodic reconstruction mechanism are introduced to segment the crack data, so that the GAF image can better reflect the influence of daily cycle and seasonal temperature changes on crack evolution and improve the subsequent classification accuracy. When the GAF image is generated, the crack time series is first normalized to the interval of -1 to 1 and converted into polar coordinates. Then, the angle matrix is ​​constructed using the cosine function, and the angle inner product relationship between any pair of time points is calculated to generate a symmetrical two-dimensional image matrix. The image size is uniformly 144×144 pixels. During the generation process, interpolation smoothing and noise filtering are used, and the GAF image is re-encoded according to the daily cycle in combination with the periodic reconstruction strategy.

5. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, In step S3, the crack response patterns are divided into five categories. Among them, the Drift and Square categories expand the number of samples through data augmentation and image synthesis methods to alleviate the sample imbalance problem. The CNN model used adopts a double convolutional layer structure combined with Softmax classification output. The convolutional neural network classifier in step S3 includes two convolutional layers, two max pooling layers, one fully connected layer, and a Softmax output layer. The convolutional kernel size is 3×3, the activation function is ReLU, and the dropout rate is 0.

2. The loss function during training is cross-entropy loss, the optimizer is Adam, the initial learning rate is 0.001, the batch size is 64, and the number of training epochs is 100. For the few sample categories, GAF graphs with noise are used for data augmentation.

6. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, Step S3 uses accuracy, precision, recall, and F1 score as performance metrics for model classification. Let TP be the number of correctly predicted samples in a given class, FP be the number of samples that are incorrectly predicted as belonging to that class, and FN be the number of samples that are incorrectly predicted as belonging to another class. Then: The accuracy calculation formula is: ; The formula for calculating accuracy is: ; The formula for calculating recall rate is: ; The formula for calculating the F1 score is: .

7. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, In step S4, the gated recurrent unit neural network model adopts a two-layer stacked structure, with each layer containing 128 neurons. Dropout layers are set between layers with a dropout rate of 0.

2. The Adam optimizer is used for training, with an initial learning rate of 0.0001 and a loss function of root mean square error. An EarlyStopping strategy is also introduced. The training data is constructed using a sliding window strategy with a window length of 144 time steps and a step size of 1.

8. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, In step S4, the model performance evaluation metrics include the coefficient of determination R², root mean square error RMSE, mean absolute error MAE, and the interval prediction evaluation metric Winkler score SWS; let... This represents the true value of the crack width. This is the predicted crack width. The average of the true values ​​is n, where n is the sample size. For confidence level, and Confidence levels The upper and lower limits of the prediction interval If the true value is: The formula for calculating the coefficient of determination is: ； The formula for calculating the root mean square error is: ； The formula for calculating the mean absolute error is: ； The Winkler score is calculated using the following formula: 。 9. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, In step S5, the predicted residual is defined as the difference between the predicted crack width and the measured value. The mean μ and standard deviation σ of the residual are updated in real time through a 24-hour sliding window. The anomaly discrimination interval is set to [μ-3σ, μ+3σa]. The early warning mechanism supports multi-point deployment and multi-channel fusion. After each monitoring channel is independently evaluated, a structural overall status assessment report is generated through weighted fusion. The early warning information can be pushed to the maintenance personnel terminal and automatically record the abnormal time, severity and trend indicator metadata.

10. The intelligent monitoring and early warning method for abnormal cracks in concrete bridges based on GAF image classification and GRU prediction according to claim 1, characterized in that, The early warning mechanism constructed in step S5 dynamically updates the early warning threshold range based on the residual distribution, and sets the risk classification standard as follows: no exceedance is normal, a single mutation is an early warning, and continuous abnormality is severe, which is used to support bridge health assessment and operation and maintenance response.

Images