Machine Learning-Based Method for Predicting Crack Resistance of Circumferential Prestressed Steel Bars in Concrete Main Towers

Abstract

This invention relates to the field of intelligent engineering performance testing technology, specifically to a machine learning-based method for predicting crack resistance of circumferential prestressed steel bars in concrete main towers. Specifically, fiber optic strain sensors are deployed on the cross-section of the circumferential prestressed steel bars in the main tower to collect strain data and environmental parameters during tensioning and operation, and to construct a sample dataset by labeling cracking states. Adaptive segmentation and outlier removal are performed on the strain sequences to clean the data. A cracking probability prediction model is constructed to process the cleaned data and obtain the cracking prediction probability. The total loss of the model is calculated and the model is trained. Newly collected data is input into the trained model to output the cracking prediction probability and determine whether to issue an early warning. This invention enables online, real-time, intelligent, and accurate prediction of cracking risk in the circumferential prestressed steel bars of concrete main towers, effectively reducing the false alarm rate and providing reliable technical support for bridge operation, maintenance, and safety management.

Description

Technical Field

[0001] This invention relates to the field of intelligent testing technology for engineering performance, and in particular to a method for predicting the crack resistance of circumferential prestressed steel bars in concrete main towers based on machine learning. Background Technology

[0002] In large-scale cable-stayed bridge projects, the concrete main tower, as the core load-bearing component, is of paramount importance for its structural safety. Circumferential prestressed steel reinforcement is a crucial measure to ensure the main tower resists circumferential tensile stress generated by concrete shrinkage and creep, temperature stress, and external loads, directly affecting the tower's durability and service life. However, circumferential prestressed steel reinforcement is under extremely high stress during actual service. Coupled with factors such as uneven concrete pouring quality, tensioning deviations, environmental erosion, and the long-term fatigue effects of vehicle loads, micro-cracks are highly likely to initiate on the surface of the reinforcement or in the anchorage area. If these initial micro-cracks are not detected in time, they will gradually expand during the operational phase, leading to prestress loss, accelerated steel corrosion, and even spalling of the concrete cover, seriously threatening the structural safety and long-term durability of the bridge.

[0003] Existing technologies objectively suffer from the following shortcomings: Conventional data cleaning methods, such as moving average smoothing or fixed threshold filtering, easily smooth out subtle fluctuations in pre-cracking precursors as noise, or fail to adapt to local data changes, leading to the accidental deletion of normal fluctuations while retaining abnormal spikes, thus losing sensitive features related to damage evolution; Conventional prediction methods only utilize the original strain amplitude or simple statistical features, ignoring the fluctuation characteristics of pre-cracking precursors at different time scales, and failing to comprehensively depict the complete evolution process from normal state to damage initiation and crack development; Relying solely on strain data while ignoring the coupled effects of environmental temperature and humidity changes on concrete material properties and steel-concrete bond performance leads to decreased prediction accuracy and poor generalization ability of the model when environmental conditions change; Existing methods typically process each sensor measurement point independently, failing to utilize the characteristics of continuous spatial distribution of circumferential prestressed steel bars and strong correlation between strain evolution of adjacent measurement points, resulting in false alarms easily caused by abnormal fluctuations of single-point sensors, and lacking a collaborative calibration mechanism based on spatial context information.

[0004] Therefore, this invention proposes a machine learning-based method for predicting the crack resistance of circumferential prestressed steel bars in concrete main towers to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a machine learning-based method for predicting the crack resistance of circumferential prestressed steel bars in concrete main towers. This invention enables online real-time intelligent assessment of the crack resistance performance of main towers, providing a basis for decision-making regarding bridge operation and maintenance safety.

[0006] The technical solution of this invention to solve the technical problem is a machine learning-based method for predicting the crack resistance of circumferential prestressed steel bars in concrete main towers, comprising the following steps:

[0007] S1. Fiber grating strain sensors are arranged on the circumferential prestressed steel section of the main tower to collect strain data during the tensioning and subsequent operation stages of the steel bars at a fixed frequency, forming a strain sequence and recording environmental parameters simultaneously. Based on the collected historical data, the cracking state of each measuring point at each moment is manually labeled. The strain sequence of each measuring point and its corresponding environmental parameters and labels constitute a sample, and multiple samples constitute a sample dataset.

[0008] S2. For each sample data strain sequence, an adaptive segmentation and anomaly removal method based on the combination of sliding window information entropy change rate and local outlier factor is adopted. The information entropy change rate is used to detect the mutation point of the strain data sequence for adaptive segmentation. Then, within each segment, the improved local outlier factor algorithm is used to identify and remove abnormal strain values. Then, cubic spline interpolation is used to fill the missing points to obtain the cleaned segmented strain sequence. Finally, the segments are spliced ​​together to obtain the cleaned strain sequence.

[0009] S3. Construct a cracking probability prediction model. The model integrates multi-scale time-frequency features with physical information constraints and introduces a spatiotemporal correlation calibration mechanism to extract, fuse, map, and calibrate the features of the cleaned strain sequence and corresponding environmental parameters to obtain the cracking prediction probability of each measuring point at each time.

[0010] S4. Construct a composite loss function that combines cross-entropy loss, physical consistency constraint and parameter regularization to calculate the total loss of the crack probability prediction model.

[0011] S5. Based on the data collected in step S1, iteratively train the crack probability prediction model, minimize the total loss of the model, optimize the trainable parameters in the model, and obtain a trained model.

[0012] S6. Deploy the trained model in the actual prediction of the crack resistance performance of the circumferential prestressed steel bars of the concrete main tower. Collect data through the deployed sensors, process it in step S2, and input it into the trained model. Output the crack prediction probability and compare it with the set warning threshold to determine whether to issue a warning.

[0013] S1 is as follows:

[0014] The collected strain data of the steel bars are time-series data, and the data collection period is the complete evolution stage from normal operation to damage initiation and crack development; at the same time, environmental parameters, including ambient temperature, relative humidity and wind speed, are collected synchronously through sensors.

[0015] The labeling information includes records of regular manual visual inspections, images of the rebar surface taken by a high-resolution camera, and micro-crack signals of the rebar captured by acoustic emission technology. When a visible crack or micro-signal representing cracking is first confirmed at a certain measuring point through one or more of the above methods, that moment and subsequent moments are labeled as cracked and the label value is set to 1. All moments before that moment are labeled as uncracked and the label value is set to 0. Thus, a corresponding binary true cracking label is generated for each measuring point on the rebar at each acquisition moment.

[0016] A complete sample dataset includes time-series strain data, environmental parameters, and binary real crack labels for a single measurement point during the data acquisition period.

[0017] S2 is as follows:

[0018] S2.1 The adaptive segmentation operation is as follows:

[0019] The strain sequence in each sample data corresponds to a strain measurement value at each time. The information entropy within the sliding window is defined with the current time as the center. The information entropy of the strain measurement value within the window is calculated. The larger the information entropy value, the more dispersed the distribution.

[0020] The information entropy change rate at the current moment is calculated based on the information entropy at the previous moment. The information entropy change rate is compared with the set adaptive threshold. If it exceeds the threshold, the moment is marked as a potential mutation point. Then, all mutation points are used as segmentation boundaries to divide the input strain sequence into multiple continuous segments.

[0021] S2.2 The outlier removal operation is as follows:

[0022] Within each segment, a local outlier algorithm is used to identify and remove abnormal strain values. First, the points within each segment are defined. The nearest neighbor set is used to calculate the local outlier factor based on the local reachability density. If the value exceeds the set threshold, it is identified as an outlier and removed. If missing points appear in the segment after the outlier is removed, cubic spline interpolation is used to fill the missing points to obtain the cleaned segmented strain sequence. All the cleaned segmented strain sequences are spliced ​​together in chronological order to obtain the complete cleaned strain sequence.

[0023] S3 is as follows:

[0024] S3.1. Multi-scale strain feature extraction is performed on the cleaned strain data. Multi-level wavelet packet decomposition is performed using a sliding window to calculate the multi-scale strain fluctuation energy features. A dynamic adjustment factor for the strain change rate is introduced to obtain the fluctuation energy feature values ​​at each time step.

[0025] S3.1 is as follows:

[0026] For a given moment, a sliding window of fixed length is defined centered on that moment. The strain subsequence within the window is extracted, and the subsequence is subjected to multi-level wavelet packet decomposition. The coefficients of each frequency band are reconstructed to obtain the reconstructed signal of the corresponding frequency band. Then, each eigenvalue of the multi-scale strain wave energy eigenvector at that moment is defined as the ratio of the energy of the signal in that frequency band to the total energy of the window. A dynamic adjustment factor based on the strain change rate is introduced to obtain the wave energy eigenvalues ​​of each frequency component at that moment. The wave eigenvalues ​​of each frequency component are fused to obtain the wave energy eigenvalue at that moment.

[0027] Wavelet packet decomposition uses the Daubechies wavelet basis, and the number of wavelet packet decomposition layers is preset according to the sampling frequency and characteristic frequency band requirements.

[0028] S3.2 Perform physical information constraint feature fusion. Combine temperature, humidity, and wind speed environmental data, and realize strain and environmental feature fusion through nonlinear gating mechanism to correct the influence of the environment on material properties and obtain feature vectors of fused physical information at each time.

[0029] S3.3. Using the sequence of feature vectors formed by fusing physical information at all times as input, dilated causal temporal convolution combined with multi-head self-attention mechanism is used to extract temporal dependencies and map them to obtain the cracking probability at each time.

[0030] S3.3 is as follows:

[0031] A temporal prediction layer combining a temporal convolutional network and a self-attention mechanism is constructed. This layer first encodes the sequence of feature vectors that fuse physical information at all times in the temporal dimension by expanding the causal convolutional network to extract deep temporal features. Then, the contribution weight of features at different historical times to the current cracking risk is calculated by a multi-head self-attention mechanism, and the weighted features are mapped to the cracking probability at the current time.

[0032] S3.4. Introduce a spatiotemporal calibration layer based on graph convolutional networks to construct a spatial adjacency matrix of measurement points, where an element of 1 indicates that two measurement points are spatially adjacent. Perform collaborative optimization on the prediction probabilities of adjacent measurement points in the spatial adjacency matrix of measurement points to obtain the final cracking prediction probability of each measurement point at each time step.

[0033] S4 is as follows:

[0034] The model's total loss function minimizes the difference between the predicted probability and the true crack label, while encouraging the predicted probability to vary in tandem with physical quantities such as the strain rate of change, and applies L2 regularization to the model parameters to prevent overfitting.

[0035] S5 is detailed below:

[0036] After cleaning the sample data in step S1 in step S2, it is divided into training set, validation set and test set. During training iteration, batches of data are taken from the training set to sequentially complete multi-scale strain feature extraction, physical information fusion, temporal feature encoding and graph convolution spatiotemporal calibration to obtain the final crack probability. Then, a composite loss function is used to comprehensively evaluate the prediction bias, physical law fit and parameter complexity, and backpropagation combined with Adam optimization algorithm is used to update the trainable parameters. The model traverses the training set in multiple rounds, and the performance is monitored with the validation set in each round. When the loss on the validation set no longer decreases or the performance tends to stabilize and degrade, training stops and the optimal parameters are saved as the trained model. Finally, the trained model is tested using data from the test set.

[0037] S6 is detailed below:

[0038] The trained model is used in the real-time crack prediction system for circumferential prestressed steel bars of the concrete main tower. The system receives data from sensors on the main tower, cleans the new data in real time, processes it through the trained model, and outputs the final crack prediction probability. If the final crack prediction probability exceeds the preset threshold, an early warning is issued to reflect the risk of steel bar cracking in real time.

[0039] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0040] This invention discloses a machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers. It employs an adaptive segmented anomaly removal method combining the information entropy change rate of a sliding window with local outlier factors. This method achieves adaptive segmentation by detecting abrupt changes in the strain sequence, and then identifies outliers within each segment. This approach can remove abnormal peaks caused by sensor malfunctions or external impacts while fully preserving subtle fluctuation features related to pre-cracking precursors. Furthermore, this invention utilizes a multi-scale strain fluctuation energy feature extraction method. It obtains the energy distribution of each frequency band by performing wavelet packet decomposition on the strain sequence and introduces a dynamic adjustment factor based on the strain change rate to enhance... The strong model is sensitive to sudden fluctuations, thus capturing cracking signs earlier. This invention also employs a physical information fusion gating mechanism, dynamically fusing physical parameters such as ambient temperature, humidity, and wind speed with multi-scale strain features through learnable gating units, and introducing an environmental residual correction term, enabling the model to adaptively adjust the influence weight of environmental factors on cracking prediction. This invention also uses a spatiotemporal correlation calibration model combining temporal convolutional networks, multi-head self-attention mechanisms, and graph convolutional networks. It first captures long-term dependencies through temporal networks, and then aggregates spatially adjacent measurement point information through graph convolution to collaboratively optimize the single-point prediction probability, effectively suppressing false alarms and missed alarms. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0042] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0043] Figure 2 Figure 1 shows a performance comparison between the data cleaning method of the present invention and conventional methods. Figure 2 shows the original strain sequence, Figure 3 shows the result of conventional moving average smoothing, Figure 4 shows the result of fixed threshold filtering, and Figure 5 shows the result of the method of the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the effect of different feature combinations on the prediction accuracy of images.

[0045] Figure 4 This is a thermogram showing the actual cracked state.

[0046] Figure 5 This is a heatmap showing the probability of single-point prediction.

[0047] Figure 6 This is a heatmap of the probabilities after spatiotemporal calibration. Detailed Implementation

[0048] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0049] Example 1

[0050] like Figure 1 As shown, a machine learning-based method for predicting crack resistance of circumferential prestressed steel bars in a concrete main tower includes the following steps:

[0051] S1. Fiber grating strain sensors are arranged on the circumferential prestressed steel section of the main tower to collect strain data during the tensioning and subsequent operation stages of the steel bars at a fixed frequency, forming a strain sequence and recording environmental parameters simultaneously. Based on the collected historical data, the cracking state of each measuring point at each moment is manually labeled. The strain sequence of each measuring point and its corresponding environmental parameters and labels constitute a sample, and multiple samples constitute a sample dataset.

[0052] S2. For each sample data strain sequence, an adaptive segmentation and anomaly removal method based on the combination of sliding window information entropy change rate and local outlier factor is adopted. The information entropy change rate is used to detect the mutation point of the strain data sequence for adaptive segmentation. Then, within each segment, the improved local outlier factor algorithm is used to identify and remove abnormal strain values. Then, cubic spline interpolation is used to fill the missing points to obtain the cleaned segmented strain sequence. Finally, the segments are spliced ​​together to obtain the cleaned strain sequence.

[0053] S3. Construct a cracking probability prediction model. The model integrates multi-scale time-frequency features with physical information constraints and introduces a spatiotemporal correlation calibration mechanism to extract, fuse, map, and calibrate the features of the cleaned strain sequence and corresponding environmental parameters to obtain the cracking prediction probability of each measuring point at each time.

[0054] S4. Construct a composite loss function that combines cross-entropy loss, physical consistency constraint and parameter regularization to calculate the total loss of the crack probability prediction model.

[0055] S5. Based on the data collected in step S1, iteratively train the crack probability prediction model, minimize the total loss of the model, optimize the trainable parameters in the model, and obtain a trained model.

[0056] S6. Deploy the trained model in the actual prediction of the crack resistance performance of the circumferential prestressed steel bars of the concrete main tower. Collect data through the deployed sensors, process it in step S2, and input it into the trained model. Output the crack prediction probability and compare it with the set warning threshold to determine whether to issue a warning.

[0057] In a specific implementation, S1 is as follows:

[0058] The basic dataset used for model training and validation is collected. The strain data is collected by fiber optic strain sensors arranged on the key sections of the circumferential prestressed steel bars of the main tower. These sensors work continuously during the steel bar tensioning and subsequent operation stages, acquiring the strain changes on the steel bar surface in real time at a fixed sampling frequency.

[0059] The acquired raw strain data constitute a continuous time series, denoted as the raw strain sequence, which contains information about the entire process from the initial tensioning stage to the possible cracking.

[0060] To construct an effective cracking probability prediction model, it is necessary not only to collect strain data of the reinforcing steel, but also to simultaneously record environmental parameters closely related to cracking evolution. Therefore, at each data acquisition moment, on-site temperature sensors, humidity sensors, and anemometers will work simultaneously to record the current ambient temperature, relative humidity, and wind speed.

[0061] To train the model using supervised learning, accurate labeling of the collected historical strain data is necessary, i.e., determining the true cracking state of each measuring point at every moment. In one implementation, labeling is performed by domain experts combining multiple pieces of information, primarily including periodic visual inspection records, high-resolution camera images of the rebar surface, and microscopic cracking signals captured by acoustic emission technology. When a measuring point is confirmed to have its first visible crack or generate a microscopic signal representing cracking at a certain moment through one or more of the above methods, that moment and subsequent moments are labeled as cracked, with a label value set to 1; all moments before that moment are labeled as uncracked, with a label value set to 0. Based on this, a corresponding binary true cracking label is generated for each sensor measuring point at each acquisition moment.

[0062] The entire acquisition process covers a sufficiently long period to ensure that the complete evolutionary stage from normal operation to damage initiation and crack development can be captured, providing rich and high-quality sample data for model training.

[0063] In a specific implementation, S2 is as follows:

[0064] Because the strain data of the circumferential prestressed steel bars of the concrete main tower are collected in continuous time series, they are affected by various factors such as ambient temperature and concrete shrinkage and creep. The data exhibits nonlinearity, non-stationarity, and frequent local abrupt changes. Conventional data cleaning methods such as fixed threshold filtering or moving average smoothing are difficult to effectively identify abnormal strain peaks caused by sensor malfunctions or external impacts, and may also smooth out subtle fluctuations related to pre-cracking precursors, leading to misjudgments in subsequent models. This invention employs an adaptive segmentation and anomaly removal method based on a sliding window information entropy change rate and a local outlier factor. Adaptive segmentation is achieved by detecting abrupt changes in the strain sequence using the information entropy change rate. Then, within each segment, an improved local outlier factor algorithm is used to identify and remove abnormal strain values, thereby preserving sensitive features related to crack evolution. The specific steps are as follows:

[0065] S2.1 Adaptive segmentation based on the rate of change of information entropy

[0066] First, define the original strain sequence as Each moment corresponds to a strain measurement value; the information entropy within the sliding window is defined to quantify the complexity of the local strain distribution. Specifically, the sliding window is set with the current moment as the center, and the information entropy of the strain values ​​within the window is calculated, expressed as:

[0067]

[0068] In the formula, Indicated by time Centered on, window length is The information entropy within the sliding window is used to quantify the degree of dispersion of strain values ​​within the window; the larger the value, the more dispersed the distribution. This represents the time index, indicating the discrete time points of strain data acquisition, with a value range of [value range missing]. ; Indicates the total number of moments; This indicates that the strain value within the window falls into the first... The probability of each histogram interval is calculated by dividing the number of points in each interval within the statistical window by the total number of points in the window. get; This represents the length of the sliding window, i.e., the number of time points contained within the window. It is a hyperparameter, with an example value of 100. This represents the histogram interval index, with a value range of [value range missing]. ; Represents the total number of intervals in a histogram, defined as follows: ,when hour ; Indicates rounding up;

[0069] The rate of change of information entropy is used to detect abrupt changes in a sequence, and is expressed as:

[0070]

[0071] In the formula, Indicates time The rate of change of information entropy reflects the rate of change of information entropy between adjacent windows; Indicated by time Centered on, window length is The information entropy within the sliding window; This represents the sampling time interval, i.e., the time difference between two adjacent moments, with an example value of 1 second.

[0072] An adaptive threshold is set, and when the rate of change of information entropy exceeds the threshold, that moment is marked as a potential mutation point. In the specific implementation, all mutation points are used as segmentation boundaries to divide the original strain sequence into multiple continuous segments. The strain distribution within each segment is relatively stable, which facilitates subsequent local outlier factor analysis. Specifically, an adaptive threshold is set. ,in, Indicates an adaptive threshold. This represents the mean rate of change of information entropy for the entire sequence. The standard deviation represents the rate of change of the information entropy of the entire sequence. This represents the sensitivity coefficient, with an example value of 3.

[0073] In specific implementation, when At that time, Potential mutation points were marked, and then, using all mutation points as segmentation boundaries, the original strain sequence was divided into segments. A continuous segment Within each segment, the strain distribution is relatively stable, among which, The total number of segments is determined by the number of mutation points. This represents the k-th segment, where k represents the segment index.

[0074] S2.2 Outlier Removal and Interpolation Imputation

[0075] Within each segment, the local outlier algorithm is used to identify and remove abnormal strain values.

[0076] First, define each point. The nearest neighbor set is used to calculate the local outlier factor based on local reachability density. If the value exceeds a set threshold, it is identified as an outlier and removed. This is represented as:

[0077] ,

[0078] In the formula, Indicates strain value The local outlier factor is used to measure strain values. The degree of abnormality; the larger the value, the more likely the point is to be an outlier. Indicates segmentation The first Each strain value corresponds to a time... ; Indicates strain value The corresponding time index or moment, i.e. the discrete time points of data acquisition; Indicates the index of the data point; Indicates strain value of Nearest neighbor set, i.e., based on Euclidean distance Recent A set consisting of points Represents a set Size; Indicates the number of nearest neighbors, used to define the size of the local neighborhood of each point, for example ; Indicates strain value Locally accessible density is used to measure the density of its neighboring points; Indicates strain value Locally achievable density, Indicates segmentation The first Each strain value; Indicates difference from Data point index.

[0079] In the specific implementation process, if ,but Values ​​identified as outliers and removed, among which, This represents the threshold for anomaly detection; an example value is 1.5.

[0080] If missing points appear within a segment after outlier removal, cubic spline interpolation is used to fill the missing points, resulting in a cleaned segmented strain sequence. The same operation is performed on all segments, and then the segments are spliced ​​together in chronological order to obtain a complete cleaned strain sequence. The cleaned strain sequence is defined as follows. ,in, Indicates the first time after cleaning A segmented strain sequence, which is the continuous strain data obtained after outlier removal and cubic spline interpolation.

[0081] It should be noted that Local-Reachability-Densitym (LRD) is a strain value. of The reciprocal of the average reachability distance between nearest neighbors is used to calculate the distance between two points using Euclidean distance, thus measuring... The density of surrounding points; the cubic spline interpolation method fits known data points with piecewise cubic polynomials and ensures the continuity of the first and second derivatives at the nodes, thereby obtaining a smooth interpolation curve.

[0082] In a specific implementation, S3 is as follows:

[0083] Conventional methods for predicting the probability of steel bar cracking typically rely solely on the original strain amplitude or simple statistical characteristics, neglecting the fluctuating characteristics of crack precursors across different time scales and the coupled influence of environmental factors on material properties. This results in low prediction accuracy and poor generalization ability. This invention constructs a cracking probability prediction model by integrating multi-scale time-frequency features with physical information constraints and introducing a spatiotemporal correlation calibration mechanism. The specific steps are as follows:

[0084] S3.1 Multi-scale strain feature extraction

[0085] For the strain sequence after cleaning, considering that the precursor to cracking often manifests as abnormal strain fluctuations at different time scales, it is difficult to fully characterize the damage evolution process by simply relying on the original strain value. This invention defines multi-scale strain fluctuation energy characteristics. By decomposing the strain sequence into different frequency bands, the energy distribution and its rate of change in each frequency band are calculated, which is used as one of the key features of the model input.

[0086] Specifically, for a given moment, a fixed-length sliding window is defined centered on that moment. The strain subsequence within the window is extracted, and multi-level wavelet packet decomposition is performed on this subsequence. The coefficients of each frequency band are reconstructed to obtain the reconstructed signal for the corresponding frequency band. Then, each eigenvalue of the multi-scale strain wave energy feature vector at that moment is defined as the ratio of the energy of the signal in that frequency band to the total energy of the window. A dynamic adjustment factor based on the strain rate of change is introduced to enhance the response to sudden fluctuations, expressed as:

[0087] ,

[0088] In the formula, Indicates at time The calculated first The fluctuation energy characteristic value of each frequency component is used to quantify the energy contribution of the frequency band at the current moment. The larger the value, the higher the energy proportion of the frequency band and the more sensitive it is to sudden fluctuations. This represents the frequency band index, with a value range of [value range missing]. ; The value represents the number of wavelet packet decomposition levels; it is a hyperparameter used to control the frequency resolution. An example value is shown below. At this point, a total of One frequency band; Indicates at time Within the sliding window, the first wavelet packet coefficients are reconstructed to obtain the second wavelet packet. The frequency band within the window The signal amplitude at each point is obtained by inverse wavelet packet transform; This represents the index of a time point within the sliding window, with a value range of [value missing]. ; This represents the length of the sliding window used for feature extraction, i.e., the number of time points contained within the window. An example value is shown below. ; This represents a very small constant used to prevent the denominator from being zero. Examples of its values ​​are shown below. ; Indicates the first The sensitivity adjustment coefficient of each frequency band is a trainable parameter that controls the amplification factor of the strain change rate at current and subsequent time points on the energy characteristics of that frequency band. Represents the strain sequence after cleaning At any moment The strain value; Represents the strain sequence after cleaning At any moment The strain value; Indicates length is The maximum absolute value of strain within the sliding window is used to normalize the rate of strain change. Indicates length is The strainor sequence within the sliding window is defined as ; This indicates rounding down to the nearest integer.

[0089] In practice, wavelet packet decomposition uses the Daubechies wavelet basis, and the number of layers in the wavelet packet decomposition is... Pre-settings are required based on the sampling frequency and characteristic frequency band requirements. The choice needs to balance time resolution and frequency resolution, for example, taking 64 time points.

[0090] It should be noted that early signs of cracking are often accompanied by drastic changes in strain. The term is used as a dynamic adjustment factor, calculated based on the strain change rate at adjacent times, to enhance the model's sensitivity to sudden fluctuations, thereby dynamically amplifying the energy characteristic value of the corresponding frequency band, making the model pay more attention to the moment of strain change, and thus capturing cracking signs earlier.

[0091] S3.2 Feature Fusion of Physical Information Constraints

[0092] Strain data alone cannot fully reflect the effects of changes in ambient temperature and humidity on the properties of concrete materials and reinforcing steel. The influence of concrete bonding performance, and these factors are crucial to the cracking process, this invention combines environmental monitoring data to construct a physical information fusion gating mechanism for dynamically correcting features extracted from strain data;

[0093] The environmental feature vector is defined as consisting of ambient temperature, relative humidity, and wind speed. The enhanced feature vector after physical information fusion is obtained by fusing multi-scale strain features and environmental features through a nonlinear gated unit, and is expressed as:

[0094]

[0095] In the formula, Indicates time The feature vector that incorporates physical information has a dimension of By using a gating mechanism to fuse multi-scale strain characteristics with environmental characteristics, the main components of the original strain characteristics are preserved while residual corrections for environmental factors are added, thus characterizing the strain at time [time value missing]. Comprehensive characteristics after considering the influence of environmental factors; This represents the Hadamard product, i.e., element-wise multiplication. This represents the Sigmoid activation function, which maps the input to... Interval, generate gated weight vector; This represents a vector concatenation operation; This represents the weight matrix used to generate the gate weights, which are trainable parameters with dimensions of . ; express Dimensions express The dimension, and ; The bias term used to generate the gating weights is a trainable parameter with dimension . ; This represents the multi-scale strain wave energy eigenvector, where the a-th eigenvalue is... ,Right now ; Indicates time The environmental feature vector, derived from time... ambient temperature relative humidity Composed of wind speed : Indicates time The ambient temperature, in degrees Celsius. Indicates time The relative humidity, expressed as a percentage. Indicates time Wind speed, in meters per second; This represents the hyperbolic tangent activation function, used to map the input to... For each interval, generate the residual term; The weight matrix used to process environmental features to generate residual terms is a trainable parameter with dimension O(n). ; Indicates time Environmental feature vector Dimensions ,and ; The bias term, representing the bias used to process environmental features to generate residual terms, is a trainable parameter with dimension . .

[0096] S3.3, Split Probability Mapping Based on Temporal Convolution and Attention Mechanisms

[0097] definition This represents the sequence of fused physical information feature vectors at all times, i.e. , as input to the temporal convolutional network;

[0098] To capture long-term dependencies in feature vectors that integrate physical information and focus on historical information most relevant to cracking, a temporal prediction layer combining a temporal convolutional network and a self-attention mechanism is constructed. This layer first expands the causal convolutional network to... Encoding is performed along the time dimension to extract deep temporal features. Then, a multi-head self-attention mechanism is used to calculate the contribution weights of features from different historical moments to the current cracking risk. The weighted features are then mapped to the cracking probability at the current moment, as follows:

[0099] ,

[0100] In the formula, This indicates the model's prediction at time [time]. The probability of cracking in the circumferential prestressed steel bars of the main tower ranges from [value missing]. This value integrates the temporal characteristics of historical strain data, environmental factors, and information from spatially adjacent measuring points, and characterizes the likelihood of steel bar cracking at the current moment based on all available information. Indicates will The outputs of each attention head are concatenated along the feature dimension to obtain a dimension of ; Indicates a historical moment index; The weight vector of the output layer represents trainable parameters with dimension . The deep feature vectors aggregated by the multi-head self-attention mechanism are linearly mapped to scalars; The bias term of the output layer is a trainable parameter, a scalar, which adds a learnable offset to the linear mapping. This represents the total number of attention heads, a hyperparameter, with examples of possible values. ; This represents the index of the attention head, with a value range of 1. ; This represents the time step output by the temporal convolutional network. The corresponding deep feature representation has a dimension of The temporal convolutional network Sequence encoding is used to obtain the input sequence in the specific implementation. First through a Convolution adjusts the feature dimension to Then, it goes through multiple residual blocks (e.g., 5), each containing dilated causal convolution, weight normalization, ReLU activation, and Dropout. The dilated causal convolution ensures that future information is not leaked, while the exponential dilation rate rapidly expands the receptive field. The output of the last layer is the time step. Corresponding deep feature representation ; This represents the time step output by the temporal convolutional network. The corresponding deep feature representation;

[0101] express In terms of dimensions, in practical implementation, temporal convolutional networks employ multi-layer dilated causal convolutions with an exponentially increasing dilation rate to expand the receptive field and output... Dimensions It needs to be preset according to the model complexity, for example ;

[0102] Indicates the first The query projection matrix of each attention head is a trainable parameter with dimension . ; Indicates the first The key projection matrix of each attention head is a trainable parameter with dimension . ; Indicates the first The projection matrix of the attention heads is a trainable parameter with dimension 1. ; The key dimension of each attention head is preferably set to... The example value is when , hour, ; Indicates the transpose operation; This means applying the softmax function to the calculated attention scores to obtain the results for each historical moment. The corresponding attention weights.

[0103] It should be noted that the attention mechanism is decomposed into multiple independent subspaces, and each head learns different attention weights in different representation subspaces, thereby capturing different aspects of the dependencies in the sequence (such as short-term fluctuations, long-term trends, etc.).

[0104] S3.4 Spatiotemporal Correlation Calibration

[0105] Considering that the circumferential prestressed steel bars are continuously distributed in space, the strain evolution of adjacent sensor measuring points is strongly correlated. In order to utilize this spatial context information and reduce false alarms at single points, this invention introduces a spatiotemporal calibration layer based on graph convolutional networks after obtaining the cracking probability of each single point (where the measuring point index is used to identify different sensors) to collaboratively optimize the predicted probability of adjacent measuring points.

[0106] Specifically, a spatial adjacency matrix is ​​defined between measuring points, where an element of 1 indicates that two measuring points are spatially adjacent. The final calibrated crack probability is then expressed as:

[0107]

[0108] In the formula, This indicates that after spatiotemporal correlation calibration, the measuring point At any moment The final cracking prediction probability is based not only on its own temporal characteristics, but also on the information of spatially adjacent measuring points for collaborative optimization, which characterizes the probability that the steel bar at the measuring point will crack at the current moment after considering the spatial context. This represents the sensor measurement point index, with a value range of [value range missing]. ; Indicates the total number of measuring points; Indicates the measuring point The corresponding model predicts at time 10:00. The probability of cracking in the circumferential prestressed steel bars of the main tower. Indicates the measuring point The corresponding model predicts at time 10:00. The probability of cracking in the circumferential prestressed steel bars of the main tower; Representation and measuring points Adjacent measurement point index; This represents the balance coefficient, used to control the contribution of the space calibration term. It is a hyperparameter, and examples of its values ​​are shown below. ; Representation and measurement points The set of adjacent measurement points is predefined based on the spatial layout of the sensors; The calibration weight matrix is ​​a trainable parameter used to map the concatenated feature vectors to scalar calibration values; its dimension is 1. ; This represents the elements of the normalized adjacency matrix, reflecting the measurement points. With measuring points Spatial correlation strength, calculated as follows: In graph convolution, the influence of nodes of different degrees is balanced to prevent nodes with high degree from having excessively large feature values ​​and nodes with low degree from having excessively small feature values, thus achieving a symmetrical normalization effect and making the model training more stable. The elements of the original adjacency matrix (1 if adjacent, 0 otherwise) are predefined based on the sensor spatial layout. For example, suppose there are... If 1, 2, 3 measuring points are arranged at equal intervals on a circular ring, and each measuring point is adjacent to the two measuring points to its left and right (circular connection), then the adjacency matrix can be defined as: (Self-loop is set to 0) ( ), ( ),and , (Circular connection), the rest are 0; The degree matrix of the first degree The diagonal elements are equal to the measurement points. The number of adjacent measuring points, i.e. Continuing with the loop connection from the previous example, each measuring point has two neighbors (left and right), therefore For all Established; Indicates the measuring point corresponding time Feature vectors that incorporate physical information.

[0109] In a specific implementation, S4 is as follows:

[0110] To train the proposed crack probability prediction model so that its output probability conforms to both the statistical regularity of historical crack samples and the physical evolution characteristics of the steel reinforcement during stress, this invention constructs a composite loss function combining cross-entropy loss, physical consistency constraints, and parameter regularization. This loss function minimizes the difference between the predicted probability and the actual crack label, while encouraging the predictive probability to change in tandem with physical quantities such as the strain rate of change. Furthermore, L2 regularization is applied to the model parameters to prevent overfitting, as expressed below:

[0111]

[0112] In the formula, This represents the total loss function, used to guide the gradient update of model parameters; Indicates the measuring point At any moment The actual crack label has a value of (No cracks) or (Cracked); The weighting coefficients representing the physical consistency constraint terms are hyperparameters used to balance the degree of fit between the data fit and the physical laws. Examples of their values ​​are shown below. ; The scaling factor represents the physical mapping. It is a trainable parameter, a scalar, and controls the sensitivity of the strain rate of change to probability, i.e., how steep the probability changes when the rate of change changes. The bias term representing the physical mapping is a trainable parameter, a scalar, and is used to adjust the threshold of the mapping curve in a linear transformation. Indicates the measuring point At any moment The normalized rate of change of strain is defined as the absolute value of the strain difference between adjacent time points divided by the maximum absolute strain value within the window (plus a small constant), i.e. This is used to quantify the severity of local strain changes as a priori physical knowledge (generally, the greater the strain change rate, the higher the probability of cracking), encouraging model predictions to be consistent with this reference, thereby incorporating physical mechanisms into model training; Indicates the measuring point The corresponding length is The maximum absolute value of strain within the sliding window is used to normalize the rate of strain change. Indicates the measuring point The corresponding length is The strain gauge sequence within the sliding window; Indicates the measuring point Corresponding strain sequence after cleaning At any moment strain value, Indicates the measuring point Corresponding strain sequence after cleaning At any moment The strain value; The weight coefficients of the regularization term are hyperparameters used to control model complexity. Examples of their values ​​are shown below. ; This represents the square of the L2 norm of all trainable parameters of the model. The set of all trainable parameters of the model. This represents the L2 norm.

[0113] It should be noted that, The term, as the cross-entropy loss term, directly measures the deviation between the predicted probability and the true label. The term, serving as a physical consistency constraint, is calculated based on prior knowledge. A higher strain rate corresponds to a higher probability of cracking. A mapping from the normalized strain rate to probability is established using the sigmoid function, encouraging model predictions to closely approximate this mapping. This implicitly integrates the physical mechanism into model training. The term "regularization term" improves the model's generalization ability by penalizing excessively large parameter values.

[0114] In a specific implementation, S5 is as follows:

[0115] The constructed crack probability prediction model is trained to determine all trainable parameters in the model.

[0116] Before training begins, the collected and cleaned historical dataset is first divided into a training set, a validation set, and a test set. The training set is used for iterative updates of the model parameters, the validation set is used to monitor the model's performance on unseen data to adjust hyperparameters and prevent overfitting, and the test set is reserved for evaluating the final model performance.

[0117] In each training iteration, a batch of data is randomly selected from the training set. This batch contains the post-cleaning strain sequence of multiple measurement points within a continuous time window, the corresponding environmental feature vectors, and the actual crack labels. This data is first calculated according to the forward propagation process described in step S3: strain fluctuation energy features are extracted through multi-scale strain feature extraction, and then combined with environmental features through a physical information fusion gating mechanism to obtain a feature vector with fused physical information. Subsequently, this feature vector sequence is fed into a temporal convolutional network and a self-attention mechanism to extract deep temporal dependencies and map them to preliminary crack probabilities. Finally, a spatiotemporal calibration layer based on a graph convolutional network is used to collaboratively optimize the preliminary probabilities using information from spatially adjacent measurement points, resulting in the final calibrated crack probability for each measurement point in this batch of data at each time step.

[0118] After obtaining the predicted probability, the loss value of the batch of data is calculated according to the composite loss function defined in step S4. This loss value comprehensively measures the deviation between the predicted probability and the true label, the degree of conformity with the physical law reflected by the strain rate of change, and the complexity of the model parameters.

[0119] Then, the gradient of the loss function with respect to each trainable parameter is calculated using the backpropagation algorithm, and the parameters are updated based on the gradient information using the Adam optimization algorithm, thereby reducing the loss value.

[0120] The entire training process iterates through the training set multiple times, meaning the entire training set is used repeatedly for parameter updates. After each round, the performance of the current model is evaluated using the validation set. Key monitoring metrics include the loss value on the validation set and the accuracy and recall of the split prediction. When the loss value on the validation set no longer decreases after several consecutive rounds, or when the model's performance metrics on the validation set stabilize or even degrade, the model is considered to have converged sufficiently. At this point, iterative training stops, and the model parameters with the best performance on the validation set are saved as the final prediction model.

[0121] In a specific implementation, S6 is as follows:

[0122] After the model has been trained and its optimal parameters have been saved, it can be applied to predict the crack resistance of the circumferential prestressed steel reinforcement in a concrete main tower. During the real-time prediction phase, the system receives data in real time from various fiber optic strain sensors, temperature sensors, humidity sensors, and anemometers deployed on the main tower.

[0123] For each newly acquired data moment, the data is first cleaned in real time using the adaptive segmentation and outlier removal method described in step S2, combined with a historical data window, to obtain the cleaned strain value at the current moment. Then, centered on the current moment, a sliding window including the current moment is constructed using the feature extraction method in step S3. Multi-scale feature extraction is performed on the cleaned strain sequence within the window to obtain the multi-scale strain fluctuation energy feature vector at the current moment. Simultaneously, the ambient temperature, relative humidity, and wind speed acquired synchronously at the current moment are stitched together to form the environmental feature vector at the current moment. Then, these two feature vectors are input into a pre-trained crack probability prediction model with fixed parameters. The model first fuses the two through a physical information fusion gating mechanism to obtain a feature vector with fused physical information at the current moment. Then, this feature vector, along with a sequence of similar feature vectors from historical moments, is fed into a pre-trained temporal convolutional network and a self-attention mechanism. Based on this, the model calculates the initial crack probability at the current moment based on historical temporal information. Finally, the initial crack probability of the measuring point, the feature vector with fused physical information, and the corresponding information of its spatially adjacent measuring points are input into a pre-trained spatiotemporal calibration layer. A graph convolutional network is used to aggregate the spatial context information, and finally, the final crack probability of the measuring point at the current moment after spatiotemporal correlation calibration is output. This probability value reflects in real time the risk of cracking of the circumferential prestressed steel bars at the measuring point under the current environment and stress state.

[0124] Engineering managers can issue timely warnings based on preset warning thresholds when the predicted probability exceeds the threshold, thereby enabling online, real-time, and intelligent assessment and prediction of the crack resistance performance of the circumferential prestressed steel bars of the main tower, providing key decision-making basis for the operation, maintenance, and safety management of the bridge.

[0125] Example 2

[0126] like Figure 2As shown in the figure, to verify the superiority of the adaptive segmentation and outlier removal method proposed in this invention compared with conventional data cleaning techniques, the experiment used accelerated deterioration data to compare the method of this invention with two conventional methods, namely moving average smoothing and fixed threshold filtering. The moving average smoothing method uses a moving average with a window length of 11 sampling points, which is a commonly used noise reduction method in engineering, but it is prone to smoothing out subtle fluctuations related to crack precursors. The fixed threshold filtering method sets upper and lower thresholds based on the global strain mean and three times the standard deviation, and regards points exceeding the threshold as outliers and removes them before linear interpolation. This method has a certain effect on sudden spikes, but it cannot adapt to local changes and may mistakenly delete normal fluctuations. This method, on the other hand, uses the information entropy change rate of the sliding window to detect abrupt changes to achieve adaptive segmentation, then uses the local outlier factor algorithm to identify outliers in each segment, and finally uses cubic spline interpolation to fill in the missing points, thereby retaining crack-sensitive features while removing outliers.

[0127] like Figure 2 As shown in Figure (a), the original strain sequence exhibits significant abnormal peaks at multiple time points, with sudden jumps in strain values, possibly due to sensor malfunction or external impact. Simultaneously, after approximately the 1000th sampling point, the sequence begins to show gradually increasing fluctuations, especially around the 1400th point where the strain suddenly increases and then slowly decays, characterizing pre-cracking precursors and the cracking event. Figure (b) shows the result after moving average smoothing. Although the peaks are suppressed, the high-frequency fluctuations of the pre-cracking precursors are also significantly weakened, making the strain curve overly smooth and losing detailed features. Figure (c) shows the effect of fixed threshold filtering. Although some large peaks are removed, some local anomalies that do not exceed the global threshold remain, and linear interpolation causes unnatural straight line segments on the curve near the removal points. At the same time, the subtle fluctuations in the cracked area are smoothed to some extent. Figure (d) shows the results after processing by this method. It can be observed that abnormal spikes are effectively identified and removed, but the fluctuation details of the cracked area (such as the sudden increase and decrease trend after the 1400th point) are completely preserved. The curve is smooth in the stable section and maintains the proper fluctuations in the fluctuating section, which reflects the ability of this method to balance noise removal and feature preservation. Figure 2 The strain values ​​on the vertical axis are in microstrain, and the horizontal axis represents time sampling points.

[0128] Example 3

[0129] like Figure 3As shown, the effectiveness of the gated fusion features designed in this invention is demonstrated by comparing the accuracy of crack prediction models under different combinations of input features. The feature combinations compared include: using only the original strain amplitude, using basic statistical features of strain (including mean, variance, maximum and minimum values, skewness, and kurtosis), using only environmental features (temperature, humidity, and wind speed), using the multi-scale strain fluctuation energy features proposed in this invention, inputting a simple concatenation of multi-scale features and environmental features, and the physical information fusion features with a gated mechanism adopted in this invention. All models were trained on the same training and validation sets, and the accuracy was calculated using the validation set. Experimental results show that the gated fusion method proposed in this invention, by dynamically adjusting the fusion weights of strain features and environmental features through a learnable gating mechanism and introducing environmental residual correction, achieves significantly higher accuracy than other combinations. This demonstrates the advantages of this technology in multi-source information fusion, enabling more comprehensive utilization of data and improving the reliability of crack prediction.

[0130] Example 4

[0131] like Figures 4 to 6 As shown, the spatiotemporal calibration mechanism based on graph convolutional networks proposed in this invention is verified to optimize the single-point prediction probability. The experiment simulates the cracking evolution process of 8 spatially adjacent measurement points (arranged in a ring) over 50 time steps at the same interval. Figure 4 The actual cracking state shown begins with cracking at some measuring points starting from the 20th time step, resulting in spatial non-uniformity. Figure 5 The single-point prediction probabilities shown are generated by the basic time series model, but there are obvious false positives and false negatives. Figure 6 The spatiotemporal calibration probabilities are shown to be collaboratively optimized by aggregating neighborhood information through a graph convolutional network. As can be seen from the figure, the false alarm regions of measurement points 1 and 4 are significantly lighter in color (probability decreases), while the false alarm region of measurement point 2 is darker in color (probability increases), and the overall color distribution is more uniform. Figure 4 The actual cracking state is shown. Figures 4 to 6 The color intensity of the heatmap represents the cracking probability or state value, with the color bar on the left showing a gradient from zero to one. The horizontal axis represents the time step, and the vertical axis represents the measurement point number. Experimental results show that this invention effectively suppresses single-point abnormal fluctuations by utilizing spatial correlation, thus improving the robustness and accuracy of prediction.

[0132] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A machine learning-based method for predicting the crack resistance of circumferentially prestressed reinforcement in a concrete main tower, characterized by, Includes the following steps: S1. Fiber grating strain sensors are arranged on the circumferential prestressed steel section of the main tower to collect strain data during the tensioning and subsequent operation stages of the steel bars at a fixed frequency, forming a strain sequence and recording environmental parameters simultaneously. Based on the collected historical data, the cracking state of each measuring point at each moment is manually labeled. The strain sequence of each measuring point and its corresponding environmental parameters and labels constitute a sample, and multiple samples constitute a sample dataset. S2. For each sample data strain sequence, an adaptive segmentation and anomaly removal method based on the combination of sliding window information entropy change rate and local outlier factor is adopted. The information entropy change rate is used to detect the mutation point of the strain data sequence for adaptive segmentation. Then, within each segment, the improved local outlier factor algorithm is used to identify and remove abnormal strain values. Then, cubic spline interpolation is used to fill the missing points to obtain the cleaned segmented strain sequence. Finally, the segments are spliced ​​together to obtain the cleaned strain sequence. S3. Construct a cracking probability prediction model. The model integrates multi-scale time-frequency features with physical information constraints and introduces a spatiotemporal correlation calibration mechanism to extract, fuse, map, and calibrate the features of the cleaned strain sequence and corresponding environmental parameters to obtain the cracking prediction probability of each measuring point at each time. S3 is as follows: S3.

1. Multi-scale strain feature extraction is performed on the cleaned strain data. Multi-level wavelet packet decomposition is performed using a sliding window to calculate the multi-scale strain fluctuation energy features. A dynamic adjustment factor for the strain change rate is introduced to obtain the fluctuation energy feature values ​​at each time step. S3.2 Perform physical information constraint feature fusion. Combine temperature, humidity, and wind speed environmental data, and realize strain and environmental feature fusion through nonlinear gating mechanism to correct the influence of the environment on material properties and obtain feature vectors of fused physical information at each time. S3.

3. Using the sequence of feature vectors formed by fusing physical information at all times as input, dilated causal temporal convolution combined with multi-head self-attention mechanism is used to extract temporal dependencies and map them to obtain the cracking probability at each time. S3.

4. Introduce a spatiotemporal calibration layer based on graph convolutional network to construct a spatial adjacency matrix of measurement points, where an element of 1 indicates that two measurement points are spatially adjacent. Perform collaborative optimization on the prediction probabilities of adjacent measurement points in the spatial adjacency matrix of measurement points to obtain the final cracking prediction probability of each measurement point at each time step. S4. Construct a composite loss function that combines cross-entropy loss, physical consistency constraint and parameter regularization to calculate the total loss of the crack probability prediction model. S5. Based on the data collected in step S1, iteratively train the crack probability prediction model, minimize the total loss of the model, optimize the trainable parameters in the model, and obtain a trained model. S6. Deploy the trained model in the actual prediction of the crack resistance performance of the circumferential prestressed steel bars of the concrete main tower. Collect data through the deployed sensors, process it in step S2, and input it into the trained model. Output the crack prediction probability and compare it with the set warning threshold to determine whether to issue a warning.

2. The method for predicting the crack resistance of the machine learning-based concrete main tower hoop prestressed reinforcement according to claim 1, characterized in that, S1 is as follows: The collected strain data of the steel bars are time-series data, and the data collection period is the complete evolution stage from normal operation to damage initiation and crack development; at the same time, environmental parameters, including ambient temperature, relative humidity and wind speed, are collected synchronously through sensors. The labeling information includes records of regular manual visual inspections, images of the rebar surface taken by a high-resolution camera, and micro-crack signals of the rebar captured by acoustic emission technology. When a visible crack or micro-signal representing cracking is first confirmed at a certain measuring point through one or more of the above methods, that moment and subsequent moments are labeled as cracked and the label value is set to 1. All moments before that moment are labeled as uncracked and the label value is set to 0. Thus, a corresponding binary true cracking label is generated for each measuring point on the rebar at each acquisition moment. A complete sample dataset includes time-series strain data, environmental parameters, and binary real crack labels for a single measurement point during the data acquisition period.

3. The machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers according to claim 1, characterized in that it is adaptive. The segmented operation is as follows: The strain sequence in each sample data corresponds to a strain measurement value at each time. The information entropy within the sliding window is defined with the current time as the center. The information entropy of the strain measurement value within the window is calculated. The larger the information entropy value, the more dispersed the distribution. The information entropy change rate at the current moment is calculated based on the information entropy at the previous moment. The information entropy change rate is compared with the set adaptive threshold. If it exceeds the threshold, the moment is marked as a potential mutation point. Then, all mutation points are used as segmentation boundaries to divide the input strain sequence into multiple continuous segments.

4. The machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers according to claim 1, characterized in that, The outlier removal process is as follows: Within each segment, a local outlier algorithm is used to identify and remove abnormal strain values. First, the points within each segment are defined. The nearest neighbor set is used to calculate the local outlier factor based on the local reachability density. If the value exceeds the set threshold, it is identified as an outlier and removed. If missing points appear in the segment after the outlier is removed, cubic spline interpolation is used to fill the missing points to obtain the cleaned segmented strain sequence. All the cleaned segmented strain sequences are spliced ​​together in chronological order to obtain the complete cleaned strain sequence.

5. The machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers according to claim 1, characterized in that, S3.1 is as follows: For a given moment, a fixed-length sliding window is defined centered on that moment. The strain subsequence within the window is extracted, and multi-level wavelet packet decomposition is performed on the subsequence. The coefficients of each frequency band are reconstructed to obtain the reconstructed signal of the corresponding frequency band. Then, each eigenvalue of the multi-scale strain wave energy eigenvector at that moment is defined as the ratio of the energy of the signal in that frequency band to the total energy of the window. A dynamic adjustment factor based on the strain rate of change is introduced to obtain the wave energy eigenvalues ​​of each frequency component at that moment. The wave energy eigenvalues ​​of each frequency component are fused to obtain the wave energy eigenvalue at that moment. The wavelet packet decomposition uses the Daubechies wavelet basis, and the number of wavelet packet decomposition layers is preset according to the sampling frequency and characteristic frequency band requirements.

6. The machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers according to claim 1, characterized in that, S3.3 is as follows: A temporal prediction layer combining a temporal convolutional network and a self-attention mechanism is constructed. This layer first encodes the sequence of feature vectors that fuse physical information at all times in the temporal dimension by expanding the causal convolutional network to extract deep temporal features. Then, the contribution weight of features at different historical times to the current cracking risk is calculated by a multi-head self-attention mechanism, and the weighted features are mapped to the cracking probability at the current time.

7. The machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers according to claim 1, characterized in that, S4 is as follows: The model's total loss function minimizes the difference between the predicted probability and the true crack label, while encouraging the predicted probability to vary in tandem with physical quantities such as the strain rate of change, and applies L2 regularization to the model parameters to prevent overfitting.

8. The machine learning-based method for predicting crack resistance of circumferential prestressed steel reinforcement in concrete main towers according to claim 1, characterized in that, S5 is detailed below: After cleaning the sample data in step S1 in step S2, it is divided into training set, validation set and test set. During training iteration, batches of data are taken from the training set to sequentially complete multi-scale strain feature extraction, physical information fusion, temporal feature encoding and graph convolution spatiotemporal calibration to obtain the final crack probability. Then, a composite loss function is used to comprehensively evaluate the prediction bias, physical law fit and parameter complexity, and backpropagation combined with Adam optimization algorithm is used to update the trainable parameters. The model traverses the training set in multiple rounds, and the performance is monitored with the validation set in each round. When the loss on the validation set no longer decreases or the performance tends to stabilize and degrade, training stops and the optimal parameters are saved as the trained model. Finally, the trained model is tested using data from the test set.

9. The machine learning-based method for predicting crack resistance of circumferential prestressed steel bars in concrete main towers according to claim 1, characterized in that, S6 Specifically as follows: The trained model is used in the real-time crack prediction system for circumferential prestressed steel bars of the concrete main tower. The system receives data from sensors on the main tower, cleans the new data in real time, processes it through the trained model, and outputs the final crack prediction probability. If the final crack prediction probability exceeds the preset threshold, an early warning is issued to reflect the risk of steel bar cracking in real time.

Images